Abstract
Introduction: Multimodal motor fault diagnosis is affected by scale inconsistency among heterogeneous signals, the coexistence of short-time fault features and long-sequence dependency features, and differences in modality availability. This study aims to develop an induction motor fault-diagnosis method based on principal component analysis (PCA) and a Multimodal Completeness- Aware Fusion Network (MCAFNet), which explicitly uses modality-availability information as a completeness condition for high-level gated fusion, rather than treating zero-masked placeholders of unavailable modalities as reliable observations. Methods: Raw vibration, acoustic, temperature, and current signals are uniformly preprocessed. Modality- wise PCA compression and modality-mask construction are performed to generate unifiedlength compressed temporal representations and modality-availability indicators. A dual-unit collaborative modeling framework is then developed using a Short-Time Fault Feature Unit (STFFU) and a Long-Sequence Dependency Unit (LSDU) to extract short-time fault features and long-sequence dependency features, respectively. A Completeness-Aware Gated Fusion Module (CAGFM) encodes modality masks into completeness embeddings to condition the gated fusion of high-level features according to the modality-availability state. Results: On the full five-class test set under the observed acoustic-availability distribution, the proposed method achieves an accuracy of 99.64% and a macro-F1 score of 99.62%, exceeding the strongest in-study comparator, PCA-MCNN, by 3.42 percentage points in accuracy (99.64% vs. 96.22%). macro-F1: By encoding the modality-availability mask, CAGFM conditions the gated fusion process on the modalities present in each sample. Discussion: The results indicate that explicitly modeling modality availability helps prevent zeromasked placeholder features associated with unavailable acoustic measurements from being treated as reliable observations during feature fusion. The evidence supports the effectiveness of the proposed availability-conditioned fusion under the acoustic-availability distribution represented in the current dataset. Conclusion: The proposed method effectively integrates heterogeneous multimodal information and modality-completeness conditions and achieves high diagnostic performance under the evaluated dataset and test protocol. Its robustness under cross-load, variable-speed, and strong-noise conditions, as well as under arbitrary missing rates, temporally contiguous missing blocks, simultaneous loss of multiple modalities, and single-modality-only inputs, remains to be evaluated.