Application of Lightweight CNN-LSTM-Attention Model in Defect Detection During Thread Tightening Process

Robotic thread tightening is a critical assembly process in automobile manufacturing, and tightening defects (such as under-tightening or thread damage) directly impact product safety and reliability. Existing deep learning detection methods generally suffer from high computational complexity, making them difficult to deploy on assembly line edge devices. To address this issue, this paper proposes a lightweight CNN-LSTM-Attention (LW-CLA) model, which achieves real-time defect detection based on torque-angle-time sequences collected from tightening guns. Firstly, a lightweight data preprocessing pipeline is designed, including dynamic time warping (DTW) sequence alignment with reduced search radius, feature engineering retaining only five core features, and efficient data augmentation. Secondly, the model adopts a simplified network structure: convolutional channels are reduced from 32/64 to 16/32, a single-layer bidirectional LSTM with 32 hidden units is used, along with a two-head lightweight attention mechanism. Finally, SMOTE oversampling is applied to alleviate class imbalance, and CPU-oriented optimization is performed. Experiments on the Atlas tightening gun dataset (264 normal samples, 80 defective samples) demonstrate that the proposed model achieves 97.12% accuracy and a 94.12% F1 score on the test set, with a model size controlled under 1 MB. Compared with standard deep learning models, its computational complexity is reduced by 68% while maintaining high detection performance, meeting the real-time deployment requirements at the edge. Pending ablation and comparative experiments will further validate the effectiveness of each lightweight module and demonstrate its superiority over mainstream methods.