1 Introduction

1.1 Motivation

Reaching and maintaining companies’ strategic goals is always of highest priority to managers and can be monitored and managed through key performance indicators (KPIs). KPIs offer effective measurements to translate business models into well-defined and comparable parameters [1]. Nevertheless, the selection of suitable KPIs remains a complex problem [2], in particular regarding interrelations between certain indicators and the need to cover all levels sufficiently, from machinery [3], production systems [4] to entire companies [1]. In the field of logistics, production and operations management, a set of 34 important KPIs is formulated in form of the ISO-22400-2:2014 standard [5], was conceived. However, the interrelations within such KPI systems are hardly understood [6]. To mitigate this lack of understanding, a hierarchical approach of a KPI graph model recently became presented by Kang et al. [4] and extended by Stricker et al. [2]. This permits further understandability, but also leads to imperfect shopfloor optimization measures in the case of missing KPIs [7], due to sensor failure, manual intervention [8] or overall missing data acquisition [9]. In addition to a hierarchical approach, graphical KPI models help to address a better and precise understanding [10]. However, the challenge of accurately reflecting production system behavior, given missing or additional relevant KPIs, still prevail. Brundage et al. additionally mention the need to take special care of the KPIs’ dependencies [11]. Against this background, this paper proposes a novel, value optimization based method to make use of the interrelations between KPIs in a KPI system to reconstruct incomplete KPI systems stemming from sensor failure, manual interventions or missing KPIs. This is extend by a random forest to initialize the value optimization and improve convergence, as shown in the case study.

Fig. 1
figure 1

Exemplary KPI set with missing (red) and relevant (blue) KPIs

The problem can be exemplified with Fig. 1, where a few KPIs are missing (red), due to data inconsistencies, sensor failures, incomplete KPI capturing and other reasons. KPIs that neighbor missing KPIs (blue), i.e. that are relevant for its calculation or vice versa, should remain at their known value while the red KPI values should be identified in order to fully reconstruct the incomplete KPI system.

1.2 Research questions and main contribution

In this paper the following research questions are addressed:

  1. 1.

    How can an incomplete KPI system be accurately completed, so that both dependencies and mutual influences are respected during the completion?

  2. 2.

    How can knowledge about a complex production system and its previous behavior be used to properly facilitate the KPI system completion?

2 Literature review

Defining KPIs and KPI systems is a common topic in literature, where KPIs describe a set of parameters [11] and KPI systems describe a set of KPIs with inherent relationships [2]. The latter can have a holistic structure [12] and is often used to improve the robustness of entire production systems. Other KPI systems use a hierarchical structure [4] and focus on the continuous improvement and management of smart manufacturing systems.

First, KPI systems are often used to observe and manage production, ranging from single processes [13] and machines to large-scale systems [12]. Hence, a major problem is the selection of the KPIs such that they not only monitor low-level processes [14] but are also suitable for overall performance measurements [15]. Nevertheless, such graph-based systems are rarely regarded holistically [10].

Secondly, the derivation of KPIs and their relationships in KPI systems commonly follow a mathematical approach. Most commonly, matrices [16], pairwise relations [11] or the direct incorporation of standardized ISO 2240 [5] are used. Alternatively, systematic relations are conveyed in graph-based approaches, e.g. Stricker et al. [12] use nodes and links to express the interconnections within KPI sets. This brings along the advantage of a clear relationship definition [17].

Table 1 Comparison of different KPI relationship descriptions

Given the complex interrelations between KPIs [4] it is important to make the hidden mathematical interrelations between KPIs understandable by a dimensional analysis [19]. For a similar goal, but focused on sustainable manufacturing, Amrina et al. [21] use structural modeling for KPI relationship analysis. In order to incorporate the different information content of KPIs under different circumstances Stricker et al. [2] select KPIs through a linear programming approach. Approaches that focus on describing KPI system interrelations are presented in Table 1.

In spite of optimizing the production, the relevance of predicting future or unknown values of KPIs is of particular importance [25]. While the application of Bayesian networks (BN) [26] to obtain optimal KPI values and predict their development or a locally weighed regression for KPI prognosis [27] recently gained interest, the majority, i.e. Zhang et al. [28], focus on an analysis of static and dynamic relationships between process parameters and KPIs. Such approaches are enhanced by factorization [29] for fault prognosis or Modified Kernel Ridge Regression [30] for process parameter decomposition. Thus, the integration of Machine Learning approaches, such as Wu et al. [31], who apply Hidden Markov Models for machine deterioration estimation and remaining useful life prediction based on multiple KPIs, show promising results. A holistic overview is shown in Table 2.

