Chinese Researchers Advance Inverter Fault Diagnosis for Electric Vehicles Using CGAN-CNN Hybrid Model.
As the global automotive industry accelerates its transition toward electrification, reliability and safety in electric vehicle (EV) powertrains have become critical engineering frontiers. Among the core components of EV drivetrains, the three-phase inverter plays a pivotal role in converting direct current from the battery into alternating current that drives the motor. However, the operational complexity and high-stress environment in which inverters function make them susceptible to failures—particularly in power semiconductor switches. A new study published in the Journal of Power Supply demonstrates a significant breakthrough in diagnosing inverter faults under real-world constraints where fault data is inherently scarce and imbalanced.
The research, led by Sun Quan, Peng Fei, Li Hongsheng, Yu Xianghai from the School of Automation at Nanjing Institute of Technology, in collaboration with Sun Guodong from the College of Automation Engineering at Nanjing University of Aeronautics and Astronautics, introduces a novel diagnostic framework that combines conditional generative adversarial networks (CGAN) with convolutional neural networks (CNN). This hybrid approach directly addresses one of the most persistent challenges in automotive predictive maintenance: the lack of sufficient fault samples.
In practical EV operations, inverter faults—especially open-circuit failures in insulated-gate bipolar transistors (IGBTs)—occur infrequently and often last only milliseconds before protective circuits intervene. This brevity severely limits opportunities to collect representative fault data. Consequently, machine learning models trained on such datasets suffer from sample imbalance, where normal operation vastly outweighs fault instances. Traditional methods like SMOTE (Synthetic Minority Over-sampling Technique) or standard generative adversarial networks (GANs) have attempted to mitigate this issue but often fall short in preserving the nuanced distributional characteristics of high-dimensional fault signals.
The team’s innovation lies in leveraging CGAN’s ability to generate high-fidelity synthetic fault samples conditioned on specific failure modes. Unlike conventional GANs, which produce data without explicit control over output class, CGAN incorporates class labels during both generation and discrimination. In this study, seven distinct operational states were defined: one normal condition and six single-switch open-circuit fault scenarios corresponding to the six IGBTs (T1 through T6) in a standard three-phase bridge inverter. By feeding raw three-phase current signals—captured under varying motor loads (550, 650, and 750 rpm)—into the system, the researchers first transformed the time-domain data into the frequency domain using Fast Fourier Transform (FFT). This step concentrated fault-relevant information into a compact spectral representation, which was then normalized for consistent input scaling.
The CGAN model was trained on these labeled spectral features. The generator, structured with three convolutional layers and a fully connected output, learned to synthesize new fault samples that mirrored the statistical and spectral properties of real data. Simultaneously, the discriminator—comprising two convolutional layers, two pooling layers, and two fully connected layers—was tasked with distinguishing real from synthetic samples. Both networks were optimized using the Adam optimizer, and batch normalization was applied to stabilize training dynamics. After 100 training epochs, the CGAN achieved a stable equilibrium, producing synthetic data that closely aligned with the original distribution, as confirmed by t-SNE visualization of feature-space clustering.
These augmented datasets were then used to train a custom one-dimensional CNN architecture inspired by AlexNet but adapted for sequential signal classification. The CNN featured two convolutional layers (with 8 and 16 filters of size 3×1, respectively), two max-pooling layers (2×1 windows), a fully connected layer, and a Softmax output layer for seven-class classification. Notably, the model processed the frequency-domain vectors directly, bypassing complex handcrafted feature engineering.
To validate the method’s robustness, the researchers constructed four experimental datasets representing varying degrees of sample imbalance. Dataset A1 contained 100 balanced samples per class. Dataset A2 reduced the T6 open-circuit fault class to 50 samples, while A3 further reduced it to only 30—a severe imbalance reflecting real-world scarcity. Dataset A4 restored balance by augmenting the T6 class to 100 samples using three different techniques: SMOTE, standard GAN, and the proposed CGAN.
The results were striking. On the balanced A1 set, the CNN achieved 99.2% accuracy. Performance dropped to 95.7% on A2 and further to 91.4% on A3, underscoring the detrimental impact of data scarcity. However, when synthetic samples were introduced in A4, accuracy rebounded dramatically—but only with high-quality augmentation. SMOTE-augmented data yielded 95.31% accuracy, standard GAN reached 96.88%, while the CGAN-enhanced model attained 98.97%—nearly matching the performance on the original balanced dataset.
