XGBoost-LSTM Collaborative Wind Power Converter Fault Diagnosis for Sensor and Power Tube Faults

Abstract This study constructs a collaborative diagnosis method built on XGBoost-LSTM for two common and serious faults: current sensor and IGBT open-circuit. This study first builds a doubly-fed WPC model on the simulation platform, carries out a deep analysis of the fault mechanism and characteristic manifestations, extracts key indicators through data preprocessing, and builds a “fault detection-fault location” two-level collaborative diagnosis framework. Unlike conventional hybrid methods that merely combine multiple models, the proposed approach integrates the probabilistic outputs of XGBoost into the input of the LSTM to achieve feature-level fusion, thereby enhancing the model's capability to capture complex temporal fault characteristics. Among them, the extreme gradient boosting model uses its powerful multi-dimensional feature classification capabilities to achieve rapid detection and preliminary type identification of fault states. The LSTM focuses on processing timing characteristics to accurately locate the power tube OCF location. Experiments show that the average performance index of sensor fault diagnosis reaches 0.93, of which the constant value fault reaches more than 0.97. In power tube OCF diagnosis, the average performance indicators of the proposed method are 0.98 and 0.95 under ideal and complex working conditions, and the average time-consuming for a single diagnosis is only 3.8 ms. The results show that this method can effectively achieve rapid and accurate fault diagnosis, has strong robustness and engineering application potential, and provides a reliable solution for intelligent operation and maintenance of WPCs.