Extremely imbalanced data intelligent fault diagnosis of rotating impeller with improved denoising diffusion probabilistic models

The operating time of rotating impellers in different faulty conditions is much shorter than its normal operating condition, and obtaining signals of faulty impellers is extremely difficult due to safety and cost. Therefore, the issue of highly imbalanced data between normal and fault status samples of rotating impellers frequently arises, which significantly obstructs the application of intelligent fault recognition and classification models. This paper addresses this challenge by proposing a novel approach that integrates improved denoising diffusion probabilistic models (IDDPM) with improved convolutional neural networks (ICNN) to solve extremely imbalanced data in fault diagnosis. IDDPM is built to synthesize one-dimensional time-domain signals stably to augment the original training set. Meanwhile, a nonlinear denoising strategy and a non-autoregressive network suitable for time-domain signal generation were introduced to enhance the quality of the generated signal. In addition, an ICNN with strong feature extraction and identification capability is designed to classify mixed datasets. The experimental results indicate that the IDDPM stably generates data that closely resembles actual data, and the ICNN can more effectively extract features from the mixed dataset, leading to a significant improvement in fault diagnosis performance. Accordingly, the proposed approach can effectively solve the low accuracy issue in extremely imbalanced data fault diagnosis of rotating impellers.