Abstract
Background: Feature extraction methods such as statistical features and image features based on traditional partial discharge spectrograms require the support of expert experience and prior knowledge, making their engineering applications have significant limitations. Currently, electrical equipment fault diagnosis is moving towards intelligent and scientific directions, and how to apply deep learning methods for more accurate, intelligent, and autonomous diagnosis has become a hot issue in the current field.
Methods: The construction and optimal selection of high-voltage cable partial discharge feature space is an important means to improve recognition accuracy. This paper proposes a new method that can independently and deeply mine partial discharge image features to quickly identify partial discharge types. Firstly, a self encoder is used to encode and decode partial discharge images, and a multi-level stack structure is used to achieve the goal of deep feature mining; then, the partial discharge data features decoded based on a stack self encoder are imported into a random forest network for type recognition; Finally, based on the random forest algorithm, the redundant feature space is optimized and compared with other traditional recognition methods.
Results: The recognition accuracy of the proposed method is about 10% higher than that of traditional methods.
Conclusion: The optimized feature parameters can maintain more than 90% recognition accuracy in both random forest algorithms, common support vector machines, and BP neural networks, proving the feasibility and effectiveness of the method in this paper.
Related Journals
Recent Patents on Engineering
Related Books
Voltammetry for Sensing Applications
