Induction motor fault classification using multimodal fusion and contrastive learning

Induction motor fault diagnosis is critical for reliable industrial systems. To tackle this challenge, this paper proposes a motor fault classification framework that combines multimodal fusion with supervised contrastive learning. The proposed approach consists of two key components. First, it converts vibration and acoustic signals into short-time Fourier transform spectrograms and performs image-level fusion to form a unified multimodal representation. Then, it applies a two-stage training pipeline: supervised contrastive pretraining of a convolutional encoder and a projection head, followed by a downstream classifier trained for fault classification. To validate the design, this paper conducts a systematic ablation study across four axes: sensor modality, signal representation, multimodal fusion strategy, and training strategy. More specifically, this paper evaluates two sensing modalities from three sensor channels (i.e., single-accelerometer, multiple-accelerometer, microphone-only), contrasts two representations (i.e., raw time-series, short-time Fourier transform spectrograms), assesses two accelerometer–microphone fusion strategies (i.e., feature-level fusion, image-level fusion), and compares two training strategies (i.e., supervised training, contrastive pretraining followed by supervised fine-tuning). Across these extensive evaluations on the UOEMD-VAFCVS benchmark dataset, the proposed approach achieves the best performance, demonstrating that time-frequency representation, multimodal image-level fusion, and contrastive representation learning jointly improve induction motor fault classification.