TY - JOUR
T1 - Explainable CNN-Based Cavitation Detection in Axial Piston Pumps Using Grad-CAM Visualization
AU - Michael Roberts
AU - Sarah Thompson
AU - James Anderson
AU - Robert Smith
JF - Journal of Trustworthy and Explainable Intelligent Systems
JO - J Trustworthy Explain Intell Syst
Y1 - 2024
VL - 3
IS - 2
SP - 55
N2 - The reliability and safety of an entire hydraulic system are compromised by the occurrence of cavitation in axial piston pumps. Because vibration signals can indicate cavitation states within these pumps, they have been increasingly coupled with machine learning techniques for detection purposes. Nevertheless, raw vibration signals captured under actual operating conditions are frequently corrupted by noise, complicating the precise identification of cavitation within noisy environments. To address this issue, this study introduces an intelligent methodology designed to identify cavitation in axial piston pumps under high-noise conditions. Initially, a convolutional neural network (CNN) is trained utilizing spectrogram images that are converted from raw, unstructured vibration datasets across varying cavitation states. Subsequently, the gradient-weighted class activation mapping (Grad-CAM) technique is applied to highlight and visualize the class-discriminative zones within these spectrogram images. Ultimately, a novel image processing approach leveraging the Grad-CAM heatmap is introduced to automatically filter out background noise and amplify the key class features in the spectrograms. The empirical outcomes demonstrate that this strategy substantially boosts the diagnostic efficacy of the CNN framework under noisy conditions. Specifically, the accuracy of cavitation condition classification is elevated from 0.50 to 0.89 at a signal-to-noise ratio of 4 dB, and from 0.80 to 0.92 at 6 dB.
UR - https://tipublications.com/v586540710
ER -