An improved isolation forest based on kernel principal component analysis for anomaly detection in helicopter-based power-line inspection data

For helicopter inspection data, including flight trajectories, task status, and equipment operating status, conventional anomaly detection methods still suffer from difficulties in separating high-dimensional nonlinear data, insufficient anomaly isolation capability, and low anomaly discrimination. To address these problems, this paper proposes an improved isolation forest anomaly detection method based on kernel principal component analysis (KPCA). The proposed method extracts five categories of features from inspection data, including standard deviation, coefficient of variation, centroid, resilience, and scale features, and constructs a fused feature matrix through sliding windows. KPCA is then used to perform nonlinear reconstruction and dimensionality reduction on the feature matrix to obtain a low-dimensional subspace. In this subspace, the single axis-aligned splitting strategy of the conventional isolation forest is improved into multi-directional splitting along principal-component directions and their combined directions. In addition, the path length is corrected according to the distance between each sample and the normal center .Experimental results based on 7027 real records from a helicopter inspection platform show that the proposed method outperforms conventional methods in terms of Precision, Recall, F1-score, FAR, and PR-AUC. The results demonstrate that the proposed method can effectively identify flight-trajectory deviations, equipment operation anomalies, and task-execution anomalies, providing a feasible data-driven technical approach for improving the intelligence and reliability of power-line inspection.