An Experiment on Anomaly Detection for Fault Vibration Signals Using Autoencoder-Based N-Segmentation Algorithm

Most manufacturing facilities driven by motors generate vibration and noise representing critical symptoms against facility malfunctioning conditions in the manufacturing industry. Due to the difficulty of obtaining abnormal data from facilities in manufacturing sites, many prior researchers who have studied predicting facility faults have adopted unsupervised learning-based anomaly detection approaches. Although these approaches have a strength requiring only data on from facility normal behaviors, it is not clear that the anomalies detected by an anomaly detection model are due to the real component faults. Also, the model performance is likely to change according to the diverse abnormal conditions of the given facility. In this paper, we took an experiment with a fault vibration simulator to measure the anomaly detection performance of a one-dimensional convolutional autoencoder model with different fault conditions. In the experiment, we used four different abnormal conditions: imbalance, misalignment, looseness, and bearing faults, which are the most frequently occurring facility component failures from the rotating machineries. Data were gathered from the simulator with the IEPE(Integrated Electronics Piezo-Electric) type sensor. We proposed the N-Segmentation algorithm that performs anomaly detection in segmented frequency region according to corresponding component faults for better anomaly detection performance. In conclusion, the proposed algorithm showed about 15 times better anomaly detection rate than not applying it.