An innovative neural network architecture designed for industrial fault diagnosis with hierarchical adaptive attention mechanism

Fault diagnosis research is crucial for guaranteeing the reliability and safety of modern industrial systems. Given the challenges of strong temporal dependencies, nonlinear dynamics, and noisy high-dimensional data inherent in chemical processes, this paper proposes a novel Multi-Scale Attention Bidirectional LSTM (MA-BLSTM) framework. Specifically, a multi-scale convolutional front-end is employed to capture temporal dependencies at different resolutions, and the extracted representations are then refined through bidirectional LSTM layers with residual connections to effectively model long-range dynamics. An adaptive attention mechanism is further introduced to emphasize fault-indicative temporal segments while suppressing irrelevant fluctuations, thereby enhancing diagnostic robustness under noisy and imbalanced conditions. The framework is systematically validated on the Tennessee Eastman Process (TEP) benchmark through experiments covering representative faults, complex dynamic scenarios, stress tests, and large-scale multi-fault diagnosis. The results confirm that the framework achieves superior performance in accuracy, robustness, and convergence efficiency, providing a generalizable solution for intelligent chemical process monitoring.