A Reinforcement Learning-Driven Physics-Informed Neural Network Framework for Multi-Parameter Inversion in the Enclosure Vibration System of Gas-Insulated Switchgear
Li Yuhang, Bai Jin, Yan Yingjie, Liu Yadong, Jiang Xiuchen
School of Electrical Engineering Shanghai Jiao Tong University Shanghai 200240 China
Abstract：To enhance the condition monitoring and diagnostic transparency of power equipment, methods based on inverse problem theory are increasingly important. For gas-insulated switchgear (GIS), where mechanical faults account for a significant portion of failures, parameter inversion of vibration systems is a key step toward interpretable fault diagnosis. However, existing methods, including Physics-Informed Neural Networks (PINNs), face significant challenges in this domain. Specifically, they struggle to accurately reconstruct high-frequency dynamic responses and to effectively coordinate the convergence of multiple, often conflicting, physical loss functions during training. This difficulty arises because the fixed or heuristically adjusted weights used in conventional PINNs are often insufficient to balance loss terms with different physical scales and convergence rates, leading to training instability and suboptimal accuracy. This study aims to address these limitations by developing a new framework that enables adaptive, intelligent control over the PINN training process for robust and accurate multi-parameter inversion in GIS vibration systems. A reinforcement learning-enhanced PINN (RL-PINN) framework with a nested dual-network architecture is proposed. The inner network is a PINN that models the GIS shell's vibration dynamics using a 3D continuous medium partial differential equation (PDE), guided by a composite loss function. The outer network is a reinforcement learning (RL) agent, structured as an Actor-Critic model, which dynamically adjusts the weights of each term in the PINN's loss function. The RL agent observes the PINN's training state—including loss values and high-frequency residuals—and selects weights to maximize a long-term cumulative reward that balances convergence and accuracy. This transforms the manual weight-tuning problem into a solved sequential decision-making task. The framework was validated using a dedicated experimental platform with a simplified GIS shell. A 100 Hz harmonic excitation was applied to simulate dominant operational vibrations, and response data was collected using a 12-accelerometer array under various spatial configurations to test for robustness. The RL-PINN framework demonstrated superior performance compared to a conventional deep neural network (DNN) and a DWA-PINN (a PINN using a dynamic weight averaging method). In parameter inversion, RL-PINN achieved the lowest relative errors for equivalent damping (3.71%) and stiffness (2.11%), representing a maximum error reduction of 47.0% and 57.1%, respectively. For response reconstruction, it yielded a linear regression R2 value of 99.02%, compared to 92.53% for DWA-PINN and 89.12% for DNN. The framework also showed higher robustness to noise, maintaining an error of only 4.61% at a 5% noise level, significantly outperforming the comparative methods. This study validates that the RL-PINN framework effectively overcomes key challenges in PINN-based parameter inversion by employing an RL agent for intelligent, dynamic loss weighting. This approach yields more accurate and stable results, providing a robust inverse modeling tool essential for developing physics-based fault diagnosis systems for power equipment. Future work will focus on extending the framework to complex, multi-component GIS models, advancing from parameter identification to precise fault source localization, and exploring its use in optimizing sensor placement strategies.
李宇航, 柏津, 严英杰, 刘亚东, 江秀臣. 面向GIS外壳振动系统多参数反演的强化学习驱动物理信息神经网络框架[J]. 电工技术学报, 2026, 41(15): 5249-5264.
Li Yuhang, Bai Jin, Yan Yingjie, Liu Yadong, Jiang Xiuchen. A Reinforcement Learning-Driven Physics-Informed Neural Network Framework for Multi-Parameter Inversion in the Enclosure Vibration System of Gas-Insulated Switchgear. Transactions of China Electrotechnical Society, 2026, 41(15): 5249-5264.
