ISDAE Model Based Operating Performance Assessment For Complex Industrial Processes

The detailed and robust operating performance assessment and non-optimal factor identification method of complex industrial process is one of the effective means to ensure its safe,green and efficient production and improve the comprehensive economic benefits of enterprises.Aiming at the practical problems of strong nonlinearity,data redundancy,and insufficient label data(more unlabeled data and less labeled data)in complex industrial processes,this paper proposes two types of auto-encoder-based operating performance assessment and non-optimal factor identification methods.The specific work content is as follows,(1)Aiming at the problems of non-linearity,information redundancy and uncertainty in complex industrial process,which make it difficult to establish a robust and reliable assessment model,this paper introduces a sparse denoising autoencoder(SDAE)model in deep learning theory,and on this basis,proposes an integrated economic indicator regularized SDAE(ISDAE)model is proposed.Based on the unsupervised feature learning of SDAE,the constraint term of comprehensive economic index is introduced to force the autoencoder to learn the raw data feature expression related to comprehensive economic index,and use it as input to train the operating performance recognition model.Then,cascade the feature extraction model and the operating performance recognition model,and fine-tune the parameters to obtain the final operating performance assessment model.In online application,sliding window and trust weight are introduced to improve the reliability of online assessment results.When the operating performance is not optimal,this paper proposes a non-optimal factor identification method based on the contribution graph algorithm of autoencoder,which can identify the non-optimal factor by calculating the contribution rate of the corresponding variables.Finally,the proposed method is applied to the operating performance assessment of dense medium coal preparation process,and its effectiveness is verified by simulation experiments.(2)Aiming at the problem of strong non-linearity in complex industrial processes,this paper expands the ISDAE model to the deep structure model by stacking,that is,a deep ISDAE(DISDAE)feature extraction model,to learn more abstract strong non-linear feature expression.In addition,affected by factors such as production conditions,environmental disturbances,and human errors,information related to comprehensive economic indicators during the process cannot be recorded in real time,which limits the ability of the DISDAE model to learn feature expression related to some extent.At the same time,the lack of such information will also lead to insufficient state-level label data,thus influencing the learning effect of the operating performance recognition model.To this end,this paper introduces a semi-supervised learning mechanism and proposes a pseudo-label based semi-supervised learning algorithm.This method utilizes a large amount of distribution information of unlabeled data,and continuously adds high-confidence pseudo-labeled data to the labeled training data set through iterative learning,so as to continuously improve the model's learning ability with a small labeling cost.What's more,in view of the inevitable existence of some low-quality and irrelevant training data in the training data set,this paper also gives a data culling criterion based on assessment model accuracy to reduce the impact of such data on model performance.Finally,the proposed method is applied to the operating performance assessment of the dense medium coal preparation process,and the simulation results verify the effectiveness and practicability of the proposed algorithm.