A Self-attention-Based Optimized Temporal Convolutional Network for Weaving Machine Faults Prediction

Weaving machines play a vital role in the rapid development of the textile industry. Accurate prediction of loom and component failure times can enhance their operational reliability and productivity while effectively reducing maintenance costs. However, the time-series data on loom failures usually exhibit nonlinear and non-smooth characteristics, which complicate accurate prediction. Therefore, this paper proposes an improved temporal convolutional network (TCN) based on the attention mechanism for predicting loom faults. First, the self-attention mechanism is employed to weight input features at each time step, enhancing feature extraction efficiency; then, a TCN is constructed to capture long-term dependencies in the data and derive feature representations from the weighted features. Subsequently, a tuna optimizer is introduced to fine-tune TCN hyperparameters, further enhancing model prediction accuracy. The proposed method is validated using a dataset of textile factory equipment failures, with comparative experiments conducted across several mainstream neural networks including RNN, GRU, LSTM, and the original TCN. The results indicate that the proposed method demonstrates superior performance in loom fault prediction when compared to existing methods.