A third body of knowledge is focused on monitoring and improving the performance of production systems by regarding its corresponding KPIs. For instance, Cao et al. [26] and Kornas et al. [22] add a feedback obtained by bottleneck analysis to overcome the production system’s weakness. Furthermore, Tahir, Mahmoodpour, and Lobov [18] use an extensible visible language to allow decision makers to identify improvement potential within the defined KPI system.

Table 2 Comparison of different KPI prediction approaches

The application of KPI systems to accurately steer production systems has still not been widely accepted within industry [4]. One major drawback of KPI systems, however, is their instability towards missing, tampered or inaccurate signals. Thus, a research gap in filling missing information within KPI Systems arises, and, additionally, the integration as well as usage of previous knowledge about KPI systems is not yet regarded holistically [6].

3 Selected KPI system

The paper considers the KPI system presented by Kang et al. [4], as it encompasses all KPIs mentionend in the ISO standard [5] and additionally is the largest regarded holistic KPI set. It is widely known [2] and for reference the KPI list is introduced in Table 5 in the appendix. Nevertheless, certain corrections, with respect to the KPI’s calculations as presented by Kang et al. [4] have to be made:

  1. 1.

    \(CMR = CMT/PMT\) is corrected to

    \(CMR = CMT / (CMT+PMT)\) according to the definition of CMR: ”ratio of the total corrective maintenance time CMT to the sum of CMT and the preventive maintenance time PMT” [4].

  2. 2.

    Some KPIs such as \(AUIT=ADET\), \(TTF=TBF\) or \(MTTF=MTBF\) have the same definition and mathematical calculation expressions in the regarded production system. In order to keep the prediction system stable and avoid unnecessary problems by formulating the optimization problem, only \(ADET,\, TBF,\, MTBF\) are kept in the KPI system.

The following requirements for missing KPI sets can derived from the literature: According to ISO-222400 [5] there are 34 relevant manufacturing KPIs, divided hierarchically into 3 groups, which we distinguish: low level KPIs (directly measured from production lines), basic KPIs (calculated from low level KPIs) and comprehensive KPIs (calculated by using basic and low level KPIs). Kang et al. [4] and Stricker et al. [2] have illustrated the interrelationships between each KPI set on all levels and extended the set to 59 interrelated KPIs. KPIs can be easily calculated by means of the formal definition from ISO-22400 [5], Stricker et al. [2] and Kang et al. [4]. Thus, if only few KPI values are missing and these are mathematically independent from one another, the same calculation logic can be applied. Hence, there is no need to use the developed comprehensive combination of a random forest and subsequent value optimization under low dimension scales. However, the more complex the interaction between missing KPIs and the larger the set and relationships among these is, the more difficult it is to find the optimal solution. In particular, this counts for KPIs that result in equations with more than one unknown impede convergence. For instance, if CMR is known and the given relationship is: \(CMR = CMT/PMT\), there are infinitely many possible solutions to satisfy this one condition. Thus, KPI sets with few missing KPIs can be solved easily and, therefore, in this case study there are only KPI sets with more than ten missing KPIs and a graph based distance larger than one between two missing KPIs considered. Otherwise, solely the less accurate random forest prediction approach can be used, which means that no subsequent value optimization is performed.

Furthermore, the KPIs to be predicted are randomly selected from the whole KPI system according to ISO-22400 [5] and the extensions from both Kang et al. [4] and Stricker et al. [2], obeying the selection rules mentioned before: mathematically independent from each other and graphically distanced at least two edges, which are easily identifiable through the corresponding adjacency matrix.

4 Methods

The ability to find the values of missing KPIs within a KPI system is targeted by using mathematical relationships between the nodes of the corresponding KPI graph. Hence, the core of the approach is twofold: (1) predicting missing KPIs based on previous observations and (2) defining an objective function that minimizes the error between contradicting KPI definitions which is optimized to obtain the value of missing KPIs. The initial values for the KPI value optimization can be obtained either by a random initialization (referred as cold start) or by a random forest based KPI prediction (referred to as warm start). Random forests are selected as bagging and bootstrapping allow successful handling of high dimensional data with low bias and moderate variance as well their robustness towards outliers and nonlinearity [35]. The applied methods are presented in detail in the following.

