A Visualization Method of Bi-Kernel t-Distributed Stochastic Neighbor Embedding and Its Application in Wind Turbine Gearboxes

This paper proposes a visualization monitoring method based on Bi-kernel t-distributed Stochastic Neighbor Embedding (Bi-kernel t-SNE). The method introduces the input data and feature kernel matrices to construct a dual kernel mapping from low-dimensional to high-dimensional spaces, improving the identification accuracy of anomalous outliers. Subsequently, Kernel Entropy Component Analysis (KECA) is applied to transform the feature kernel matrix, so the algorithm can distinguish between normal and abnormal values even when the feature kernel matrix values of the new sample data approach zero and thereby achieve accurate fault warnings. Finally, a confidence boundary combining the fault detection Composite Index (CI) and Kernel Density Estimation (KDE) is designed to perform fault detection in a two-dimensional (2D) scatter plot, with fault monitoring and warning realized by evaluating whether the characteristic data exceed the limits. The algorithm's effectiveness is validated through case studies using the faulty and healthy Supervisory Control And Data Acquisition (SCADA) data from a wind farm. The experimental results demonstrate that the proposed method can accurately distinguish faulty data points and significantly reduce the number of alarms. The fault sensitivity reaches 91.3%, representing an 11.89% improvement in monitoring accuracy over conventional t-SNE.