Cumulative and Rolling Horizon Prediction of Overall Equipment Effectiveness (OEE) with Machine Learning

Abstract

Nowadays, one of the important and indispensable conditions for the effectiveness and competitiveness of industrial companies is the high efficiency of manufacturing and assembly. These enterprises based on different methods and tools systematically monitor their efficiency metrics with Key Performance Indicators (KPIs). One of these most frequently used metrics is Overall Equipment Effectiveness (OEE), the product of availability, performance and quality. In addition to monitoring, it is also necessary to predict efficiency, which can be implemented with the support of machine learning techniques. This paper presents and compares several supervised machine learning techniques amongst other polynomial regression, lasso regression, ridge regression and gradient boost regression. The aim of this article is to determine the best estimation method for semiautomatic assembly line and large batch size. The case study presented with a real industrial example gives the answer as to which of the cumulative or rolling horizon prediction methods is more accurate.

1. Introduction

Nowadays, various forecasting tools and techniques are playing an increasingly important role in industrial manufacturing companies in order to fulfill customer orders on time. In addition to traditional estimation methods, such as various trends, simulations of more effective techniques supported by machine learning have also appeared [1,2,3,4]. This is also true for Overall Equipment Effectiveness (OEE), the most frequently used efficiency Key Performance Indicator (KPI) in the domain of assembly operations [5]. The prediction of manufacturing and assembly efficiency is relevant, among others, in the fields of production planning, scheduling, investments and management decisions [6,7]. Accurate forecasting generates profit, reliability and competitive advantage for industrial enterprises.

Due to the development of information technology, many systems provide support for the recording, processing and storage of production-related data. The most frequently applied systems are Manufacturing Execution System (MES), Enterprise Resource Planning (ERP) and Customer Relationship Management (CRM) [8,9]. Currently, it is a constant challenge to reveal the patterns and relationships behind the data, which is made more difficult by the turbulent industrial environment and changes in product variances [10,11]. With the help of machine learning, it is possible to determine the expected and searched values faster, more efficiently and more accurately, including the OEE percentages. Due to the huge amount of data, machine learning is one of the best methods for processing production data.

Overall Equipment Effectiveness is measured by the machines’ capability of performing a task to produce a finished product as per customer needs in a timely manner. The prediction of OEE can also be conducted in the case of machine learning in several ways, even by predicting each component (availability, performance and quality) separately. All three types of machine learning, such as supervised learning, unsupervised learning and reinforcements learning, can be used for forecasting in the field of production [12,13,14].

The aim of this paper is to determine the best estimation method for semiautomated assembly line and large lot size by finding out whether cumulative or rolling horizon forecasting is more accurate. In recent years, this has not been analyzed in detail, so several machine learning methods are examined and compared in this article. Considering that there are many machine learning approaches, this paper does not examine all of them but, rather, tries to outline a general state.

2. Materials and Methods

A production system and, within it, the efficiency of assembly operations can be evaluated in many ways [15]. The most common method in automotive practice is to use Key Performance Indicators (KPIs) [16]. Overall Equipment Effectiveness (OEE) as a standard and best practice indicator was introduced within the Total Productive Maintenance concept by Nakajima [17]. The original formula for calculation of OEE is written as:

OEE = a p q

(1)

where:

• a—availability (%);
• p—performance (%);
• q—quality (%).

At the domain of press-hardening process in manufacturing, Lejon et al. used three machine learning methods, such as autoencoder neural network (ANN), one-class support vector machine (OCSVM) and isolation forest (IF), for anomaly detection. The ANN method was the best performing candidate in a similar dataset based on precision, recall and accuracy [18]. Fast angle-based outlier detection (FABOD) and K-nearest neighbor (KNN) methods performed better compared to histogram-based outlier score (HBOD), local outlier factor (LOF), isolation forest (IF) and one-class support vector machine (OCSVM) techniques when examining and detecting anomalies occurring on the assembly lines [19].

A quality factor affecting OEE was investigated by Peres et al. with machine learning classifiers. In the frame of multistage quality control, the following algorithms were examined: Gaussian naïve Bayes, K-nearest neighbor, XGBoost, random forest, support vector machine and logistic regression. After the evaluation of accuracy, recall, precision and F1 score, the XGBoost tuned model performed the best [20].

