Keywords

1 Introduction

Artificial intelligence (AI) has gained much space in the industrial environment by providing increasingly accurate analysis of a large amount of data. Furthermore, with the evolution of Industry 4.0, in which it is possible to implement, collect information and analyze various sensors more quickly, artificial intelligence is favorably viewed as it enables the application of methods that generate more knowledge about the production system.

One of these applications has been in the maintenance of equipment and systems through predictive maintenance. Predictive maintenance (PM) is conceptualized as “estimations that identify potential machine breakdown, allowing for the setback source to be eliminated or maintained” [13]. With this, we can identify the use of AI as an enabler for developing and applying predictive maintenance in the industry. Pagano [11] applied Long Short-Term Memory Neural Networks and Bayesian inference in the heavy industry. Ayvaz and Alpay [2] applied predictive maintenance in a baby diapers assembly line, and for that purpose tested the Random Forest, XGBoost, Gradient Boosting, AdaBoost, Multilayer Perceptron (MLP) Regressor, Neural Network and Support Vector Regression (SVR).

However, to obtain accurate results that support predictive maintenance analyses and decisions, artificial intelligence needs a large amount of data to train the algorithms and thus obtain reliable predictions [8]. Thus, the accuracy and reliability of AI results are compromised when it comes to companies that do not have this large amount of data. Another point that some companies are currently facing regarding data accuracy is that in addition to a large amount of data, these must have a good sampling of real data. Thus, to remedy the lack of good sampling and the absence of a large amount of data, the development of a digital Twin can enable the understanding of the entire system by collecting data from the physical system and simulating various scenarios in the virtual system.

A digital twin (DT), as defined by NASA [1] is: “an integrated multi-physics, multi-scale, probabilistic simulation of a vehicle or system that uses the best available physical models, sensor updates, fleet history, etc., to mirror the life of its flying twin” [15].

As the term got adopted more widely in different industries beyond the aerospace sector, different viewpoints emerged. The understanding of the term started to shift depending on the industry and its specific application. However, the idea of having a bi-directional connection between a physical asset and its virtual counterpart is a common vision for the implementation of a digital twin [14]. As a promising key tool for the success of digital transformation, DT is being used in industries to improve the manufacturing process’s operation and maintenance [6]. However, companies and the literature are still discussing how to structure and implement the digital twin for predictive maintenance [4]. Haghshenas et al. [7] applied DT to offshore wind farms, but the simulation developed in Unity was used to simulate the as-is scenarios and not to generate what-if scenarios to provide more data for the AI. With that, the AI used only historical data to perform the forecast, in addition, they applied the forecast to the sensors individually, not looking at the system as a whole.

This paper aims to propose a framework for how the development of a digital twin for predictive maintenance in an industrial conveyor system can be structured. The framework is part of a bigger project that involves the development of an AI-driven software framework that merges and utilizes heterogeneous technologies to identify a better trade-off between performance and energy consumption. In this way, we present and discuss how the framework was developed and the beginning of the implementation of this digital twin framework for a conveyor belt predictive maintenance system since this work is still in progress.

2 Related Work

Despite the few publications regarding the use of DT for conveyor belt predictive maintenance, some papers can be highlighted. Răileanu et al. [12] developed an architecture of an embedded DT for a shop-floor material integrated with the manufacturing, planning, scheduling, and control architecture where the DT monitors and forecasts the conveyor’s operating parameters like pallet traveling time. The focus was on predicting failures based on operational data collected from the conveyor system such as transportation time, current pallet, batch identification number, and timestamp, and not by sensors measuring the performance of each conveyor component such as conveyor speed, engine, conveyor tension, among others.

Bondon et al. [3] also presented an application of the learning and identifying phases of a LIVE Digital Twin for a roller conveyor system consisting of a motor, frame, belt, rollers, and two sensors. However, as presented by the authors, data analytics and the maintenance plan were still missing. Despite having presented a methodology to design a model-based Digital Twin for prognostic and diagnostic based on four stages of analysis, this methodology did not address the issue of how information flow was implemented, what software and hardware requirements, and especially regarding the connection of equipment and sensors in the Digital Twin system.

