Fault diagnosis of air conditioning compressor bearings using wavelet packet decomposition and improved 1D-CNN

Abstract

This paper presents an innovative fault diagnosis method for air conditioning compressor rolling bearings, employing acoustic signals through a tailored integration of wavelet packet decomposition (WPD) and an advanced one-dimensional convolutional neural network (1D-CNN). Traditional methods typically rely on vibration signals and contact sensors, which are often impractical for compressor bearings. Despite their weakness and noise susceptibility, the study leverages acoustic signals as a noncontact alternative. The paper utilizes WPD to extract multiresolution features from acoustic signals to tackle these challenges, effectively capturing subtle fault signatures across frequency bands. These features are processed by an improved 1D-CNN, optimized with an attention mechanism, residual networks, and domain adaptive learning, achieving 100% recognition accuracy on original signals and 95.49% under a −10 dB signal-to-noise ratio (SNR), compared to 81.6% for the baseline 1D-CNN (a 13.89% improvement). This approach outperforms alternative methods, with WPD and the improved 1D-CNN yielding up to 3.2% higher accuracy than empirical mode decomposition (EMD), variational mode decomposition (VMD), and fast multiscale decomposition (FMD) on original signals and maintaining robustness in noisy conditions where these methods falter. By exceeding the performance of alternative feature extraction methods like VMD, EMD, and FMD, particularly in adverse environments, the study provides a robust and adaptable solution for reliable bearing fault diagnosis, facilitating preventive maintenance and enhancing system longevity.