1 Introduction

The lack of skilled labor due to demographic change in many societies poses a major risk to maintaining the productivity of local industries. According to a survey conducted by the European Commission more than 54% of small and mid-sized enterprises (SMEs) in the EU identify the shortage of skilled labor as the most serious problem for their business development [1]. However, these challenges are not limited to companies operating in the EU. Globally, Japan, Germany, Brazil and Indonesia are predicted to suffer the greatest economic loss in 2030 due to the shortage of skilled labour in the manufacturing sector [2].

Manufacturing technology often consists of complex mechanical processes. As an example, roll forming is a continuous cold forming process for the production of profiles at high production rates of up to 230 m/min for automotive, furniture and construction applications [3]. The forming takes place in progressive forming stages consisting of tool rolls. Producing complex profile geometries, the sheet metal passes through up to 70 forming stages, which must be precisely adjusted to each other in the preceding setup [4]. Usually, the adjustment is done manually and requires many iterations to align the neighboring tool rolls. Therefore, it is dependent on experience. Due to the high degree of complexity in combination with short product cycles, the initial setup and adjustment of roll forming lines results in long downtimes, which have a major impact on system efficiency [3]. However, data on downtimes is neither collected nor shared publicly. There are also no statistics from profiled product manufacturers available regarding the requirement of digitalizing their human-bound knowledge and experience. From this point of view, two less technical but crucial questions remain unanswered in order to introduce the topic of assistance systems in the profile manufacturing sector:

  1. 1.

    Are companies aware of the challenges ahead, and is there a need for digital assistance in profile manufacturing technology?

  2. 2.

    Which functions should an assistance system offer for the support of inexperienced operators to counteract the shortage of skilled labor?

1.1 Study on the shortage of skilled labor and the use of digital assistance systems

For this introducing study, several companies were asked if and how digital assistance systems could support skilled and especially unskilled operators. The target group consists of specialists and managers with personnel responsibility in the roll forming industry, but also in process areas linked to the production of complex profiled components, such as stamping and bending technology. The questionnaire was distributed digitally at the “Blechexpo Stuttgart 2023” trade fair. A total of 58 participants were acquired for the study. This limits the generalizability of the results to other industries and sectors but allows for a perception of the participating companies that can be set in relation to documented social developments. The participants in the study describe their own position in the company as “senior management” (41%), as “specialist” (41%), as “middle management” (9%) or as “other” (9%). They are distributed as follows: 15% for companies with less than 50 employees, 35% for companies with 50–250 employees, 24% for companies with 250–1,000 employees and 24% for companies with more than 1,000 employees. The answers to the questions are based on a four-point Likert scale. It is also possible to answer open questions at certain points. Figure 1 shows all results of the study. In Question 1, respondents rate the importance of experience for the manual setup and adjustment of tools. Here, 95% of respondents stated that experience in their company was “important” or “rather important” for such operations. When asked to what extent the company’s internal processes for securing expertise are effective, an ambiguous picture emerges (Question 2). Here, 40% of respondents stated that the securing of knowledge was “ineffective” in their company, while 60% of participants answered “effectively” or „very effectively”.

In line with these results, most respondents stated that the acquisition of new skilled labor is currently “rather challenging” (58%) or “very challenging” (35%) for their company. Participants also expect these challenges to increase in the future, as the proportion of respondents who consider this acquisition to be “very challenging” in the future (53%) has increased compared to the current situation (Question 3). In accordance with these expectations, the respondents stated that their companies expect to lose in-house expertise in the future, despite the fact that knowledge transfer within the companies works well in some cases (“rather yes” 56%, “yes” 30%, Question 4). As a result, respondents attach great importance to the increasing use of assistance systems for their business. Overall, 90% of respondents stated that the future use of assistance systems would be “rather important” or “important” to their company’s business model (Question 5). Despite the expectation of high relevance to their own organization, 69% of respondents say they have no experience using digital assistants to support manual setup and adjustment processes (Question 6).

In this paper, the decision-making process for the assistance system is designed to be based on the use of Artificial Intelligence (AI), which is why the acceptance for AI is also surveyed on a scale from 0 (no acceptance) to 5 (high acceptance). Here, new employees are more likely to accept the assistance of AI-based systems compared to those with longer experience (Question 7). However, acceptance is still high for both groups surveyed.

