A Mechanical Fault Identification Method Based on 3D Feature Construction and 3D Dilated Residual Network
Most existing methods for mechanical fault recognition based on audio signals often use two-dimensional neural networks and a single audio feature(such as power spectrum)for fault detection.However,the extraction of a single audio feature may lead to the loss of critical information during the process.Moreover,it typically captures only a single dimension of audio feature information(such as spatial dimension),greatly limiting the effectiveness of current algorithms for device fault audio analysis.To address these issues,this study proposes a three-dimensional feature construction method that includes various audio features to compensate for the loss of critical information during feature extraction.Furthermore,a three-dimensional dilated residual network model(DR-3 DCNN)is constructed,employing dilated convolutions to enhance the model's global attention and capture features at different scales.By fully exploiting the correlations among different features,a deep-level association between features and original audio data is established.Finally,experiments are conducted using the publicly available industrial machine investigation and inspection dataset(MIMII).The experimental results demonstrate that the combination of three-dimensional features and DR-3DCNN significantly improves the classification performance of mechanical fault recognition,achieving higher classification accuracy than previous single audio feature recognition algorithms.
万方数据
维普
