Robust neural networks with random weights based on generalized M-estimation and PLS for imperfect industrial data modeling

Actual industrial data inevitably contain a variety of outliers for various reasons. Even a single outlier may have a large distortion effect on modeling performance with conventional algorithms, not to mention the complicated process modeling by the imperfect industrial data existing various outliers both in input direction and output direction. Therefore, the robustness of the algorithm must be fully considered in modeling of complicated industrial processes. Aiming at this, the robust neural network with random weights based on generalized M-estimation and PLS (GM-R-NNRW) is proposed for data modeling of complicated industrial process, whose samples coexist input and output outliers and have multicollinearity problem. Firstly, the input weights and biases of the proposed GM-R-NNRW are randomly assigned within their respective given ranges. Secondly, the GM-R-NNRW determines the weights of the sample by the residual size of the model and the distance information of the input vector in the high-dimensional space according to the generalized M-estimation. Then these weights were combined to determine the final model contribution of each sample, solving the problem that the samples exist both the input direction and the output direction outliers. Moreover, the improved PLS is used to solve the multicollinearity problem existing in data samples. Finally, both data experiment and actual industrial application have showed that the general approximation performance of the algorithm is greatly improved, and an easy-to-use model with better accuracy and robust performance can be obtained.