A Novel Multiscale Petal Pattern and CNN-BiLSTM Structure for Acoustic Signal-Based Rolling Bearing Fault Diagnosis

Abstract Rolling bearings play a critical role in industries such as wind power, rail transit, and aerospace, where their health directly impacts equipment reliability. Despite the growing use of machine learning for fault diagnosis, challenges remain under nonstationary conditions—particularly in robust feature extraction under variable speeds and generalization across transacoustic modes. While deep learning models like convolutional neural network (CNN) and long short-term memory (LSTM) can learn directly from raw signals, they require large labeled datasets and suffer from speed fluctuations. In this article, a novel multiscale petal pattern (MSPP) method is proposed which reconstructs the original one-dimensional time-domain signal into a two-dimensional image signal. By selecting the key scale parameter L in the MSPP reconstruction process, a loudness-based characteristic extraction (LCE) method is introduced to identify the impact features of the signal. Obtaining the features by MSPP, a bidirectional long short-term memory network (BiLSTM) method is introduced to diagnose the fault automatically. As the MSPP method can supply rich fault information and has a stronger discrimination ability in neural network models compared to traditional time-frequency methods, the result of BiLSTM combining MSPP shows high ability in diagnosis. The experiment validated that the diagnostic accuracy of the proposed method can reach 98.44% under typical operating conditions, while other methods such as CWT + CNN reach 86.5%.