A novel heterogeneous coupled neuron network and its application in mechanical fault feature enhancement

Abstract

In practical engineering, fault characteristic signals are often overwhelmed by strong background noise and complex working conditions, which poses a serious challenge for timely detection and accurate identification of faults. Therefore, a novel heterogeneous coupled neuron network (HCNN) is proposed and the network using the alternating deep reinforcement learning algorithm (ADRL) is optimized to ultimately enhance the feature frequency of bearing fault diagnosis and identify faults. Gaussian neurons and hyperbolic tangent neurons are coupled by a coupling factor of δ to form coupled neurons, associating with the same type of neurons using a coupling discharge mechanism among multiple coupled neurons, respectively. The optimal output of each neuron is regarded as the model's final output, and theoretical analysis is carried out to assess the performance of HCNN. In the on-site measured data of wind turbine bearings, the HCNN enhancement method proposed achieved a signal correct classification rate of 98%. Compared with the original signal and the signals processed by EMD, SR, 1D-CNN, and LSTM, the correct classification rates were improved by 54%, 3%, 4%, 39%, and 31%, respectively. In the bearing data of DDS experimental platform, the correct classification rates of the signal processed by the proposed HCNN enhancement method is 99%, which is 57%, 1%, 14%, 2% and 15% higher than the original signal and the signal processed by EMD, SR, 1D-CNN and LSTM, respectively. The data shows that HCNN based on ADRL algorithm has strong enhancement of feature frequency and fault recognition ability, providing a new approach for mechanical fault feature enhancement.