Interpretable coal mine fault detection via orthogonal Kolmogorov-Arnold networks

The continuous monitoring of safety within underground coal mines generates multivariate sensor data of high dimensionality, posing substantial challenges for accurate and interpretable fault detection in safety-critical industrial environments. Existing machine-learning classifiers often suffer from performance degradation when applied to raw high-dimensional inputs, whereas black-box deep models lack the symbolic transparency required for operational trust. Kolmogorov-Arnold Networks (KANs) provide a promising route toward interpretable nonlinear modeling, but standard KANs rely on non-orthogonal B-spline basis functions that may introduce basis-function overlap and unstable optimization in dense feature spaces. Building on recent orthogonal-polynomial KAN developments, this study proposes a PCA-coupled Orthogonal Kolmogorov-Arnold Network, termed OrthoKAN, for high-dimensional coal mine fault detection. The novelty of the proposed framework lies in the integration of three components: PCA-based feature orthogonalization, Legendre-polynomial edge-function mapping, and a strictly additive symbolic topology for industrial fault classification. Specifically, 1,680 engineered sensor features are first projected into an orthogonal principal-component space, after which each retained component is independently mapped through Legendre polynomial edge functions. This dual-orthogonality design reduces redundancy at both the feature-representation and function-approximation levels while preserving principal-component-level symbolic traceability. When applied to the benchmark dataset, the proposed pipeline utilizing 300 principal components achieves a classification accuracy of 94.74% and a weighted F1-score of 0.9497. The predictive performance of the proposed method outperforms the standard B-spline KAN baseline, which achieves an accuracy of 90.56%, and also exceeds the tested conventional machine-learning and deep-learning baselines. Furthermore, the additive architecture extracts human-readable symbolic formulas involving linear, polynomial, sinusoidal, and exponential terms. These results indicate that coupling PCA with Legendre-based KAN edge functions provides an effective and interpretable framework for high-dimensional industrial fault detection in coal mine safety monitoring.