Novel local density‐adaptive Wasserstein distance NPE with fast angle‐based outlier detection for fault diagnosis

Abstract

The accurate fault diagnosis of complex chemical processes is a challenge because process data is characterized by high dimensionality and strong nonlinearity, and typically exhibits non‐uniform local density and time‐varying distributions. Manifold learning can reduce the dimensionality of complex high‐dimensional features to a low‐dimensional space to extract effective features. Neighbourhood‐preserving embedding (NPE), which employs the k‐nearest neighbour rule and Euclidean distance similarity, can extract internal features of the data with relative accuracy. However, for boundary points of different data categories, NPE feature extraction results in blurred inter‐class boundaries. To overcome these limitations, this paper proposes a joint fault diagnosis methodology named Local density adaptive Wasserstein distance neighbourhood preserving embedding with fast angle‐based outlier detection (LWNPE‐FA), as follows: the neighbourhood size is firstly adaptively adjusted based on local density, using compact neighbourhoods in dense regions and expanded neighbourhoods in sparse regions; then, a composite similarity metric combining Euclidean and Wasserstein distances is employed to capture differences in locations and distribution spaces caused by faults, thereby establishing discriminative neighbourhood relationships; finally, fast angle‐based outlier detection (FastABOD) is utilized to apply intra‐class contraction and inter‐class separation updates, identifying samples with blurred boundaries in low‐dimensional spaces to clarify class boundaries. A joint diagnostic model combining LWNPE‐FA with a naive Bayes classifier is further constructed to achieve fault identification in complex industrial processes. On the Tennessee‐Eastman process (TEP) and three‐phase flow facility (TFF) benchmark datasets, LWNPE‐FA achieved average test accuracy rates of 98.22% and 98.96%, respectively, outperforming related methods by 1% ~ 5%; Ablation analysis and sensitivity studies confirm the effectiveness of local density adaptive modelling, Wasserstein distance‐based neighbourhood determination, and FastABOD inter‐class boundary detection in enhancing fault diagnosis performance.