The accumulation of toner on the contactor rail poses a significant challenge as it interconnects various operational processes. The buildup of carbon powder on the contactor rail will cause severe issues such as short circuits. Traditional maintenance approaches, like routine inspections and maintenance, not only demand considerable manpower and time but also fall short in providing accurate and real-time monitoring of the contactor rail status. Therefore, integrating state-of-the-art technologies in state monitoring and fault diagnosis is imperative to assess the carbon deposition status effectively.
This research aims to use convolutional neural networks (CNNs) for the recognition and evaluation of carbon deposits on contactor rails. While CNNs demonstrate remarkable performance in tasks such as image classification and object detection., their limited generalization capability poses challenges in health assessment and fault diagnosis domains. To address this problem, the project employs image augmentation techniques such as rotation and cropping to enhance the diversity of the dataset. Additionally, an approach of causal learning is introduced into the ResNet18 residual network, mitigating feature dependency through random Fourier features and sample weighting. These methods enhance the ability of model to generalize and interpret results effectively. Following the preprocessing of the dataset and fine-tuning of model hyperparameters, the evaluation results are generated. Finally, a comparative analysis is conducted using two key performance metrics—accuracy and interpretability—to validate the efficacy of dataset augmentation and causal learning in enhancing model performance and interpretability.