LVW‐ k NN : A local variable weighted k nearest neighbours algorithm for multimodal and nonlinear process fault detection

Abstract Modern industrial production processes frequently require continuous switching of operating modes to meet diverse process production requirements. Therefore, industrial process data exhibits distinct characteristics of multimodality, nonlinearity, and significant variance differences across modes. To address the above challenges, a multimodal process fault detection method based on local variable weighted nearest neighbours (LVW‐NN) is proposed in this paper. First, an intermediate nearest neighbour set is constructed to establish a local reference background for the test sample, providing an intra‐modal reference framework for subsequent fault detection. Second, variable weights are dynamically assigned based on the deviation between the samples and its nearest neighbours, and a local normalization strategy is employed to eliminate the effects of inter‐model variance disparities. Third, based on the preceding step, a weighted reconstruction strategy is applied at the feature level, which significantly amplifies faint fault signals and directly enhances the detection ability for minor faults. Finally, the current test data is reconstructed by local variable weighting, and a fault detection model based on the nearest neighbour rule (FD‐NN) is built to achieve fault detection. To verify the effectiveness of the proposed LVW‐NN method, comprehensive tests and evaluations are conducted using numerical case studies and the Tennessee Eastman (TE) industrial benchmark process. The experimental results show that the LVW‐NN is well‐suited for processing nonlinear and multimodal industrial process data, it exhibits significantly superior performance in fault detection, especially demonstrating a distinct advantage in detecting minor faults.