Adaptive parameter tuning stacked autoencoders for process monitoring

In process monitoring based on stacked autoencoders (SAEs), the performance of monitoring models is directly decided by the validity of the structure and parameters, which are primarily determined by time-consuming manual adjustments. This paper presents a novel method that can adaptively select parameters rather than tuning them manually. The proposed method is called adaptive parameter tuning SAE (APT-SAE). Basic SAEs aim to compress the original input data and extract simple and abstract features. Thus, the redundant information of each hidden layer output should be as small as possible. The next layer of nodes can be remarkably reduced if the amount of redundant information is large. During the pre-training stage of APT-SAE, an adaptive parameter tuning strategy is used for rapidly determining the number of layers and nodes in the paper. The cross-covariance of each AE’s input data is used to determine the node number of succeeding AE. The pre-training stage ends when the correlation is weak, which is decided by the average value of cross-variance matrix. The proposed method is applied to a benchmark problem, and it outperforms several state-of-the-art methods.