关键词: 质量预测，概率密度分布，注意力机制，可解释性，多元时间序列

Abstract: In modern product manufacturing processes, process data often exhibit high dimensionality, non-linearity, and complex temporal dependencies, significantly complicating model development. These challenges hinder traditional methods from effectively extracting key process features, thereby limiting the accuracy and interpretability of product quality predictions. To address these challenges, this paper proposes an explainable product quality prediction model leveraging a probability distribution-based attention mechanism. Initially, a dynamic feature attention mechanism based on probability density distributions is developed, enabling the model to handle the intricate relationships between multi-dimensional features and accurately identify the critical factors influencing product quality at each time point. Additionally, a time attention mechanism is incorporated to capture important time points within the temporal sequence. By integrating an encoder-decoder framework with Long Short-Term Memory (LSTM) networks, the model is capable of effectively processing the complex temporal data inherent in product manufacturing, thereby enhancing predictive performance. Lastly, the explainable dynamic attention weights offer insights into the influence of both feature and temporal factors on product quality indicators. Empirical results from practical case studies demonstrate that, compared to traditional models, the proposed approach yields significant improvements in prediction accuracy. Moreover, its inherent explainability provides a robust methodology and actionable path for quality control in product manufacturing processes.

Key words: quality prediction, probability density distribution, attention mechanism, interpretability, multivariate time series