A hybrid prediction model was proposed to estimate whether the automotive assembly process is functioning normally or abnormally. The conception utilizes an outlier detection based on density-based spatial clustering of applications with noise and random forest classification model. The presented model achieved higher accuracy than other examined models, such as naïve Bayes, logistic regression and multilayer perceptron [21].

Wang et al. applied support vector machines algorithm to estimate the quality of welding in a high-power disk layer. The proposed quality control system worked in real-time mode [22]. Regarding the quality factor, Lee et al. compared four prediction methods as decision tree, random forest, artificial neural network and support vector machine at the area of metal casting. The best accuracy was at the ANN with 93.84% [23].

Predicting the OEE value at a production line with six machines, decision tree regression algorithm was the more robust and had the best result in terms of mean square error than K-nearest neighbors, support vector machine and artificial neural network [24]. El Mazgualdi et al. presented the use of various machine learning algorithms under different configuration to predict OEE of an automotive wiring factory. Support vector regression (SVR), support vector regression cross-validation (SVRCV), support vector regression genetic algorithm (SVRGA), random forest (RF), random forest cross-validation (RFCV), extreme gradient boost (XGB), extreme gradient boost cross-validation (XGBCV) and deep learning (DL) were evaluated based on mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE). It was concluded that neither method was effective enough due to the small size of training data [25].

In terms of transmission assembly line predicting OEE, using Bayesian ridge regression shows higher accuracy (99%) than other machine learning methods among others such as adaptive naïve Bayes-based algorithm (96%), logistic regression model (84%), support vector machine (97%) and decision tree (89%) [26,27]. However, it is worth using more detailed minimum daily or shift-level basic data than the monthly OEE percentages used here. Martinez et al. attempted to estimate OEE by separately predicting availability, performance and quality. A new approach was taken in which they tried to combine the predictions of completely different machine learning algorithms. It was a new ensemble which combines the best of each method, is able to operate in all the cases, and reduces the error [28].

Khdoudi et al. compared four machine learning methods to predict welding process parameters. It was concluded that the convolutional neural network (CNN) was the most accurate for energy prediction, the support vector regression (SVR) model for amplitude prediction, and the regression model for pressure prediction. The best prediction method was a combination of machine learning techniques [29,30].

A deep reinforcement learning framework was proposed using three methods as deep Q-network, proximal policy optimization and advantage actor–critic algorithms for ensuring the product quality and minimizing the overall energy consumption of an industrial glass manufacturing process [31]. In the area of quality forecasting, 14 algorithms were compared (ridge regression, linear regression, light gradient boosting machine, lasso regression, random forest regressor, artificial neural networks, gradient boosting regressor, extra trees regressor, elastic net, Bayesian ridge, K neighbors regressor, AdaBoost regressor, least angle regression and orthogonal matching pursuit); the ridge regression algorithm presented the best overall predictive performance for the test examples [32].

Zouhri et al. used polynomial, sigmoid and (radial basic function (RBF) kernels as genetic-based SVM for chemical and rolling process quality data classification. The RBF kernel function was the recommended for classification with 87.15% (chemical data) and 99.08% (rolling process data) accuracy [33].

Based on the above-mentioned scientific literature, it can be seen that many methods can be used to predict OEE percentages; however, rolling horizon and the application of the cumulative method are not mentioned.

Selection of Independent Variables for Machine Learning

In this article, the authors examine the estimation of OEE from a prediction point of view, so this is the dependent variable. The basis for the selection of the dependent variables is a cause-and-effect diagram published in a previous article [34]. The factors affecting OEE are analyzed according to the following six aspects: man, environment, method, material, machine and measurement. A total of 150 influencing factors were revealed and described. Further examining the independent factors,

Table 1. Based on Ishikawa diagram, an excerpt of OEE contributors and measurability of attributes.

Attribution		Can Be Measured by MES, ERP, SQL and Log Files	Example	A	P	Q
Qualification	Practical experience	Yes	Variable cycle time per station per person		x
Motivation	Goals	Yes	Availability, performance, quality, OEE target		x
Organization	Improvement	Yes	Trends (OEE, scrap, etc.)		x
Production technology	Assembly process	Yes	Manual or nonmanual (automated) assembly		x
Measurement, control	Maintenance	Yes	Downtime, reason: maintenance	x
Work process	Process parameters	Yes	Time and duration data		x
Material and information flow	Available workforce	Yes	Staff (operator, setter, etc.)	x
Material failure	Material quality	Yes	Downtime, reason: quality problems			x
Material handling	Not available, not accessible	Yes	Downtime, reason: logistics problem	x
Machine and tool adjustment	SMED, OTED	Yes	Downtime, reason: changeover	x
Product control	Sampling frequency	Yes	Number of checked products			x
Checking of assembly process	SPC	Yes	In MES: SPC report			x

The elements of the complete Ishikawa diagram and effects on OEE are presented in Table A1, Table A2, Table A3, Table A4, Table A5 and Table A6.

as an excerpt shows which can be measured and recorded with MES, ERP, SQL or log files. In addition to all this, it shows an example and indicates the effect of the factors on availability, performance and quality. Appendix A contains more details.

