Hydraulic axial piston pump is broadly-used in aerospace, ocean engineering and construction machinery since it is the vital component of fluid power systems. In the light of the undiscoverability of its fault and the potential serious losses, it is valuable and challenging to complete the fault identification of a hydraulic pump accurately and effectively. Owing to the limitations of shallow machine learning methods in the intelligent fault diagnosis, more attention has been paid to deep learning methods. Hyperparameter plays an important role in a deep learning model. Although some manual tuning methods may represent good results in some cases, it is hard to reproduce due to the differences of datasets and other factors. Hence, Bayesian optimization (BO) algorithm is adopted to automatically select the hyperparameters. Firstly, the time–frequency images of vibration signals by continuous wavelet transform are taken as input data. Secondly, by setting some hyperparameters, a preliminary convolutional neural network (CNN) model is established. Thirdly, by identifying the range of each hyperparameter, BO based on Gaussian process is employed to construct an adaptive CNN model named CNN-BO. The performance of CNN-BO is verified by comparing with traditional LeNet 5 and improved LeNet 5 with manual optimization. The results indicate that CNN-BO can accomplish the intelligent fault diagnosis of a hydraulic pump accurately. •An adaptive CNN model is constructed for fault diagnosis of hydraulic piston pump.•Using pump ontology information realizes the non-destructive condition monitoring.•The selection of model hyperparameters is completed by employing Bayesian theory.•Intelligent fault diagnosis is based on the time–frequency images of vibration signals.•CNN-BO represents high fault diagnosis accuracy and good generalization ability.