Crucially, the confusion matrices revealed that T6 faults were frequently misclassified as T2 faults under imbalance, likely due to symmetrical current distortion patterns in the three-phase system. The CGAN-generated samples preserved these subtle discriminative features, enabling the CNN to distinguish between visually similar fault modes. This capability is essential for real-world deployment, where misdiagnosis could trigger unnecessary shutdowns or, worse, allow latent faults to escalate into catastrophic failures.
From an automotive engineering perspective, the implications are profound. Modern EVs increasingly rely on over-the-air (OTA) software updates and cloud-based analytics for predictive maintenance. However, such systems require accurate fault models trained on diverse operational data. The scarcity of fault data has historically constrained the development of robust diagnostic algorithms, forcing engineers to rely on rule-based systems or physics-informed simulations that lack adaptability. The CGAN-CNN framework bridges this gap by enabling data-driven diagnostics even when physical fault occurrences are rare.
Moreover, the method aligns with industry trends toward modular, software-defined vehicle architectures. By operating on standard current sensor outputs—already present in most EV motor control units—the approach requires no additional hardware. The FFT preprocessing step is computationally lightweight and can be implemented on existing microcontrollers, while the CNN inference can be offloaded to the vehicle’s domain controller or performed in the cloud during routine diagnostics.
The study also addresses key concerns under the Google EEAT (Experience, Expertise, Authoritativeness, Trustworthiness) guidelines. The authors bring deep domain expertise: Sun Quan, the lead and corresponding author, specializes in power converter fault detection and holds a Ph.D. in electrical engineering. Peng Fei and Yu Xianghai focus on machine learning applications in fault diagnosis, while Li Hongsheng, a professor with decades of experience in automation systems, provides methodological rigor. Sun Guodong, affiliated with a top-tier aerospace university, contributes advanced knowledge in power electronics reliability. Their institutional affiliations—Nanjing Institute of Technology and Nanjing University of Aeronautics and Astronautics—are well-established in China’s engineering research ecosystem, with strong ties to the automotive and energy sectors.
The research was supported by the National Natural Science Foundation of China (Grant No. 61901212), the Major Project of Natural Science Research in Jiangsu Higher Education Institutions (20KJA510007), and the Jiangsu Collaborative Innovation Center for Smart Distribution Network Technologies and Equipment (XTCX201909)—funding sources that underscore the project’s strategic importance to China’s broader electrification and intelligent manufacturing initiatives.
While the study was conducted on a brushless DC motor (BLDCM) platform—a common choice for cost-sensitive EVs and industrial applications—the underlying methodology is readily transferable to permanent magnet synchronous motors (PMSMs), which dominate the premium EV segment. Future work may extend the framework to multi-fault scenarios, partial degradation tracking, and cross-domain adaptation across different vehicle platforms and operating environments.
Critically, the team avoided common AI pitfalls that undermine real-world deployment. There is no reliance on synthetic data alone; instead, the CGAN augments real, physically measured signals. The model architecture is transparent and reproducible, with no black-box ensembles or uninterpretable deep stacks. Validation was performed across multiple operational loads, ensuring generalizability beyond a single test condition.
For automotive OEMs and Tier 1 suppliers, this research offers a scalable path toward intelligent inverter health monitoring. As vehicle fleets grow and data accumulates, such models can be continuously refined, forming the backbone of next-generation EV durability and safety systems. Regulators, too, may find the approach compelling: with fault diagnosis accuracy exceeding 98%, the method could support certification standards for functional safety (e.g., ISO 26262 ASIL levels) in power electronics.
In summary, the integration of CGAN for targeted data synthesis and CNN for discriminative classification represents a pragmatic and effective solution to a longstanding data scarcity problem in EV fault diagnosis. It exemplifies how domain-aware machine learning—grounded in physical understanding and validated through rigorous experimentation—can deliver tangible engineering value.
Sun Quan, Peng Fei, Li Hongsheng, Yu Xianghai
School of Automation, Nanjing Institute of Technology, Nanjing 211167, China
Sun Guodong
College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China
Journal of Power Supply, Vol. 22, No. 6, November 2024
DOI: 10.13234/j.issn.2095-2805.2024.6.318