Complex Fault Diagnosis of Rolling Bearings Based on Improved Residual Shrinkage Networks

Failures of mechanical components are characterized by high concurrency, high coupling, and feature frequency overlapping, etc. To solve the above problems and improve the diagnostic efficiency and accuracy, an improved residual contraction neural network for composite fault diagnosis of bearings is proposed. Null convolution and residual contraction blocks form a feature extraction layer, superimposed multiple contraction blocks form the whole residual contraction network, while migration learning is applied in the method to realize the mapping from source domain to target domain, so that the diagnostic model fault features are shared in the source and target domains, and the experimental results show that: the migration learning algorithm is combined with the DHRSN algorithm, and the composite fault diagnosis in the environment of variable operating conditions reaches 96% or more. Through theoretical deduction and bench test, the correctness and reliability of the improved residual shrinkage network method in the composite fault diagnosis of variable operating conditions are verified, promoting practical engineering applications.