BILSTM based on quality driven attention for product defect prediction

Deep learning-based soft sensors play a crucial role in modern industrial process monitoring and measurement by establishing predictive models between high-quality data and easily measurable variables. However, traditional deep learning methods encounter challenges such as the high dimensionality of process variables, fluctuations in predictor variables, and weak correlations among variables. Addressing these issues, this paper introduces a novel approach called Quality-Driven Bidirectional Attention Mechanism LSTM (QBILSTM). The model prioritizes variables related to quality indicators in the input data by calculating the maximum mutual information coefficient (MIC). During training, it further enhances the importance of key variables by integrating quality indicator-related information through an attention mechanism. Moreover, to capture trends in quality variables comprehensively, the model incorporates a two-way propagation structure, and hyperparameter selection is optimized effectively using the Bayesian optimization algorithm. Finally, the superiority of QBILSTM is demonstrated through testing in real industrial processes.