Anomaly Classification in Photovoltaic Modules Using Infrared Images with CNN and Hyperband Optimization

Photovoltaic module anomalies can reduce power generation efficiency and compromise system safety. This work proposes an automatic classification approach for infrared images of photovoltaic modules using convolutional neural networks (CNN) and automatic hyperparameter optimization with Hyperband. The objective is to identify eleven types of anomalies and one non-anomalous class using a compact model with good generalization capability.

The approach employs the InfraredSolarModules dataset, composed of 20,000 grayscale infrared images of size 40×24 pixels distributed across 12 classes. The No-Anomaly class contains 10,000 images, while minority classes such as Diode-Multi and Hot-Spot-Multi contain 175 and 247 images, respectively. The data were divided into 80% for training and 20% for testing without applying data augmentation, oversampling, or class weighting techniques.

A convolutional neural network architecture with convolutional layers, max-pooling, dense layers, ReLU activation, dropout regularization, and a softmax output layer was used. Instead of fixing a single architecture, a configurable hypermodel was defined. Hyperparameter optimization was performed using Hyperband, enabling an efficient exploration of multiple architectural configurations. During the search process, 535 configurations were evaluated, and the five best models were retrained using repeated 6-fold cross-validation with three repetitions. Performance was evaluated using accuracy, precision, recall, and the F1 score with macro-averaging.

The results showed that compact architectures with two or three convolutional layers achieved the best performance. The best model obtained an average accuracy of 98.75%, outperforming results reported for UDenseNet, EfficientNetB0, and MobileNet-V3 on the same dataset. The difference between training and validation accuracy remained below 2%. These results indicate that systematic hyperparameter optimization with Hyperband enables the construction of accurate and robust CNN classifiers for infrared images of photovoltaic modules even without relying on data augmentation or transfer learning.