Long short-term memory autoencoder-based anomaly detection system for electric motors

Abstract

Predictive maintenance (PdM) systems have the potential to detect underlying issues in electric motors, and this can allow them to prevent production downtime and loss of manufacturing yield. However, majority of the PdM systems for electric motors that have been proposed so far are unsuitable for industrial implementation, since they require hours of manual data collection and annotation, and are unable to account for more than one type of motor fault. Therefore, this thesis presents an unsupervised long short-term memory (LSTM) autoencoder-based anomaly detection system for electric motors. It analyzes the vibration and current consumption data from motors to detect anomalies, which is sufficient to account for the various motor defects. Aside from this, it can adapt to varying operating conditions. The system is created to autonomously collect vibration and current consumption data from the motor, and then use the data to train the LSTM autoencoder model and deploy it in real-time to detect anomalies. In addition to this, the system comes with several features including personal computer and web user interfaces that enable ease of access as well as remote monitoring of the motor’s conditions. To test the system, a hardware test bench using a stepper and a brushless direct current (BLDC) motor is made to simulate defective conditions. LSTM autoencoder models are trained on the data from this setup and deployed once the training is completed. If the system detects increasing rate of anomalies, the users are informed through an email or a short message service notification. The presented anomaly detection system is tested on hardware test bench. Based on the experimental results, as the simulated defect worsened, the rate of anomalies detected by the system increased, with the maximum anomaly rate reaching 7 anomalies per second. Additionally, the LSTM autoencoder technique is also compared with principal component analysis and isolation forest for validation purposes, and it proved to be the most accurate in the case of both the stepper and BLDC motors with accuracies of 66.24% and 86.43% respectively.