Dry-Type Air-Core Series Reactor Turn-to-Turn Short Circuit Fault Detection Method Based on Multi-Parameter Data Fusion

[Objective] To address the problems of weak turn-to-turn short-circuit faults in dry-type air-core series reactors， which are difficult to recognize， and the lack of an early warning mechanism in traditional methods， this study proposes a multi-dimensional feature and intelligent algorithm fusion of an early fault diagnosis method. This method can overcome the lack of sensitivity of a single fault feature as it is easily interfered with by the noise of the fault leakage judgment. [Methods] First， the unbalance degree， power factor， zero sequence voltage， and characteristic impedance of the shunt capacitor bank are extracted as fault feature quantities， and their respective evolution laws after the fault are analyzed. Second， principal component analysis （PCA） is used to reduce the dimension and denoise the original data to eliminate interfering information. Subsequently， the denoised features with high saturation are input into the k-nearest neighbors （KNN） algorithm to construct a fault identification and classification model. Based on Maxwell， a field-circuit coupling model is established to generate single-turn， slight， and multi-turn short-circuit datasets； noise-free and 5% noise conditions are considered to verify the robustness of the algorithm. [Results] Simulation results show that the proposed method can achieve 100% recognition accuracy for minor turn-to-turn short circuits under both no noise and 5% noise， and manually adjusting the action threshold is not required. [Conclusions] This study realized high-precision early identification of weak faults through the three-stage architecture of “feature extraction-data noise reduction-intelligent classification.” The innovations include four-dimensional feature synergy to improve fault sensitivity， a PCA-KNN joint anti-noise mechanism； and an adaptive non-threshold discrimination system. The results provide a new idea for power-equipment condition monitoring， and the generalization ability of the model can be optimized by incorporating field data.