The processing of rubber is determined by numerous influencing variables. Due to the large number of control variables, batch-dependent material fluctuations, especially in natural rubber, and the existing system inertia, manual process control is prone to errors. In practice, control processes are mainly based on user experience and are neither automated nor digitized. For this reason, work is being carried out at the Institute of Transport and Automation Technology (ITA) in collaboration with the DIGIT RUBBER research network on digitizing the rubber extrusion process by implementing data mining and AI systems.
Data mining to identify correlations
For the development of an AI-based control system, process-related cause-effect relationships must first be identified in order to determine correlations between the control and measurement variables. Based on these correlations, specified tolerance limits for the respective measured variables can be met. An excellent approach for identifying cause-effect relationships is so-called data mining.
Data mining is a subfield of AI in which statistical models and artificial neural networks (ANNs) are used to identify patterns and trends from large amounts of data. Applications range from classification and cluster analysis to regression analysis. A significant advantage over other correlation analyses, such as the Pearson coefficient for linear models, is that ANN also identifies non-linear correlations. Particularly for the complex rubber extrusion process, which consists of the sub-processes mixing, rolling and extrusion (see Fig. 1), this depth of analysis is highly relevant in order to be able to consider the largest possible set of correlations in the modeling. Another challenge is the high system inertia that occurs with a number of control variables, such as range temperature control. As a result, certain set temperatures can sometimes only be reached after several minutes. It is therefore necessary to predict process-immanent cause-effect relationships as accurately as possible in order to be able to react to tolerance deviations accordingly at an early stage.
Learning process of data mining and AI systems
As part of the process modeling, the recorded measurement data are first summarized and classified according to their properties. Since the accuracy of the model strongly depends on the quality of the available data, erroneous data and data irrelevant to the use case must be filtered, redundancies removed and scaling methods applied. For the further support of the machine learning process, the hyperparameter optimization of the algorithm follows. Through an iterative approach, the control parameters of the learning process are adjusted to achieve the best possible results for the prediction of the rubber extrusion process. Depending on the complexity of the value to be predicted, for example, the topology of the neural network must be extended or the number of iteration steps for the weighting (batches) must be adjusted.
In the first instance, the validation of the data mining and AI models is done by using only a part of the test data (80%) for training, whereas the remaining test data (20%) serve as reference for the prediction. For sub-processes of the rubber extrusion line, ITA has already achieved initial successes in this context (see Fig. 2). The figure shows the measured torque as a function of time for the mixing process (see Fig. 3) of the rubber extrusion chain. The training (orange) and test (blue) curves of the data mining model, which are calculated exclusively from the control variables, already show a good approximation of the real curve (black). Thus, it is possible to predict control-relevant measured variables. The prediction accuracy of the data mining model is to be increased in the following steps by additional training data and further model optimizations and transferred to other processes.
AI-based control to minimize process deviations
The objective of AI-based control is to ensure that predefined tolerance limits are adhered to on the basis of the cause-effect relationships identified in data mining. Like the data mining model, the AI-induced control of the rubber extrusion line is also based on artificial neural networks. Here, however, measured variables are used as input data for the AI model in order to calculate a default control variable as a function of the specified tolerance limits and to counteract deviations accordingly. Finally, the calculated default control variables are to be validated by the data mining model and, if sufficiently accurate, forwarded to the hardware interface.
The development and implementation of the AI-induced control system in the existing rubber extrusion plant is planned for 2023. In cooperation with the DIGIT RUBBER research association, a hardware interface will be integrated into the system to validate the AI-induced control under real production conditions. Successful implementation can reduce the reject rate of rubber extrusion lines and sustainably increase process stability.