Multimodal Feature Learning and Interpretable Feature Selection for Elevator Fault Diagnosis

Modern intelligent building systems require reliable vertical transportation solutions. Equipment malfunctions in elevator systems can significantly impact operational safety and passenger experience. Developing efficient diagnosis approaches becomes crucial for minimizing service interruptions. Current condition monitoring techniques, primarily designed for gear systems, require substantial adaptation for elevatorspecific applications. This research proposes an integrated feature learning framework combining multidomain signal analysis with interpretable machine learning. Our methodology extracts temporal characteristics, spectral energy distributions, and system-specific frequency components from triaxial vibration data. These multimodal features feed into a kernelbased classifier for condition classification. To enhance model interpretability and efficiency, we employ cooperative game theory for feature importance analysis. Experimental validation demonstrates three key advantages: (1) Triaxial signal processing captures spatial-temporal patterns more effectively than single-axis approaches, improving diagnosis precision by $3.2-5.7 \%$ across test cases. (2) Incorporating domain-specific frequency signatures boosts classification accuracy by $4.1 \%$ compared to generic feature sets. (3) Our feature selection mechanism reduces dimensionality by $60-95 \%$ while maintaining 96.7\% mean accuracy, significantly lowering computational overhead.