Acoustic emission diagnosis of lubricant water contamination fault by using adaptive multi-fidelity Bayesian optimization

In rotating machinery, water contamination in lubricants can lead to oil film failure and exacerbate equipment wear. The acoustic emission (AE) technology demonstrates superior performance over traditional detection methods, such as spectral and vibration analyses, for early fault detection due to its sensitivity to weak signals. This paper introduces an adaptive multi-fidelity Bayesian optimization approach for detecting water contamination in lubricants. The technique utilizes a detection index (DI) in conjunction with the interquartile range (IQR) statistical filtering for denoising AE signals, thereby effectively improving the signal-to-noise ratio. An adaptive penalty parameter, dynamically adjusted with each training epoch, is designed to regulate the hyperparameters of the Bayesian optimized convolutional neural network (CNN), including learning rate, mini-batch size, and training epochs. Experimental results demonstrate that the proposed method effectively denoises AE signals and achieves a high recognition accuracy of 99.15% for the lubricant water contamination fault. Compared to CNNs without hyperparameter optimization, the accuracy improves by 18.48%, while computational time reduces by 7.35% relative to the Bayesian optimization without the adaptive penalty parameter. This approach achieves an optimal balance between the rapid exploration and robust convergence, enabling more efficient hyperparameter optimization and reliable identification of lubricant water contamination across varying moisture levels.