基于姿态传感器的高压隔离开关机械故障智能诊断研究

The "health" condition of the high-voltage disconnector is very important to the safe operation of the power grid system. In order to effectively monitor the working condition of a disconnector online, this paper proposes an intelligent diagnosis method for the mechanical faults of the disconnector based on the attitude sensor. This method includes three modules: the attitude feature extraction, the sample processing and the fault diagnosis. In the feature extraction module, aiming at the problem of rough feature extraction of the disconnector attitude angle information, a feature extraction based on the characteristics of the attitude information is proposed. In the sample processing module, a GA-SMOTE-ET sample processing method is put forward to deal with the degradation of the diagnostic performance caused by the unbalanced numbers of the samples of each fault category in the current attitude information sample library and the excessive redundancy and interference information in the feature space. This method adaptively enhances the proportion of the effective information in the sample database and eliminates the negative effects of the redundant and interfering information. In the fault diagnosis module, in view of the problem that only a single algorithm is used in the field of the disconnector fault diagnosis at this stage without the fusion technology applied to integrate the advantages of each algorithm, an improved Stacking model fusion technology and two selection principles of the Stacking base learners are presented. This technology fuses multiple learning devices to achieve the complementary advantages so as to improve the diagnostic performance. Taking the GW5-35 disconnector as a prototype, this paper simulated six typical mechanical faults. The attitude sensing system is adopted to obtain the attitude data of seven mechanical states of the disconnector, and three groups of experiments are formulated to verify the effectiveness of the proposed method. Finally, the F1 score of this method is tested 0.971, significantly better than the traditional method.