Improved Mask R-CNN with Attention-Guided Feature Fusion for Fault Diagnosis of Electrical Equipment

Abstract

With the advent of the industry 4.0 era, intelligent operation and maintenance, along with fault warning, have emerged as crucial elements for ensuring the efficient and stable operation of electrical equipment. Traditional fault detection methods, which rely on manual inspections, are time-consuming, labor-intensive, and inefficient, failing to meet the demands of modern production. In this context, deep learning, particularly target detection technology, has demonstrated significant potential due to its high precision and automation capabilities. Among these, Mask R-CNN stands out as an ideal choice for electrical equipment fault diagnosis, thanks to its exceptional instance segmentation ability and high target location accuracy. This study focuses on key electrical components, including transformers, circuit breakers, and cables commonly found in the power industry, and amasses a substantial collection of high-definition visual data to serve as a training sample database. The architecture of Mask R-CNN is enhanced in this research. Initially, an attention mechanism is incorporated to augment the model's focus on details. Subsequently, a multi-scale feature fusion module is integrated to boost the recognition rate of small targets and enhance the algorithm's robustness and generalization ability in complex environments. During the experimentation phase, over 10,000 images depicting various failure modes were utilized for model training and validation. The mean average precision (mAP), representing the average detection accuracy across all test datasets, reached an impressive 90%. The specific mAP values reported in the charts (0.986, 0.915, 0.987) correspond to different test datasets and experimental conditions, ensuring consistency between the summary and detailed experimental results. The experimental data reveals that even under low-light conditions or with occlusions, the improved Mask R-CNN can accurately differentiate between normal and abnormal states. Minor damages such as oil stains, cracks, and corrosion can be promptly identified. Compared to the previous version, the algorithm's running speed has been significantly increased, with the average processing time per frame reduced to approximately 0.3 seconds, thereby greatly enhancing the fault response efficiency.