Research on Lightweight Fault Diagnosis Method Under Small Sample Conditions

To address the problem of effectively extracting multi-modal features from the non-stationary vibration signals of aero-engines, this paper proposes a lightweight multimodal fault diagnosis method for aero-engine under the condition of small samples. Firstly, the simplified residual contraction network is utilized for extract the time domain abrupt feature. Secondly, the continuous wavelet transform is employed to construct the timefrequency maps, and the attention mechanism is employed to extract the frequency domain abrupt feature from these maps; Finally, classification is achieved by integrating the dual-mode features in the fully connected layer. Experimental validation conducted on the aero-engine bearing data set of Harbin Institute of Technology demonstrates that the proposed model attains a diagnostic accuracy of $99.38 \%$ with a parameter size of only 1.00 M, confirming its merits of high accuracy and lightweight architecture under complex operating conditions.