In a paper published in the journal PLOS One, researchers introduced a novel approach to fault monitoring in high-voltage circuit breakers using a specialized monitoring device to analyze vibration data to diagnose mechanical issues. Through deep learning techniques, they trained the system to recognize fault-related patterns within this data, facilitating precise fault identification.
Study: Deep Learning-Based Fault Monitoring in High-Voltage Circuit Breakers. Image credit: Wang An Qi/Shutterstock
This systematic process allows classifying circuit breaker faults by comparing characteristics from test sets against a trained dataset. With an impressive accuracy rate of over 95% for similar circuit breakers, this method presents an efficient and practical solution for real-time fault diagnosis and pre-emptive warnings in high-voltage systems.
Background
Traditional fault detection methods need help in accurately analyzing vibration data due to the intricate internal structure of circuit breakers. The present paper's focus is on proposing an unsupervised deep learning-based fault monitoring approach to address these challenges. This method aims to simplify fault monitoring in high-voltage circuit breakers by leveraging deep learning algorithms for data analysis and fault identification. It promises a more practical and accurate solution to overcome existing limitations, setting the stage for subsequent sections covering data collection, feature extraction, fault diagnosis, and experimental validation.
Enhancing Circuit Breaker Fault Diagnosis
The "Improved deep learning algorithm" delves into pivotal aspects of enhancing circuit breaker fault diagnosis through refined feature extraction using deep belief networks. This segment focuses on data preprocessing, unsupervised feature reconstruction methods, and the subsequent phases of data classification, retraining, and vibration feature extraction leading to fault diagnosis.
It meticulously explains the mechanics of a deep belief network comprising stacked restricted Boltzmann machines (RBMs) and elucidates the energy function and joint probability density distribution between visible and hidden layers. It delves into the encoding and decoding processes, emphasizing unsupervised learning and retrieving confidential layer data. Researchers introduce a cost function to measure data reconstruction quality and propose using the BP algorithm for enhancing feature extraction.
Following this, the segment moves into data classification and retraining, highlighting the necessity for classification layers in fault diagnosis. It describes the output vectors of RBMs, probability calculations, and the cost function to minimize errors. Researchers elaborate on the retraining process, integrating tag data and unsupervised learning, showcasing a comprehensive training methodology. Lastly, the segment outlines a structured process for vibration feature extraction and fault diagnosis in six sequential steps. These steps encompass data collection, model construction, training, feature extraction, retraining, and the eventual input of test data into the trained network for fault diagnosis.
Conducting Circuit Breaker Fault Diagnosis
This segment outlines the comprehensive steps taken to develop an experimental platform for fault diagnosis in circuit breakers, using the YFGZ35 (EP) -40.5 circuit breaker as the focus of the study. The platform integrates vibration sensors, current transformers, and mobile devices to collect and process vibration data during various circuit breaker operations, simulating different fault types such as iron core jamming, spring fatigue, and component looseness. The aim is to build a robust fault database for high-voltage circuit breakers, which is essential for training and validating classifier models.
The study investigates feature extraction using a deep belief network with a hidden layer set to 5 and 16 uncoded features as inputs. For supervised fault classification experiments, researchers categorize training samples comprising normal and fault states into four operating conditions (B = [1,2,3,4]). To ensure practical applicability, researchers modify the mechanical structure parameters of the circuit breaker to simulate faults. They detail the resulting fault types and dataset compositions for training and verification.
The detailed parameters for training the deep learning method establish the significance of optimizing the layer number to balance error rate reduction and algorithm stability. The subsequent feature extraction process compares the proposed method with a traditional deep belief network and a single deep automatic encoding algorithm, demonstrating superior performance in training and testing errors while maintaining similar calculation times.
The validation of the proposed fault diagnosis method employs historical operating data from the YFGZ35 (EP) -40.5 circuit breaker, verifying the trained model's accuracy in identifying different operating states. Through comprehensive analysis, the method exhibits an average identification accuracy of 95.1%, with a 100% accuracy rate in identifying normal operating states, showcasing its effectiveness in real-world scenarios. The approach is further validated by comparing it with support vector machine and pattern recognition methods, demonstrating significantly improved identification accuracy despite a slight increase in computation time.
Utilizing unsupervised feature learning, the proposed deep learning model showcases enhanced accuracy in circuit breaker fault monitoring, adapting model parameters to feature expression and demonstrating high accuracy and resolution in identifying various fault types during actual circuit breaker operations.
Conclusion
To summarize, this paper introduces a high-voltage circuit breaker fault monitoring device and a mechanical fault monitoring method based on deep learning. The device efficiently collects and processes vibration information, achieving accurate fault diagnosis with over 95% accuracy.
Compared to other algorithms, this method demonstrates improved feature extraction accuracy and faster computation, offering a promising approach for monitoring mechanical failures in high-voltage circuit breakers. Future research may focus on refining fault localization within deep learning-based fault diagnosis methods.
Deep Learning Secures IoT with Federated Learning