Adaptive attention principal component analysis with continual learning ability for multimode process monitoring

Aimed at multimode nonlinear nonstationary processes, this paper investigates an adaptive attention principal component analysis with continual learning capability, which assumes that data from multiple modes are received in a sequential manner. To settle the nonlinearity, attention mechanism is employed to characterize the relationship between latent variables and original variables within PCA. When a new mode arrives, the monitoring model is updated by consolidating the previously learned knowledge and assimilating new significant features simultaneously. In addition, elastic weight consolidation is adopted to measure the importance of attention PCA model parameters. When the system operates normally in the same mode, the attention PCA model would be updated adaptively based on the forthcoming normal samples. This paper illustrates the effectiveness of the proposed method with a practical industrial system.