Introduction

Production planning and control (PPC) refers to the activities of loading, scheduling, sequencing, monitoring, and controlling the use of resources and materials during production. Loading concerns how much to do; scheduling concerns when to do things; sequencing concerns in what order to do things; and monitoring and control is concerned with whether activities are going to plan, and corrective actions needed to bring activities within plan (Slack et al., 2013). Commonly, these activities of PPC are carried out and coordinated using enterprise resource planning (ERP) systems (Arnold et al., 2011) and spreadsheet solutions (de Man & Strandhagen, 2018). However, ERP systems are typically unwieldy and do not support real-time decision-making that today’s market environments demand. Furthermore, manufacturing execution systems (MES) and advance planning and scheduling (APS) systems have also been developed in the last two decades to address some of these weaknesses of ERP systems (Öztürk & Ornek, 2014). While APS systems have been associated with various potential benefits, including support for real-time decision-making, the challenges associated with their implementation and integration with ERP systems render these benefits far from achievable in practice (Lupeikiene et al., 2014).

Currently, the business environment is typified by increasing market and supply chain complexity, globalization and global competition, waves of protectionism, and customer expectations of more sophisticated products. From time immemorial, notable studies have highlighted the aforementioned challenges as the impetus for increased efficiency in production systems and have devised strategies that could be used to achieve this aim – including those that involve the use of information technology (IT) and lean manufacturing techniques (Chan, 2005; Hong et al., 2010; Skinner, 1974). Additionally, production systems are generating increasingly large volumes of data and the potential for enhancement of planning systems to use this data for performance improvements has been widely promoted in industry and academia, but with limited adoption (Chavez et al., 2017; Fatorachian & Kazemi, 2020; Nagy et al., 2018). This brings one to the concept of smart manufacturing or Industry 4.0, which presents a new frontier for the advancement of manufacturing planning and control for its potential realization, spurred by the concurrent maturation of emerging ‘smart’ technologies such as cloud computing, internet of things (IoT), big-data analytics (BDA) and machine learning (ML) for improving the PPC system and processes (Bueno et al., 2020; Cadavid et al., 2020; Oluyisola et al., 2020).

For smart manufacturing (and the associated terms of industry 4.0), while many authors have addressed the potential impact, the expectations, industry implementation experiences, strategies for adoption, there is currently no clear methodological guide towards the design and development of a smart PPC system – the valley that separates conceptual literature from implementation reality. There are gaps, of architectural designs and concepts, and more importantly about how to translate the system requirements and attributes to the lower level elements – of data structures, of class definitions, of entity-relationship diagrams, of matching appropriate algorithms, etc. – in a way that supports the development of smart PPC systems that fit the near- and long- term requirements of a production system. This is particularly important for smaller companies who have more restrictive research and development budgets, and now for big industry leading companies at times of global economic crises. In this regard, Kusiak (2017) notes that:

“Academics push technological frontiers, from artificial intelligence to deep learning, without considering how they will be applied. Manufacturers want to know what types of data to sample, which sensors to use and where along the production line to install them.”

Smart PPC focuses on the ‘brain’ of manufacturing operations and aims to intelligently plan and control current industrial assets and materials as well as future, more adaptive production systems. A Smart PPC system should employ emerging technologies to: enable the reduction of forecast uncertainty by using real-time demand and production system data; offer dynamic re-planning in the sense that it enables frequently updating and the ability to re-plan when there are new developments in the production system; capture the influence of an expanded set of factors including telemetry factors especially for the process and semi-process industries; to capture the experience of the operators or the production planners over time; and predict short-term system parameter values and enable increase agility (Bueno et al., 2020; Oluyisola et al., 2020; Strandhagen et al., 2017).

Currently, there is no systematic, holistic design and development guide for the design and development of a smart PPC system. This paper presents an attempt to address these gaps by discussing the design principles for smart PPC solutions and demonstrating (with a case illustration from a semi-process industry) the use of a 5-step methodology for designing and developing smart PPC solutions. Note that in this paper, design refers to the architectural design rather than a user-interface of graphical design. The presented method details how to capture a production system’s attributes into the design and development process. Consequently, the following research question motivates this study:

  • How should a smart PPC system be designed and developed so that it is fitting with the current characteristics and the future requirements of the production system?

A design science research approach was used to address this research question. Design science, as an active problem solving research method, is useful when researchers aim to develop an artefact (Holmström et al., 2009). The case study used for illustrating the artefact – the methodology – was selected because it offers a production environment amenable to a smart process strategy where it is easier to demonstrate the benefits of smart PPC (Oluyisola et al., 2020; Tenhiälä, 2011). The proposed smart PPC design and development methodology is developed from a combination of extant primary and secondary literature and thereafter illustrated by the case study. While the methodology is designed to be generic, the illustration highlights the importance of case context during application. Furthermore, while an attempt has been made at providing details for the implementation, the complete technical implementation tools are not within the scope of this paper. This is because the technology stack or platform each company chooses could differ based on company policy and currently held expertise within the company and thus excessive implementation details will be of little cross platform value. Therefore, this study emphasizes only those details that present the elements which any sufficiently trained IT person can understand and use in designing a smart PPC solution for their case.

The remaining sections of the paper are structured as follows. In Sect. 2, a theoretical background is presented which covers PPC, emerging technologies applications for PPC processes, design considerations for complex information systems and data and systems’ architecture requirements. In Sect. 3, a derived method is presented. In Sect. 4, the method is illustrated in a case to show its strengths and weaknesses. In Sect. 5, the implications of the proposed methodology for the future development of smart PPC within both research and practice are discussed. In the final section, conclusions, limitations, and further research ideas are presented.

Theoretical background

In this section, the practice and the fundamental challenges in current applications of PPC and the limitations of the enterprise planning systems to address most of the requirements of smart PPC are outlined. Thereafter, the uses of smart, emerging technologies within the PPC domain are evaluated, followed by design and development considerations.

Production planning and control theory and the limitations of classical PPC systems

Fundamentally, PPC is tasked with the problem of managing uncertainty in production systems, either through stabilizing the system (common with lean approaches) or through predicting and reacting effectively and speedily to events and changes in state of the production system. The latter requires few or frequent rescheduling depending on the kind of operation and the stability of the production environment (Vieira et al., 2003). In achieving these goals, various process logics and methods have been developed at different levels of detail and time (hierarchical systems) and at different domains, for example, algorithmic research, strategic selection of PPC systems and implementation challenges and limitations. This diversity of topics and issues have led to different streams of research.