Mahmoodian et al. [10] present a novel DT architecture design for an intelligent civil infrastructure maintenance system to monitor an offshore jetty conveyor of a port terminal. The authors also highlight the need to monitor an offshore jetty conveyor of a port terminal and include simulating maintenance measures focusing on time and cost savings in asset management. In this architecture, the virtual twin is used to simulate, based on historical performance data, the current and future structural infra-structural condition, and the what-if scenarios are used after analyzing the calibration model just to evaluate the maintenance measures alternatives. Virtual Twin in this case was not used to generate data, through what-if scenarios, to improve the artificial intelligence models’ accuracy but was used as a tool to choose the best scenario for decision-making.

In this way, studies are being developed to use the Digital Twin to improve predictive maintenance, but there is still room for improvements and applications, as it is possible to identify a gap in the literature regarding the proposition of developing a DT framework for conveyor belts predictive maintenance in which it addresses the software, hardware, and communication requirements that evaluate the conveyor system components, as well as a way to connect both the Physical Twin and the Virtual Twin to a complete Digital Twin system. In addition, there is a lack of studies that use DT as a tool for generating new data about the system.

3 Framework

To identify how the DT for predictive maintenance of a conveyor system can be structured, this paper was based on the 5C architecture proposed by Lee et al. [9] and adopted by van Dinter, Tekinerdogan, and Catal [5], in which it presents the development of a DT-based to predictive maintenance based on 5 features, namely Connection, Conversion, Cyber, Cognition, and Configuration.

Connection is related to the Physical twin development, the Conversion deal with the system security, the Cyber consists of the digital twin development, the Cognition layer consists of the visualization feature and the Configuration is concerned with machine optimization. Thus, to develop the framework was structured into three topics, namely Data Connectivity and collection, PLC, and Sensors, and Virtual Twin, to ensure the feasibility of development and future implementation of this system. Subsequently, they were combined into this single framework presented in Fig. 1. Then, each topic is addressed in the following subsections.

Fig. 1
A flow diagram of a predictive maintenance framework. An operator sends fault generation to H M I which sends an input user to P L C. The physical system has a conveyor, linear actuator, I O link sensors and master, A C motor, I o T broker, and software. Cloud has a virtual system, A I, and storage.

Conveyor belts predictive maintenance framework

Full size image

3.1 Data Flow

Fig. 1 is divided into two parts, with the bottom rectangle representing the physical system and the top part standing for the cloud. The physical twin consists of the conveyor and the IO-Link sensors, as well as the AC motor and a linear actuator that can be controlled over Profinet and Serial communication respectively. Data will be collected from the IO-Link sensors and the frequency converter used for the AC motor. The HMI will serve as the user interface for the operator to interact with the equipment.

The gathered data is combined in a CSV (Comma-separated values) file and sent to the Cloud from the PLC using node-red. Here it is stored in a cloud-computing platform like Microsoft Azure and used to run the digital twin simulation. The saved data from the physical twin, as well as the newly generated data from the digital twin, is now made available to the AI, increasing the overall amount of available data for the training process Subsequent paragraphs, however, are indented.

Fig. 2
A photograph and a 3-D rendering of conveyor systems. Left. A photo of a conveyor system model on a rack, connected to a control system, in a room. Right. A 3-D model of the conveyor system with the control unit beside it.

