TRANSFER LEARNING ROLLING BEARING FAULT DIAGNOSIS METHOD BASED ON DEEP DOMAIN ADAPTIVE NETWORK

. The variation in motor operating conditions in nonsmooth dynamical systems leads to discrepancies in the distribution of condition monitoring data and label acquisition. To address this issue, we employ the transfer learning domain adaptation method and propose a deep domain adaptive network for assessing the state of motor rolling bearings in cross-working conditions. Initially, the vibration signals are preprocessed by FFT and SincNet to extract key features related to state assessment and perform feature fusion. Subsequently, a convolutional neural network is built to map the known label features of both the source and target domains into a shared space, enabling the extraction of common features. Employing the CBAM attention mechanism to enhance the network's feature expression capability aids in comprehending and utilizing input information more effectively. To enhance domain confusion, we combined the Maximum Mean Difference (MMD) and Correlation Alignment (CORAL) as novel measures for reducing discrepancies in the feature distribution of the same state under varying working conditions. Lastly, I-Softmax loss is utilized to improve network evaluation capability. Experimental results from multiple migration tasks conducted on two datasets demonstrate the excellent performance and adaptability of the proposed method in cross-condition fault diagnosis, surpassing the transfer learning method's focus on global domain adaptation.