Based on the measurability and occurrence of the factor characteristics of assembly lines, the following 12 independent variables were selected: process failure downtime, break downtime, technical downtime, changeover downtime, quality reason downtime, logistics reason downtime, not planned downtime, other downtime reason, number of changeover, average cycle time, number of assembled units and number of scrap units. OEE percentage, availability percentage, performance percentage and quality percentage are considered as dependent variables. These variables will be used in the next chapters. The basis for selecting the independent variables was that they should be characteristic of each assembly line, be objectively measurable and occur in large quantities. In addition, it was important not to select too many independent variables.

3. OEE Prediction with Machine Learning

After presenting the real work environment and the data used, this chapter examines the possibilities of OEE prediction using different machine learning methods. The authors are aware that there are many machine learning techniques; however, the described elements give a complete picture of the prediction processes.

3.1. The Real Work Environment of Applied Machine Learning

This article illustrates the presented methods through a real industrial example. The selected semiautomatic assembly line is a metal seat structure assembly line for the automotive industry located in Central Europe. Figure 1 shows the simplified layout of the hybrid line.

The assembly line consists of 16 work stations named with letters from A to P. The blue arrow shows the direction of material flow. The cycle time of each station is marked with different colors. Assembly operations are performed by human or machines according to

Table 2. Human and machine operations at the hybrid assembly line.

Assembly Station	Operation by	Assembly Station	Operation by
A	human	I	human
B	human	J	machine
C	human	K	human
D	machine	L	human
E	machine	M	machine
F	human	N	machine
G	human	O	human
H	human	P	human

.

This production unit has been operating for more than 10 years and, in two shifts, produces generally 380–440 products per shift. Since the beginning of the production, the assembly line has produced the same main product; there is no significant difference between the product variations.

In order to accurately describe the production environment, it is first necessary to calculate and determine the type of production process based on Equation (2) and

Table 3. Production types and rates.

Types of Production Process	Production Rate
Mass production	0.8 < Tf ≤ 1.0
Large batch size	0.6 < Tf ≤ 0.8
Medium batch size	0.4 < Tf ≤ 0.6
Small batch size	0.2 < Tf ≤ 0.4
Job-shop or project production	0 ≤ Tf < ≤ 0.2

[35].

$${\mathrm{T}}_{\mathrm{f}}=\frac{\mathrm{Q}\mathrm{T}}{{\mathrm{I}}_{\mathrm{p}\mathrm{r}}}$$

(2)

where:

• T_f—production rate;
• Q—annual volume of the production task (unit/year);
• T—working time requirement of the production task assigned to the production unit (hour/unit);
• I_pr—actually available productive time base (hour/year).

Based on these, the analyzed hybrid assembly line with T_f values (0.76, year 2021; 0.78, year 2022) can be classified as large batch class.

3.2. Production Data for Machine Learning

All the data related to the semiautomatic assembly line mentioned in the previous subsection are available with the support of the factory MES and the SQL database system. The production data used are real, accurate and continuous. A set of assembly data can be considered as Big Data. Every second, hundreds of data are generated on the line, which are stored in different systems, for example, a separate system handles process data and a separate system handles product-specific data or logistics barcode data. During the processing and sorting of the data of these systems, 769 records are created and one record contains the data of an entire eight-hour shift. Assembly operations run in two shifts, so that means 10 records per week. The data of the examined period come from the years 2021 and 2022 and are scaled with no factors. In addition, the extreme or outlier values are not excluded so that the real industrial environment is depicted as well as possible. In the following, the goal is only to predict the OEE values; the individual components (availability, performance and quality) are not examined separately. The values of the dependent variable are shown in Figure 2. The OEE values are shown on the vertical axis and the individual records are shown in chronological order on the horizontal axis.