One stream of research has focused on investigating the effectiveness of enterprise resource planning (ERP) systems for PPC in different industrial environments, e.g., in dynamic market environments (Tenhiälä & Helkiö, 2015), in make-to-order manufacturing environments (Aslan et al., 2012, 2015), in small and medium enterprises (Ahmad & Cuenca, 2013), etc. The research within this stream has often been triggered by perceived limitations and inadequacies of ERP systems in supporting manufacturing planning and control activities. The most frequently mentioned limitation of ERP systems is generating unrealistic or infeasible production schedules due to infinite capacity scheduling (Arica & Powell, 2014; Steger-Jensen et al., 2011). Meanwhile, these limitations of ERP systems have paved way for the second research stream, which concerns auxiliary planning and control systems such as MES and APS. Consequently, the infeasibility of production schedules generated by ERP systems and the inability to tightly control operations have led to some large manufacturers using APS systems for planning and MES for production control respectively (Saenz de Ugarte et al., 2009; Steger-Jensen et al., 2011).

While MES and APS systems can address some limitations of ERP systems, these planning and control systems are known to have their own limitations. The processes within these systems have remained simplistic or too rigid, which limits the factors that can be considered within production planning and control decisions. Adjustment to schedules based on real-time or near-real-time data is infeasible and commonly avoided by production planners. It is also expensive to integrate additional software (called ‘add-ons’) with the large, monolithic systems, often making it difficult to adapt to changing business needs and leading many manufacturing managers and planners to build simpler, easier to manage, but disparate tools outside their PPC systems (Carvalho et al., 2014; Shaikh et al., 2011). Consequently, another stream of research has looked at the development of complementary decision support systems for addressing some of the challenges being faced by companies implementing ERP, APS and MES systems. Indeed, it is commonly reported that planners and supervisors, in many instances, tend to prefer simpler and more flexible tools and are more likely to avoid more complex, albeit theoretically performance-improved methods for addressing many of the production planning and control needs (de Man & Strandhagen, 2018; Tenhiälä, 2011). Therefore, while enterprise planning systems inhibited high efficiency for PPC processes by being unwieldy and not including additional real-time system data, flexible approaches have been limited in that they are often very manual, dependent on the availability of specific people and also not holistic (Oluyisola et al., 2020).

Emerging uses of smart technologies for smarter PPC

The adoption of smart technologies has seen tremendous increase in recent decades due to increased availability and affordability of computing power (Guha & Kumar, 2018). Generally, there are two ways in which companies adopt a technology: either a company (or its leadership) is pushed by its industry peers in the form of a market trend, or a business need leads to a search for a technology solution (Beckman & Rosenfield, 2008). In either case, the technology’s potential value is fully harnessed only when there is a fit between the requirements and application of the technology, and the firm’s strategy, processes – both manufacturing and support – and its planning environment (Bharadwaj, 2000; Buer et al., 2020). With the enormous hype that came with Industry 4.0, it could be said that technology push has been the driver for most of its recent research and applications thus far.

Within the last two decades, there have been huge interest in research exploring the use of smart technologies to improve the performance of production systems and these studies can be grouped into three categories. The first group consists of studies where smart technologies are used individually. For instance, there are studies on the use of radio frequency identification (RFID) or other IoT technologies for tracking of materials and goods within a manufacturing system to provide data for evaluation and optimization of material flows and layouts (Lee & Özer, 2007; Ngai et al., 2008). An example is Zhong et al. (2013) who used RFID in a mass-customisation production environment to track and trace items on the shop floor, collecting real-time production data to identify and control shop floor disturbances through an MES. In another example, Ngai et al. (2007) report on a case study on the development of an RFID-based traceability system for tracing repairable items in aircraft maintenance operations.

The second group are those that build upon the use IoT technology and other tracking and tracing technology, adding the power of cloud computing to these solutions. This addition typically enables the management of several thousands more IoT sensors thereby allowing for a more nuanced tracking of materials and resources on the shopfloor and in the wider supply-chain. The concept of digital twin falls within this category of research and application especially when applied to a factory or individual machines in the factory. For example, Qu et al. (2016) develop a concept and system for IoT-based dynamic logistics control with cloud manufacturing and demonstrate their approach within a paint-manufacturing company in China which uses the make-to-order strategy. The solution concept offers real-time tracking and dynamic re-planning based on changes to the state of the system. In another example, Tao et al. (2018), in their conceptual study on data-driven smart manufacturing, discuss the distribution and tracking of materials, and the integration of data from the production process into production plans using an example in wafer production. The paper raises several points that can be useful in the design of smart PPC systems (such as the integration of digital twins and IoT technologies like edge gateways and edge computing) but does not address this explicitly. In a related study, Sun et al. (2020) propose a visual analytics approach to production planning, to address the need for solutions that will enable a quick response to sudden changes in the operations and market environment, and with the ability to handle the deluge of data in emerging industry 4.0 manufacturing systems.

The third group is newly emerging, with the recent interest in advanced analytics tools and artificial intelligence and its derivatives/subsets – i.e., machine learning and deep learning (Bueno et al., 2020; Cadavid et al., 2020). The interest in using machine learning in PPC by itself is not entirely new. Garetti and Taisch (1999) long ago investigated the application of machine learning in production scheduling problems. However, as with several studies of its type, their approach to the use of machine learning to improve manufacturing through smart PPC suffers from the solution linearity problem (Cadavid et al., 2020). The solution linearity problem is the issue that most of these studies are linear from data cleaning, to data exploration, and so on until insights generation and retraining, typically carried out through desktop operations. However, for production scenarios where scalability and autonomous system operation is desirable, these linear solutions are inadequate and will require continuous, often expensive human expert management to use in production. Thus, there is a need for self-sustaining solutions.

These studies have raised, although indirectly, some pertinent issues as regards the design and development of smart PPC systems. These issues are perhaps best synthesized in Bueno et al. (2020), where the authors identified several gaps and suggestions for future research in the smart PPC research domain. First, (on p.15), they highlight a scarcity in extant literature regarding the question of fit of industry 4.0 solutions and the integration of PPC in different environments. This question determines whether a solution, even if well executed, will deliver any real and lasting value to a manufacturing operation. Secondly, they emphasize on the need for research within development of intelligent decision support systems, frameworks and architectures that can advance smart PPC. And thirdly, they highlight that there is the need to determine the types of data to collect and use, the types of sensors to use and where in the production system to deploy them. This paper aims to address these gaps.

Design and development considerations

Concerning the application of smart technologies for PPC processes, the common cases reported in the literature can be categorized according to whether they address the strategic or long-term, tactical or medium-term, and operational or short-term scope within the PPC domain. The strategic use cases remain scarce in the literature (Bueno et al., 2020). This could be due to the immaturity of the emerging technologies to handle such broad data types and sources that typically feed into the strategic process, currently typified by use of managerial judgement who are able to also include those data sources that are difficult – but not impossible – to codify or assign a numerical value to. Meanwhile the tactical and operational PPC domains have seen increasing use of data with big data and machine learning for decision-making especially because of greater automation in operations processes. Furthermore, the distinction in the application of emerging technologies at the tactical and operational levels is not always clear, and use cases often overlap. Examples of use cases include real-time visualization and scenarios’ simulations (Sun et al., 2020), product quality control, and integrated production-maintenance scheduling (Biondi et al., 2017).

