Anomaly Detection with Convolutional Autoencoder for Predictive Maintenance

Abstract

Predictive maintenance is set to prevent downtime and failures of equipment and processes to meet the quality and availability requirements of several industrial, commercial, and even residential activities. This paper proposes CAE-AD, a novel convolutional autoencoder anomaly detection method that relies only on normal operation data for training the intelligent classi-fier. The method also accommodates a sliding window algorithm for generating the input from sensor readings, which accounts for the dynamic characteristics of the data. The anomaly detection is accomplished by comparing the convolutional autoencoder reconstruction error to a threshold value to segregate between normal and anomalous predictions. The threshold value is found by minimizing the False Positive Rates and False Negative Rates. Using a benchmark water pump sensor time-series data, the model successfully classified all water pump breakdowns and correctly identified 98.8% of anomalous data and 94.8 % of normal data using a chosen best window length of the past 37 sensor readings.