A light deep adaptive framework toward fault diagnosis of a hydraulic piston pump

•A light deep framework is constructed for fault diagnosis of a piston pump.•Multi-sensor and multiple channel signals are monitored for information mining.•Adaptive learning of model hyperparameters is achieved by Bayesian algorithm.•The design of batch normalization can improve the stability of the model.•The proposed diagnosis method can accurately achieve the fault identification. The health condition of hydraulic axial piston pumps is crucial for the safety and reliability of hydraulic transmission systems. Diagnosis results of traditional methods indicate the high reliance on the experience, and current deep model-based methods are confronted with the difficulty of parameter tuning. A light adaptive deep framework is therefore constructed to reduce reliance of diagnosis on expert experience and still achieve the automatic screening of model hyperparameters. First, multi-sensor and multiple-channel signals of a piston pump are acquired for comprehensive raw data input. The raw one-dimensional signals are then converted into two-dimensional images using continuous wavelet transform. Then, a light deep model is built based on the convolutional operation and batch normalization techniques. The final model is obtained via global optimization of Bayesian algorithm. Next, the improved deep model is adopted for failure recognition of the essential components in the piston pump based on the transformed images. The average diagnosis accuracy achieves 97.46%, 98.71%, and 99.94% based on vibration signal, acoustic signal, and pressure signal respectively. The results reveal that the typical fault of the piston pump can be recognized intelligently and accurately with the proposed diagnosis method.