In addition to the respondents’ expectations regarding the importance of assistance systems in production, the requirements for the use of these assistance systems were asked. Here, “process reliability” was cited as the most important criterion, followed by “integrability into existing processes” and “intuitive operability” of the assistance system (Question 8).

In summary, skilled operators are of great importance for the manual adjustment of manufacturing processes and the respondents consider digital assistance systems to be highly relevant for their company’s success in the future. This highlights the need for assisting solutions to support operators in manual tasks in this sector.

Fig. 1
figure 1

Results of the study on the shortage of skilled labor

Full size image

2 State of the art

In a fundamental position paper on AI in manufacturing technology, the German Scientific Association for Production Engineering (WGP) identifies assistance systems as one of the main areas for the application of AI in production technology. Furthermore, the need for methodologies for the development and use of these assistance systems in manufacturing processes is emphasized [5]. Existing approaches for the systematic processing of production data are considered and evaluated in terms of their feasibility as operator assistance in roll forming.

2.1 Methodologies for the model-based generation of knowledge from data

A systematic and automatic extraction of knowledge from data is the foundation for supporting operators in the manual setting of roll forming processes. This forms the core of a machine learning based (ML-based) assistance system. In the literature, there are various methodologies for deriving knowledge from data sets. The individual process steps in data processing are similar, but have different priorities depending on the application.

Fayyad et al. present an initial methodological framework. The so-called KDD (Knowledge Discovery in Databases) aims to extract knowledge from the increasing amounts of data generated during the digitalization of our society [6]. The methodology suggests the use of data mining techniques to derive patterns from data sets, but does not provide instructions for a data-specific transformation and selection of algorithms. This does not ensure that the pre-processed data is suitable for the selected algorithm.

CRISP-DM (CRoss Industry Standard Process for Data Mining), presented by Chapman et al., is a methodology for the implementation of data mining projects with a corporate focus [7]. CRISP-DM is based on the KDD and used in engineering applications as well as in the healthcare industry and the public sector [8]. However, when applied to production processes, this general focus does not provide proper guidance on how to correctly aggregate data on a domain-specific basis. Since process knowledge is important for the setup and adjustment of production plants, it is necessary to understand the link between the data and the physical process behavior.

Based on the KDD, Kubik et al. present KDT-EA (Knowledge Discovery in Engineering Applications), a methodology with focus on data acquisition and data analysis in engineering applications [9]. It considers the boundary conditions of manufacturing processes by placing a high priority on the accurate sensory acquisition of data in production processes (phase I), the data preparation (phase II) and the data transformation (phase III). This is followed by modelling (phase IV) and evaluation (phase V). In contrast to the previous methodologies, the inclusion of domain-specific process knowledge is linked to each process phase. This is important as different process-specific challenges arise in the production context in the form of vibrations, accessibility of measuring points and high production rates. The aim is to select the sensory measurement chain and the standardized data processing procedure to correctly map the actual process state in terms of the optimization objective [9]. However, an interface between the model and operators in terms of given suggestions or received feedback is not provided by KDT-EA. As none of these methodologies from the literature offer a complete framework for machine learning based operator assistance, a new development is necessary. In order to consider the boundary conditions of roll forming, it is necessary to review the existing research on data-based modeling in roll forming.

2.2 Data-based modelling in roll forming

In roll forming literature, there are various approaches for the acquisition and modelling based on process data. Studies focus on different optimization tasks like energy demands, fluctuations within semi-finished products, process load behavior and the correct adjustment of the machinery.

In 1987, Bhattacharyya et al. presented the first semi-empirical approach for estimating the process force during roll forming. For two materials with yield strengths between 137.7 MPa and 265 MPa the modelled correlations achieve an error of up to 20% [10]. By adapting empirical equations, Lindgren et al. model loads in roll forming processes based on experimental data to calibrate the required forming loads in finite element models. An estimation for torques is evaluated by power comparisons, which are conducted for seven different sheet materials with tensile strengths from 193 MPa to 1129 MPa [11]. These results show that numerical simulations in roll forming are able to extend analytical considerations and are suitable to estimate the expected measuring range for variables such as forces and torques.

