Real-time Vibration Signals Anomaly Detection of Load Motor Based on Deep Learning and One-Class Classification

The objective of this research is to build a robust fault detection system that is applicable to load motor for various working conditions. What’s more, during the model training stage, only vibration signals under normal conditions are used, as it can be difficult to obtain abnormal vibration signals in practical scenarios in the manufacturing industry. In this paper, the aging phenomenon of the load motor is simulated by adjusting stiffness. The research methodology is comprised of two stages: vibration signal prediction and anomaly detection. Three different networks were used for vibration prediction: Recurrent Neural Network (RNN), Gate Recurrent Unit (GRU) a combination of One-Dimensional Convolutional Neural Network (1DCNN) and GRU. Calculating statistical features, such as Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), to assess predicted results and served as input features for anomaly detection. In the stage of anomaly detection, The Support Vector Data Description (SVDD) is a method used to determine a damage threshold, indicating that machines exceeding this threshold are considered damaged.