Robust Modeling of Machine Vibration Based on Control Loop Data for Predictive Maintenance

Predictive Maintenance is a crucial technique to reduce machine downtime. One challenge is the absence of labeled run-to-failure data. This thesis explores whether predicting a system’s physical response to external impulses can help detect anomalies, especially when only the machine’s normal behavior is known. For that, we show that the simple mass-spring-damper model explains most of the axis vibrations of an exemplary production machine. With this finding, we propose a semi-supervised anomaly detection approach, TCN-AD, which predicts a sensor signal based on parallel recorded signals. Our study reveals that the TCN-AD sensor signal prediction is quite accurate for normal data. The prediction error of TCN-AD serves as a score that is useful for identifying anomalies, as supported by our comparison with other anomaly detection algorithms. For one of the datasets that we looked at, we show that a simple FFT-based score gives similar results to the anomaly scores produced by LSTM-AD and our own TCN-AD. This similarity is reflected by the Pearson correlation coefficient, which is larger than 0.94 for all combinations of TCN-AD, LSTM-AD, and FFT-Score.