Process data in roll forming is also used to investigate uncertainty due to semi-finished product properties. Abeyrathna et al. use regression and variance analyses to model the relationship between semi-finished product properties, process forces and drive torques. Based on the correlations with the final profile geometry, they suggest an in-line compensation of fluctuating semi-finished products using adjustable tooling [12]. A lot of recent approaches focus on the exploitation of process data for tool adjustment, as a correct tool setup has a significant impact on process reliability. If several neighboring forming gaps are not correctly aligned with each other, this may result in undesirable bending leading to deflections in the longitudinal direction of the profile [3]. Also Traub et al. [13] show existing correlations between tool positions, profile deflections and in-line data of forming forces and torques. Müller shows numerically and experimentally that the mechanical behavior of several forming stands differs greatly from one to another, and that these differences in play and compliance are reflected in the profile quality. By measuring forming forces and tool positions, Müller generates force-displacement curves for the roll forming stands under process load and shows the displacement of the tools in the load-free setup state compared to the process state with forming load [14]. Sáenz de Argandoña et al. measure process forces for three gap heights between top and bottom roll of a forming stage and show that the forces increase as the gap height decreases. Based on this, they state that the measurement of forces could be used for a more efficient adjustment of machinery [15]. Bleicher et al. repeat setup processes using the same force data every time and show that this sensor-supported adjustment process achieves a higher reproducibility of settings in a four-stage roll forming process. In this context, the adjustment procedure is not described in detail and it is not specified whether the roll positions are meant to be adjusted under load-free or loaded conditions [16]. Leonhartsberger et al. present a methodology that uses numerical simulations as well as experimental measurements to describe the varying mechanical load behavior of roll forming stands. They quantify the influence of stiffness, play, misalignment and spacers on the reproducibility of measured forces of roll forming processes and quantify the uncertainties up to 66.9% [17]. As these factors are not or only simplified considered in conventional roll forming simulations. Finite element analysis is currently not able to map the full complexity of the roll forming setup. The authors state, that experimental force and displacement measurements for off-line characterization of each stand lead to an accurate initial roll position setup and eliminate the need for permanent integration of in-line sensors [17]. However, due to the lack of in-line process monitoring, this procedure does not provide the ability to detect changes that may occur after the initial setup, such as fluctuations in mechanical behavior or semi-finished product properties.

For this purpose, Traub et al. present a methodology for decision-making using operator assistance systems in roll forming processes. This framework is based on the derivation of a diagnostic matrix that links sensor-based process information with setup and control variables of the roll forming process. In addition to improving energy efficiency, the method is used to suggest forming gap adjustments qualitatively. Based on a pattern analysis of the relationships between tool position, force data and torque data the assistance system assesses forming gaps as too large or too small compared to a pre-defined reference [13]. The authors extend the approach by using support vector machines, but the limitations are small experimental datasets and a visual evaluation without a quantification of the model’s prediction accuracy [18]. Becker et al. advance these approaches by using a random subspace algorithm based on the idea of ensemble learning. This enables a classification of misalignments of a single roll forming stand based on force and torque mean values, but is also limited by discrete prediction [19]. 

In terms of literature review, there are several approaches for the setup and adjustment of roll forming tools supported by data. But none of the discussed approaches is capable of estimating and monitoring exact setup values for the tool positions in roll forming so far. The reasons for this status quo are diverse and reach from the lack of in-line data over small data sets to the general discrete limitation of classification models. With regard to machine learning in roll forming, the only known regressive prediction model is used for tool wear which achieves a worn tool radius accuracy of 22.214 ± 1.765 μm. By comparing wear in shear cutting and roll forming, this approach also aims to prove that methodologies for the model-based generation of knowledge like KDT-EA work similar for different production processes [20].

In summary, the state of the art in roll forming shows that the use of machine learning models for the analysis of process data has great potential but is not fully exploited so far. The main reason is a lack of combination between application-oriented data transformation and model selection with a knowledge-based integration of the operator into the whole process flow. Applied research in the setup and process optimization of roll forming is based on manual interpretation of thresholds or allows only qualitative statements about process states so far. As some of the discussed approaches show, the possibility of exploiting in-line data for the adjustment of roll forming processes is given in theory. Therefore, the main objective of this study is to develop a methodology for the machine learning based generation of knowledge from in-line data in roll forming which includes operator assistance. For this purpose, the focus is on suggestions of quantitative measures and an immediate evaluation of the proposed actions done by the operators.

