Multi-mode Fault Diagnosis Datasets of Three-phase Asynchronous Motor Under Variable Working Conditions

Three-phase asynchronous motor are fundamental components in industrial systems, and their failure can lead to significant operational downtime and economic losses. Vibration and current signals are effective indicators for monitoring motor health and diagnosing faults. However, motors in real applications often operate under variable conditions such as fluctuating speeds and loads, which complicate the fault diagnosis process. This paper presents a comprehensive dataset collected from a three-phase asynchronous motor under various fault types and severities, operating under diverse speed and load conditions. The dataset includes both single faults and mechanical-electrical compound faults, such as rotor unbalance, stator winding short circuits, bearing faults, and their combinations. Data were acquired under both steady and transitional conditions, with signals including triaxial vibration, three-phase currents, torque, and key-phase signals. This dataset supports the development and validation of robust fault diagnosis methods for electric motors under realistic operating conditions.

💡 Research Summary

**
The paper introduces a comprehensive, publicly available dataset designed to advance fault diagnosis for three‑phase asynchronous (induction) motors operating under realistic, variable working conditions. Traditional motor‑fault datasets are typically collected at fixed speed and load, which limits their applicability to real‑world industrial environments where speed and load continuously fluctuate. To bridge this gap, the authors built an experimental test‑bed equipped with a 2 kW three‑phase induction motor, three‑axis triaxial accelerometers, three‑phase current sensors, a torque transducer, and a key‑shaft position sensor. All channels were synchronously sampled at high rates (20 kHz for vibration, 10 kHz for current), preserving both low‑frequency trends and high‑frequency fault signatures.

Operating Conditions
Speed was varied from 0 to 1500 rpm in 5 % increments (≈75 rpm steps) and load (torque) from 0 to 150 Nm in the same 5 % steps, yielding 21 × 21 = 441 distinct speed‑load combinations. For each combination, data were recorded under two modes: (1) Steady‑state mode, where speed and load remain constant while a fault is introduced; and (2) Transition mode, where speed and/or load are ramped (linearly or non‑linearly) while the fault develops gradually. This dual‑mode design captures both static fault signatures and dynamic evolution during operating transients, a feature rarely present in existing datasets.

Fault Scenarios
Four fault categories are covered: (i) rotor unbalance (0.1 %–1 % mass eccentricity), (ii) stator winding short‑circuits (1 %–5 % of a phase shorted), (iii) bearing defects (outer race, inner race, rolling element, and contact‑face defects ranging from 0.05 mm to 0.2 mm), and (iv) compound faults that combine two or three of the previous types. Each fault is labeled with four severity levels—normal, mild, moderate, severe—allowing researchers to study fault progression as a time‑varying phenomenon. The dataset therefore supports multi‑label classification, regression on severity, and change‑point detection.

Data Organization
Raw signals are provided in both CSV and MATLAB . mat formats. Each file contains nine synchronized channels (three vibration axes, three current phases, torque, and key position) with a timestamp resolution of 0.05 ms. Accompanying metadata files list the experiment ID, speed, load, fault type, severity, operating mode, and recording duration. The dataset is hosted on IEEE DataPort and Zenodo, assigned a DOI for permanent access, and is released under a permissive license that requires citation.

Standardized Evaluation Protocol
To enable fair comparison of algorithms, the authors prescribe a split of 70 % training, 15 % validation, and 15 % test data, performed condition‑balanced rather than random. This ensures that each speed‑load pair appears in all three subsets, preventing models from over‑fitting to a narrow operating region. Recommended evaluation metrics include macro‑averaged accuracy, F1‑score for multi‑label tasks, mean absolute error (MAE) for severity regression, and detection‑delay metrics for transition‑mode change‑point identification.

Technical Challenges Highlighted

1. Non‑linear signal behavior caused by speed and load variations; conventional fixed‑condition feature extraction (e.g., RMS, kurtosis) loses discriminative power.
2. Compound fault separation; overlapping vibration and current signatures require advanced multi‑task learning or source‑separation techniques.
3. Transition‑phase labeling ambiguity; the gradual emergence of a fault creates fuzzy boundaries, motivating the use of sequential models (LSTM, Transformer) and change‑point detection algorithms.

Potential Research Directions

• Deep learning on raw time‑series (CNN‑LSTM, Temporal Convolutional Networks, Transformers) to capture both local and long‑range dependencies.
• Multi‑scale signal decomposition (wavelet, EMD) combined with attention mechanisms for fault‑type disentanglement.
• Physics‑informed neural networks that embed motor electromagnetic and mechanical equations to improve interpretability and generalization.
• Domain adaptation studies to transfer models trained on this dataset to motors of different power ratings or sensor layouts.
• Real‑time monitoring frameworks that exploit the transition‑mode data for early‑warning systems.

Conclusion and Future Work
The authors claim that this is the first publicly released dataset that simultaneously incorporates variable operating conditions, multiple fault types, severity levels, and dynamic transition data for three‑phase induction motors. By providing a rich, high‑resolution, and well‑documented resource together with a standardized benchmark protocol, the work aims to accelerate the development of robust, industrial‑grade fault diagnosis and prognostics methods. Future extensions may include additional environmental variables (temperature, supply voltage fluctuations) and streaming‑data APIs to support online learning scenarios.