Axial piston pump fault diagnosis under variable working conditions based on feature transfer learning

Fault data can always be varied with the working conditions of the axial piston pump. It is necessary to monitor the distributions of fault data for accurate fault diagnosis. Much effort has been made on fault diagnosis using transfer learning. But, most signals are collected using a single sensor with one specific function and measurement range. The all-around signals are still lacking for the operating state of the equipment from various aspects. In this study, a subspace embedded distribution adaptation with coupled classifiers (SEDACC) method was proposed to improve the fault diagnosis performance of the axial piston pump under variable working conditions. The multiple sensors were first used to collect the time-domain signalsand then convert them into the frequency spectrum data by fast Fourier transform (FFT). The main and auxiliary datasets were then constructed, where the spectrum data usually presented the high dimensions with redundant information. A subspace alignment (SA) was used to project the main data of the source domain and the target domain into a common subspace. As such, the dimension reduction was realized for the less discrepancy of marginal probability distribution between the main data in the two domains. A principal component analysis (PCA) was directly used for the dimensionality reduction in the auxiliary data in the target domain. The weighted conditional maximum mean discrepancy (WCMMD) was used as a metric standard to adapt the conditional distribution for the low dimensional feature of the main data in both domains. Furthermore, the WCMMD provided the different weights to the different categories in the process of adaptation, particularly different from the conditional maximum mean discrepancy (CMMD). A structural risk minimization (SRM) was selected as a main classifier to learn from the labelled data in the source domain. Meanwhile, an auxiliary classifier was learned from the auxiliary data in the target domain, according to the prediction from the main classifier. A weighted fusion of two classifiers was performed to obtain the final diagnosis after the continuous optimization of the classifier parameters using alternate and iterative strategies. The SEDACC fully realized the decision level fusion of multiple sensor information. The fault diagnosis experiment was conducted to adjust the outlet pressure of the axial piston pump under three working conditions of 5, 10, and 15 MPa. The signal acquisition was implemented using three mutually and vertically arranged vibration acceleration sensors (x, y, and z direction acceleration sensor) and a sound level meter. The z-direction vibration signal was then used to construct the main dataset, while the x-direction vibration signal or sound signal was for the auxiliary dataset. A higher recognition accuracy was achieved than before, particularly when there were only a few samples under the target working condition, indicating that the proposed SEDACC greatly improved the fault diagnosis performance under variable working conditions.