The reliable operation of metal-enclosed gas-insulated equipment is crucial to high-voltage power systems, and accurate diagnosis of its status is the key to ensuring safe and efficient operation of the power grid. Accurate state diagnostics help to prevent gas-insulated equipment failures, ensuring stable and efficient operation of the electricity grid. The existing methods are difficult to effectively capture the temporal and sequential characteristics of partial discharge (PD) signals, resulting in insufficient recognition accuracy for complex defects such as conductor protrusions (C-type), surface contamination (S-type), and floating electrodes (F-type). At the same time, the on-site environmental noise interference is large, and the fault samples, especially the severe defect samples, are rare and unbalanced. In addition, the gas state is easily affected by temperature and pressure, which makes the existing diagnostic models weak in generalization ability and poor in real-time performance, making it difficult to meet the urgent needs of high-voltage power grids for accurate, online, and efficient diagnosis of equipment status. This limitation restricts timely and precise fault detection, leads to equipment failure, and operational disruptions pave the way for deep learning algorithms. Thus, the research proposes a Long Short-Term Memory based Gas Insulation State Diagnostics (LSTM-GISD) method for accurate fault detection and state diagnostics in gas insulation equipment. The study utilizes sensor data from critical vital components, including gas pressure, temperature, partial discharge, and electrical parameters, to accurately assess the operational state of gas insulated as a dielectric medium in high-voltage equipment. The methodology includes data preprocessing feature extraction to identify the relevant characteristics of PD signals. The LSTM-based Recurrent Neural Network (RNN) algorithm is used for model training to improve the classification of PD patterns and diagnose the operational state of gas insulation switch gear equipment, enhancing fault detection accuracy and supporting maintenance strategies. The model’s effectiveness is evaluated using various metrics, including accuracy, precision, and recall. It focuses on their ability to identify both normal and fault states (C, S, F types) of equipment with improved reliability. The study results demonstrate improved fault detection speed in terms of computational time, accuracy and robustness in detecting faults and predicting potential failure states in equipment maintenance practices.
Wang et al. (Fri,) studied this question.