Research And Application Of Soft Sensing Modeling Method In Chemical Process

In the chemical production process,in order to ensure product quality and production safety,it is necessary to perform soft-sensing on some key quality indicators that cannot be directly measured.Traditional first-principle methods,which develop soft-sensing models based on reaction principle,are powerless to get precise prediction results because of the characteristics of highly nonlinear,dynamic and time-delay of chemical production process.In contrast,the data-driven modeling methods are developed based on the actual measured data in the factory,which can describe the real reaction process.Compared with firstprinciple models,data-driven models have more flexibility and realistic relevance.With their powerful learning and representation capabilities,they can fully dig out important information in historical data,establish accurate forecast models for key quality indicators,and save a lot of time and resources.This paper focuses on the research on the data-driven soft-sensing methods of chemical process,and the methods proposed in this paper are studied and applied on the actual production process of polypropylene.In particular,the main contributions of this article mainly include:(1)Obtaining an effective data set is a prerequisite for establishing an accurate soft-sensing model.Aiming at the problem of varying degrees of delay between offline data and process variables,this paper proposes a delay correlation analysis algorithm based on adaptive genetic algorithm,which is then applied to the delay correlation detection of chemical time series data.It can be concluded from the experiment results that the proposed method can bring benefit to reduce time overhead of traditional delay correlation detection methods with a guaranteed accuracy,and is useful for establishing accurate chemical process soft-sensing models,which is of great significance.(2)Standard Extreme Learning Machine(ELM)is often difficult to establish accurate soft-sensing models when facing high-dimensional,nonlinear and noisy chemical process data.Therefore,this paper proposes a supervised Auto-Encoder-based ELM(SAE-ELM)soft-sensing modeling method.SAE-ELM introduces supervision signals to constrain the feature learning process of the auto-encoder,thereby important features related to the target space are learned.Then the features are used as the input of the ELM.It can be concluded from the experiment results that the SAE-ELM can obviously improve the modeling accuracy and robustness of standard ELM.(3)Traditional soft-sensing modeling methods usually only focus on the current state of the system limited by structural factors.They can't fully excavate the timing characteristics in the chemical process data,and have trouble in establishing effective multi-step forecasting models.In response to this problem,this paper proposes a soft-sensing model based on Gated Convolutional neural network-based Transformer(GCT),which is able to perform multi-step prediction of key quality indicators.The GCT model uses a gated convolutional neural network to mine short-term time patterns in time series data and screen important features,and uses attention mechanism to model the correlation of time series data between any two moments in a parallel manner.This paper uses two UCI time series data sets and polypropylene actual production data set as examples to verify the effectiveness of the GCT model.Finally,the above methods are applied to polypropylene soft-sensing modeling in the form of microservices,which can provide guidance for production optimization control and product quality management.