A Hierarchical Fault Inversion Framework for Complex Marine Systems Based on Attention-fused CNN-BiLSTM

Against the backdrop of increasingly complex ship systems, accurate fault diagnosis and localization have become core requirements for ensuring stable operation. Commonly used fault diagnosis models have distinct characteristics: CNNs excel at extracting spatial features but tend to overlook temporal information; RNNs can capture dynamic processes but fail to handle multidimensional spatial information at the same moment. This study proposes a hierarchical diagnostic procedure based on an attention-fused CNN-BiLSTM: (1) establishing a dual-channel structure with parallel CNN and BiLSTM to synchronously extract spatiotemporal features; (2) introducing an attention mechanism to dynamically weight and fuse features, focusing on critical information; (3) extending the fused model into a multilevel framework to achieve layered localization from system phenomena to specific faults. The effectiveness and inversion accuracy of this method have been fully validated through case studies on a publicly available ship fault dataset.