Deep Learning Based Fault Recognition and its Source Analysis for Dynamic Industrial Processes

A spatially organized input sequence and a branching LSTM-based encoder-decoder model are used to anticipate production defects and identify their origins in this study. This technique uses sensor data from many manufacturing phases to detect abnormalities more precisely than standard time-series models. This approach improves prediction performance and root-cause analysis by using Attention Mechanisms (AM) to manage lengthy sequence inputs and focus on the most important data, boosting fault detection accuracy and interpretability. This method outperforms existing models in empirical assessment using the Tennessee Eastman Process (TEP) dataset with high macro-average F1 scores across many fault categories. The model detects 20 fault kinds using 41 sensor inputs, achieving high accuracy and recall. The Attention Mechanism also illuminates production defect causes, making the model quantitatively and qualitatively useful. The approach yields outstanding results, but its computing needs may hinder real-time applications. Optimizing the model for lower resource usage without reducing accuracy will make it better for real-time industrial deployment. This work shows that LSTM-based architectures with attention mechanisms can improve manufacturing defect identification and root-cause investigation.