ADG-KAE: novel adaptive dynamic graph embedded kernel autoencoder algorithm for industrial process monitoring and fault detection

Abstract As a classical machine learning technique, the kernel autoencoder (KAE) algorithm exhibits outstanding capabilities of nonlinear data reconstruction, making it highly suitable for industrial process monitoring and early fault detection. However, its fault detection rate (FDR) will decrease significantly when KAE is applied to complex industrial process monitoring with dynamic characteristics. In response to this challenge, a novel adaptive dynamic graph embedded KAE (ADG-KAE) algorithm is proposed in this paper, which is a novel graph representation learning based fault detection approach. In ADG-KAE, a newly designed ADG computation strategy can endow traditional KAE with the capability to extract dynamic information from time-series data, and significantly improve its the FDR of dynamic process monitoring. Unlike conventional dynamic graph computation strategies rely on experience or intuition, the weights of the time-lagged nodes in the ADG are computed adaptively using the least squares (LSs) theory from a new perspective of temporal manifold information reconstruction. Meanwhile, the weight ratio between a node and its time-lagged nodes within ADG is no longer fixed at 1 or a constant value. Instead, it adaptively adjusts based on the approximation error of LS. Afterwards, ADG is combined with the dynamic augmentation matrix method to improve the KAE and construct the ADG-KAE, which can extract dynamic information from adjacent time-series samples, and meanwhile it can retain unique information from its own sample. The effectiveness and superiority of ADG-KAE are validated through experiments on a simple dynamic numerical example, the Tennessee Eastman industrial process dataset, and the real multiphase flow facility industrial fault dataset. It is demonstrated that ADG-KAE achieves outstanding performance in industrial process monitoring and fault detection scenarios.