3.3. Cumulative and Rolling Horizon Prediction

During machine learning prediction of OEE, the following two main cases can be distinguished according to how the training period is selected:

• Cumulative approach;
• Rolling horizon approach;

  ○

  Fix rolling horizon;

  ○

  Changing rolling horizon.

In the cumulative case, the amount of training data increases continuously over time so more and more data are used as part of the training set. This has the advantage that all occurring data are taken into account; however, overfitting can occur. In the rolling horizon approach, the selected training data use a different time window for each prediction, which moves forward continuously. In general, rolling horizon prediction pushes the time window; cumulative does not.

3.4. Applied Machine Learning Methods

The authors used R and RStudio program for the entire research work. The following machine learning methods were used for OEE prediction:

• Multiple linear regression (MLR);
• Polynomial regression (simple) (Pol 1);
• Polynomial regression (complex) (Pol 2);
• Lasso regression (Lasso);
• Ridge regression (Ridge);
• Random forest regression (RF);
• Gradient boost regression (GB);
• Mixed GAM computation vehicle with automatic smoothness estimation regression (MGCV).

The basis of the selection was to choose methods within the R environment that can also be used in industrial conditions without requiring a large computing capacity. Two main considerations guided the authors’ choice of machine learning methods. The first aspect was the selection of the most suitable and widely used learning methods for the analysis of production time series. The selection was based on an extensive literature analysis. The second main criterion was the applicability in an industrial environment; where possible, the authors preferred procedures with a relatively low computational demand and a predictable time requirement for their application in industrial practice. A separate Rstudio program was written for each of the selected machine learning methods, and these contain the most frequently used settings. The following R packages were used: xlsx, xml, readxl, tidyverse, ggplot2, readr, openxls, chron, lubridate, dplyr, caTools, corrplot, doParallel, randomForest, MASS, mlbench, Amelia, plotly, reshape2, caret, moments and mgcv.

In the following chapter, the cumulative and fix rolling horizon approaches were used for each machine learning method.

4. Discussion

OEE was predicted for two periods (testing set), a shorter period (6 weeks, 60 records), which is the simulation of production planning, and a longer period (15 weeks, 150 records), which can help estimate the necessary industrial investments. In the first case, the rolling is performed on a weekly basis (10 records), while, in the second case, it is performed on a monthly basis (50 records). These periods are derived from automotive practice, because assembly plans are reviewed and production scheduling is conducted weekly. On the other hand, it is sufficient to monitor the OEE data required for industrial investments every month for the period of the next three months. Industrial investment means whether it is necessary to convert, rent, build or buy a new machine or production line in order to meet customer needs. In this article, several training periods were examined; in the case of production planning, the range is from 10 to 100 in increments of 10, while, in the case of industrial investment, the range is from 100 to 200 also scaled by 10.

In each case, the evaluation and comparing of the predicted results were carried out using root mean squared error (RMSE) based on the following equation:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\sum _{ i=1}^{ n}\frac{{\left({\overline{\mathrm{y}}}_i-{\mathrm{y}}_i\right)}^2}{n}}$$

(3)

where:

• n—number of fitted points;
• ${\mathrm{y}}_i$—actual value;
• ${\overline{\mathrm{y}}}_i$—predicted value [36].

For the different regressions applied by machine learning, the RMSE values for production planning are shown in

Table 4. RMSE values for production planning (short time prediction).

			RMSE
Step	Training	Testing	MLR	Pol 1	Pol 2	Lasso	Ridge	RF	GB	MGCV	Min
10	10	60	0.3737	0.0587	0.1294	0.0786	0.0721	0.0701	-	-	0.0587
10	20	60	0.0881	0.0553	0.0924	0.0682	0.0676	0.0655	-	0.0882	0.0553
10	30	60	0.0635	0.0501	0.0777	0.0569	0.0582	0.0621	-	0.0635	0.0501
10	40	60	0.0545	0.0543	0.0644	0.0519	0.0534	0.0599	-	0.0545	0.0519
10	50	60	0.0516	0.0541	0.0611	0.0513	0.0524	0.0577	-	0.0516	0.0513
10	60	60	0.0493	0.0541	0.0580	0.0498	0.0517	0.0564	0.0587	0.0493	0.0493
10	70	60	0.0490	0.0543	0.0573	0.0489	0.0507	0.0549	0.0570	0.0490	0.0489
10	80	60	0.0489	0.0546	0.0569	0.0489	0.0513	0.0540	0.0559	0.0489	0.0489
10	90	60	0.0510	0.0547	0.0580	0.0497	0.0513	0.0536	0.0540	0.0491	0.0491
10	100	60	0.0514	0.0550	0.0581	0.0505	0.0516	0.0529	0.0524	0.0492	0.0492
Avg.			0.0881	0.0545	0.0713	0.0555	0.0560	0.0587	0.0556	0.0559	0.0545
10	Cum.	60	0.0496	0.0603	0.0604	0.0503	0.0507	0.0498	0.0472	0.0496	0.0472

, while

Table 5. RMSE values for industrial investments (long time prediction).

