Enhanced feature extraction method for motor fault diagnosis using low-quality vibration data from wireless sensor networks

Abstract

Wireless sensor networks (WSNs), which are usually powered by batteries, have been extensively used in condition monitoring and fault diagnosis of motors. To extend the battery service life, the length of the acquired and transmitted signal should be short and the sampling resolution should be reduced. In this case, the motor signal quality is low, which affects the fault diagnosis accuracy. To address this issue, this study proposes an enhanced feature extraction method for motor fault diagnosis using low-quality vibration signals acquired from a battery-powered WSN node. First, the vibration signal is converted to an image using a wavelet synchrosqueezed transform technique. Second, the constructed image is enhanced using a histogram equalization. Finally, the enhanced image is inputted into a convolutional neural network (CNN) model, and the motor fault type can be recognized from the CNN output. The effectiveness and efficiency of the proposed method are validated by comparing its performance in the brushless direct motor test rig with the performance of several traditional methods. The relationship between the fault diagnosis accuracy and WSN performances is investigated and discussed. The proposed method shows potential applications for remote motor fault diagnosis using the low-quality vibration signal acquired from a WSN node with limited battery capacity.