Fault diagnosis of agricultural machinery bearings based on multimodal feature fusion

Highlights Analyzing the modals with different frequencies in the bearing vibration signal through mode decomposition. Generating time-frequency domain feature for each modal through continuous wavelet transform. Proposing a multi-scale convolutional neural network to extract and fuse multimodal features. This method improves the diagnostic performance. The innovative feature visualization technique indicates that the proposed method can effectively extract multimodal features. Abstract. A bearing is the precise part in agricultural machinery equipment whose function is to ensure rotation accuracy and reduce wear. Diagnosing bearing faults is a strategy to maintain the mechanical equipment since faults will cause abnormal vibration and affect the operation of the machine. The existing diagnostic models lack sufficient supervised information for training, which leads to limited diagnostic accuracy. In response to this challenge, this research proposed a method based on multimodal feature fusion to diagnose bearing fault. Firstly, mode decomposition decomposes the original bearing vibration signal into multiple modals with different frequencies. Modals can characterize the frequency domain information of the original signal. Then, transform each modal into a time-frequency graph by using continuous wavelet transform. Establish a convolutional neural network to extract multiple modal features in the frequency domain and time-frequency domain. Finally, fuse multimodal features and diagnose bearing fault types. Through conducting experiments and verification on the bearings inside the driving part of the GK80 harvester, the accuracy of the proposed method in this research reached 99.75%. Compared with the existing methods, multimodal feature fusion technology can extract and represent features more effectively. The proposed method can monitor the state of the bearing inside the driving part of the operating agricultural machinery equipment intelligently; it can also provide a theoretical basis for applied engineering in related fields. Keywords: Bearing fault diagnosis, Frequency domain, Mode decomposition, Multiple modal features, Time-frequency domain.