When using machine learning, the choice of appropriate algorithms and the system features to be used in training models can both be critical factors on project outcomes because different algorithms fit or perform better depending on the use case, features’ data quality and data architecture, and system architecture (O’Mahony et al., 2008; Pineda-Jaramillo, 2019). As noted by James et al. (2013), “on a particular data set, one method may work best, but some other method may work better on a similar but different data set” [p. 29]. Therefore, it is crucial to find a fitting method to fit the use-case when using ML. An overview of machine learning algorithms in PPC use-cases and some architecture considerations follows.

Choosing an appropriate machine learning algorithm

As there are several ML tools and algorithms in the public domain currently, it can be a daunting task in finding one appropriate for a PPC use case. Within each of the three general categories of machine learning–that is, supervised, unsupervised, and reinforcement–new and more efficient algorithms and hybrids are being created continually, encouraged by the deluge of data, geometric reduction in computing cost that cloud computing brought about in the last decade, and advances in algorithm development and transference across multiple domains (Cadavid et al., 2020; Risi & Togelius, 2020).

Supervised learning concerns the approximation of a function based on a given set of input–output pairs. In this learning paradigm, the learning algorithm is provided (training) data which provides both, input values and output values, and the algorithm approximates the function that relates the inputs to the outputs. The approximated function can then be used to predict the outputs, given a set of inputs from outside the training set. The second machine learning paradigm, i.e., unsupervised learning is more exploratory in nature. Unlike supervised learning, there is no requirement for predefined input–output relationships in the training data that is used in unsupervised learning. Instead, the learning algorithm explores the data to find patterns and structures in the dataset, revealing which data-elements can be used as predictors of other elements. The third paradigm, i.e., reinforcement learning involves the use of iterative trial-and-error logic to train an algorithm to generate responses to inputs, that are expected to yield the highest reward (Monostori et al., 1996). Some use cases for the different machine learning types are presented in the following paragraphs and a summary in Table 1.

Table 1 Analytics and ML algorithms applied to PPC use cases
Full size table

Examples of supervised in the literature include Gyulai et al. (2014) who report on a case where supervised learning is used in optimizing the allocation of different products to two types of assembly lines, namely, reconfigurable and dedicated assembly lines. They use a random forest algorithm for predicting production costs for given order volumes and resource pools. In subsequent work, the authors use multivariate linear regression for predicting capacity requirements for future production scenarios on a flexible assembly line based on data from the manufacturing execution system (Gyulai et al., 2015). Heger et al. (2016) use Gaussian process regression for estimating the effect of different parameter settings on dispatching rules for scheduling. Examples of the use of unsupervised learning includes Pillania and Khan (2008) who applied k-means cluster analysis for categorizing firms in a supply chain according to each firms agility. Huang et al. (2019) propose the use of deep neural network for predicting future bottlenecks in a flexible manufacturing system, which is a use case for unsupervised learning. In another example, Shiue et al. (2012) propose the use of self-organising maps for selection of scheduling rules in semiconductor wafer fabrication.

Reinforcement learning, despite its huge potential for systems such as manufacturing systems, has only seen timid interest for PPC applications. Of interest within PPC research is the type of reinforcement learning called inverse reinforcement learning (IRL). According to Ng and Russell (2000), IRL may be useful when an agent is learning a “skilled behaviour,” such as the planning optimal scheduling process, and for which the reward function being optimized is determined by “a natural system”, such as a production system. Li et al. (2012) propose the use of Q-learning algorithm-based reinforcement learning for joint pricing and lead time decisions in a make-to-order system, where the decision problem is modelled as a semi-Markov decision problem. Tuncel et al. (2014) propose a Monte Carlo reinforcement learning algorithm for line balancing in disassembly operations under uncertain demand. Aissani et al. (2012) use a multi-agent based SARSA (state-action-reward-state-action) algorithm for production and distribution scheduling in a multi-site production network of a clothing company. Palombarini and Martínez (2012) use relational reinforcement learning for real-time (re)scheduling of extrusion operations in a secondary case study, i.e., the problem formulation is taken from literature. Lin et al. (2019) demonstrated an adaptation of the deep-Q network using an edge computing framework with multiple dispatching rules to demonstrate improved simulation results for job shop scheduling problems compared to methods using singular dispatching rules.

A related topic which has also seen significant recent developments is the use of ML algorithms in conjunction with heuristics and metaheuristics to address planning and control problems, especially at the operational level. Due to the mathematical intractability of production scheduling problems, using exact algorithms is often infeasible in practice, and heuristic policies are sometimes more pragmatic alternatives (Ðurasević & Jakobović, 2018). Metaheuristics such as genetic algorithms, tabu search, particle swarm optimization, etc. provide better results than heuristics for some scheduling applications (Maoudj et al., 2019; Ouelhadj & Petrovic, 2009; Xiong & Fu, 2018). However, most metaheuristic algorithms are non-deterministic and require long solution times for large problem sizes (Maoudj et al., 2019). Recent studies explore the use of ML to address these limitations and to support more efficient use of metaheuristics, for example, by using ML for the reduction of the solution space for metaheuristics (Bouzary et al., 2021) or for identifying when it is beneficial to rerun the metaheuristic (Li et al., 2020). Bouzary et al. (2021) propose a combination of support vector machines and genetic algorithm for addressing the service composition problem in a cloud manufacturing context, where they use ML for identifying the solution space for the metaheuristic. Li et al. (2020) use tabu search and genetic algorithm for schedule optimization, and a random forest classifier for identifying instances when production should be rescheduled based on whether the metaheuristic is likely to yield a more efficient schedule than the one available.

In all these developments, one important area that has seen little overage in the smart PPC literature is about the management of data acquisition and integration, data exploration, and a process to continually update and retrain ML models during use (Cadavid et al., 2020). In practice, the absence of a complete (or “circular”) workflow leads to changes to the system going undetected over time, a phenomenon known as concept drift. This is a major shortcoming of extant data analytics and machine learning research in general, and especially with regards to application within the PPC domain (Cadavid et al., 2020; Hammami et al., 2017).

