19 February 2024 in IEEE Transactions on Industrial Informatics
Fault diagnosis of dynamic multivariate systems is a challenging problem. In this article, a novel fault diagnosis scheme based on variable-wise stacked temporal autoencoder (VW-STAE) is proposed. First, a variable-wise strategy is proposed on the raw industrial data, which sorts the variables for a specific fault by its deviation factor and introduces fault label information during pretraining procedure. Then, temporal autoencoder (TAE) is designed to capture the temporal and spatial feature synchronously and model the complex dependencies of dynamic samples. The stacked TAE is built to enhance the ability of feature extraction by combining multiple TAEs. By inputting the sorted variables sequentially, the VW-STAE is trained as a binary classifier for a specific fault; thereby its input variables and the corresponding network parameters are ultimately selected according to the VW-STAE with the optimal diagnosis performance. Finally, a bank of VW-STAEs is adopted for all faults, which is followed by a fully connected layer to achieve comprehensive fault diagnosis result. The effectiveness of the proposed method is demonstrated in the sensorless drive diagnosis example. The results indicate that the proposed method outperforms other existing deep learning methods.
4 June 2025 in IEEE Sensors Journal
Fault diagnosis is critical to flight safety for multi-rotor unmanned aerial vehicles (UAVs). While Graph Neural Networks (GNNs) have shown validity in modeling complex structural data, conventional approaches relying on undirected graphs and local message passing struggle with graph heterophily and limited receptive fields. To overcome these limitations, this paper proposes DIGPGNN—a novel GNN architecture that integrates Directional Information and Global Perception Fusion for enhanced UAV fault diagnosis. Initially, a directed graph is constructed based on feature similarity to reduce graph heterophily. Subsequently, an Adaptive Bi-directional Aggregation GNN (ABAGNN) is employed to adaptively aggregate the feature information of bidirectional neighbor nodes, thereby enhancing the model’s focus on homophilic neighborhood nodes. Finally, the Efficient Graph Transformer (EGT) incorporates a global self-attention strategy to learn the optimal graph structure by connecting distant homophilic nodes. Experimental results validate the effectiveness of DIGPGNN, achieving a 4.01% and 1.88% improvement in classification accuracy on public and self-built UAV datasets, respectively.
13 September 2025 in Knowledge-Based Systems
12 March 2026 in IEEE Transactions on Artificial Intelligence
Hypergraph neural network (HGNN) excels at representing intricate relationships within non-Euclidean infinitedimensional spaces. However, traditional hypergraph frameworks are grounded in homophily assumption and encounter challenges in handling heterophily problem in semi-supervised classification tasks. To tackle this problem, this paper first analyzes the underlying reasons that limit homophily-based hypergraphs, and proposes a heterophily-based dynamic structure. Under this structure, a Heterophily-based Dynamic HGNN (HDHGNN) is then proposed for semi-supervised classification. HDHGNN decomposes the hypergraph’s triple pattern iterative process into two causally related sub-procedures, within which attention mechanisms are incorporated to enable dynamic interactions. To evaluate the effectiveness of HDHGNN, a relative entropybased heterophily degree evaluation method is first introduced, followed by comprehensive fault classification experiments on two real-world flight datasets from rotary-wing and fixed-wing unmanned aerial vehicles, with comparisons against 11 state-of-the-art methods.