3 Framework for machine learning based operator assistance (MLbOA)

After the evaluation of the needs and the discussion of the state of the art, this section answers the second question about the functions of an assistance system in roll forming. The newly developed framework MLbOA is shown in Fig. 2. It distinguishes between the three subsystems Roll Forming Process, machine learning based Data Exploitation and Operator. Those systems are necessary to support inexperienced operators and thus counteract the shortage of skilled labor in profile manufacturing technology.

Roll forming process

The goal of the MLbOA framework is to influence the Roll Forming Process subsystem, so that the desired nominal process state ynom,i defined by the operator at time i corresponds to the actual process state yact,i. Discrepancies from ynom,i occur in the roll forming process due to the uncertainty related process deviations \(\mathop \sum \nolimits^ {{\bf{d}}_{j,i}}j\varepsilon {\rm{}}\left\{ {{\rm{environment}},{\rm{}}\,{\rm{process}},\,{\rm{human}}} \right\}\). The deviations are divided into impacts caused by temperature and humidity de,i which can be traced back to the process environment [21]. Furthermore, human intervention of the operator dh,i occur due to different empirical knowledge and daily fluctuations in mood and behavior. Process deviations inherent in the roll forming process are summarized in dp,i. An example for such deviations are geometric errors of the profile due to inhomogeneous longitudinal strain distributions in the cross-section [3]. In addition, dp,i includes fluctuations in semi-finished product properties and fluctuations in the geometric dimensions of semi-finished products, which can occur between coils but also within coils [22]. Further factors belonging to the process deviations dp,i are unknown motion behavior due to tolerances in the mechanical chain as well as the wear of the tools and other mechanical components such as bearings [19]. In conventional roll forming processes, the actual process state can be characterized at each iteration step i by.

$$\:{\mathbf{y}}_{\text{a}\text{c}\text{t},i}={\mathbf{y}}_{\text{n}\text{o}\text{m},i}+{\sum\:\mathbf{d}}_{j,i}$$
(1)

The goal of the assistance system is now to generate possible actions \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\:\)for the operator to apply.

$$\eqalign{& {{\bf{y}}_{{\rm{act}},i}} = {{\bf{y}}_{{\rm{nom}},i}} + {{\bf{y}}_{{\rm{adj}},i}} + \mathop \sum \nolimits^ {{\bf{d}}_{j,i}}{\rm{with}} \cr & {{\bf{y}}_{{\rm{adj}},i}} + \mathop \sum \nolimits^ {{\bf{d}}_{j,i}} \le l \cr}$$
(2)

The limit vector l describes the application-specific tolerance limits that must be satisfied. Due to the large number of quantifiable and non-quantifiable influencing factors in \(\:{\sum\:\mathbf{d}}_{j,i}\:\), the relationship between the desired nominal process vector \(\:{\mathbf{y}}_{\text{n}\text{o}\text{m},i}\) and the adjustment vector \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\) is non-linear, dependent on stochastic randomness and cannot be fully described by analytical considerations and numerical simulations (see Chap. 2).

Data based modelling

The framework utilizes the established data processing steps of the KDT-EA methodology [9] to develop an ML-based model. Additionally, it extends the workflow of the ML model for the use after the initial implementation phase. In order to support operators with instructions, the goal is to move from the mere description of process states to the integration of the mechanical load behavior for the application of countermeasures. This allows it to derive actions in the form of a vector \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\) for process misalignments.

The initial implementation phase for the ML model is based on a process and target specific data acquisition. In this phase, a comprehensive understanding of the roll forming process is required in order to select the sensors, the measurement chain, the measurement location, and to critically evaluate the generated signals for their information content with regard to the physical phenomena depicted [20].

