A spectral clustering approach for multivariate industrial streaming data based on local linear embedding

In real industrial applications, large amounts of unlabeled streaming data are typically generated, presenting a series of challenges in their effective utilization. This study uses unsupervised clustering methods to analyze these data sets and reveal potential patterns and structures, thereby providing effective support and decision points for optimization of industrial production processes. Given the spatial nonlinearity and complex geometric shapes of the streaming data, local linear embedding is used to map data from high-dimensional, non-linear, concave-convex characteristics to lower-dimensional, relatively linear characteristics, thereby achieving feature extraction. Conventional spectral clustering algorithms use Euclidean distance to measure similarity, which does not adequately capture the non-linear nature of the data. In contrast, our method combines the Minkowski distance and cosine similarity to more accurately measure the similarity between data points and thereby improve clustering effectiveness. The method was validated on real industrial coal gasification data, industrial wastewater treatment data, and two public datasets. Comparison with other clustering algorithms demonstrated the superior clustering performance of our method.