Dual-Dimensional Adaptive Extraction Method for Mechanical Equipment Acoustic Features Under Variable Working Conditions

In industrial environments with limited measurement space, noncontact measurement methods and feature extraction techniques based on sound signals are increasingly emphasized. However, variable working conditions lead to inconsistent distribution of equipment acoustic features and domain shifts. The existing noncontact acoustic measurement methods have limitations, such as an insufficient ability to express acoustic features and poor adaptability. To overcome these disadvantages, a dual-dimensional adaptive method for extracting mechanical equipment acoustic features is proposed in this article. An innovative multidomain feature vector saliency and feature reconstruction method is proposed to generate a multidomain spectrogram (MDS), and a space–spectrogram joint feature set based on the MDS, Mel spectrograms, and Mel-frequency cepstral coefficients (MFCCs) is constructed to nonlinearly map the high- and low-frequency acoustic features of mechanical equipment. A multigranularity feature dimension-adaptive extraction (MGFAE) method based on the difference of Gaussians (DoG) pyramid and multihead attention mechanism (MHA) is proposed to realize the significant feature representations of multispectrogram of geometric and semantic information under variable working conditions. A dynamic convolution dimension-adaptive extraction (DCAE) method is proposed to capture the dynamic temporal and channel correlations, realize the dynamic adjustment of the convolution kernel and set in depth and width, and weaken the influence of working conditions. The applicability of dual-dimensional adaptive extraction method for mechanical equipment acoustic (MSODDAE) was validated through the iFLYTEK rotating machinery challenge dataset and our laboratory dataset. The final experiments revealed that this method has outstanding superiority under various unknown speeds and loads, contributing a new perspective for noncontact mechanical equipment measurement and acoustic feature extraction.