Dual-channel 1D-CNN-BiLSTM transfer learning for diagnosing tooth breakage across circular saw blade wear stages

As the core cutting tool for dismantling stainless steel cladding of spent fuel assemblies, the circular saw blade is prone to tooth breakage failure, which can directly lead to reduced machining accuracy, decreased production efficiency, and even equipment safety accidents. Furthermore, dynamic changes in the wear stage of the circular saw blade cause shifts in the distribution characteristics of cutting signals, posing a challenge for domain adaptation in cross-wear stage tooth breakage diagnosis. To address the issues of scarce fault annotation data for each wear stage in actual production and the insufficient generalization ability of traditional models, this paper proposes a transfer learning fault diagnosis method based on a dual-channel 1D-CNN-BiLSTM network. First, the method effectively preprocesses dual-channel vibration signals during circular saw blade cutting, including effective signal extraction and sliding window framing, to construct a source-domain fault dataset for new saw blade tooth breakage and target-domain datasets for each wear stage. Subsequently, a hybrid 1D-CNN-BiLSTM network is designed, where 1D-CNN extracts multi-scale local spatial features of the saw blade cutting signals, and BiLSTM captures bidirectional temporal dependency features. This network is pre-trained in the source domain as a general fault feature extraction model. Finally, a fine-tuning strategy combining layer replacement and adaptive learning rate factors is adopted, freezing the bottom feature extraction layers and retraining the top classification layer, thereby achieving efficient transfer of fault knowledge from the source domain (new saw blade tooth breakage) to the target domain (various wear stages). Using three types of circular saw blade tooth breakage as the research object and comparing with methods such as SVM, 1D-CNN, LSTM, ResNet and DANN, experimental results show that the proposed method achieves an accuracy of 97.9% in cross-wear stage tooth breakage fault diagnosis, significantly outperforming other methods. It effectively solves the transfer diagnosis problem of tooth breakage faults across wear stages and improves the cutting efficiency of stainless steel cladding of spent fuel assemblies.