A novel deep learning–based fault diagnosis algorithm for preventing protection malfunction. (January 2023)

• Record Type: Journal Article
• Authors: Hu, Jiaxiang; Liu, Zhou; Chen, Jianjun; Hu, Weihao; Zhang, Zhenyuan; Chen, Zhe
• Is Part Of: International journal of electrical power & energy systems. Volume 144(2023)
• Publication Date: 2023-01
• Subjects: DL -- Deep auto-encoder -- Fault diagnosis -- Unsupervised and supervised learning -- Disturbance -- Malfunction prevention; Electrical engineering -- Periodicals; Electric power systems -- Periodicals; Électrotechnique -- Périodiques; Réseaux électriques (Énergie) -- Périodiques; Electric power systems; Electrical engineering; Periodicals; 621.3
• DOI: 10.1016/j.ijepes.2022.108622
• Languages: English
• ISSNs: 0142-0615

Abstract

Highlights: DAE (deep auto-encoder) is implemented to extract key fault feature and re-divide sample dataset. Facing complex power system operation data, this process can be regarded as automatic data cleaning for system operation information. Supervised deep networks learn key fault samples extraction through the knowledge of unsupervised process. In this way, the samples containing key fault features are automatically extracted by the fault identification network and indicate the system status. Instead of increasing the complexity of the model, this article improves the performance of framework from the perspective of sample space and the union of multiple special networks. The samples with key features required by the tasks ensure that the models can learn the correct knowledge and have better performance. The proposed framework can resist the influence of disturbance and distinguish fault state and fault types. In addition, it provides a kind of idea based on samples distribution for diagnosis tasks. Abstract: To prevent serious malfunctions and reduce the impact of faults during an emergency state of a power system, protection systems are required to have disturbance and fault state identification abilities. In this study, a novel fault diagnosis framework based on deep learning with anti-disturbance ability is proposed to identify the fault state and fault type information, even under the influence of system disturbance. The framework consists of two parts: unsupervised and supervised learning. Specifically, an unsupervised deep auto-encoder (DAE) is applied for offline feature selection and data cleaning. The DAE can extract key fault features and significantly improve the fault detection accuracy. Furthermore, two supervised convolutional neural networks are used to learn key fault feature extraction online from complex operation information in power systems and assess the fault situation and type. Using case studies, the proposed method was implemented and compared with existing intelligent methods. The results indicate that the proposed framework has a better performance in terms of fault state identification and protection malfunction prevention.