Data-driven airflow velocity prediction for air-jet loom main nozzles using an interpretable INRBO-optimized CNN-BiLSTM-Attention model

Air-jet weaving is a highly efficient textile manufacturing process in which compressed air propels the weft yarn, and the aerodynamic performance of the main nozzle directly affects weaving efficiency and fabric quality. However, strong nonlinearity between nozzle structural parameters and internal airflow behavior makes accurate velocity prediction difficult. To address this problem, this study proposes an interpretable INRBO-optimized CNN-BiLSTM-Attention framework for data-driven airflow velocity prediction in air-jet loom main nozzles. The framework combines CNN-based local feature extraction, BiLSTM-based modeling of ordered structural-parameter dependencies, and attention-based adaptive feature weighting. An improved Newton-Raphson-based optimizer (INRBO), incorporating Tent chaotic initialization, adaptive inertia weighting, and logistic chaotic perturbation, is introduced to automatically tune key hyperparameters and enhance convergence stability under limited-sample conditions. Airflow velocity measurements at 12 positions in the main-nozzle jet field were used for model construction and validation. In repeated tests on Measurement Point 1, the proposed model achieved the best predictive performance, with an RMSE of 5.6247 ± 0.1065, an MAE of 4.5428 ± 0.1016, and an R 2 of 0.9875 ± 0.0010, outperforming conventional machine learning, deep learning, and other optimization-based hybrid models. Statistical tests confirmed the significance of these improvements. SHAP analysis identified the conical section length (Lz1) and the major diameter of the nozzle core cone (Dz1) as the dominant structural factors, with average SHAP values of 40.46 and 20.80, respectively. The proposed approach provides an accurate, automatically optimized, and interpretable tool for rapid main-nozzle structural evaluation and optimization.