Complex multi‐stage process fault detection based on t‐ <scp>SNE</scp> ‐ <scp>BPNN</scp> combined with spatiotemporal neighbour center distance

Abstract To address the difficulty of fault detection in nonlinear, dynamic, and multi‐stage processes, a spatiotemporal neighbour centre distance (SNCD) statistic is proposed. SNCD is combined with t‐distributed stochastic neighbour embedding (t‐SNE) and back propagation neural network (BPNN) to develop the t‐SNE‐BPNN‐SNCD (tB‐SNCD) fault detection method. The t‐SNE‐BPNN leverages BPNN to learn the nonlinear implicit mapping relationships during the t‐SNE feature extraction and dimensionality reduction process, solving the problem of embedding new samples in t‐SNE. SNCD utilizes not only the spatial neighbour information of samples but also their temporal neighbour information, providing a more comprehensive extraction of process features, eliminating the autocorrelation of process data, and overcoming the difficulties posed by the dynamics of the process for fault detection. Since SNCD makes decisions based on the neighbourhood of samples, it is applicable to nonlinear, multi‐stage processes. The performance of tB‐SNCD is tested through numerical simulation processes and the Tennessee Eastman process, showing a higher fault detection rate compared to KPCA, DPCA, DKPCA, KNN, PC‐WKNN, and LOF methods. Particularly, when faults are time‐related, the fault detection rate of tB‐SNCD is significantly higher than that of classical methods.