Structural state identification based on multi-channel convolutional neural network and homology analysis

Traditional deep learning-based structural state identification methods are typically built upon a closed-set assumption, which limits their ability to handle previously unseen structural states in structural health monitoring. To address this limitation, this paper proposes a state identification framework based on homology analysis using a convolutional neural network, which reformulates the conventional multi-class classification task into a multi-channel state consistency learning problem. Normalized amplitude-frequency spectra are first extracted from time-domain vibration responses and then concatenated along the channel dimension (corresponding to different sensors or different time windows) to construct multi-channel samples. Each sample is then assigned a binary label of homologous or non-homologous depending on whether its constituent channels originate from the same structural state. Instead of learning a direct mapping from samples to fixed categories, the CNN learns inter-channel consistency relationships. As a result, the model can identify and issue warnings for previously unseen structural states based on patterns of inter-channel inconsistency. On the IASC-ASCE Benchmark, an overall open-set identification accuracy of 91.67% is achieved by the proposed method, as is demonstrated through experiments. The applicability of the framework to single-sensor monitoring is further confirmed by the CWRU bearing experiment. When the model is trained with three known states, an overall open-set accuracy of 98.67% is obtained. The potential of homology-based analysis for structural damage identification in cases where labeled damage data are scarce is highlighted by these findings, and relational homology learning is suggested as a promising alternative for structural state identification under incomplete state coverage.