			RMSE
Step	Training	Testing	MLR	Pol 1	Pol 2	Lasso	Ridge	RF	GB	MGCV	Min
50	100	150	0.0597	0.0573	0.0588	0.0555	0.0593	0.0573	0.0542	0.0597	0.0542
50	110	150	0.0619	0.0576	0.0593	0.0600	0.0631	0.0560	0.0551	0.0619	0.0551
50	120	150	0.0524	0.0565	0.0573	0.0525	0.0567	0.0528	0.0518	0.0524	0.0518
50	130	150	0.0513	0.0570	0.0578	0.0516	0.0557	0.0514	0.0494	0.0513	0.0494
50	140	150	0.0498	0.0563	0.0565	0.0513	0.0539	0.0514	0.0495	0.0498	0.0495
50	150	150	0.0496	0.0571	0.0565	0.0513	0.0540	0.0514	0.0515	0.0496	0.0496
50	160	150	0.0506	0.0585	0.0585	0.0524	0.0546	0.0522	0.0507	0.0506	0.0506
50	170	150	0.0504	0.0576	0.0576	0.0519	0.0540	0.0488	0.0472	0.0504	0.0472
50	180	150	0.0497	0.0571	0.0569	0.0513	0.0531	0.0471	0.0456	0.0497	0.0456
50	190	150	0.0482	0.0562	0.0559	0.0503	0.0512	0.0476	0.0450	0.0482	0.0450
50	200	150	0.0489	0.0570	0.0566	0.0513	0.0522	0.0476	0.0456	0.0489	0.0456
Avg.			0.0520	0.0571	0.0574	0.0527	0.0553	0.0512	0.0496	0.0520	0.0496
50	Cum.	150	0.0544	0.0617	0.0619	0.0562	0.0558	0.0562	0.0562	0.0544	0.0544

is for industrial investments.

The following conclusions can be made about predictions in the case of production planning:

• In terms of the examined training periods, simple polynomial regression (Pol 1) showed the best average RMSE result with a fix rolling horizon (0.0545), followed by lasso (0.0555) and gradient boost regression (0.0556), although, with GB, the 10–50 training period cannot be interpreted due to the specificity of the method;
• In the case of a rolling horizon, the lowest RMSE value (0.0489) can be achieved with lasso regression set to the training parameters 70 and 80;
• Analyzing the cumulative methods, the gradient boost regression showed the best RMSE result (0.0472), followed by multiple linear regression (0.0496) and MGCV (0.0496);
• Considering the average speed of calculation, the fastest method is simple polynomial regression (4.7 s), followed by multiple linear regression (5.1 s) and random forest (9.9 s). The individual values are shown in

  Table 6. Average calculation time of each method.

  Applied Regression	Production Planning		Industrial Investments
  	Roll. Hor.	Cum.	Roll. Hor.	Cum.
  Multiple linear regression	5.1 s	7.0 s	2.0 s	5.0 s
  Polynomial regression (simple)	4.7 s	6.0 s	2.7 s	3.9 s
  Polynomial regression (complex)	56.3 s	70.3 s	14.4 s	15.0 s
  Lasso regression	546.9 s	765.9 s	86.3 s	117.4 s
  Ridge regression	17,727.3 s	18,906.0 s	2758.2 s	3506.2 s
  Random forest regression	9.9 s	33.9 s	4.0 s	6.8 s
  Gradient boost regression	511.2 s	1016.7 s	94.1 s	152.8 s
  MGCV	10.5 s	13.3 s	2.9 s	3.7 s

  . (The laptop used with Intel ^® Celeron ^® CPU N2840 @ 2.16 GHz 2.16 GHz, 4.00 GB RAM);
• The RMSE values of the examined cumulated methods are generally better than those of the fix rolling horizon methods;
• In the case of production planning, the recommended OEE prediction method is the gradient boost cumulated approach due to the most accurate RMSE value and the moderately long calculation requirement.

