Data Mining Technology and Its Application in Chemical Process

Technological Progress Data Mining Technology and Its Application in Chemical Process Yu Zhangyi 1 Xia Luyue 2 Pan Haitian 2 1. Zhejiang Zhongyuan Chemical Co., Ltd., Jinhua 321000; 2. School of Chemical Engineering, Zhejiang University of Technology, Hangzhou 310032) Advanced technologies such as statistics and statistics Mining and discovering valuable and implicit knowledge in data information. This paper introduces some data mining techniques that are mainly used in chemical processes, such as rough set theory, artificial neural networks, genetic algorithms, principal component analysis and partial least squares, and their modeling, control, and fault diagnosis in chemical processes. The application of optimization is reviewed.

With the development of computer storage technology and network technology, a large number of chemical production process data are collected and stored in various databases, and in these databases there are a large number of technologies that can improve chemical process modeling, control, fault diagnosis, etc. Valuable information, however, because these data sets have characteristics such as multi-factor, nonlinearity, high noise, and non-homogeneous distribution, if traditional methods are used to analyze and process these huge amounts of data, it is not only time-consuming but also difficult to effectively mine and Discover the hidden information in the data.

Information technology represented by computers has played an ever-improving role in the development of the chemical industry since it was created. It has promoted the further integration of information technology and the chemical industry. Therefore, people are overwhelmed by large amounts of data, but they are eager to be implicitly In the case of information in the data, Data Mining (DataMmmg) technology came into being I and related data mining standards have been formed. H. This paper focuses on data mining technology and its application in chemical process modeling, control, fault diagnosis and optimization. Application review.

1 Data mining technology Data mining technology is an important research content of today's intelligent system theory and technology. It comprehensively uses advanced technologies such as artificial intelligence, computational intelligence, pattern recognition, and mathematical statistics, from a large number of incomplete, noisy, fuzzy In the random data, we extracted the information and knowledge implicit in it that people didn't know beforehand, but were potentially useful. In recent years, we have received great attention at home and abroad and studied the fields of finance, medicine, and administrative management. Such as: fuzzy controller modeling, fault diagnosis modeling, DNA sequence analysis, process data prediction, correlation analysis, engineering design%.

system.

1.1 Rough Set Theory Rough Set Theory (RS) is a new type of mathematical tool for dealing with ambiguities and inaccuracies. Rough set theory is characterized by the fact that there is no need to predetermine the quantitative description of certain features or attributes. Instead, it directly proceeds from the set of descriptions for a given problem, determines the approximate domain of a given problem through indiscernibility and indiscernibility, and thus The inherent law in the problem is an important data mining tool, and the rough set theory is very appropriate when dealing with the uncertainty in different types of data.

1.2 Neural Network Artificial Neural Network (ANN) Structurally imitating biological neural networks, a large number of simple neurons connect to form a network system in accordance with certain rules in order to achieve the goal of simulating the human intuitive image thinking. The neural network method has greater advantages when it is used for nonlinear data and noise-containing data, and it can perform various data mining tasks such as classification, clustering, and feature mining.

1.3 Genetic Algorithms Genetic Algorithms GA) Originated from evolutionary and population genetics, the adaptive phenomenon originally used to model natural systems was later introduced to a wide range of engineering problems. Genetic algorithm is a kind of bionic global optimization technology. It simulates the life evolution mechanism and passes the inferior initial solution through a set of genetic operators. It searches iteratively according to certain random rules in the solution space until the optimal solution of the problem is obtained.

1.4 Statistical Analysis Methods Statistical analysis methods are one of the most basic data mining techniques. Among them, principal components analysis and partial least squares methods are widely used in chemical engineering. Principal component analysis (PCA) can achieve the following goals: data reduction, data compression, modeling, variable selection, classification, and forecasting. Partial Least Squares (PLS) mainly focuses on the multivariate regression modelling of multiple independent variables. Partial Least Squares regression is more effective when the internal variables are highly linearly correlated.

1.5 Data Mining Integration Techniques Since each data mining technology has the disadvantages that some methods cannot avoid, it is possible to combine these data mining technologies to achieve each other's advantages and achieve better than using a single data mining technology. effect.

1.5.1 Neural network and genetic algorithm BP algorithm is the most widely used neural network training algorithm, but because BP algorithm uses error derivative to guide the learning process, it is essentially a local optimization algorithm, especially for modeling complex problems. At the minimum, it is easy to fall into the local minimum point, and the combination of neural network and genetic algorithm can solve this problem well. In the process of establishing a neural network model, a genetic algorithm can be used to optimize the neural network structure and learning parameters mm. 1.5.2 Principal component analysis and genetic algorithm Principal component analysis is a standard linear statistical technique, often used at large heights In the analysis of related data, however, this has limited the application of PCA in non-linear chemical systems. Therefore, PCA can be combined with genetic algorithms to extend the application of principal component analysis to the processing of nonlinear data sets.

