Computer Engineering Development Research and Design Technology. Article ID: 1000-3428(2003)17―0193―02:A:TP389.1 Application of Neural Network and Genetic Algorithm in Traditional Chemical Enterprises Kong Wenjun, Liu Shiyuan, Du Runsheng (Department of Mechanical and Electronic Information Engineering, College of Mechanical Engineering, Huazhong University of Science and Technology, Wuhan, China) 430074) 1 Overview After entering the information age, various data are accumulating at an alarming rate. In addition to the continuous accumulation of chemical and chemical research work, there are thousands of sensors continuously recorded in a modern chemical plant. In recent decades, with the advancement of computer technology, a large number of databases have been established. This data contains a lot of valuable information. The data mining objects in chemistry and chemical industry are mostly multi-factor, high-noise and non-linear data sets. This information mining of complex data is a big problem.
Chemical reactions are generally affected by many factors, such as temperature, pressure, and material ratio. The chemical process has more influencing factors. In addition to the above factors, it is also related to the properties of the material and the process conditions. Therefore, chemical and chemical data processing is mostly a multi-factor problem.
Classical statistics mainly deal with the linear relationship, but the multi-factorial problems in chemistry and chemical industry do not obey the linear relationship, and the chemical reaction obeys the fixed-ratio law. Therefore, chemical and chemical data processing must deal with nonlinear relationships.
Noise artifacts distort the target value or argument. The noise may be white noise or colored noise. Many chemical and chemical phenomena are sensitive to many factors. Therefore, when only a limited number of independent variables are used to describe an object, those that are ignored become noise.
In chemical production, the technical and economic indicators of the product (eg, yield, conversion, energy consumption, etc.) are all influenced by the factors of the raw material formulation (eg, raw material composition, catalyst composition, etc.). The fluctuation of these factors will cause the products of different batches and different times to vary greatly. In order to achieve a satisfactory result, people have used various search technologies to try to find a correct direction to guide the production. For example, optimization methods such as analog tuning and statistical tuning are widely used in production. in. This paper proposes a method of formula optimization - neural network and genetic algorithm method (ANN-GA).
Compared with these optimization techniques, one of the features of ANN-GA optimization technology is that it is particularly suitable for a variety of complex situations. Many industrial processes have many influencing factors, the mechanism of the process has not yet been ascertained, and traditional computer simulations are powerless; many process influencing factors are not linearly related to the optimization goals, and traditional regression methods have not been satisfactory, but these complex situations can be Find the direction of optimization through the ANN-GA method.
2 Artificial neural network and genetic algorithm principle The basic idea of ​​artificial neural network (ANN) is to simulate the human brain's nervous system from the perspective of bionics, so that it has the ability of human brain to perceive learning and reasoning. Its advantages include: highly non-linear mapping capabilities, ability to process incomplete data and noisy data, self-learning, adaptive and generalization capabilities, fast and efficient computing capabilities, fault tolerance, flexibility, and ease of maintenance. Neurons are basic processing units in neural networks.
This paper adopts the most commonly used error back-propagation algorithm (BP). At the same time, the batch learning method and the additive method are used to accelerate the convergence speed.
Genetic algorithm (GA) is a global optimization adaptive probability search algorithm developed by referring to the natural selection and genetic evolution mechanism of organisms. It uses a group search technique to generate a new generation of populations by applying a series of genetic manipulations such as selection, crossover, and mutation to the current population, and gradually enables the population to evolve into a state containing or close to the optimal solution. P. 3 Genetic Algorithms and Neural Networks Network Collaborative Solution Formula Optimization Problem The optimization problem is to determine the ratio of raw materials under the constraints of given raw material quality indicators and optimal product performance parameters. There are many factors that affect the major performance indicators of a product. The proportion of various materials is one of the factors, such as A, B, C, and D. On the other hand, the performance parameters of the material (or other factors such as temperature, etc.) also have a great influence on the performance of the product.