4.1 Optimization

Value optimization is based on the application of mathematical search to find the convergence point of an objective function. It offers significant efficiency on convex functions and can be divided into constrained and unconstrained optimization. In this paper the performance and suitability to solve the described KPI optimization problem with the following two different optimization methods is investigated: unconstrained optimization and constrained optimization, i.e. constraining the solution space to meaningful KPI values. Due to their superior results in application the following algorithms are applied: Nelder-Mead as an unconstrained algorithm [36, 37] and L-BFGS-B (Limited-memory (L-) Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm with bound constraints (-B)) for the constrained optimization [38].

The problem of finding the missing value of a KPI that is defined by given and known KPIs is trivial by application of the describing functions. However, in a KPI system if more than one KPI is missing, a KPI value is defined twofold:

  1. 1.

    The numerical, known (i.e. measured) value of the KPI i (denoted as \(A_i\)), and

  2. 2.

    the value obtained by the interrelationships with it’s defining neighboring KPIs x and their values: \(v_i = v_i (x)\), where x is the vector of relevant neighboring KPIs, excluding i itself.

Thus, finding the true value of an unknown KPI i is similar to ensuring that the known or estimated value \(A_i\) equals the defined value \(v_i\). Since several KPIs must be missing to regard non trivial cases, the following total objective function can be set up by using a quadratic derivation punishment:

$$\begin{aligned} \hbox {min}\ f(\cdot )_{i}= \sum (\frac{A_{i}-v_{i}(x)}{A_{i}} )^{2} \end{aligned}$$

where \(f(\cdot )\) signifies the total objective function, which is accumulated from each optimization expression of KPIs that are relevant for finding the missing KPIs. Hence, i refers to the index of the KPI to be optimized for and A is the value of the KPI that is either known (A remains constant) or unknown (A converges over time). The parameter v(x) is the new value of the KPI in the current optimization step, which usually is used to improve A within each iteration.

Based on the newly defined objective function f, the applicable solution space is subject to the constraints \(g_i \, \text {and}\, h_j\). Given that \(f, g_i \, \text {and}\, h_j\) are real-valued functions on X including at least one nonlinear term, a nonlinear value optimization problem is formulated as:

$$\begin{aligned} \begin{array}{rl} min &{} \; f(X)\; \\ s.t. &{} \; g_{i}(X)\le 0 \,\forall i \in (1,2,3,...,m), \\ &{} \;h_{j}(X)= 0 \,\forall j \in (1,2,3,...,p),\\ &{} \;where \; X = \left[ x_{1},x_{2},..., x_{n} \right] ^{T} \in R^{n}. \end{array} \end{aligned}$$
(1)

The variable \(x_{i} \in R\) refers to measurable variables, i.e. KPIs, with n missing KPIs. Equality constraints (\(h_{j}\)) and inequality constraints (\(g_{i}\)) can be used to integrate further domain or system specific knowledge about the KPI system. For instance are upper bounds (\(A_i \le 100\%\)) for an individual KPI i implemented through the inequality constraint \(g_i(x)=v_i(x) \le 100\%\). In application the problem in general appears not only once but whenever the KPI system was recorded, so that each individual KPI system solution according to the above describe value optimization can be regarded as a sample of the overall problem.

The optimization problem, when considering known bounds of the missing values, can efficiently be solved by the application of L-BFGS-B, which operates on an approximated Hessian matrix and belongs to the field of Newton methods. Thus, in cases with vastly differing boundaries, the optimization can stop immediately as the second order partial derivatives might be, falsely, considered to be equal to zero. Without additional boundaries, the optimization is solved through a more stable Nelder-Mead solver, implemented in Python by [39].

4.2 Optimization initialization

Solving the above described optimization problem with constrained or unconstrained value optimization requires feasible initial values for the missing KPIs. Better initialization values decrease the number of value iterations and, hence, can significantly improve the speed of convergence. To that end, random initialization values for unknown KPIs can be used, i.e. standard implementations, or alternatively knowledge acquired by observation of previous KPI values can be used. The former is regarded as the base case, while the latter can be approached by taking the known or optimized KPI values of KPI system at time t as the initialization value of the succesor at time \(t+1\). Due to the sensitivity of minor derivations between the values of each KPI at time t and \(t+1\), this simple, time invariant approach approach is rejected. However, the dynamics of KPI systems and the individual KPI values could be estimated based on a large observed time frame, to predict accurate, well-performing initialization values.

