Curve-Registration-Based Feature Extraction for Predictive Maintenance of Industrial Equipment

With the growing adoption of Internet of Things (IoT), predictive maintenance is gaining momentum for ensuring the reliability of industrial equipment. A common practice of predictive maintenance is to conduct feature extraction on the original sensor data, then conduct deep learning to train predictive maintenance model with the extracted data and finally, conduct prediction by model. Because of the low value density of industrial sensor data stream, feature extraction is usually based on dimensionality reduction. However, traditional methods for dimensionality reduction seldom consider time-lagged correlations which are very common among industrial sensor data streams. More importantly, time-lagged correlations are less sensitive to the traditional dimensionality reduction methods, leading to poor effect of feature extraction. In this paper, we propose a feature extraction method based on curve registration to deal with the time-lagged correlation problem. Our experimental results indicate that our method can: (1) effectively improve the accuracy of prediction; and (2) improve the performance of the prediction model.