Data architecture considerations

The data architecture describes the design, structure and control of the data generating and collection elements. As data is the foundation for smart manufacturing and related concepts including smart PPC, the data architecture plays a vital role in the implementation and long-term viability and flexibility (to adjust to change) for any such system. For convenience and for hierarchical analysis, data from the manufacturing system should be amenable to grouping according the familial associations. This can be achieved using classes and objects belonging to those classes, in fitting with the object-oriented architecture. The objects that are members of the same class with similar attributes such as usage area, etc. For example, a ‘Sugarproducts’ class can have members such as ‘Orangemix’, ‘Gingercandy’ (all random names) which comprise that class. The machines can also be grouped into classes for instance the ‘Driers’ class could comprise all the driers in a factory’s production line.

Furthermore, data quality played a key role in the value companies were able to derive from enterprise planning systems like ERP and MES systems before the emergence of smart PPC systems (Gustavsson & Wänström, 2009). The importance of data quality is now more crucial because of the data intensiveness of smart PPC systems which use data from a wide range of sources including from within the plant, (potentially) from other partner systems, and from the production system’s environment (Oluyisola et al., 2020). And while current enterprise systems collect sales transaction data from external customers and transactions generated directly from operations such as materials consumption in warehouses and factory floor production data (Koh et al., 2011; Mantravadi et al., 2019), the capacity to derive value from the abundant data in real-life environments has been a challenge (Kusiak, 2017).

Furthermore, there are different types of data available to any PPC system. Based on the temporal proximity of the data generation and collection processes, they can be classified as being either batch data, where data is collected and updated periodically, or stream data, where data is being generated, collected, and potentially analyzed in real-time. In production environments, many data processing systems implement some kind of runner using the Apache Beam model (Li et al., 2018). Most of data from the factory’s environment and some of the machines in the production lines are time-series, stream data. An example of the time-series data snippet from an IoT device on a production machine in the JavaScript Object Notation (JSON) format is as shown in Fig. 1 below. But there are also batch data which are seldom revised, for example the setup cost, and are input to the PPC processes.

Fig. 1
figure 1

Example of the telemetry data generated by an IoT sensor on a production line

Full size image

Systems architecture considerations

An information system’s architecture can be defined as a collection of artefacts, namely a definition of constituent components of the IS, a specification of the properties of those components, and a description of the relationship among those components and their interactions during operation (Bass et al., 2013; Goepp et al., 2006). The use of the term design in this paper generally refers to the creation of the architecture of the smart PPC system. Because smart PPC systems are information systems, its developers must follow similar principles. This design must be made early in the overall development process, and in a way that allows for enough detail so that the developers have enough guidance, while at the same time it must allow sufficient freedom for the developers to make decisions during the actual development stage (Bass et al., 2013).

Within the broader Industry 4.0 research domain, generic architectural models have been proposed for the industry 4.0 production system and these can provide inspiration for the smart PPC solutions designers and architects. Common examples include the Reference Architecture Model for Industrie 4.0 (RAMI 4.0), the Industrial Internet Reference Architecture (IIRA) and the internet-of-things reference architectures (IoT RA) standard in the ISO/IEC 30,141:2018 (Standardization, 2018). Nevertheless, these models should only serve as reference due to their generic nature and the fact that they do not cater for the context each production manager must address.

Typically, enterprise planning systems are designed as hierarchical control systems using a monolithic architecture (Themistocleous et al., 2005). This is the case in which the system is built on a single, large, highly powered computer hardware. Such an architecture has several benefits, not least its speed due to its low natural latency, its limited need to manage integrations with several units, and that there is only a single hardware device to be managed instead of potentially several. And this was important for many decades before the advent of cloud computing, since companies had to create a physical datacenter with server hardware and all the attendant management requirements. But this architecture also hard several shortcomings. It limited flexibility to, for example, add new tools and functionalities as this needed to be upgraded every time the main server itself received an upgrade from its suppliers who are often very big software vendors and who made upgrades based on general needs, and not on the specific needs of each customer. It is more costly to start-up, manage and run with a savings of up to 50% in terms of total cost of ownership (Mattison & Raj, 2012). This contrasts with the emerging smart technologies which are changing so fast, that there is an intrinsic need to design for flexibility and frequent changes.

These design considerations are addressed by the modular-by-design microservices architecture instead of a monolithic architecture. The microservices architecture presents a considerable benefit for several reasons: it can scale easily, and it is highly adaptable. It has been reported that self-adapting and self-optimizing multi-agent distributed production control systems have been demonstrated to perform better during transitions when used in job-shop environments where hierarchical systems are too rigid to adjust to the flexibility requirements of such environments (Ma et al., 2020). Thus, smart PPC systems initiatives have a better chance at success if they employ a microservices architecture.

Finally, most research on the use of ML in PPC suffer from workflow-design linearity in addition to being based on artificial or historic, sampled data (Cadavid et al., 2020). While these conditions make testing specific models for confined problems easy, they are not feasible in real-life industrial practice. The challenge with linear design is that to use it in practice, a human operator needs to administer the intelligence creation process of the system, as seem in reported case literature, for example, Garetti and Taisch (1999) and Brintrup et al. (2019). In real-life industrial scenarios however, the smart PPC system should be able to collect data, clean it, prepare it for analysis, retrain its models, and offer refreshed insights without human intervention potentially self-adjusting its control parameters (Oluyisola et al., 2020; Rojas & Garcia, 2020). It should address the risk of concept drift, for instance by using adaptive time windows (Bifet & Gavalda, 2007). This could be achieved using data processing pipelines and monitoring scripts connected to a version control system for managing model versioning, a concept referred to as MLOps – that is, machine learning operations, which is a derivation of DevOps for ML.

In summary, there are two main perspectives in the literature through which topics related to smart PPC have been viewed. First, in the puristic production and operations management perspective, information and communication technologies (ICTs) are viewed as add-ons or auxiliary that can enable or improve information flow but are usually considered exogenous (Slack et al., 2013). A contrasting view is that of information systems-centered research within the context of manufacturing, that considers ICTs as an integral variable and focuses on opportunities for performance improvements by employing ICTs – for example, Huang (2017). In the methodology proposed in this paper, an attempt is made at using a more balanced, multi-disciplinary view. In this context, material flow is controlled and monitored with ICT-enabled information flow, thus making ICTs integral components of the industrial system. Smart technologies or advanced ICTs are thus viewed as intrinsic elements of the smart production system as opposed to being add-ons.

A methodology for designing and developing a smart PPC Solution

Having already established the need for a systematic methodology and guide for manufacturing firms who may want to develop a smart PPC solution, the key steps that such an initiative could follow and the elements that should be given proper consideration are presented in this section. Here, the presented ‘steps’ suggests ‘sequence’ suggesting which steps should precede what. However, as it will be explained in the case study, the process does not have to be linear. In practice, it is often necessary to revisit preceding steps while at another as the requirements become clearer to the stakeholders of the project. The following steps can be followed in developing a smart PPC solution:

  • Stem 1. Preliminary study: determine objectives and priorities in fitting with the planning environment variables.