4.3 Random forest for initialization improvement

A random forest [35] can be used as a method for classification and regression problems. Random forests do bagging on decision trees [40], i.e. by creating many random regression trees. These types of decision trees, with leaves representing numeric output values, search for a function \(H(X) : X_i \longrightarrow y_i\), that maps from a given KPI system \(X_i\) excluding the values \(y_i\) to the missing values \(y_i\). (\(X_i\),\(y_i\)) is referred to as a sample and denotes the incomplete KPI system at a certain time. Within the bagging process, random forests introduce two sources of randomness for each regression tree: First, a bootstrap sample (i.e. sampling with replacement from the training data set) is used to construct each individual regression tree. Second, during splitting a node when constructing the tree, the best split is randomly drawn from all input feature (here \(X_i\)) or a random subset. The size of the random subset can be controlled, yet, in the regarded case equal to 90% of the input feature size.

Random forests use bagging and, as outlined, is a simple machine learning algorithm with few parameters, (partial) interpretability, good generalizability, non-linearity [35] that is well established [40]. For the subsequent validation, the random forest implementation of sci-kit learn [41] is used. Note, that in contrast to original work, which averages on each classifier vote or regression [35], their average probabilistic prediction, or regression, is used [41].

In this paper a random forest is applied as a warm start tool for the optimization based prediction of certain missing KPIs delivering start values in a very likely region. In contrast to simple cold start optimization, where the initial values are randomly selected or based on previous values the idea is to boost the optimization by selecting more likely well performing initial optimization values. Random Forest algorithms have many advantageous properties: Training speed is quick, even with high dimensional attributes, results are acceptable concerning stability and, thus, it is not required to frequently retrain from scratch, while storing is also easy to perform. To compare the performance, a cold start refers to a random initial value generation. Thus, the searching space in the cold start is larger assuming a sufficiently well-performing random forest prediction. Using the warm start method, the optimizations obtains certain benefits:

  • The value overflow errors exhibited by Nelder-Mead can be widely reduced.

  • The value optimization process can be significantly shortened due to a faster convergence resulting in a prediction system with quicker responses.

  • Contribution to the stability by high dimension scale KPI prediction model.

  • Partially restrain the local minimum problem.

5 Results

In order to compare the proposed methods, four control test are performed. The performance and the number of iteration steps of the value optimization process are compared regarding the following ceteris paribus tests:

  • Test 1: Warm Start vs. Cold Start with the same number of prediction KPIs

  • Test 2: Warm Start vs. Cold Start with the same number of samples

  • Test 3: Nelder-Mead vs. L-BFGS-B with the same number of prediction KPIs

  • Test 4: Nelder-Mead vs. L-BFGS-B with the same number of samples

Each iteration number in the case study refers to the average result from 20 identical tests with random missing KPI set selections, where the number of KPIs to be predicted, i.e. prediction KPIs, is controlled. In addition, a performance check of the objective function under the influences of the number of prediction KPIs and samples is carried out. This targets the performance of the optimization methods under a high dimension scale and with plenty of data samples. For all of these investigations large KPI data sets based on random realizations of various production systems are used. At last, the performance of the prediction methods is validated on simulation-based data. For this purpose, simulated KPI sets of a real production system are provided by running an ontology based discrete event simulation [42] initialized with a real-world production system data set. Given the incompleteness, random noise and time variant KPI tracking and calculation in reality, applying the OntologySim is beneficial as the KPI sets introduced by both [5] and [4] are tracked [42].

5.1 Initialization with warm start (Tests 1 & 2)

In this section the influence of the initialization methods (i.e. warm start and cold start) on the performance of KPI predicting system, with respect to the number of samples and prediction KPIs, is discussed.

