AI System Improves Induction Motor Fault Detection

Published in Discover Electronics, the study introduces a method named MFC², which integrates vibration and acoustic signals into a unified representation to distinguish between various electrical and mechanical failures. The approach achieves a test accuracy of 84.31% on a benchmark dataset, outperforming traditional supervised learning models that rely on direct cross-entropy training. This advancement addresses a critical challenge in industrial maintenance, where undetected motor faults can lead to substantial operational downtime and financial losses.

Induction motors are essential components in many industrial systems, and the global market for these devices was valued at USD 26.42 billion in 2024, with projections to reach USD 46.09 billion by 2032. Despite their widespread use, accurate fault diagnosis remains difficult because motors operate under diverse speeds, load conditions, and installation environments. Commercial condition monitoring systems often rely on single sensing modalities, such as vibration accelerometers or acoustic microphones, which may not capture the full spectrum of fault signatures. Electrical faults, such as stator winding faults and broken rotor bars, arise from electromagnetic stresses, while mechanical faults, like bearing degradation, originate from physical defects in moving components. These different fault types can produce overlapping vibration and acoustic signatures, making them difficult to separate using a single modality or a narrowly trained classifier.

The proposed MFC² framework consists of two main components: feature engineering and a two-stage training pipeline. In the feature engineering stage, raw sensor signals from accelerometers and microphones are segmented and transformed into short-time Fourier transform (STFT) spectrograms. These spectrograms provide a two-dimensional time-frequency representation that captures both temporal and spectral characteristics of the signals. The accelerometer and microphone spectrograms are then fused at the image level by assigning them to the red and green channels of a three-channel image, respectively, while the blue channel is set to zero. This encoding preserves modality-specific information in separate image channels while satisfying the standard input requirements of convolutional neural network architectures.

The training pipeline begins with a pretext stage where a ResNet-18 encoder and a projection head are trained using supervised contrastive learning. Unlike standard contrastive learning that relies on data augmentations, supervised contrastive learning uses class labels to construct positive and negative relationships within each training batch. Embeddings from segments belonging to the same fault class are pulled closer together, while embeddings from different classes are pushed apart. This encourages the model to learn class-discriminative representations that are less dependent on specific operating conditions. In the downstream classification stage, the pretrained encoder is fine-tuned on a classification head trained using cross-entropy loss to predict the specific fault class from the learned representations.

To validate the design, the researchers conducted a systematic ablation study across four axes: sensor modality, signal representation, multimodal fusion strategy, and training strategy. The study utilized the UOEMD-VAFCVS benchmark dataset, which contains vibration and acoustic measurements collected from a motor test rig under diverse operating conditions, including four constant-speed settings, four dynamic-speed profiles, and two load conditions. The dataset covers seven fault classes and one healthy baseline condition, with signals sampled at 42,000 Hz. The experimental design included eight configuration sets to isolate the effects of each technical factor, ranging from raw time-series baselines to multimodal fusion with contrastive pretraining.

The results demonstrated that time-frequency representations provide more transferable fault features than raw time-domain signals. A baseline using raw time-series data achieved a test accuracy of 41.53%, whereas the corresponding STFT spectrogram baseline achieved 56.39%. Further improvements were observed when combining accelerometer data from multiple placements, increasing test accuracy to 72.07%. The microphone-only baseline reached 60.18% test accuracy, indicating that acoustic signals contain complementary fault-related information. Feature-level fusion of accelerometer and microphone embeddings achieved a test accuracy of 68.59%, confirming that the two modalities contain complementary fault-discriminative information.

Image-level fusion, where spectrograms are stacked as separate channels of an image, yielded the best supervised result with a test accuracy of 80.33%. However, the broken rotor bar class remained the most persistently misclassified, with 45% of segments still predicted as voltage unbalance. The proposed MFC² configuration, which added supervised contrastive pretraining to the image-level fusion setup, achieved the best overall performance with a test accuracy of 84.31%. This represents a 3.98 percentage-point improvement over the best supervised configuration, demonstrating that contrastive pretraining improves the fused representation beyond direct cross-entropy training.

Statistical analyses of twenty repeated runs confirmed that the performance improvements of MFC² were consistent and not attributable solely to random training variability. The method achieved a statistically significant improvement in macro recall and macro precision compared to the supervised baseline. These findings provide evidence that the combination of time-frequency representation, multimodal image-level fusion, and contrastive representation learning jointly improves induction motor fault classification. The approach effectively reduces inter-class confusion, particularly for faults with overlapping spectral characteristics under unseen operating speeds.

Despite the overall improvements, certain fault classes remained challenging. Rotor misalignment and broken rotor bars continued to be the two worst-performing classes across all experimental sets. This pattern suggests that these classes produce time-frequency signatures that are particularly sensitive to the speed shift between training speeds and the held-out evaluation speeds. The researchers noted that further improvements are needed to distinguish fault conditions whose signal characteristics may overlap under varying operating speeds. The study highlights the importance of jointly designing signal representation, sensor fusion, and representation learning strategies for robust motor fault diagnosis.

