Issue 3 (July)

With the extensive use of intelligent monitoring systems in contemporary power system, it is becoming a significant challenge to make the power grid operate stably in achieving high robust fault diagnosis of transformers across operating conditions and environmental domains. In this work, a transfer learning enhanced deep CNN-LSTM cross-domain transformer fault diagnosis model (TL-DCNN-LSTM) is presented. The model, built on the classical CNN-LSTM, applies Convolutional Neural Network (CNN) to capture multi-scale spatial features, which are from the original current (or vibration) signal, between the circuit and gearbox, and employs Long Short-Term Memory Network (LSTM) to learn the time-dependent patterns of the fault signal. To address the distribution discrepancy between the source domain and the target domain, the paper presents a transfer learning mechanism, where the Maximum Mean Discrepancy (MMD) loss is embedded in the deep feature space to reduce feature distribution differences between different domains. Meanwhile, a domain-wise batch normalization (DSBN) block is introduced to maintain the discrepancy in domain. Furthermore, an auxiliary domain discriminator and adversarial training strategy are proposed to enhance the domain irrelevance and discrimination of the feature representation. The TL-DCNN-LSTM model obtains a cross-domain diagnosis accuracy of over 96.7%, which is substantially higher than traditional CNN, LSTM and common fine-tuning methods.

As one of the most important parts of rotating machinery, the early prediction of rolling bearing is very important for the safety of operation and cost reduction of maintenance. To improve the intelligent level of bearing fault recognition, this paper proposes an intelligent bearing fault diagnosis frame work based on a hybrid CNN-LSTM-GRU network, which combines the advantages of CNN, LSTM, and GRU, and makes structural innovation and optimization design on the basis of the existing deep learning model. The proposed model follows a three-segment hybrid structure, in the first segment, 1DCNN is employed to extract local time-domain feature from original vibration signal. Also, We feed the features that have been extracted by the CNN network as the inputs to the both LSTM network and GRU network in parallel in order to take care of the multiple dependencies between the sequence of data items and make the model optimize more efficiently, respectively. 5.fusion multi-head attention module is introduced to perform feature weighing for the dual-channel outputs, thus further enhancing the capability of identifying complicated failure modes. To prevent overfitting and coalescence while enhancing the generalization capability of the model, the final classification module is built upon Dropout regularized fully connected design with residual connection. Experiments on the CWRU bearing fault dataset demonstrate superior classification performance of the proposed network under different load and fault severities, providing an average accuracy of more than 94.2%.