Photovoltaic Panel Failure Detection Using Class-Conditioned Generative Adversarial Networks

Technical Summary: Photovoltaic Panel Failure Detection Using Class-Conditioned Generative Adversarial Networks

Problem Statement
The automated inspection of solar photovoltaic (PV) systems relies heavily on infrared (IR) thermography to detect defects such as hot spots, cell cracks, and bypass diode failures. While deep learning models have shown promise in automating this process, their effectiveness is severely constrained by the class imbalance inherent in real-world datasets. In operational solar farms, the vast majority of modules are healthy, while critical failure modes (e.g., diode failures, multi-cell hot spots) are rare. This scarcity of minority class samples leads to supervised learning models that are biased toward the dominant "normal" class, resulting in poor recall for critical faults. Furthermore, models trained on such skewed data often suffer from poor probabilistic calibration, producing overconfident misclassifications that are dangerous in safety-critical autonomous operations. Collecting more real-world data for rare faults is operationally complex and costly, as it requires specific environmental conditions and often destructive fault induction.

Methodology
To address these challenges, the authors propose an Auxiliary Classifier Generative Adversarial Network (AC-GAN) framework designed specifically for the class-conditioned synthesis of high-fidelity thermal images of PV defects.

• Architecture:
  • Generator: The generator accepts a latent noise vector ($z$) and a discrete class label ($c$). The label is embedded into a continuous representation and concatenated with the noise. The network employs a progressive upsampling structure with two blocks, each containing batch normalization, bilinear upsampling, convolution, and LeakyReLU activation. Crucially, a self-attention mechanism is integrated between upsampling blocks to capture long-range spatial dependencies characteristic of thermal anomalies across the module. The final output is a single-channel grayscale thermal image constrained to $[-1, 1]$ via a $tanh$ activation.
  • Discriminator: The discriminator performs a dual task: distinguishing real from fake images (adversarial task) and classifying the image into the correct anomaly category (auxiliary task). It utilizes four convolutional blocks with spectral normalization to constrain the Lipschitz constant and ensure training stability. The output is split into an adversarial head (real/fake probability) and an auxiliary classification head (class probabilities).
• Training Strategy:
  • The model is trained using a composite loss function combining adversarial loss and auxiliary classification loss.
  • To ensure stability, the authors employ label smoothing (targets 0.9 for real, 0.1 for fake), Gaussian noise injection into real images, and gradient clipping.
  • Optimization is performed using the Adam optimizer with distinct learning rates for the generator and discriminator.
• Downstream Evaluation:
  • The synthetic images generated by the AC-GAN are used to balance the training dataset.
  • A MobileNetV3-Large classifier is trained on the augmented data using a three-phase fine-tuning strategy (frozen backbone, partial fine-tuning, and full fine-tuning) to adapt to the photovoltaic anomaly task.
  • Performance is evaluated against baselines including class-weighted loss, oversampling, and conventional geometric augmentation. Metrics include top-1 accuracy, validation loss, and Expected Calibration Error (ECE).

Key Contributions

1. Fault-Conditioned Synthesis: The paper introduces a specialized AC-GAN for infrared PV imagery that enables the controlled generation of rare fault signatures without destructive fault induction, directly addressing severe dataset imbalance.
2. Rigorous Comparative Evaluation: The study systematically compares generative augmentation against conventional methods (oversampling, class weighting, and geometric augmentation), quantifying their impact on both classification performance and uncertainty estimation.
3. Calibration Focus: The work demonstrates that while synthetic data may yield marginally lower top-1 accuracy compared to simple augmentation in some configurations, it significantly improves probabilistic calibration. The proposed framework achieves an ECE of 1.5%, indicating that the model's confidence scores align much more closely with actual correctness, a critical factor for safety-sensitive diagnostics.

Results

• Generative Quality: The proposed AC-GAN achieved a Fréchet Inception Distance (FID) of 35.4, outperforming ablated versions lacking self-attention (FID 52.3) or spectral normalization (FID 68.7). This indicates high fidelity in the generated thermal signatures.
• Classification Performance:
  • The AC-GAN-based approach achieved a classification accuracy of 80.45% with a validation loss of 0.6987.
  • This significantly outperformed the class-weighted baseline (67.97% accuracy, 2.1747 loss) and the oversampling baseline (70.50% accuracy, 1.0260 loss).
  • While conventional geometric augmentation achieved a slightly higher accuracy (82.12%), it resulted in a higher validation loss (0.9064) and poorer calibration.
• Calibration and Reliability: The AC-GAN model demonstrated superior uncertainty awareness. Reliability diagrams showed a strong alignment between predicted confidence and empirical correctness. The model exhibited an ECE of 1.5%, compared to 2.56% for the augmentation-based model. The GAN-based model was significantly more confident when correct and appropriately less confident when incorrect, whereas the augmentation-based model showed greater overlap and overconfidence.

Significance and Claims
The authors claim that this methodology offers a scalable solution for enhancing the reliability of autonomous solar operations and maintenance (O&M). The primary significance lies in the trade-off between raw accuracy and reliability: the paper argues that for safety-critical applications like autonomous aerial inspections, a model with superior probabilistic calibration (low ECE) is fundamentally more reliable than one with marginally higher accuracy but poor calibration. By generating functionally informative, high-quality synthetic samples, the AC-GAN framework ensures that diagnostic models remain robust even when field-acquired data for rare failure modes is unavailable, thereby reducing the risk of undetected critical faults and overconfident misclassifications. The study acknowledges limitations regarding generalizability across different camera resolutions and climatic conditions, and notes that diffusion models were not explored due to computational constraints, suggesting these as areas for future work.