The study has several limitations that must be considered when interpreting the results. The evaluation was based on a single laboratory dataset, UOEMD-VAFCVS, and did not include cross-dataset validation or testing under industrial operating conditions. Therefore, the results demonstrate classification performance under the evaluated dataset and operating conditions but do not establish generalization across different motor platforms or industrial environments. Laboratory-constructed datasets address the scarcity of labelled fault data by simulating faults under controlled conditions, but they introduce a lab-to-field generalizability gap where models may fail to generalize to the operating diversity of real-world environments.

Future work will investigate cross-dataset evaluation, validation using industrial data, and improved feature representations for distinguishing difficult fault classes under varying operating conditions. The code supporting the findings is available on a public GitHub repository, allowing other researchers to replicate and extend the work. This study contributes to the broader trend toward advanced learning architectures and data representations for intelligent manufacturing, offering a viable path for enhancing the reliability and maintenance efficiency of industrial induction motors.

The UOEMD-VAFCVS dataset used for this study was collected from a SpectraQuest motor test rig, providing a controlled environment for benchmarking fault diagnosis models. The dataset comprises 128 data files with signals sampled at 42,000 Hz, covering seven distinct fault classes alongside a healthy baseline condition. These faults are categorized into electrical types, such as stator winding faults, voltage unbalance with single phasing, and broken rotor bars, and mechanical types, including rotor unbalance, rotor misalignment, bowed rotor, and faulty bearings. The data collection process incorporated diverse operating conditions to test model robustness, specifically including four constant-speed settings at 15, 30, 45, and 60 Hz, as well as four dynamic-speed profiles that transition between these speeds. Additionally, the dataset includes recordings under two different load conditions, allowing for an evaluation of how varying mechanical loads influence fault signatures.

To ensure the model’s ability to generalize to unseen operating speeds, the researchers employed a specific data partitioning strategy. The 15 Hz and 60 Hz constant-speed conditions were held out entirely from the training phase and reserved exclusively for validation and testing. Each fault class was assigned to one held-out speed for validation and the complementary held-out speed for testing, with this assignment alternating across classes and reversing between load conditions. This approach ensured that each class contributed exactly two files to validation and two to test, including one unloaded and one loaded recording in each split. To prevent data leakage from dynamic-speed recordings that might overlap with the held-out constant-speed regimes, a one-second boundary trim was applied to the affected ends of specific training files. This trim removed the portion of the speed trajectory closest to the held-out conditions, creating an approximately 3-Hz separation between the retained training data and the evaluation points.

The experimental design included eight distinct configuration sets to systematically isolate the effects of sensor modality, signal representation, fusion strategy, and training methodology. Sets A and B established raw time-series baselines using 1D-CNN classifiers, with Set A utilizing only the near-motor accelerometer and Set B using data from either the near-motor or far-motor sensor. Sets C, D, and E served as unimodal STFT spectrogram baselines, where Set C compared raw time-series inputs directly against STFT spectrograms under the same near-motor accelerometer condition. Set D extended this by pooling accelerometer data from both placements, while Set E evaluated the microphone signal as a standalone acoustic input. Sets F and G assessed multimodal fusion strategies, with Set F applying feature-level fusion through independent feature extraction branches and Set G applying image-level fusion by stacking spectrograms as separate channels. Finally, Set H represented the proposed MFC² configuration, which utilized the same image-level fused input as Set G but incorporated supervised contrastive pretraining followed by fine-tuning.

Implementation details for the experiments were standardized to ensure fair comparison across all configurations. All models were implemented in PyTorch and executed on an NVIDIA RTX 4090 GPU with 16 GB of VRAM using CUDA 12.6. Mixed-precision training was employed across all sets, utilizing BF16 autocast where supported and enabling TF32 for matrix multiplication operations. The supervised training configurations used cross-entropy loss as the primary objective, with a weight decay of 1e-4 applied unless otherwise specified. Hyperparameters, including optimizer selection, learning rate, batch size, and learning-rate scheduler, were adjusted manually through iterative model training and comparison of validation performance. For the MFC² configuration, the encoder and projection head were trained using supervised contrastive loss with a temperature parameter of 0.2 for 100 epochs, followed by downstream fine-tuning with a learning rate of 1e-3 and a batch size of 64.

Subject of Research: Machine learning for induction motor fault diagnosis

Article Title: Induction motor fault classification using multimodal fusion and contrastive learning

Article References: Chan, Y. T., Soon, F. M., Teo, T. Y., Tian, J., & Xie, X. Z. (2026). Induction motor fault classification using multimodal fusion and contrastive learning. Discover Electronics, 3(1), Article 156. https://doi.org/10.1007/s44291-026-00309-5

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00309-5

Keywords: induction motor, fault classification, multimodal fusion, contrastive learning, STFT spectrograms, machine learning, vibration analysis, acoustic monitoring, ResNet-18, industrial maintenance, Induction, motor