Figure 2a illustrates, that the parameters iteration steps and number of samples show a nearly linear relationship using both cold start and warm start. Note, that the number of iteration steps is used to compare performance, as time measurements depend greatly on the used hardware. Clearly, the warm start offers a significant advantage in terms of average required iteration steps compared to the cold start. It is obvious that with an increasing number of missing KPIs, both optimization procedures require more iteration steps. Nevertheless, the number of needed iterations till convergence increases slower using the warm start. However, this speed up in optimization only becomes significant dealing with more than 12 missing KPIs, as shown in Fig. 2b.

Overall, the warm start delivers a better performance than the cold start on both, iteration steps and convergence time respectively.

Fig. 2
figure 2

Comparison of the performance between Warm Start and Cold Start in iteration steps for different numbers of missing KPIs and samples, showing the improvements in runtime achieved through warm start

5.2 Comparison of the value optimization algorithms (Tests 3 & 4)

In this section, the performance of the two value optimization algorithms (Nelder-Mead and L-BFGS-B) is compared. Figure 3a shows again a nearly linear relationship between the required iteration steps and the number of samples. However, the L-BFGS-B needs significantly less iteration step to converge towards a solution. Hence, the quasi newton method L-BFGS-B provides a more efficient optimization algorithm for this application. Additionally, this accentuates the worth of including domain knowledge, i.e. reasonable bounds for the unknown KPIs.

Fig. 3
figure 3

Comparison of the performance between Nelder-Mead and L-BFGS-B

Regarding the number of predicted KPIs, the L-BFGS-B still delivers the better results compared to Nelder-Mead, as demonstrated in Fig. 3b. This performance gap widens for higher numbers of prediction KPIs. However, Nelder-Mead has a relative stable performance when the number of prediction KPIs increases using the improved high dimension adaptive function mentioned in [37].

In addition to the two above visible effects, the iteration steps also vary due to the settings of the objective function tolerance: the smaller the tolerance of the iteration process is, the more calculation steps are needed.

With respect to the value optimization process we recall that each number of missing KPI trials consists of 20 tests with different KPI sets. Figure 4 illustrates, how many times the value optimization method successfully reaches the convergence point.

Nelder-Mead shows a very stable optimization process, when the dimension scale of the objective function, i.e. the number of missing KPIs, is under 13. Beyond that the optimization process easily becomes unstable, and in some cases Nelder-Mead even failed to find the optimal solution before the process reached the termination criterion. On the other hand L-BFGS-B less often provides a stable result even if the dimension scale of the objective function is at a low level. The reason might be an invalid Hessian matrix during the value optimization process, which terminates the process improperly. Another typical optimization problem occurs in cases of (partly) non-convex objective functions.

Fig. 4
figure 4

The number of valid value optimization tests

5.3 Performance regarding the objective function value

In this section the comprehensive performance of the KPI prediction system is analyzed, considering the influence of the number of samples and KPIs that are predicted. In Fig. 5, which show the normalized value of the objective function for an exemplary study of the influence of sample size and missing KPI number, there are two trends clearly visible:

  1. 1.

    The value of the objective function is nearly independent from the number of samples. Once the number of prediction KPIs is set, the value of objective function keeps almost the same level despite some small oscillations, as the number of samples is (almost) irrelevant.

  2. 2.

    The number of prediction KPIs affects the accuracy of the KPI prediction system. If the number of prediction KPIs increases, the value of the objective function rises significantly. However, the value oscillation is not dependent on the number of missing KPIs. Otherwise, the number of samples has a relative small contribution to the oscillation on the value of objective function or the accuracy of the KPI prediction system.

Fig. 5
figure 5

Influence of sample number and missing KPI number on function value

5.4 Individual KPI performance comparison

In the following, a comparison of the predictability, i.e. the deviation from the original value for a missing KPI, is presented. For this experiment, 500 randomly selected sets of missing KPIs per missing number, each consisting out of 20 samples, are optimized. The number of missing KPIs is varied to identify the effects of larger numbers of missing effects on the prediction quality. The resulting deviation is reported in a \(log_{10}\) scale, which is illustrated in Fig. 6. Apparently, there are certain KPIs that are prone to deviate largely, especially the earlier mentioned problematic KPIs, as well as ratios. These KPIs can easily result in sub-optimal points satisfying a reasonably low optimization function. With a higher number of missing KPIs, in general, the error remains similarly small or increases, as expected. The overall performance is consistently high, even for larger numbers of missing KPIs. Thus, the proposed method can be effectively applied to complement missing KPIs in KPI systems that are standardized and widely applied in industry.