[1] 韩先才, 孙昕, 陈海波, 等. 中国特高压交流输电工程技术发展综述[J]. 中国电机工程学报, 2020, 40(14): 4371-4386. Han Xiancai, Sun Xin, Chen Haibo, et al.The overview of development of UHV AC transmission technology in China[J]. Proceedings of the CSEE, 2020, 40(14): 4371-4386. [2] 李鹏, 钟瀚明, 马红伟, 等. 基于深度强化学习的有源配电网多时间尺度源荷储协同优化调控[J]. 电工技术学报, 2025, 40(5): 1487-1502. Li Peng, Zhong Hanming, Ma Hongwei, et al.Multi-timescale optimal dispatch of source-load-storage coordination in active distribution network based on deep reinforcement learning[J]. Transactions of China Electrotechnical Society, 2025, 40(5): 1487-1502. [3] 向超群, 尹雪瑶, 伍珣, 等. 基于物理信息神经网络的牵引变流器直流支撑电容参数辨识方法[J]. 电工技术学报, 2024, 39(15): 4654-4667. Xiang Chaoqun, Yin Xueyao, Wu Xun, et al.Parameter identification of DC-link capacitor in traction converter based on physical information neural network[J]. Transactions of China Electro-technical Society, 2024, 39(15): 4654-4667. [4] 盛戈皞, 钱勇, 罗林根, 等. 面向新型电力系统的数字化电力设备关键技术及其发展趋势[J]. 高电压技术, 2023, 49(5): 1765-1778. Sheng Gehao, Qian Yong, Luo Lingen, et al.Key technologies and development trends of digital power equipment for new type power system[J]. High Voltage Engineering, 2023, 49(5): 1765-1778. [5] 胡应宏, 李雨, 李阳, 等. 面向柔直换流站高频谐振问题的换流变压器建模与参数辨识[J]. 电工技术学报, 2024, 39(22): 7154-7166. Hu Yinghong, Li Yu, Li Yang, et al.Modeling and parameter identification of converter transformer for high-frequency resonance problem of flexible DC converter station[J]. Transactions of China Electrotech-nical Society, 2024, 39(22): 7154-7166. [6] 牛硕丰, 李涛, 徐珂, 等. GIS开断设备典型机械故障模拟及运行状态智能诊断研究[J]. 高压电器, 2023, 59(12): 36-47. Niu Shuofeng, Li Tao, Xu Ke, et al.Study on typical mechanical fault simulation and intelligent operation condition diagnosis of GIS switching equipment[J]. High Voltage Apparatus, 2023, 59(12): 36-47. [7] 李国栋, 俞华, 刘宏, 等. GIS设备内外部紧固松动机械缺陷振动特性对比分析研究[J]. 高压电器, 2024, 60(5): 12-19, 30. Li Guodong, Yu Hua, Liu Hong, et al.Comparative analysis and research on vibration characteristics of internal and external fastening loose mechanical defects of GIS equipment[J]. High Voltage Apparatus, 2024, 60(5): 12-19, 30. [8] Zhong Yao, Hao Jian, Ding Yilin, et al.Novel GIS mechanical defect simulation and detection method based on large current excitation with variable frequency[J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 3519815. [9] Zhong Yao, Hao Jian, Lliu Q, et al.Novel diagnosis method for GIS mechanical defects based on an improved lightweight CNN model with load adaptive matching[J]. IEEE Transactions on Industrial Informatics, 2023, 19(11): 11041-11051. [10] 冯英, 李旭, 钟尧, 等. 基于多层融合振动数据分析的GIS设备机械缺陷诊断方法研究[J]. 中国电机工程学报, 2024, 44(14): 5797-5810. Feng Ying, Li Xu, Zhong Yao, et al.Research on the GIS mechanical defects diagnosis method based on multi-layer fusion vibration data analysis[J]. Proceedings of the CSEE, 2024, 44(14): 5797-5810. [11] 郝建, 李旭, 邵子琦, 等. 基于多图谱融合分析的GIS设备机械振动缺陷负载电流自适应诊断方法[J]. 中国电机工程学报, 2025, 45(24): 9764-9776. Hao Jian, Li Xu, Shao Ziqi, et al.Load current adaptive diagnosis method for mechanical vibration defects of GIS equipment based on multi-graph fusion analysis[J]. Proceedings of the CSEE, 2025, 45(24): 9764-9776. [12] 刘亚东, 严英杰, 严波, 等. 电力装备内部状态反演重构研究框架与应用展望[J]. 高电压技术, 2022, 48(8): 2883-2896. Liu Yadong, Yan Yingjie, Yan Bo, et al.Framework and application prospect of internal state inversion and reconstruction of power equipment[J]. High Voltage Engineering, 2022, 48(8): 2883-2896. [13] 陈思, 刘亚东, 严英杰, 等. 10 kV油浸式变压器匝间短路故障反演与定位[J]. 高电压技术, 2023, 49(5): 1870-1881. Chen Si, Liu Yadong, Yan Yingjie, et al.Inversion and localization of turn-to-turn short-circuit faults in 10 kV oil-immersed transformers[J]. High Voltage Engineering, 2023, 49(5): 1870-1881. [14] 廖瑞金, 罗豪, 成立, 等. 面向数字孪生变压器的计算机辅助运维(CAM+)构架与关键技术[J]. 