Frequency-enhanced network with self-supervised learning for anomaly detection of hydraulic piston pumps

Abstract

This paper presents a novel fault diagnosis method to address the challenge posed by the variable operating conditions of hydraulic piston pumps. The proposed method incorporates three key features. First, a frequency-enhanced network (FENet) is developed in this research to consider characteristics from both time and frequency domains simultaneously. FENet overcomes the limitations of conventional convolution operations through its frequency-aware convolution architecture, which enhances feature-extraction capabilities in the frequency domain. Second, a new CutMix-based self-supervised learning approach is proposed to achieve robust generalization performance across varying operating conditions and equipment variations. Third, a novel health index that combines the Mahalanobis distance and the Fisher discriminant ratio is developed to enhance the training stability of high-dimensional latent vector representations during model training. To validate the proposed method, experiments are conducted on industrial hydraulic piston pumps, focusing on wear defects between the slipper shoe and the piston ball under different temperature and pressure conditions. The generalization performance of the proposed method is evaluated through cross-validation across multiple equipment units. Feature map analysis reveals that vibration intensity increases in lower-frequency regions associated with fault severity, while it decreases in a relatively higher-frequency range. These analytical results demonstrate that the proposed method can serve as an effective condition monitoring tool for hydraulic piston pumps in industrial applications. © 2025 Elsevier Ltd