  • Stem 2. System requirements specification: validate the operations’ problems and identify performance indicators.

  • Stem 3. Identify data sources and select relevant analytics and machine learning algorithms that fits the problem.

  • Stem 4. Design system and data architecture with consideration for integration with extant systems and IoT telemetry.

  • Stem 5. Implement with considerations for development methodologies, continuous innovation and long-term adaptability.

Step 1: preliminary study: determine objectives and priorities in fitting with the planning environment variables

Most digitalization projects are driven by either an identified business problem or a perceived market opportunity. And since they are innovation projects, the immediate goals of the smart PPC solution must be determined ex ante to reduce the risk of scope creep and to increase the chance of success. The preliminary study takes a high-level view of the problem or opportunity, and with particular emphasis on how the market, product, and process attributes inhibits or promotes the problem. Also, the management sets the objective regarding how much of the opportunity the company is willing to pursue or to what extent an issue needs to be addressed. For example, if a production planning process is having a fulfillment rate of 75 percent on average and leading to unacceptable underutilization of booked operator hours and wastage of materials, management could set an objective to improve this key performance indicator (KPI) with the use of smart technologies to, say, 90 percent in a year’s time. These objectives and priorities must be weighed against the constraints imposed by the planning environment attributes of the operation.

It is also a common occurrence for there to be a need to make tradeoffs over which elements of the solution requirements to prioritize in the short and long term. For instance, a company in the process industry manufacturing, say, industrial paint, may see several opportunities and use cases for digitalizing its operations and PPC processes. Easily, managers could be interested in digitalizing the production line with IoT sensors that will collect various kinds of data about the production processes and send these data to the cloud for analytics and predictions, or on an edge device for real-time response. Another use case could involve attaching sensors to the packaging containers (which may be a bucket) or pallet, enabling a full tracking and tracing of the inventory coming out of the production line; yet another could involve the tracking of weather or climate factors and how this affect demand or sales at the stores; and so on.

Now, if this were a large multinational with millions of euros in research and development budget, then the company could start with and run multiple projects simultaneously, bearing in mind that results will be mixed. However, for a smaller company with a tighter budget, it will be critical to prioritize, focusing only on projects with a high expected return be it financial or digital competence gains for the company. In the example, following the argument that a process strategy is has great potential in this type of production environment, and the budget-constrained producer will prioritize those initiatives that lead to a smart process, for instance digitalizing the production line with IoT sensors capturing parameters that affect the yield of the operations. This could also be combined with other telemetry data from the production line’s immediate environment.

Step 2: system requirements specification: validate the operations’ problems and identify performance indicators

Step 2 takes the preliminary objectives and initial assessment from the top-management horizon in step 1 down to the detailed, solution-specific design requirements that could be used for the technical design and the actual development of the solution. As previously explained in step 1, the objectives are typically the prerogative of the company’s management team and often represent their interpretation of the problems that must be addressed from a top-down view of the operation. However, a lot of the data driven decisions and insights affect or are affected by junior managers and operators on the factory floor. Therefore, there is a need to validate the objectives of the solution from the perspectives of persons directly interacting with the production system before specifying the functional and non-functional requirements of the proposed solution. One way to achieve this is to formalize the requirements using user stories. User stories are written in the format “As an [role/persona], I want to [action] so that [why],” and each user story should be clear and descriptive. For example, a user story could go as follows: “As a production planner, I want to be able to upload productions orders for the next two weeks into the solution with approved production orders from the ERP system so that I do not have to copy this manually.” User stories should be independent, negotiable, value-focused, estimable, small, and testable. Later during implementation, production managers and the system developers will determine how to prioritize the user stories for development. In addition, there should be flexibility in terms of which elements of the system remain on the list of functionalities to be developed, while allowing for future adjustments (Pressman & Maxim, 2015).

In addition, performance indicators (PIs) are needed to monitor both the quality of analysis and predictions being generated by the smart PPC system and of the reliability of the system. The PIs relating to the quality of the results can include the standard deviation and errors for individual predictions determined through random spot measurements. Those relating to the performance of the smart PPC system can include prediction lag, simulation request processing time, and general indicators like availability/downtime hours and the like. While PIs relating to the quality of the analysis and predictions will be context specific, most of the system PIs are generic and common to service-oriented, cloud-based ICT systems.

Step 3: identify data sources and select relevant analytics and machine learning algorithms that fits the problem

The user stories give an indication of the services that primary users – production planners and operations managers – require the smart PPC system to fulfill. After identifying these services, the next step is to determine the relevant data sources from the production system and identify the appropriate analytics tools and machine learning algorithms that works best for the kind of insight or prediction required. This determination and identification can be done by a small technical team involving a machine learning engineer or data scientist with a good understanding of not just the technical problem but also the business problem.

In many manufacturing use-cases, pilot projects could start with simpler ML algorithms such as Gaussian linear regression and logistic regression (supervised), and with PCA and k-means clustering (unsupervised) with an acceptable level of success. However, after the pilot phase of such projects–that is, during the real-life implementation–there will be a need to improve the performance of the solution and which can be achieved using hybrid models which combine multiple features of the basic algorithms. For example, when the use case involves sparse data inputs and an extensive feature list, the hybrid algorithm called the DNNCombinedLinearRegression can be used in place of the common supervised learning to combine the strengths of neural networks (generalization) and linear regression models (memorization of feature interactions) (Cheng et al., 2016).

Step 4: design system and data architecture with consideration for integration with extant systems and IoT technology

Many large manufacturing organizations, in addition to having an ERP system also have full-fledged solutions for the control of manufacturing operations on the factory floor–the MES. Some MES systems have basic analytics capabilities built in such as statistical process control charts that allows process-tracking, and most collect time-series stream data from the discrete units of production lines to which they are connected. Alone, using the MES for manufacturing control misses the opportunity that a holistic, connected smart system affords. Therefore, the system architecture should cater for the introduction of IoT sensors to the factory even for factories are already automated. The MES and ERP systems provide a good starting point for developing smart PPC solutions. The data from these systems and other factory IT systems might however require extensive transformations before they can be used in combination with newly installed IoT technology in the smart factory.

In general, modular smart PPC solution design would perform better than a monolithic solution since it will allow for future improvements within each module independent of others and will also ensures that failure in one service does not break the entire system. Furthermore, when the solution is built on a service architecture from the onset, the it is easier to add more modules in the future and to update individual modules that are already in use. This is achieved by designing the modules as services and building application programming interfaces (APIs) to manage interaction among services. The data processing, model development, and prediction processes can be carried out without manual human interaction by automating the data preparation and prediction processes using ML pipelines.

