Multirate Signal based Fault Diagnosis of Hydraulic Systems with Residual Shrinkage Separable Convolutional Network

For the multi-rate problem of data caused by different sampling rates of sensors in the hydraulic system, the multi-rate data is difficult to be directly input into a single neural network for inference training because of the difficulty of aligning its time dimension. Therefore, this paper proposes a diagnostic framework for hydraulic systems based on multi-rate signals. A residual shrinkage separable convolutional network is utilized to capture the fault characteristics of the system's high-rate data and the network has fewer parameters compared with the traditional convolution, and combined with a statistical feature module, the time and frequency domain features of the low-rate data are extracted. Secondly, a multi-rate feature fusion module based on gated attention mechanism is designed to capture the fault features at different levels, and the adaptive weighting mechanism is used to aggregate the feature information at different levels and obtain the fusion features. The experimental results show that the framework can still obtain high diagnostic accuracy in the more extreme case of data sampling rate imbalance ratio of 1: 100, and the average fault recognition accuracy also reaches 99.55% and 99.55% in the diagnosis of complex faults such as pumps and accumulators, respectively.