Outlier Detection via SVDD with Mixed Kernel in Distributed Sensors

In this study, a strategy for denoising nonstationary signals and detecting outliers is proposed for distributed fiber optic sensors (DFOS). Bi-dimensional Empirical mode decomposition adaptively decomposes bi-dimensional detection signals with nonlinear attenuation into multiple intrinsic mode functions to suppress confusion of noise and outliers. The unsupervised learning method, Support Vector Data Description (SVDD), is employed for classification. SVDD excels in handling imbalanced samples by converting them into outliers, thereby designing an optimal separation hypersphere with the smallest volume. This feature is particularly advantageous in unsupervised scenarios where labeled data is scarce or unavailable, as it allows for robust outlier detection without the need for prior knowledge. A calculation module integrates classification outputs to obtain the location of high-speed trains. Finally, an experiment is conducted based on the basis of the data collected from a high-speed line in China. Results validate the feasibility and accuracy of the proposed strategy.