| Summary: | In the rapidly evolving Industry 4.0 space, predictive maintenance is shifting towards data-driven techniques. This shift is driven by advanced computing, reduced costs of sensing, abundantly available data as well as maturing machine learning algorithms. In particular, deep learning, a subset of machine learning, has been rapidly growing compared to traditional machine learning approaches. This is mainly due to its ability to automatically extract features and its high performance in tackling complex problems. Furthermore, predictive maintenance data are often unlabeled because labeling relies heavily on expensive domain expertise. As such, unsupervised techniques are gaining more popularity compared to their supervised counterpart. Therefore, the work at hand focuses on deep unsupervised anomaly detection.
Our work capitalizes on previous work conducted at MIT to develop an automated fault detection algorithm that was shown to work with high detection accuracy across diverse applications such as rolling element bearings, plasma etching machines, and milling machines.
To further explore the ability of the algorithm to generalize to new applications, we consider the problem of anomaly detection in belt-driven blower-motor units due to variable belt tension. To this end, we instrument a belt-driven motor-blower testbed and generate a dataset featuring electrical and vibration time series data. The dataset contains nominal and anomalous instances at different belt tension and motor speed values.
Applying the automated fault detection model to the dataset initially shows that belt problems can be detected but only with a limited accuracy of 12.5%. Upon further experimentation, an accuracy of 64.22% is achieved using a tuned set of hyperparameters on a subset of the data that contains nominal and no-belt conditions only. We conclude that additional hyperparameter tuning may be required in order for the existing algorithm to generalize well to our application. Finally, the dataset and testbed presented here will contribute to exploration of future anomaly detection techniques for time series data.
|
|---|