If the formulation mathematical model is described as a mapping relationship with multiple inputs and multiple outputs /F =/(magic, the matching optimization problem can be described as: the known input vector X is the performance index of the product), and the component in the input vector X is required to be calculated. q ",. "1cn is a fund project: National Natural Science Fund Project (59985005) and Genetic Algorithm; Liu Shiyuan, Associate Professor, Ph.D.; Du Runsheng, Prof. n Proportional ratio of the main materials, a is the main material that has a great influence on formula performance. Performance parameters or other influencing factors.
The process of formula optimization using BP neural network combined with GA is as follows.
Optimizing Performance Proportioning Optimization Optimization Processes Based on Neural Network and Genetic Algorithms for Optimizing Proportional Solutions Using genetic algorithms to solve optimization problems requires first defining a fitness function to represent the adaptability of a feasible solution. This paper uses the strong nonlinear mapping ability of ANN to establish the mapping between decision variables and objective function values. Then the genetic algorithm is used to search the decision variables under constraint conditions. Each set of decision variable values ​​is searched and input. In a well-trained network, the network automatically matches it with the learned knowledge and infers and predicts the corresponding fitness function value. From the fitness function value, the adaptability of the set of decision variables is obtained. According to its adaptability, the mutation operation is performed to find the global optimal solution. The solution steps are as follows: (1) Initialize the group. M individuals are randomly generated (ie, the coding of M matching schemes) as the initial group.
(2) Calculate the objective function. The neural network is used to forecast the performance parameters of products as the objective function of each individual in the group.
(3) Fitness evaluation. Calculate a transformation of the objective function to evaluate the fitness of all individuals in the population: If the total fitness distribution meets the optimization requirements or the population evolution algebra has reached the preset value, the calculation is terminated and the optimal solution is output (ie, the most Excellent ratio); otherwise, the following 3 genetic operations are performed.
(4) Select the operation. According to the individual's fitness, according to certain rules and methods, select good individuals from the group to inherit to the next generation of groups.
(5) Crossover operation. Pairs of individuals within a group are randomly paired, and for each pair of individuals, some of the chromosomes between them are exchanged for a certain probability (called crossover probability).
(6) mutation operation. For each individual in the population, the genetic value at one or more loci is changed with a certain probability (called the mutation probability).
(7) Create new groups. After group selection, crossover, and mutation operations, the next-generation group is obtained, and the process proceeds to step 2 for calculation.
4 The realization system of the system can be decomposed into two modules: the main product performance forecasting system and the matching scheme optimization system.
The specific functions of each module are shown in Table 1.
Table 1 Specific functions of each module Product performance forecasting system database Knowledge discovery preprocessing module extracts the data pair set that reflects the product performance parameters and its influencing factors from the formula database, uses the data to select and refine the set, and forms a neural network learning sample. Neural network based data mining and formula modeling module to build a neural network model to determine the topology of the network, use training samples to train the network, determine the connection weight of the network based on neural network product performance forecasting module using trained nerves Network, after given raw material ratio and raw material quality index (or other influencing factors), quickly calculate the product's performance parameters. Proportioning scheme optimization system. Genetic algorithm-based ratioing scheme. Optimization module using genetic algorithm combined with trained Neural network, under the constraints of given raw material quality indicators (or other influencing factors) and product performance, to determine the optimal proportioning scheme neural networks and genetic algorithms "in addition to optimization techniques can be used to optimize the formula, you can also use In industrial process control, equivalent to excellent The expert system is characterized by optimizing the parameters that influence the target according to the objectives, and selecting parameters that have a more important influence on the target in many influencing parameters; after network training, finding out the optimized region, finding out the optimization direction, and quantifying the forecast results, The production operating conditions are always kept in an optimized state and the production potential is tapped as far as possible. It has a wide application prospect in the process industry and can be widely used in ceramics, rubber, solid propellants, metallurgy, building materials and other industrial production.
5 Conclusion The use of genetic algorithms to solve the traditional optimization methods for complex optimization problems easily fall into the local minimum problem, the use of genetic algorithms combined with neural networks, both the use of the neural network nonlinear mapping network reasoning and prediction functions, but also the use of genetic algorithms Global optimization features combine the two to complement each other.