The following conclusions can be made about predictions in the case of industrial investments:

• Regarding the examined training periods, the gradient boost regression showed the best average RMSE result with fix rolling horizon (0.0496), followed by random forest (0.0512), multiple linear regression (0.0520) and MGCV with the same result (0.0520);
• In the case of fix rolling horizon, the lowest RMSE value (0.0450) can be achieved with gradient boost regression set to the training parameters 190;
• Analyzing the cumulative methods, the multiple linear regression showed the best RMSE result (0.0544), followed by MGCV (0.0544) and ridge regression (0.0558);
• Considering the average speed of calculation, the fastest method is multiple linear regression (2.0 s), followed by simple polynomial regression (2.7 s) and MGCV (2.9 s);
• The RMSE values of the examined fix rolling horizon methods are always better than the cumulated approach;
• In the case of industrial investment, the recommended prediction method is the gradient boost fix rolling horizon approach with 180–200 training sets, due to the most accurate RMSE value and the medium–long calculation requirement.

The machine learning methods used have the advantage of being able to estimate the real production time series under study with an average RMSE rate of approximately 4–7%. Based on our analyses, it is clear that the so-called over-learning does not occur for the methods under investigation. A weakness identified is the inaccuracy of the ability to react to sudden changes in the real world, which is an important and pronounced feature in the real environment. Based on our observations, the range of the appropriate learning set is quite wide, and it is difficult to clearly determine based on these real-world data sets how many shift cycles are required for machine learning methods to achieve an RMSE prediction value of 5% or below.

During the regressions presented so far, in addition to the fix rolling horizon (10 records for production planning, 50 records for investment) and fix testing period (60 records for production planning, 150 records for investment), only the training periods were modified. It follows that, in addition to these, there are many variation options and even smaller RMSE values can be achieved if the testing periods are changed. As an example, for multiple linear regression, Figure 3 shows the additional options.

The predicted or testing records are shown on the horizontal axis and the RMSE values on the vertical axis. The different colored lines show the number of used training records. Figure 3 shows that, in all cases, if the amount of testing data is reduced, the RMSE value decreases. For production planning, the use of the figure provides guidelines as to how much prediction accuracy the used training set will result in during the testing period. Based on Figure 3, the main conclusion is that multiple linear regression fits better with more training records and less testing data. However, it is important to monitor the RMSE values, depending on how accurately it is necessary to predict the OEE values.

5. Conclusions

At the area of the manufacturing industry, it is essential that the follow up of assembly operations’ performance and efficiency takes in real-time mode and the estimation works quickly and reliably. Based on supervised machine learning, assembly efficiency metrics, e.g., Overall Equipment Effectiveness (OEE) can be predicted, but it is not clear by which method.

This paper collected and evaluated the factors as initial information or data affecting OEE from different perspectives. The individual elements were examined according to whether they affect availability, performance or quality and how they can be measured by Manufacturing Execution System (MES), SQL query or another way. Based on the measurability and occurrence of the factor characteristic of assembly lines, 12 independent variables were selected among others, including process failure downtime, technical downtime, changeover downtime, average cycle time and number of assembled units. OEE percentage, availability percentage, performance percentage and quality percentage are considered as dependent variables.

Based on the selected data, the prediction of OEE was analyzed using different machine learning methods, such as multiple linear regression, simple and complex polynomial regression, lasso regression, ridge regression, random forest regression, gradient boost regression and MGCV regression. Using real industrial assembly line data, each machine learning method has been demonstrated for production planning and investment. Two approaches were used, fix rolling horizon and cumulative way. The results were evaluated using root mean squared error (RMSE) and the computation time.

In the case of production planning, the recommended OEE prediction method is the gradient boost cumulated approach due to the most accurate RMSE value and the moderately long calculation requirement. In the case of industrial investment, the recommended prediction method is the gradient boost fix rolling horizon approach with 180–200 training sets, due to the most accurate RMSE value and the medium–long calculation requirement. In addition to these, further optimization possibilities were outlined using the extended testing period, which was demonstrated with multiple linear regression.

The authors plan to examine further machine learning methods and search for additional optimization options to predict the dependent variable even more accurately. In the future, it will be possible to combine the cumulative approach with the fixed rolling horizon way for a lower value of RMSE.