Fault diagnosis plays an important role in the safe and stable operation of rotating machinery, which is conducive to industrial development and economic improvement. However, effective feature extraction of rotating machinery fault diagnosis is difficult in the complex sound field with characteristics of reverberation and multi-dimensional signals. Therefore, this paper proposes a novel acoustic feature extraction method of the rotating machinery based on the Weighted Prediction Error (WPE) integrating the Linear Constrained Minimum Variance (LCMV). The de-reverberation signal is obtained by inputting multi-channel signals into the WPE algorithm using an adaptive optimal parameters selection function with the sound field changes. Then, the incident angle going from the fault source to the center of the microphone array is calculated from the full-band sound field distribution, and the signal is de-noised and fused using the LCMV. Finally, the fault feature frequency is extracted from the fused signal envelope spectrum. The results of fault data analysis of the centrifugal pump test bench show that the Envelope Harmonic Noise Ratio (EHNR) is more than twice that of the original signal after the WPE-LCMV processing. Compared to the Recursive Least Squares and the Resonance Sparse Signal Decomposition (RLS-RSSD) and the parameter optimized Variational Mode Decomposition (VMD), the EHNR has a higher value for all types of faults after applying the WPE-LCMV processing. Furthermore, the proposed method can effectively extract the frequency of bearing faults.
18 June 2025 in IEEE Transactions on Instrumentation and Measurement
The clean signal is used as the reference signal in acoustic signal denoising methods based on the supervised deep learning, but the clean signal of the operating state of the rotating machinery is difficult to obtain, which leads to difficulties in constructing the denoising model. Therefore, the novel acoustic signal denoising method based on the clean signal virtual sample (CSVS) integrating the deep neural network (DNN) (CSVS-DNN) is proposed. The frequency band that contains the most fault information is selected based on the sideband characteristics of the modulation signal. Then, the CSVS dataset is generated based on the distribution of amplitude variations of the acoustic signal, the fault characteristic frequencies, and the transmission paths in the mechanical structure and air. Moreover, the generated CSVS dataset and the actual device operation signal are used to construct the acoustic signal denoising model based on the DNN. Finally, the denoising effect is evaluated using the signal-to-noise ratio (SNR) and the prominence degree coefficient (PDC) through the experimental data and the actual industrial data. The analysis results indicate that the average SNR of the proposed method is improved by at least 0.6 dB, and the average PDC is enhanced by at least 0.05 compared to other denoising methods.
6 February 2026 in IEEE Transactions on Industrial Informatics
In response to the dual bottleneck of difficulty in obtaining the clean signal and extracting key features, which leads to poor denoising effect in mechanical equipment acoustic signal, a novel denoising method based on clean signal virtual sample (CSVS) integrating the dual-path convolutional augmented Transformer (DpConformer) is proposed. The frequency bands that contain more fault information are selected based on the cyclic spectral correlation. Then, the CSVS is generated through the fault characteristic frequency, amplitude distribution, as well as the transmission paths in the mechanical structure and air. Moreover, considering the distinct characteristics of the high- and low-frequency time-frequency features, the convolution module of the Conformer network is enhanced to a dual-path convolution module for facilitating the extraction of the deep global features, the deep low- and high-frequency features. The denoising model is subsequently constructed through the generated CSVS and the noisy signal. Finally, the effectiveness of the proposed method is evaluated using the signal-to-noise ratio (SNR), the prominence degree coefficient (PDC), and the warning accuracy through the experimental data and the actual industrial data. The analysis results show that the SNR, the PDC and the warning accuracy of the proposed method are improved by at least 0.25 dB, 0.04, and 2.5% compared to other denoising methods.