Fig. 6
figure 6

Deviation between optimized and true values in \(log_{10}\) per individual KPI depending on number of missing KPIs

5.5 Validation on industrial use-case simulation data

Since the combination of random forest initialization and subsequent optimization shows significantly superior results, the validation focuses on the warm start method comparing the two different optimization algorithms. The KPI data sets are generated using a multi-machine simulation of a real production system implemented with OntologySim [42]. The comparison itself targets the prediction accuracy and the convergence speed per iteration. Concerning the optimization result, the mean squared relative error (MSRE) is regarded, where every single KPI of the presented KPI system is included in 100 missing KPI sets varying the prediction number. Tables 3 and  4 show the results of the iteration speed comparison and the accuracy respectively.

Table 3 Speed comparison of both optimization methods w.r.t. the number of missing KPIs (n)
Table 4 Accuracy comparison of both optimization algorithms w.r.t. the number of missing KPIs (n)

In both categories the Nelder-Mead algorithm reveals its strengths compared to the L-BFGS-B. Despite a significantly larger number of iteration steps, the Nelder-Mead predicts the KPI-system faster. For 20 missing KPIs it even converges more than four times faster than the L-BFGS-B. L-BFGS-B may have got the more efficient algorithm but also is way more complex to calculate, resulting in a slower optimization convergence. The comparison of the accuracy demonstrates the second advantage of using Nelder-Mead. The maximum MSRE stays close to zero even up to 15 missing KPIs. Considering high scaled problems with 20 missing KPIs, the Nelder-Mead still predicts 54 KPIs with a MSRE below one percent. The L-BFGS-B on the other hand cannot compete with this performance.Moreover, the stability of Nelder-Mead is higher, yet it lacks the ability of L-BFGS-B to incorporate boundary values for KPIs.

Besides the two presented comparisons, the effects pointed out in previous sections are also visible during the validation with real simulation data. This especially includes the stability and the needed amount of iterations steps per algorithm. Table ?? summarizes the validation of both optimization strategies. Overall, the presented prediction technique allows to reconstruct almost half of KPI system, which is a remarkable performance.

6 Conclusion and outlook

6.1 Conclusion

This paper presents a KPI system prediction model that can be used with KPI systems that are calculated based on a graphical approach as introduced by Kang et al. [4] for missing data in a manufacturing environment. Given the mathematical relationships of KPIs, a random forest can be trained to initialize the value optimization, dubbed warm start. A subsequent optimization predicts the missing KPI values, handling different sizes of missing KPI sets. The proposed method performs suitably well and allows the identification of KPI values that are only indirectly observed in the form of its neighboring KPIs in the KPI system. The method is robust when increasing the sample size and applying different optimizers, however, best results are obtained in the warm start scenario with Nelder-Mead. The presented approach is validated based on simulation data of a real production system and is able to accurately predict almost half of the comprehensive KPI structure.

6.2 Outlook

In future work, it is possible to extend the method by integrating a very small \(\epsilon > 0\) in the prediction formula for KPIs in order to improve the accuracy of the prediction and in rare cases avoid an invalid Hessian matrix. For dealing with the stochastical influences in real-world applications, where sensors or gathered KPIs might not perfectly fit to one another, this can alleviate the shortcoming of hardly being able to reduce the actual prediction error to zero. Furthermore, the integration of knowledge about the previous behavior of the KPI system can be extended, as the training model of the random forest can become more comprehensive and abundant to provide the value initialization with a better start point. The internal interdependence within the KPI system follows the patterns effectively used by the random forest. In future, the interdependence within a KPI system should be further evaluated. Additionally, beyond random forests alternative predictors can be evaluated, however, the core of the approach lies not in the predictor application. Also, it can be considered to effectively use L-BFGS-B, as this should avoid large value ranges of missing KPIs, that could help in calculating the Hessian matrix significantly better and faster. However, a more comprehensive approach requires the transition from this simulation based-data gathering approach towards a real world application. Further research shall also focus on the actual effects of implementing a more detailed KPI system that can be completed in the case of missing KPIs. Managerial implications and value adding tasks have to be rigorously analyzed. Nonetheless, this research can enable a suitable comparison of different products according to their, then completed, KPI representation. Including more product-based KPIs for multi-product production systems can amplify the methods real-world applicability.