高电压技术, 2024, 50(3): 924-940. Liao Ruijin, Luo Hao, Cheng Li, et al.Computer aided maintenance (CAM+) framework and key technologies towards digital twin transformers[J]. High Voltage Engineering, 2024, 50(3): 924-940. [15] Nguyen T T, Ta Q B, Ho D D, et al.A method for automated bolt-loosening monitoring and assessment using impedance technique and deep learning[J]. Developments in the Built Environment, 2023, 14: 100122. [16] Khoshmanesh S, Watson S J, Zarouchas D.The effect of the fatigue damage accumulation process on the damping and stiffness properties of adhesively bonded composite structures[J]. Composite Structures, 2022, 287: 115328. [17] Qiu Hao, Li Faxin.Bolt looseness monitoring based on damping measurement by using a quantitative electro-mechanical impedance method[J]. Smart Materials and Structures, 2022, 31(9): 095022. [18] Tamatam L R, Botto D, Zucca S.A novel test rig to study the effect of fretting wear on the forced response dynamics with a friction contact[J]. Nonlinear Dynamics, 2021, 105(2): 1405-1426. [19] Bajrić A, Høgsberg J.Identification of damping and complex modes in structural vibrations[J]. Journal of Sound and Vibration, 2018, 431: 367-389. [20] Safari S, Londoño Monsalve J M. Data-driven structural identification of nonlinear assemblies: Structures with bolted joints[J]. Mechanical Systems and Signal Processing, 2023, 195: 110296. [21] Xue Lian, Zhao Xue, Li Hu, et al.Genetic algorithm-based parameter inversion and pipeline subsidence prediction[J]. Journal of Applied Geophysics, 2023, 215: 105133. [22] Karniadakis G E, Kevrekidis I G, Lu Lu, et al.Physics-informed machine learning[J]. Nature Reviews Physics, 2021, 3(6): 422-440. [23] Raissi M, Perdikaris P, Karniadakis G E.Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations[J]. Journal of Computational Physics, 2019, 378: 686-707. [24] Haghighat E, Amini D, Juanes R.Physics-informed neural network simulation of multiphase poroelasticity using stress-split sequential training[J]. Computer Methods in Applied Mechanics and Engineering, 2022, 397: 115141. [25] Haghighat E, Raissi M, Moure A, et al.A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics[J]. Computer Methods in Applied Mechanics and Engineering, 2021, 379: 113741. [26] Tang Pengfei, Zhang Zhonghao, Tong Jie, et al.Predicting transformer temperature field based on physics-informed neural networks[J]. High Voltage, 2024, 9(4): 839-852. [27] 张宇娇, 孙宏达, 赵志涛, 等. 基于物理信息神经网络的电磁场计算方法[J]. 电工技术学报, 2024, 39(17): 5251-5261. Zhang Yujiao, Sun Hongda, Zhao Zhitao, et al.Electromagnetic field calculation method based on physical informed neural network[J]. Transactions of China Electrotechnical Society, 2024, 39(17): 5251-5261. [28] 金立军, 王飞龙, 王金宇. 基于物理信息网络的开关柜三维温升场域重建与缺陷热源反演[J]. 中国电机工程学报, 2026, 46(4): 1685-1697. Jin Lijun, Wang Feilong, Wang Jinyu.Reconstruction of 3D temperature rise field and defect heat source inversion for switchgear based on physics-informed networks[J]. Proceedings of the CSEE, 2026, 46(4): 1685-1697. [29] Li Yuhang, Lv Yuhui, Yan Yingjie, et al.Physics-informed neural networks for robust equivalent damping parameter inversion and fault diagnosis in gas-insulated switchgear vibration systems[J]. Scientific Reports, 2025, 15: 24521. [30] Chen Zhao, Badrinarayanan V, Lee C Y, et al.GradNorm: gradient normalization for adaptive loss balancing in deep multitask networks[C]//Proceedings of the 35th International Conference on Machine Learning, PMLR, Stockholm, Sweden, 2018: 794-803. [31] 杜志叶, 郝兆扬, 赵鹏飞, 等. 适用于声源定位的气体绝缘输电线路超声导波传播特性研究[J]. 电工技术学报, 2024, 39(3): 852-862. Du Zhiye, Hao Zhaoyang, Zhao Pengfei, et al.Research on propagation characteristics of gas-insulated transmission line ultrasonic guided wave for sound source localization[J]. Transactions of China Electrotechnical Society, 2024, 39(3): 852-862. [32] 丁登伟, 何良, 龙伟, 等. GIS设备运行状态下振动机理及检测诊断技术研究[J]. 高压电器, 2019, 55(11): 93-99. Ding Dengwei, He Liang, Long Wei, et al.Analysis on vibration mechanisms and detection diagnosis technology of GIS in operating condition[J]. High Voltage Apparatus, 2019, 55(11): 93-99. [33] 郝金鹏. 外壳振动信号在GIS故障诊断中的应用研究[D]. 北京: 华北电力大学, 2016. Hao Jinpeng.The applied research of shell vibration signal in GIS fault diagnosis[D]. Beijing: North China Electric Power University, 2016. [34] 党永亮, 张玉焜, 祝令瑜, 等. 基于麦克斯韦力-振动耦合的高压油浸式并联电抗器铁心多倍频振动机理研究[J]. 电工技术学报, 2024, 39(增刊1): 117-126. Dang Yongliang, Zhang Yukun, Zhu Lingyu, et al.Study on multi-harmonic vibration mechanism of high-voltage shunt reactor core based on coupling between Maxwell force and vibration[J]. Transactions of China Electrotechnical Society, 2024, 39(S1): 117-126.
2026, Vol. 41 
