Modified residual network architecture for early-stage piston-cylinder wear detection in hydraulic axial piston pumps

Abstract

Axial piston pumps are widely used in aerospace and construction machinery because of their high volumetric efficiency, compact design, and high-pressure, variable-flow capability. However, their closely mating components are susceptible to wear, making early detection essential. This paper presents an artificial intelligence-based modified residual network (ResNet18) model incorporating a Convolutional Block Attention Module (CBAM) and Generalized Mean (GeM) pooling for early-stage wear detection at 50 μm in the piston-cylinder arrangement of an axial piston pump (APP). The experimental dataset comprises one healthy reference condition with a clearance of 10 μm and five mechanically induced piston-cylinder wear conditions of 50, 80, 100, 180, and 400 μm, resulting in a six-class wear classification problem. For each wear condition, 25 min of discharge pressure signal data were recorded at a sampling frequency of 1 kHz. Five minutes were used for model training, while four independent runs of 5 min each, totalling 20 min, were reserved for unseen testing. The discharge pressure signals were transformed into Continuous Wavelet Transform, Gramian Angular Difference Field, and Markov Transition Field representations. Each representation was processed through a dedicated ResNet18-CBAM branch, and the extracted features were combined via feature-level fusion, followed by GeM pooling and classification. On the independent 20-min unseen dataset, the proposed model achieved an average accuracy of 98.52% across five independent seeds and a best accuracy of 99.32%. The results demonstrate strong early-wear detection with computational efficiency, highlighting the model's potential for real-world intelligent wear monitoring in safety-critical applications, such as aircraft hydraulic systems.