Moreover, in cases where an active control (rather than just a monitoring) of the production process is required, it is advisable to have the trained machine learning model interacting with the production machines and processes on the “edge” without the need to send to the cloud and send instructions back to the plant. However, because the real-time data processing occurs at the edge, this creates a challenge due to the limited processing power at the edge and need for continuously monitoring the performance of the model to avert model drift. Furthermore, edge devices may lose their connection to the cloud and thus the solution must cater for offline operations. Otherwise, where there is no need for any serious computing at the edge, it suffices to send all data generated from the production system to the cloud.

Step 5: implement with considerations for development methodologies, continuous innovation and long-term adaptability

For the implementation of smart PPC solutions, there are at least four key considerations: whether to outsource or develop in-house, which software development methodology to adopt, whether to choose managed-cloud services or to use completely open-source technologies, and how to design the system so that it supports continuous innovation. It is possible to develop in-house or to establish joint development partnership arrangements with service providers for small-scale functionalities. But it is more likely to outsource major system development activities to established IT firms if the needed project execution competence is lacking in-house. In addition, the development of the solution will often require the choice of building almost from scratch with the use of open source technologies, or – if faster deployment is desired – the use of any one or a combination of the several managed-cloud services for a faster development process, while allowing for agile development.

Smart PPC systems need to support continuous innovation. Continuous innovation in this sense relates to how the established IT infrastructure and software development processes eliminates tedious manual processes for making changes and improvements to working system, and allow a seamless, continuous integration, testing, and deployment of those changes without any downtime. Therefore, non-agile methodologies will generally be insufficient for their development because of the relative rigidity of such methods. And because many of the technologies being used in smart PPC systems are experiencing constant, fast-paced advancements, the success of any smart PPC solution requires that there is a smooth and simple process in place for its continuous improvement. Moreover, the alignment or integration of the workflows and processes of the both the machine learning engineers and software developers will enable the streamlining of continuous innovation and the refinement of ML models as new data becomes available from the production system being monitored.

Finally, for information systems’ developers, the concept of DevOps has emerged as a preferred way to manage the continuous, version-controlled, code development cycle – that is, write, test, (revise,) build, (revise,) deploy (revise). While machine learning engineers and data scientists take the ML cycle – that is, experimentation, model-creation, testing, operations, and maintenance. By integrating these two workflows – to have what is now referred to as DevOps for machine learning (MLOps) – productivity can be improved significantly through software development process automation, allowing machine learning engineers and data scientists to focus on the model performance rather than being bogged down in tedious software development operations’ activities. One way this is achieved is by using infrastructure-as-code and process automation in managing the system’s improvements and the revisions’ process. Process automation could be achieved using Bash or Python scripts, or through robotic process automation software that allows automation using drag-and-drop tools. The latter, less programming-intensive option can be managed by a trained production planner, thereby lowering the cost of development.

Case study

In this section, the method and processes presented above are illustrated with an application within a case, which will be referred to as Sweets and Nuts ASA (not real name) or SNASA for short. The company manufactures sweets and nuts-based products in its factory based in Norway from where it supplies grocery chains, kiosks, and petrol stations’ mini-stores in the Scandinavian region. The nuts production section of the factory is isolated from the rest of the factory in line with regulations concerning the control of allergens. The rest of the factory produces chocolate-based and non-chocolate sweets such as pastilles. The unit of analysis in this case study is the non-chocolate-based (henceforth, NCB) section of the operation.

Determine objectives and priorities in fitting with the planning environment variables

The NCB production is serviced by two production lines. The production process for the NCB products is as shown in the Fig. 2 below.

Fig. 2
figure 2

The NCB production process at SNASA

Full size image

The operations at the NCB section falls into the semi-process class. Raw materials are fed into the cooking drums in amounts determined by the recipe for the batch to be produced. When the cooking process is completed, the output is temporarily stored in a cooking buffer before molding, using mold trays with the shapes of the sweets engraved. The trays are thereafter arranged in racks which are loaded into one of the seven drying chambers in the drying section of the factory. The production data currently used in the production planning process includes the estimated lead time for all processes, stock levels of the different stock keeping units (SKUs) in the finished goods warehouse, recipes (which also provide a bill of materials). The maximum batch size the line can produce for each product is pre-calculated based on the capacity of the production processes.

The challenges of this current PPC system can be described in three categories namely, market (demand and supply) related, product related, and process related. First, market related demand related challenges stem from the high competitiveness of the industry and the fickle nature of human taste preferences. A popular product can sometimes loose its spark with consumers or get overshadowed by new trending products. For this reason, the NCB industry witnesses a lot of promotions and discount sales to drive and sustain demand. Secondly, the product related challenges are minimal in this case because the products are neither complex nor have any deep bill-of-materials which could have required extensive materials requirements planning tools. Furthermore, the simplicity of products made by this case company (packed sweets) and the price per unit implies that the product itself will not benefit from a smart product strategy. Rather, a smart process strategy will be for fitting for this type of case (Oluyisola et al., 2020). Such process approach must be able to track the remaining life for any product or batch in the finished goods storage and in the various warehouse within the company’s value chain and must also be able to trace its journey through the value (Høyer et al., 2019).

Lastly, the process related challenges are generally due to the nature of the materials being processed and the level of maturity of the process technologies. Currently, there is a long set-up and changeover time due to the need to wash the machines and equipment producing every new batch. This is also required to meet regulatory requirements for cleanliness and food safety. There is also a yield uncertainty that planners currently must guess when issuing production orders and this causes additional variability in the production system. Also related to the process is the operator-planning related challenges relate to how labour is planned in the company. Over several years, the company has developed a practice of planning batch sizes that can be completed within a production shift. This is a suboptimal constraint on the planning process. Therefore, with the attributes of this production environment, this company’s approach to smart manufacturing should take a smart process strategy, rather than a smart product strategy since the product is simple and the unit price is very small.

System requirements specification: operations’ problems and performance indicators

Problem specification

The company SNASA faces an immediate challenge: finding an optimal production schedule and managing the scheduling process to minimize variation. Thus, the production planning problem for this case comprises two main elements, namely: the determination of the optimal plan, which maximizes throughput through the bottleneck drying process and assumes no yield variation (that is, yield = 1.0 or 100%); and an estimation of the yield uncertainty factor, to improve the accuracy of production plans. Currently, planners must guess the what the yield will be and add some buffer to the amount that is produced so that at least the final production output for each batch exceeds the planned amount required to meet order forecasts. This leads to overproduction, and it particularly expensive for products which serve as inputs into ‘mix’-type products. The mix-type products are made by combining three to five different types of products into one assortment.

