A strong noise-bearing fault diagnosis method based on A-CPO dual-stage associated parameter optimization

Abstract

In practical rolling bearing fault diagnosis, the bearing vibration signals collected by sensors are often mixed with a large amount of noise, which blurs key fault features and reduces the distinguishability among different types of faults. This issue not only undermines the reliability of feature extraction but also limits the identification accuracy of subsequent fault classification method. To address these issues, this article proposes a strong noise-bearing fault diagnosis method based on adaptive crested porcupine optimization (A-CPO) dual-stage associated parameter optimization. First, a dynamic multi-scale convolutional denoising autoencoder is constructed to undergo denoising preprocessing on the collected vibration signals. Second, to enhance the adaptability of the CPO algorithm in fault diagnosis, the A-CPO is proposed by designing and embedding a fault feature-driven operator. This algorithm can convert physical prior knowledge from signal processing into guiding information for the optimization process, thereby effectively driving the search process to converge rapidly toward fault-sensitive regions. Finally, a dual-stage collaborative optimization strategy based on A-CPO is designed: In the first stage, A-CPO’s global exploration capability is leveraged to perform a global coarse search for the key parameters K and α of variational mode decomposition (VMD), thereby determining the optimal combination of decomposition parameters. The denoised signals are then input into the optimized VMD to decompose them into multiple intrinsic modal functions (IMFs). In the second stage, leveraging A-CPO’s local exploitation capability, the support vector machine (SVM) penalty factor C and kernel parameter g are fine-tuned. The VMD parameters obtained in the first stage are fixed to maintain feature space stability. The optimal IMF is selected, and multi-dimensional time-domain indicators are extracted as feature vectors, which are then input into the optimized SVM to complete fault classification. Multi-model comparative experiments conducted on the Case Western Reserve University and Jiangnan University fault datasets demonstrate that the proposed method achieves average diagnostic accuracies of 99.31, 99.25, 95.14, and 91.92% at −2, −4, −6, and −8 dB, respectively, validating its high accuracy and reliability in rolling bearing fault diagnosis under strong noise conditions.