Fig. 2
figure 2

Framework of the machine learning based operator assistance

Full size image

The results of the acquisition form the raw dataset matrix \(\:{\mathbf{X}}_{\text{r}\text{a}\text{w}}\in\:{\text{R}}^{\text{m}\times\:\text{p}}\), where \(\:\text{m}\) is the number of signals and \(\:\text{p}\) is the number of measurement points per signal. The use of supervised machine learning approaches requires the acquisition of a target vector containing the correct process adjustments \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\). Both can be acquired by preceding iteration loops or a targeted data acquisition for training data. The use of data from other domains to reduce the training data [23], the artificial expansion of data sets [24] and the use of domain adaptation techniques in production technology is part of current research [25]. In order to use the data \(\:{\mathbf{X}}_{\text{r}\text{a}\text{w}}\) for initial training and validating of the ML models, it is necessary to filter erroneous information caused by outliers or data drift during preprocessing. In addition, the quality of the information content may be increased by denoising [26]. When applying these steps, process knowledge is essential in order to correctly interpret and filter incomplete and erroneous data. If \(\:{\mathbf{X}}_{\text{r}\text{a}\text{w}}\) originates from different signal types and is therefore linked to different scales, a normalization procedure is required. To prepare the data for processing the model, a final synchronization and interpolation is necessary. The result is the data matrix \(\:\mathbf{X}\in\:{\text{R}}^{\text{m}\times\:\text{p}}\), which contains the processed data set. In the phase of data transformation, the use of domain-specific knowledge for the extraction of features is not decisive anymore due to model-based transformation methods such as Principal Component Analysis (PCA) [27] or Uniform Manifold Approximation and Projection (UMAP) [28] as well as parameter extractions in the time, frequency and time-frequency domain by python libraries such as “TSFEL” [29]. Furthermore, deep learning methods provide the capability to process entire time series or automatically extract features. For instance, convolutional neural networks can replace the need for manual feature extraction [30]. Also Modelling is becoming more independent of process knowledge due to software such as “sklearn” [31] and “tensorflow” [32], which allow standardized model optimization through hyperparameter optimization methods. The correct model is selected by comparing different algorithms using standardized evaluation methods such as cross-validation and metrics such as the mean absolute error (MAE) or the root mean squared error (RMSE) for regression, accuracy, precision and recall for classification. Once a suitable ML model has been trained and evaluated, all the in-line acquired process signals \(\:{\mathbf{x}}_{\text{r}\text{a}\text{w},i}\:\) can process the prediction pipeline. The data pass through the equivalent preprocessing and transformation steps in order to be correctly assigned to an appropriate process adjustment \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\).

Operator

The operator is provided with the associated process adjustment and the model’s prediction quality Operators retain decision-making authority in the knowledge generation process through plausibility checks based on the information provided and their domain knowledge. If the operator deems the specified setting plausible, it is implemented in the roll forming process. This step is similar to the procedure of Traub et al. [13] but in this case the operator benefits from the precise adjustment values provided by the model. If the operator does not agree with the assisted suggestion, this adjustment is adapted and implemented as a unique solution in the roll forming subsystem. The decision authority can help filter out implausible process settings and can ensure greater acceptance among skilled operators [33]. Additionally, implementing new correct decisions as targets into the knowledge generation cycle extends the solution space of the model. New data adjustment pairs can lead to an adaptation of the ML model to the changing conditions by means of continuous retraining. If the operator’s implementation does not lead to the desired approximation of and, the iteration loop continues. As soon as the process adjustments meet the condition in Eq. (2), no further iteration loops are required until the system’s boundary conditions change to the point where limit l is exceeded.

4 Application for MLbOA on the adjustment of a roll forming line

The newly developed MLbOA is experimentally applied on a roll forming line with \(\:\text{n}=4\:\)forming stages (Fig. 3). Each forming stage has a top and a bottom forming tool roll driven on shafts. The contour of the rolls determines the shape of the sheet metal by forming it incrementally. According to the frameworks sensory acquisition step, relevant process variables which depict important characteristic features of the roll forming process need to be selected. The correct setup of the tools is very complex due to the multiple degrees of freedom and therefore highly prone to errors. The forming gap provided by each pair of rolls acts as the negative contour for the desired cross-sectional geometry of the profile. If this gap is set incorrectly, the tolerances for angles and radii may be missed.

Fig. 3
figure 3

Experimental set up of the sensorial equipped roll forming line

Full size image

Usually, during the setup process of the forming tools, the operator uses a setup plan that specifies the horizontal and vertical tool positions to form the given profile. The challenge is that this load-free initial setup only serves as a reference point. As play and compliance occur under load, the desired nominal setup state \(\:{\mathbf{y}}_{\text{n}\text{o}\text{m},i}\) is not fulfilled during the forming process. In current practice, the load behavior is considered iteratively and experience-based by applying the process load several times and readjusting especially the vertical positions until Eq. (2) is fulfilled. This procedure is neither time nor labor efficient in the short-term, nor is it sustainable for future setups.