The planning problem can be summarized as follows:

  1. o

    Given a set of firmed customer orders, and master production scheduling orders (MPS orders are those generated by the ERP system based on demand and supply forecasts), with each order characterized by: its drying time (which is an indication of the throughput time), its due date, its volume or amount; and

  2. o

    Given a set of drying sections or rooms, each with a fixed drying capacity, and given a set of packaging lines, each characterized by a fixed capacity;

  3. o

    Find the schedule of orders that maximizes the number of completed orders at the two stages drying and packaging.

Furthermore, the planning problem can also be viewed as a multi-stage or multi-echelon scheduling problem for which although the drying stage, which is used for all products from the production line, is not always the bottleneck. This is because the average speed of the packaging machines is low enough that they can cause delays if poorly scheduled and depending on the product. This is partly because there are several packaging lines with varying speeds and no single product has a dedicated packaging machine. After production schedule is made, the plan must be adjusted for reality by estimating a yield uncertainty factor. This yield uncertainty is a factor of environmental parameters such as humidity and temperature.

Requirements specification

The requirements, shown in Table 2, were gathered from the production managers and planners of SNASA during this research-based improvement project towards smart manufacturing. An overview of the solution concept is presented in Fig. 3. KPI result data going into recommender system will include actual production performance (lateness, earliness, on-time, etc.), specific operator working the process (this shows how specific operators affect performance), etc. The newly added elements of this smart PPC system are described in Sect. 4.3. A description of each step in Fig. 3 is provided in Table 3.

Table 2 SNASA's requirements for the smart PPC solution
Full size table
Fig. 3
figure 3

Conceptual overview of the as-is compared to the to-be smart PPC solution

Full size image
Table 3 Comparison of the as-is and to-be processes (reference to Fig. 3)
Full size table

Performance indicators

It is important to have predetermined how the performance of the system will be measured. In the selection of performance measures or indicators for this case, there are two categories namely, operations reliability and services quality. The operations reliability measure has to do with how the software system is designed, architected, and developed. It is measured by reliability measures such as up-time (a maximizing measure) or downtime (a minimizing measure), percent failed schedule requests from the user interface and waiting time between schedule launch and results presentation on the dashboard. The services quality refers to the quality of the results, estimates and recommendations offered by the smart PPC solution. Measures include the amount of deviation of the estimated yield from the actual yield, the average performance of the recommended schedule logic over period.

Step 3: identify relevant tools and algorithms

There are two choices to be made regarding the two applications of machine learning within this SNASA case: one for estimating the yield and the other for recommending which schedule logic alternative will perform best for each planning scenario. The yield estimation (or prediction) can be hypothesized to be the dependent variable of a linear or non-linear system. As such, a simple linear regression model is a good start for this use case. Once the system is built and in place, other variants of the linear regression can be tested in a development environment to see how much improvement in performance is possible. Examples of those are models combining basic models with more performant neural networks such as the wide and deep DNNCombinedLinearRegression algorithm or similar. This model will be fitting for this purpose due to the potential sparseness of the features. The data fields that will be used in the model are shown (without telemetry) in the class diagram in Fig. 5 and a detailed list (with telemetry) is provided in the table in Fig. 5. Meanwhile, the subsystem for recommending which planning logic option to choose appears amenable to inverse reinforcement learning.

Step 4: solution architecture –data and systems architecture design

While academic projects on the use of ML in PPC tend to use linear development processes, live production projects require the use of recyclable, reproducible machine learning pipelines which can be automated. For the SNASA case, an illustrative system architecture for the yield estimator use-case is presented in Fig. 4.

Fig. 4
figure 4

An example smart PPC solution architecture for the yield estimator use-case

Full size image

In the illustrative smart PPC architecture in Fig. 4, the ‘connected’ production system which is connected by IoT sensors sends data via secure connections to a cloud data-ingestion service. This service can use a distributed commit log technology such as the open-source Apache Kafka or one of the easier-to-use IT vendor solutions. This is to be configured so that it guarantees that every data sent by a sensor is delivered, and so that the data are in the right sequence when they arrive at the Analytics Service. The real-time analytics output from the analytics service could be made available on a dashboard on the factory floor or in the production supervisor’s office for real-time monitoring of the factory. This data then flows from the analytics service to the data warehouse where it is accessed by the ML solution. The ML solution, which also runs in the cloud, will continuously check for model drift and it will activate retraining when set KPI thresholds are met. The ML model works within a web app with a graphical user interface for the production planner to interact with. The planner will input the actual yield and production order data after every production order is fulfilled. The web app will also ingest production plan data from the ERP system for the computation of the yield, and it will continually send both the estimated and actual yields to the data warehouse for later use during model retraining. (Fig. 4).

Fig. 5
figure 5

Class diagram for a demo in UML notation

Full size image

In this case study, the assumption of a one week “fixed” planning window is made in line with current practices by the production planners, during which the list of orders to be processed is assumed to be deterministic – except if a major disruption or urgent firmed customer order is received. However, during this one-week period, the forecasts for some of model feature variables (for example, environmental data) are only precise for two days into the future at any given point. Therefore, there will be a need for re-scheduling at least once every two days to take advantage of the trained model. In the future, when a lengthy historical data has been gathered, it will be possible to train the model using only the historical data without the need to use the weather forecast data whose accuracy diminishes materially beyond a 48-h from the reference point.

Steps 5: implementation considerations and performance assessment

This use-case is illustrated using open source technologies for the sake of demonstration. However, for production, the company might be better served by using managed services on any of the major cloud platforms. One could start with a small pilot to test an idea, or go big, with a large-scale project and iterate on improvements. The latter approach can lead to faster business impact. There are also pilot versus real-life production implementation considerations. The nature of this production planning is such as that the properties of the system of interest changes frequently relative to the target precision of prediction results. Furthermore, as the data scientist and the developers working on this project will need close collaboration, and there is a requirement to be able to scale the solution to address other PPC use cases as the companies gains organizational competence with PPC. These factors strengthen the need for MLOps.

Insights from the case study and implications for research and practice

In the preceding sections, a methodology for designing and developing smart PPC systems was described and the application of this methodology was illustrated through a case study. In this section, the application of the methodology within the case study is reviewed, after which insights gained from the case study and the implications for research and practice are discussed.

The objectives and priorities identified in the first step of the methodology were used as basis for formalizing the problem and specifying the requirements and relevant performance indicators. This step helped refine the requirements that were put forth by production managers and planners, who are the intended beneficiaries of the smart PPC system. These requirements included having multiple schedule logic options, integration with existing ERP system, dynamic rescheduling or more frequent scheduling updates, yield estimation using telemetry factors and capturing the experience of managers. While these were the functional requirements for the PPC system, non-functional requirements such as ease of use and readability of the user-interface layout were also identified although the latter non-functional requirements were not the subject of this case study. Consequently, operations reliability and services quality were deemed as the relevant performance measures for the smart PPC solution design.

