An intelligent fault diagnosis method based on L2 regularization and deep transfer LSTM

With the advancement of the modern manufacturing industry, fault diagnosis has become increasingly critical, especially for rotating machines operating under varying working conditions (WCs). While numerous deep learning methods have been proposed, they often require extensive labeled data for training, which is challenging due to data scarcity and limited label availability. Moreover, their performance tends to deteriorate when applied to different domains. To overcome these issues, this paper introduces an intelligent fault diagnosis technique that leverages L2 regularization and deep transfer learning with LSTM networks, capable of adapting to different environments. The approach involves a pre-training phase followed by fine-tuning, where knowledge from the pre-trained model is transferred and adjusted for new working conditions. The study finds that fine-tuning all layers of the model results in minimal variation, with accuracy within 0.05%, indicating high consistency. In contrast, fine-tuning only the final classification layers shows a broader range of accuracies, approximately within 6%, indicating moderate consistency—a conclusion further supported by t-SNE feature visualization.