The process variables force, torque and tool position are suitable to record relevant digital process information and feed a setup assistance system, as they physically describe the dynamic characteristics of the roll forming process very well. For this purpose, the roll forming line is equipped with torque sensors in all eight driving shafts. Two force sensors are integrated into the upper part of each forming stand to measure vertical forming forces, resulting in eight force acquisition points. Magnetic absolute measuring sensors are integrated into all eight top bearing blocks as position sensors for the top tool rolls. For straight vertical movement with no tilt of the top rolls, the two opposite sensors of each forming stand have the same position value. The lower tool bearing blocks are not equipped with sensors, as their adjustment remains unchanged during these investigations.

To support the operator with quantifiable adjustment measures, it is necessary to know the absolute positions under load and provide the operator with adjustments in the load-free state. For this purpose, the load-free forming gaps build the target vector, which consists of\(\:\:\:{\mathbf{y}}_{\text{n}\text{o}\text{m}}={\left[{\text{s}}_{1,\text{s}\text{e}\text{t}\text{u}\text{p}},\:{\text{s}}_{2,\text{s}\text{e}\text{t}\text{u}\text{p}},{\text{s}}_{3,\text{s}\text{e}\text{t}\text{u}\text{p}},\:{\text{s}}_{4,\text{s}\text{e}\text{t}\text{u}\text{p}}\right]}^{\text{T}}\). For sheet thicknesses of 1 mm, a load-free gap of around 0.9 mm has proven to be a good starting point. Therefore, the position \(\:{\text{s}}_{\text{n},\text{s}\text{e}\text{t}\text{u}\text{p}}\:\) of the top roll is varied around this value in seven classes by +/- 0.05 mm steps to:

\(\:{\text{s}}_{\text{n},\text{s}\text{e}\text{t}\text{u}\text{p}}\)\(\:\in\:\) {0.75 mm, 0.80 mm, 0.85 mm, 0.90 mm, 0.95 mm, 1.00 mm, 1.05 mm}

The distribution of these gap setups within the total number of 200 experiments is illustrated in Table 1. The experimental plan was conducted changing one factor at a time. For example, during the setup of 0.75 mm at stage 1 all other stages stay at the initial gap size of 0.90 mm and six metal sheets were roll formed. All metal sheets belong to the same batch of dual phase steel ‘DP600’ with a yield strength of approximately 330 N/mm² and a tensile strength of approximately 650 N/mm². The size of the sheets is 2000 × 100 × 1 mm. The initial reference setup of 0.90 mm for all stages was set at the beginning and at the end of the test series and therefore contains twelve instead of six process runs. As in industrial processes erroneous process states occur less often than well performing setups, this kind of inhomogeneous data distribution is prestigious. To investigate if this heterogeneity affects the model’s accuracy, random setups were run once more as well. Figure 4 shows an example of the measured forces (top), torques (middle) and vertical top shaft positions (bottom). These time series illustrate the typical process behavior, which is characterized by the run-in phase, the cyclostationary phase and the run-out phase. The force data, torque data and the loaded tool positions of the top rolls in the cyclostationary process phase are available as inputs for the model. The vector containing the process state\(\:\:{\mathbf{y}}_{\text{n}\text{o}\text{m}}\:\)is measured before the sheet metal is fed in. The vertical tool position is calculated as the mean value of the left and right side for each forming stage. As the force-displacement characteristic of the mechanical components is time-independent, no time labels are required and each time step is used individually. The data set \(\:{\mathbf{X}}_{\text{R}\text{a}\text{w}}\in\:{\text{R}}^{721.200\times\:20}\) consists of 20 columns for all sensors, and 721.200 rows for all data points of 200 process runs.

The first step in preprocessing is to determine the relevant offsets in all time series to achieve a homogeneous data set over all sensors and all process runs. As a result, all signals start with the value 0 at time t0.

Table 1 Number of experiments for each process variation
Full size table
Fig. 4
figure 4

Sensorial acquired force, torque and position data

Full size image

Due to a non-uniform measurement rate between 10 and 50 Hz caused by the controller’s performance, the raw data is resampled to 50 Hz. This is achieved by interpolation followed by discrete sampling with a constant step size. The three characteristic phases are automatically identified and windowed based on the significant positive and negative force gradients. The run-in and run-out phases are eliminated and only the cyclostationary phase is used, as this is predominant in continuous roll forming processes with long sheets or coils. The lasting \(\:\text{p}=20\) columns of 8 force, 8 torque and 4 position time series are shortened to \(\:\text{m}=300.000\) row entries. For the comparability of signals with different units, a normalization to standard deviation 1 and mean value 0 is implemented.

