Predictive Maintenance with Deep Learning Algorithms in Industry 4.0

The key to reliably predicting machine errors lies in identifying recurring data patterns prior to historically documented machine malfunctions and failures. In this case, the data science specialists relied on a deep learning model. Deep learning, a method from the field of machine learning, uses artificial neural networks to analyze large amounts of data of varying complexity at multiple levels to identify complex relationships. The more data used to train the deep learning model, the more precise the results will be, as the learning algorithm optimizes itself during the analysis process.

Based on this method, eoda developed a multi-label classification model to register the occurrence of specific error messages in sensor data. Several error messages were divided into classes and analyzed over specific time intervals for descriptive analysis. The programming language R was chosen for this purpose, which is particularly suitable for conducting a proof of concept when there are a large number of individual variables.

Data preprocessing was performed beforehand: Sensor data measurements from different sources were aggregated, scaled, and transformed to create a uniform standard for data comparison. Only after this step could the data be analyzed, since each sensor measures at different intervals and times. These measurements had to be processed in such a way that the data from different information sources could be transferred into a uniform format and also be used on a consistent basis across machines.

The next step was to create a meaningful division of information into training, test, and validation data. Due to the uneven distribution of errors in the data, it was necessary not only to take random samples but also to use down- and upsampling to obtain a meaningful distribution of errors and non-errors. The resulting learning effect from training, testing, and verification is the most important step in model building – especially to avoid under- and overfitting.

Due to the big data scenario and the long computation time, the cloud-based workstation Amazon Web Services (AWS) was integrated to accelerate the process. The first prototype enabled the identification and evaluation of various layered architectures. Deviations resulting from the preprocessing phase were thus improved. The technical framework for the deep learning framework in this case was provided by a Keras API written in Python, which provides an interface for the TensorFlow deep learning library.

In the final step, the evaluation of a classifier based on relative frequency was calculated. For this purpose, a confusion matrix was created with a focus on cost minimization. The values ​​derived from the matrix indicate the frequency of occurrence of feature combinations.