Chain breakage fault diagnosis in scraper conveyors based on CNN-BiLSTM-Attention

Existing deep learning approaches to chain fault diagnosis in scraper conveyors mostly use a single neural network model. They have limitations including a narrow range of extracted features, insufficient extraction of deep latent fault features, and inadequate fault representation and generalization, making it difficult to fully capture complex fault characteristics during chain operation. To address these problems, a hybrid neural network model integrating a Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and an Attention mechanism (Attention) for chain breakage fault diagnosis in scraper conveyors was proposed and applied to chain breakage fault diagnosis in scraper conveyors. First, operating data, including the rotational speed, torque, and current of the head and tail motors, and chain tension data were collected. The raw data were then preprocessed by outlier removal, linear interpolation, low-pass filtering, and downsampling to obtain high-quality continuous time series. Finally, the preprocessed data were input into the CNN-BiLSTM-Attention model. The CNN extracted local fault features from the signals, BiLSTM captured long-range temporal dependencies, and Attention adaptively increased the weights of key features, enabling accurate identification of three states: normal operation, incipient chain breakage, and chain breakage. Experimental results showed that the CNN-BiLSTM-Attention model outperformed the single neural network models in accuracy, recall, and F₁ score, achieving 96.92%, 96.77%, and 96.75%, respectively. It also achieved the highest classification accuracy for each of the three chain states. These results indicate that the model can effectively reduce the false-alarm rate in scraper chain breakage fault diagnosis.