For transformation, different approaches are selected to compare different data subsets as inputs in the model and to obtain the best transformation rule. This allows the information quality of the sensor signals to be compared with each other. The following subsets are created:

  1. 1.

    force and position data \(\:{\mathbf{X}}_{\text{F}\text{S}}\in\:{\text{R}}^{300.000\times\:12}\)

  2. 2.

    torque and position data \(\:{\mathbf{X}}_{\text{T}\text{S}}\in\:{\text{R}}^{300.000\times\:12}\)

  3. 3.

    force, torque and position data \(\:{\mathbf{X}}_{\text{F}\text{T}\text{S}}\in\:{\text{R}}^{300.000\text{x}20}\)

  4. 4.

    PCA features of force, torque and position data \(\:{\mathbf{F}}_{\text{F}\text{T}\text{S}}\in\:{\text{R}}^{300.000\times\:\text{N}}\)

For dataset 4 a PCA is performed. Six components explain 99.54% of the variance of the Dataset \(\:{\mathbf{X}}_{\text{F}\text{T}\text{S}}\) and are therefore defined as a suitable number N of features. The target vector of the model is the vertical load-free setup position for a correct load-anticipated setting of the top rolls of all four forming stages \(\:{\mathbf{y}}_{\text{n}\text{o}\text{m}}\in\:{\text{R}}^{300.000\times\:4}\).

In modelling, a linear regression (LR), a random forest regressor (RFR) and a neural network (NN) are compared. Training and validation are based on a five-fold cross-validation. The LR and the RFR were implemented using sklearns default parameters [31]. The LR does not have significant hyperparameters that need tuning. The RFRs hyperparameter configuration includes an architecture of 100 decision trees, whose maximum depth, as well as the maximum number of nodes is not restricted by a fixed threshold. The NNs hyperparameters are optimized for each dataset using a bayesian optimization algorithm. The optimized hyperparameters, including the parameter range are illustrated in Table 2. The learning rate is adapted during the training process using the Adam optimizer [34]. Furthermore, an early stopping algorithm was used in the training process to prevent overfitting.

Table 2 Hyperparameter range of the NN
Full size table

The regression qualities are compared using the Mean Squared Error (MSE) and the Mean Absolute Error (MAE). Due to the squaring, the MSE responds comparatively stronger to large outliers and thus takes the fluctuation range of the predictions into greater account than the MAE. The regression qualities shown in Table 3 enable a quantitative evaluation of the results given by the different transformations and models.

The MAE of the LR is in the range of 20–25 μm for all input data subsets. As the forming gap is adjusted in steps of 50 μm, these linear correlation-based results cannot provide valid statements for improved setup processes. This proves that a linear relationship is not sufficient to approximate the local load behavior of the forming stages within their individual mechanical play and compliance.

Linear modelling therefore does not meet the complex functional context of the data set and the roll forming setup procedure. The cascade-like decision limits in the form of the RFR enable a significant improvement in the prediction quality. For the PCA subset, the best MAE result is 0.57 ± 0.21 μm. Nevertheless, all RFR results contain a high MSE. This may be explained by the discontinuous regression based on rigid decision boundaries. As a result, a high number of correct predictions without errors are produced (low MAE), but an incorrect prediction ends up in a different class and thus represents a strong outlier (high MSE).

The NN provides the best combination of MAE and MSE, which can be seen in the low level of estimation errors with good accuracy over the entire data set. The neurons in the fully connected layers of the NN transform the input vector \(\:\mathbf{x}\:\in\:{\text{R}}^{1\times\:\text{D}}\:\)for the k-th layer based on the weight Matrix \(\:\mathbf{W}\) and the bias factor \(\:\text{b}\), which are adapted in the training process into an activation. This serves as an input for the activation function \(\:f\).

$$\:\text{y}({\mathbf{x}}^{k-1},\mathbf{W})=\text{f}\left(\sum\:_{j=1}^{\text{D}}{\text{w}}_{j}^{k}\cdot\:{\text{x}}_{\:j}^{k-1}+{\text{b}}^{k}\right)$$
(3)

This NN uses the continuous nonlinear activation function \(\:f\left(\text{x}\right)=\text{max}(0,\text{x})\) (ReLU function) in the input and hidden layers. A linear activation function is chosen for the output layer. This allows for the mapping of the individual nonlinear load behavior of the mechanical chain during the forming process (low MAE), while predicting values in a continuous regression without rigid decision boundaries. This explains the low MSE compared to the RFR.

