Complex chemical process operation evaluations using a novel analytic hierarchy process model integrating deep residual network with principal component analysis

Abstract

The analytic hierarchy process (AHP) is usually employed to evaluate complex decisions, in which the assigned weights are specified artificially. In response, neural networks instead of manual selections can be used to optimize the assigned weights to avoid subjective errors. However, traditional neural networks usually suffer local extremums and cannot extract the deep features in the data. At the same time, variables that do not necessarily participate in the evaluation will also affect the results. In this context, a novel deep residual network based analytic hierarchy process which integrates principal component analysis algorithm (PIDRN-AHP) is suggested in this paper. Firstly, the operation variables most related to the target variables are extracted as the input data by principal component analysis (PCA) method. By fitting the residuals between input data and mapped output data with multiple continuous stacked non-linear computational layers, a deep residuals network is constructed with residual block local depth neural network structure units. Taking advantage of the deep network, the PIDRN-AHP can achieve optimal solutions of AHP global weights, avoid subjective errors and avoid possible gradient disappearance of deep networks due to the layer increase. The proposed method is applied to operational evaluations of Tennessee Eastman Process (TEP). Compared with the traditional AHP, BP-AHP and DRN-AHP experiments show that the PIDRN-AHP can achieve the most satisfied performances.