1.5.3 Neural Networks and Principal Component Analysis, Partial Least-Squares Since principal component analysis and partial least-squares methods both have the ability to compress data and extract information, principal component analysis and partial least squares methods can be used to perform sample data sets. Processing to form a new training and test sample set, reducing the number of inputs in the neural network modeling network, while eliminating the correlation of the input factors and simplifying the network structure, greatly improving the neural network learning rate.

2 Chemical Process Applications 2.1 Chemical Process Modeling Neural networks are widely used in chemical process modeling and often use a combination of prior knowledge and neural networks because prior knowledge is a form of training when sparse and noisy data are trained. Means to improve the predictive power of neural networks. This method has been used in the prediction of the number of cell and secondary metabolites during the penicillin fermentation process. Genetic programming is used to model the input-output process of experimental data. The chemical reactor system obtained by using this method simplifies the model structure accurately. Reflect the corresponding process characteristics.

Most chemical processes are nonlinear systems, so fuzzy neural networks can be used to model nonlinear chemical processes. M studies the fuzzy neural network modeling of a pH reactor. The results show that the fuzzy neural network model is easier to understand than the ordinary neural network model, and can use the process information to explain the trained fuzzy neural network weights.

The use of combined neural networks instead of a single neural network can improve the generalization performance of neural networks. Therefore, Sridhar et al.19 studied the dynamic modeling of the combined neural network of nonlinear chemical processes, and illustrated the application of the method in three examples. It can improve the performance of chemical process neural network models.

2.2 The possibility of combining chemical process control strategies, thus proposes a predictive control strategy based on neural network model, and uses this control strategy in a nonlinear distillation system. The application results show that based on the neural network control system potential.

Model feedback controllers are generally based on linear models, and many chemical processes are nonlinear. m facilitates the use of the neural network's nonlinear mapping property, applies the neural network to nonlinear chemical process control, and demonstrates the performance of the method through a continuous stirred reactor control example.

Because the principal component analysis and partial least squares method have the characteristics of data dimension compression, it is proposed to be used in the feedforward control design of multivariable systems. Three research examples illustrate that this method can be applied to multivariable linear and nonlinear systems. In control.

2.3 Chemical process monitoring and fault diagnosis Through the multivariate statistical analysis of the operating data of the chemical production process, fault diagnosis of the chemical process can be achieved. The essence of these methods is to make use of the highly relevant characteristics of chemical process data, use PCA and PLS methods to reduce the dimension of process operation data, thus establishing a low-dimensional data model, so that process monitoring can be performed in this low-dimensional space.

Analytical methods are used for batch process analysis and monitoring. This method not only allows the process variables to be significantly reduced, but also eliminates time-varying problems in intermittent operations. M uses the projection technology of principal component analysis to achieve the goal of reducing the dimensionality of the data set. The principal component analysis method is used to process the data set so that the process monitoring of the polymerization reactor is performed in a low dimensional space.

The method is applied to the on-line fault diagnosis of a CSTR system, in which signals of different process faults are determined by performing multivariate statistical analysis on the online monitoring data.

2.4 Chemical Process Optimization The advantage of optimizing neural networks as an optimization tool is that the full mapping of the objective function allows easy identification of multiple optimal states, with important features not found in other traditional optimization methods, and the constraints are easy to handle. A backpropagation neural network based on feedforward structure is applied to a separation process optimization. The application results prove the feasibility of neural network as an optimization tool. Neural networks as an optimization tool have also been successfully applied to a nylon-6,6 polymerization process and an acetic anhydride production process. In addition to neural networks, genetic algorithms can also be used as an optimization tool and have been successfully applied to batches. Distillation optimization.

3 Conclusion Data mining technology can extract valuable and implicit knowledge from a large number of chemical process data, so it not only has important research and application value in chemical process modeling, control, optimization, fault diagnosis, etc., but also in the chemical process. Design, process planning and operation, and product design will also play an increasingly important role.

Data mining is a brand-new technology. It is currently at the initial stage both at home and abroad. Various theories, methods, tools, software, and applications are not mature, and from a general point of view, these data mining technologies are for specific backgrounds. There are still some limitations in the application of chemical processes, but data mining is a hot research field. With the development of chemical process monitoring and control hardware equipment technologies, computer control systems can collect and store large amounts of process measurement data. New technologies and new applications will continue to emerge. It can be expected that in the future data mining technology will be further developed and improved, and its application prospects in the chemical process will be even broader.