The use of force and torque data performs better than the dedicated selection of one single variable. Even if the vertical force component is intuitively perceived as more important, the horizontal load behavior represented by the torque appears to have a recognizable influence on the mechanical compliance of the forming stands as well. The model results prove that the information given by both variables is not completely redundant, despite their physical coupling via the friction mechanisms within the tribological system. Instead, the information content from both working directions of the applied process energy (vertical force and horizontal torque) is meaningful for the prediction quality.

Table 3 Evaluation of the models with MAE and MSE
Full size table

After the initial creation, training, and validation process, the model can be used within the MLbOA for the setup processes. Also, if a maladjustment suddenly appears in the Roll Forming Process at time \(\:i\) and the limit vector l is exceeded, the data acquired from the force, torque and position sensors in the input vector \(\:{\mathbf{x}}_{i}\in\:{\text{R}}^{1\times\:20}\) is analyzed by the pretrained NN. The model output \(\:{\mathbf{y}}_{\text{a}\text{d}\text{j},i}\) will be passed to the Operator along with the MAE and the MSE of the evaluation process. The Operator can identify the misalignment in the process after the plausibility check by initiating a direct countermeasure, which reduces the expertise required to adjust the process correctly. Furthermore if the Operator judges the models output as unplausible and implements an new solution, it can be implemented as a new target for the input vector \(\:{\mathbf{x}}_{i}\) for a continuous retraining of the model implemented. This enables to gradually include the Operators knowledge into the model and avoid the deterioration of model accuracy due to continuous datashifts over time.

5 Conclusions and future work

To address the challenges posed by the shortage of skilled labor in the roll forming industry, this paper answers two main questions.

1. Are companies aware of the challenges ahead and is there a need for digitalized assistance in profile manufacturing?

This paper presents a study that examines the challenges posed by the increasing shortage of skilled labor for companies in the profile manufacturing sector and their expectations regarding the future use of digital assistance systems. The evaluation of the acquired responses shows that companies are increasingly expecting a loss of in-house expertise due to the shortage of skilled labor, which is seen as essential for the adjustment of production processes. Based on these expectations, 90% of respondents consider the use of digital assistance systems to be “rather important” or “important” for companies’ business model in the future.

2. Which functions should an assistance system offer to support inexperienced operators and thus counteract the shortage of skilled labor?

A framework for the generation and use of machine learning based operator assistance is presented and evaluated. This allows the generation of specific instructions for process set-ups according to a standardized procedure, which are applied to the process after a plausibility check by the operator. The MLbOA is implemented within an example for the assisted setup procedure of forming tools in a roll forming process based on force, torque and position data. During the setup process, the system suggests an ideal load-anticipated tool position to the operator, with a mean absolute error of 1.26 ± 0.36 μm. This is achieved by training an NN with the characteristic play and compliance behaviour based on the tool movement of the system under process load. The use of the assistance system shortens the conventional iterative adjustment process with load-free setup according to the setup plan and iterative adjustment after load application by digital knowledge generation. Since the NN is trained during the very first initial setup process of a new tool set and during ongoing production, this advantage is already realized from the second setup process. These results suggest that ML-based operator assistance systems are a promising approach to meet the challenges posed by the shortage of skilled labor in the manufacturing of profiled products. Before industrial implementation, a full-factorial data acquisition of the different gap setups should be conducted, to explore the models capability of predicting more than one changed adjustment simultaneously, as interactions between multiple factors are not investigated so far.

Due to the already established in-line capability of the model, the assistance system is able to locate misalignments during the process, even after the setup procedure. Future investigations will extend this opportunity to avoid downtimes due to misalignments caused by suddenly occurring influences such as fluctuations in semi-finished products, effects of wear and other process impacts. A further application is intended to be the integration of in-line profile geometry adjustment by a sensor-equipped straightener. This enhancement will enable the operator to use the framework beyond the presented setup procedure even for the profile geometry as the assistance system’s target.