Physical Twin (left picture) and Virtual Twin model (right picture)

Full size image

3.2 PLC and Sensors—Physical Twin

In the below written, we refer to the physical twin as [5]; being the physical system and the components herein. The predictive machine setup consists of a conveyor in a 1:2 gearing configuration driven by a 750-W 3-phase electrical AC motor, as depicted in Fig. 1 and Fig. 2. To this configuration, several sensors were added based on the parameters that influence the systems. These include:

  • A humidity and temperature sensor for the motor, bearings, and environment

  • An encoder for measuring the RPMs

  • A distance sensor for measuring the position of the items on the belt

  • A microphone to detect abnormal noises from the belt and motor

  • A vibration sensor to detect abnormal vibrations

  • A linear actuator to measure and apply force onto the belt

  • A variable frequency drive to log speed, voltage, current, frequency

  • A camera to perform image processing (object detection) and object placement

These sensor modules are connected to a Siemens ET200SP PLC, except for the camera, which is connected to a Xilinx board to preprocess the data before sending it into the cloud. The Siemens ET200SP was chosen as it consists of four-core CPUs in which one is dedicated to the machine operations while the three other CPUs operate the Windows operating system. This configuration makes it possible to create a PLC server that can be connected to, e.g., a client PC. The communication to and from the PLC happens through ethernet by the protocol Profinet or TCP/IP. This allows communicating quickly and safely with both Siemens and third-party equipment.

3.3 Data Connectivity and Collection—Cyber-Physical System

This section presents the connectivity and data transfer system consisting of LANs and wireless communication through TCP messages to send all sensor data to the cloud. The PLC is connected through LAN to the wireless hub, where the communication protocol is developed to transfer the data.

Continuous transfer of the data to the cloud is important to keep system memory usage lower and the performance efficiency better of the PLC. The TCP protocol is used because there are fewer chances of losing any part of data as our data. There is a possibility to also have redundant data while generating the data from sensors. To avoid sending redundant data, the preprocessing method is used to read the encoded data obtained from sensors and decode it to save the clean data in a CSV file for each day having 288 records (interval of 5 min) and then send it to the cloud server for further processing. Saving the data to CSV format help to deal with any network disturbance. In the preprocessing step, the data collected from all other sensors is directly read in PLC instead of the camera sensor. For running the object detection algorithm on the camera sensor, we used Xilinx board, which runs the real-time deep learning object detection model with 20 FPS (Frames per second). The transferred data is then used in digital twin to further assess each sensor’s performance and use it for machine learning models on the cloud-based servers.

3.4 Virtual Twin

Following the 5C framework, this section deals with the development of the digital twin. The virtual counterpart of the physical conveyor belt is executed as a digital shadow, with data flow only from the physical to the digital model [1], and the ex-tend model to a fully integrated bi-directional digital will be implemented with the AI output. The virtual twin is used to mirror the physical conveyor as well as generate new data using a simulation approach as presented in Fig. 2. Therefore, the model helps acquire large amounts of data for predictive maintenance purposes by reducing the necessary time for the physical model data collection. As the accuracy of the predictive maintenance algorithm is highly dependent on the simulated data from the digital twin, the accuracy of the virtual model plays a fundamental part in the success of the predictions. Therefore, it is necessary to simulate the entire conveyor system to explore the correlations between its components.

Based on the model parameters and the incoming sensor data from the physical twin, the digital twin simulates failures/ non-failures in the system. The simulation outputs an array of sensor data, along with the label of “failure/ non-failure” within the system, indicating whether the current parameters lead to a fault in the conveyor. This data can then be processed further for predictive maintenance purposes.

The digital twin will be implemented as an actual multi-domain model that is described by mathematical equations. The conveyor will be modeled using OpenModelica as a modeling language and using predefined components to describe the system and the relationships between its individual parts. After modeling the entire system, the input variables can be configured as needed using the data gathered from the physical twin. Based on this data the digital twin can then extrapolate new data by solving the equations describing the model. This approach will let the digital twin be able to answer questions like What will happen to the conveyor belt temperature if the belt is running at a high speed with low tension?

4 Discussion and Future Work

This paper presents a framework for developing a digital twin for generating data based on a simulation model using the collected data from a physical conveyor system. The newly generated data will then be used as input for a predictive maintenance Machine Learning algorithm to improve the model’s accuracy.

For future work, we suggest using the model predictions to automatically adjust the physical system parameters, therefore closing the loop for a bi-directional digital twin model.