Of interest in this case study was the problem of yield of estimation at the drying station. This was important in this case because the yield, which affects the precision of the entire planning process, is highly influenced by exogenous factors, e.g., temperature, humidity, etc., factors which can be modelled and predicted using analytics and ML tools. This again reemphasizes the importance of fitting smart technologies to production systems according to fit as pointed out in Oluyisola et al. (2020). By the same principle, a smart product strategy would not be beneficial in this case company. In addition, the formalized problem and specified requirements were used to identify candidate tools and algorithms to address the problem and fulfill the requirements. For the purpose of this case study, this selection of tools and algorithms was based on extant literature on smart PPC (reviewed in Sect. 2.3.1). While this was lightly covered in this paper, this as an area that future research need to address for the ML value in PPC to be realizable. The final step in the methodology focuses on continuous innovation and/or development, i.e., the system should be adaptable when weaknesses are identified during use or as opportunities for utilization of better or more mature technologies become available. The performance of the current, as-is process is compared with the improvements that can be achieved by the proposed smart PPC (when fully operational) in Table 4. These are measured against the general goals of the smart PPC system established in the literature. By reason of the capacity, consistency and flexibility that the smart PPC system affords, as the case illustration highlights, the improvements are such that the manufacturing organization will be able to anticipate and react more precisely to changes in the production environment.

Table 4 A comparison of as-is and to-be PPC systems
Full size table

The general implications of having a methodology such as the one presented in this paper are significant for research and practice. By having a methodology which starts with the determination of fit according to the planning and control environment variables, it will be possible to streamline smart PPC initiatives and increase their chances of success. Based on the PPC environment characteristics, it was possible to determine early in the process that the case company would benefit more from a smart process strategy rather than a smart product strategy. And while the issues of interest in the case study are primarily operational, the methodology itself is not constrained vis-à-vis the application context or decision levels and can be applied for initiatives pursuing strategic and tactical decision support.

Furthermore, due to the current rate of innovation within the disciplines of big-data analytics and machine learning, the availability of tools and algorithms for a given set of problems is constrained by the state of art at any point of time and may change as time progresses. Therefore, this step of the methodology could be reviewed after an interval, which should be decided during the initial or pilot implementation. The next step in the methodology concerns architectural considerations for the implementation of the solution. This step not only considers the architectural design for the proposed solution itself, but also considers the integration of the solution with the existing enterprise systems, thus re-emphasizing the focus of the methodology on ensuring fit of the smart PPC system with the planning environment. Furthermore, while designing the data architecture in this step, due consideration must be given to future scenarios, such that the developed system is scalable and amenable to future operational demands.

Additionally, as Cadavid et al. (2020) highlight in a recent review paper, there is a need to address the linearity limitations of extant research on ML-enhanced PPC and also a need to link tools, techniques and activities for industry get real benefits from research on the subject. The architectural considerations prescribed in this methodology addresses this key issue and should be a major consideration for future applied research on the subject. This cannot be overemphasized considering how small and medium sized manufacturing companies must grapple with the uncertainties of a pandemic-battered global economy and the post-pandemic global market.

Additionally, anecdotal evidence with manufacturing companies in the Scandinavia region shows that while increasing automation and digitalization has led to the creation of massive volumes of big data in manufacturing systems, a lot of the data is neither used nor is useful. The reasons vary for each case, but a recurring theme is that the data architectures are often designed primarily as a logging system for use in maintenance activities and many manufacturing companies still are yet to fully adoption an IoT strategy. All these factors then make it more challenging to derive value using analytics or machine learning to build intelligence into these production environments.

From the foregoing, the several considerations to be made when developing a smart PPC solution include the planning environment challenges which are often relatively consistent in the long-term, and the technology-related challenges which are related to the fast-paced evolution. And due to the significant uncertainty involved in the innovation process, and the high risk of project failure, the selection of use cases cannot be done randomly or based sole on what is trending with competitors. Indeed, while over 60 percent of IT projects fail outrightly or when defined by one of the performance metrics of timeliness, cost or quality (Mark, 2016), anecdotal evidence suggests that this may be even worse for projects involving emerging technologies. In one example, a major distribution and logistics center recently had an innovation project where it tried to deploy autonomous robots with machine learning capabilities in one its warehouses. The project failed both technologically and operationally, and the company did not share information about this failure publicly potentially because it does not help the company’s brand posture as a technology savvy organization.

It can therefore be assumed that there is a greater likelihood or perhaps a tendency for companies to want to report only successful digitalization projects. This may, over time, lead to a ‘survivorship bias’, as researchers would only have cases of successful projects to extract knowledge from, while losing access to the valuable knowledge that could be extracted from the failed implementations. Furthermore, this creates a lacuna because while there may be ‘local learning’ within each company, there is a global loss due to several companies repeating pilot projects that many others previously tried and failed at. Therefore, a systematic method of the type proposed in this study can help reduce the risk of smart PPC project failure and can reduce the variation amongst several subsequent smart PPC initiatives, thus enabling easier shared learning.

Conclusions, limitations, and further research

The question of how a smart PPC system should be designed and developed for an environment has been addressed in this paper through a five-step methodology. The steps of the methodology have been formulated and structured with the consideration that the resulting PPC system should fit the characteristics of the environment in question. Furthermore, the importance of contextual fit in algorithm selection, solution scalability and amenability of the smart PPC system to address future demands have been highlighted. In summary, the principles and considerations that guide the design in a smart PPC system are as follows:

  • The design of the smart PPC system should fit the characteristics of planning environment. This highlights an issue that has been observed in numerous ERP and APS implementation case studies – expensive monolithic systems forcing managers to modify the production system to fit an inflexible PPC system. The proposed methodology can guide the design and development of such a fitting smart PPC system.

  • The design and architecture of the PPC system should be scalable and amenable to variations in future demand volumes, demand patterns, product portfolios, number of users, etc. Since these parameters cannot be controlled or accurately predicted in advance, it is important to have provisions in the architecture to adapt as these parameters change during drift.

  • The implementation plan of a smart PPC system should also include a period of ‘incubation’ where data can be collected to train the ML models, if the data is not already available. Simultaneously, the models can be tested for accuracy, such that the estimation errors can be accounted for in the planning activities.

However, this study has the following limitations. First, implementing the methodology requires experience and judgement to ensure that the relevant contextual variables have been considered in assessing the fit of objectives and priorities with the planning environment variables. A framework of contextual variables could provide an exhaustive reference and reduce the requirements for experience in implementing the methodology effectively and will be addressed by future research. Finally, in future studies, this methodology will be tested in other types of production environments and industry sectors to assess its weaknesses and improve its robustness and generalizability.