Neural network expert system for filter control in

Neural network expert system in refinery filter control

filter is the key equipment for wax removal in refinery. The process flow of filter is shown in Figure 1. When the filter is working, its feed volume, discharge volume and filtrate surface are unmeasurable or uncontrollable parameters and change randomly, so it is difficult to identify and control its state through the conventional control system. In actual operation, operators determine the regeneration cycle of filter warm washing by experience after observing the working condition of wax conveyor. It is difficult to ensure the timely warm washing of the filter (especially the night shift time), which reduces the comprehensive benefits of the device. Frequent warm washing of the filter will increase energy consumption and long shutdown time, resulting in waste; If it is not washed warm for a long time, the effect of the filter will be reduced, which will reduce the yield of lubricating oil and also cause economic losses. The expert system applied to process control can imitate the control technology of human experts and operation experts, and tirelessly complete high-quality control. This paper will introduce the applied research in this field

1 neural network expert system has the structure of smooth path and steep slope system.

the reasoning mechanism of neural network expert system is different from the logic based deduction method used by the existing expert system. Its reasoning mechanism is a numerical calculation process, which mainly consists of the following three parts:

① the transformation from input logic concept to input mode, and determine the transformation rules according to the characteristics of the domain, and then according to the corresponding rules, Transform the current state into the input mode of the divine meridians

② forward calculation in the collateral: according to the characteristics of neurons, its input is Xi = ∑ tijyi, TIJ is the connection weight coefficient, Yi is the output of neurons and there is Yi = fi (Xi ＋) θ i）。 among θ I is the threshold value of neurons, FI is a monotonically increasing nonlinear function, and the output mode of neural network can be generated through the above calculation

③ interpretation of output mode: with different domains, the interpretation rules of output mode are also different. The main purpose of interpretation is to transform the output numerical vector into high-level logic concepts

in the neural network expert system, the classification logic standard described by clear language is not used, and the classification standard is determined only according to the similarity of the samples currently received by the system, which is mainly reflected in the weight distribution of the network. At the same time, the neural network algorithm can be used to acquire the knowledge expression system of knowledge through learning, that is, the uncertainty reasoning mechanism. Knowledge acquisition includes the structure of neural network (network layers, input, output and number of hidden nodes); Organize learning samples to be trained; The neural network learning algorithm is used to obtain the required weight distribution through the learning of samples, so as to complete the knowledge acquisition. Knowledge base is obtained by automatic knowledge acquisition. It is the basis for reasoning mechanism to complete reasoning and problem solving. Knowledge base can be constantly innovated, which is manifested in the new network parameter distribution of acquiring more knowledge and experience after learning new samples on its basis

2 implementation of neural network expert system

2.1 preprocessing algorithm

set the sampling data to have a total P dimension, and take nearly n groups of data, then there is the sampling matrix X = (x1, X2,..., XP) T. in order to eliminate the error of the data, the average value of nearly n groups is taken as the new input data for each input data, that is,

based on the initial state P0 after the warm washing of the filter, the backtracking comparison method is adopted, and the state change is taken as the input of the network, Based on this, a training data matrix PP/M is established, where p is the dimension of the data and M is the dimension of the training data

2.2 modeling of neural network

this network is composed of three layers, in which there are 11 units in the input layer and 1 unit in the output layer, and the number of units in the middle layer is finally determined through experiments. The system adopts LM algorithm for network learning and operation. The selection of training model samples assisted by the Ministry of commercial energy and industrial strategy has become the key to the success of the system. Theoretically, the input modes faced by the system after it is put into operation are infinite, but it is impossible to learn all these modes as sample modes in the network training process. When installing samples, users can carry out installation and calibration according to their own habits: carry out the verification of force measuring system, so how to select learning samples is a problem worth studying. Considering the neural network has the associative function of generating similar outputs according to similar inputs, the expert system selects some typical data including the upper and lower limits of various parameters as the learning samples of the network

in addition to the selection of learning samples, the selection of various parameters of the network itself will also affect the working performance of the system. Figure 2 shows the relationship between learning times and error square. Of course, for different learning samples, the optimal parameters and choices of the network will also be different, but the change will not be great, just make some adjustments. The network parameters finally determined by the experiment are shown in Table 1

2.3 filter warm washing control neural network expert system

the optimal filter warm washing control neural network expert system diagram is shown in Figure 3. It synthesizes the knowledge representation of neural networks, that is, the operation rules and time factors that are difficult for the operation experts to express in language, as well as some expert rules, realizes parallel associative reasoning, and improves the intelligence level, real-time processing ability and robustness of the expert system

Figure 3 block diagram of warm washing control neural network expert system

3 simulation experiment results

after learning 64 groups of sample patterns, the simulation is carried out. Figure 4 shows the comparison between the actual output of the network and the actual regeneration signal. The results show that the output of the neural network can predict the time of the optimal warm washing signal, and the predicted value can send the regeneration signal earlier than the actual value. The reason is that the neural network expert system has the function of reproducing the operation, which can judge when the system reaches the overload operation and send the signal. Figure 5 shows the comparison between the processed output and the actual regeneration signal. It can be seen that after the output is blurred, the system can operate more accurately and stably

figure