Intelligent Fault Diagnosis for Nonlinear Uncertain Industrial Processes Based on Simplified Interval Kernel Global–Local Feature Embedding

The data collected during industrial are uncertain due to the interference of various factors and are usually represented by intervals. However, the developed interval principal component analysis (PCA) is a linear monitoring algorithm that performs poorly in nonlinear data feature extraction and cannot extract local neighborhood information that reflects topological relationships between the internal data points. To address these issues, this article proposes an intelligent fault diagnosis method based on simplified interval kernel global-local feature embedding (SIKGLFE). First, a global-local feature embedding (GLFE) algorithm, which can simultaneously extract global and local information, is constructed by combining the advantages of PCA and local preserving projection (LPP). The algorithm is then extended to an interval kernel GLFE (IKGLFE) algorithm suitable for nonlinear interval-value data through kernel function and interval inner product estimation methods. To improve the computational speed and reduce data redundancy, a simplified interval matrix is obtained using the mutual information (MI) between samples and is used in the training of IKGLFE to obtain the SIKGLFE. In addition, the fault diagnosis function of industrial process is increased by the defined nonlinear reconstruction contribution graph. Finally, the superiority of the SIKGLFE was verified using simulation data from the Tennessee Eastman process (TEP) and actual operating data from a shearer in an inclined-ditch coal mine. The experimental results show that, compared with midpoint-radius kernel PCA (MRKPCA), kernel LPP (KLPP), and IKGLFE, the proposed method demonstrates remarkably improved diagnostic accuracy, computational efficiency, and robustness in processing nonlinear interval-value data and has better comprehensive capability.