Interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization

Abstract

The invention provides an interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization, and the system comprises an input module which is used for collecting the data of measurable variables in the system; the optimization module comprises a multi-strategy adaptive differential evolution algorithm and is used for optimizing the structure and parameters of the interval type-2 fuzzy neural network; the calculation module is used for carrying out soft measurement on an unmeasurable process variable based on the optimized interval type-2 fuzzy neural network; and the output module is used for outputting the prediction result and transmitting the prediction result to the industrial control system for feedback and adjustment. The method has high prediction precision.

Description

Technical Field

[0001] The invention belongs to the field of measurement, and in particular relates to an interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization. Background Art

[0002] In modern industrial production, many key process variables (such as temperature, pressure, concentration, etc.) are difficult to measure directly through traditional sensors due to the influence of multiple complex factors. Soft sensing technology has important application value by constructing mathematical models to infer unmeasurable variables from measurable variables. Traditional soft sensing methods, such as those based on classical neural networks and fuzzy logic, often face problems such as complex models, low computational efficiency, and susceptibility to noise interference.

[0003] With the development of differential evolution (DE) and fuzzy neural network (FNN), soft measurement systems based on intelligent optimization algorithms and fuzzy neural networks have gradually emerged in recent years. However, there are still some shortcomings in the existing technology, such as the slow convergence speed of the optimization algorithm and the easy fall into the local optimal solution; the type-II fuzzy logic system has high robustness, but the processing ability of large-scale data sets is still limited. Therefore, there is an urgent need for an improved soft measurement system that can strike a balance between optimization efficiency and model accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide an interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization to improve the accuracy and computational efficiency of the soft measurement model and overcome the shortcomings of the prior art.

[0005] The technical solution of the present invention is as follows:

[0006] An interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization, the system comprising:

[0007] Input module, used to collect data of measurable variables in the system;

[0008] Optimization module, including multi-strategy, adaptive differential evolution algorithm, used to optimize the structure and parameters of interval type-2 fuzzy neural network;

[0009] The estimation module performs soft measurement of unmeasurable process variables based on the optimized interval type-II fuzzy neural network;

[0010] The output module is used to output the prediction results and pass them to the industrial control system for feedback and adjustment.

[0011] Furthermore, the interval type-2 fuzzy neural network includes an input layer, a membership function layer, a rule layer, a reduction layer and an output layer;

[0012] Wherein, the input layer is used to receive external input data, and each node represents an input variable;

[0013] Each node of the membership function layer represents a fuzzy membership function, which is responsible for converting the input signal into a fuzzy membership value;

[0014] The rule layer uses a plurality of fuzzy rules to perform reasoning according to the membership function value output by the membership function layer;

[0015] The reduction layer completes the conversion from the interval type II fuzzy result to the traditional fuzzy set, and reduces the interval fuzzy result to the traditional fuzzy output. This layer calculates the left and right boundaries of the output interval according to the reasoning result, and obtains the final fuzzy output through the turning point calculation method.

[0016] The output layer calculates the final output result using weighting factors according to the output of the degradation layer.

[0017] Furthermore, the multi-strategy, adaptive differential evolution algorithm is specifically:

[0018] First, the initial parameters are set, including the number of rules, the number of populations, the initial generation, the maximum number of iterations and the maximum number of rules. Based on these initial parameters, the initial population is generated. Each population individual contains multiple parameter vectors, involving the center, width and consequent centroid of the antecedent membership function of the fuzzy rule.

[0019] For each generation of individuals in the population, the fitness of the individuals is evaluated by calculating the error between the output of the interval type-2 fuzzy neural network and the actual sample output;

[0020] In each generation, mutation, crossover and selection operations are performed to gradually update the individuals in the population. When the number of generations reaches the maximum number of iterations, the iteration process is stopped. In each generation, the number of rules will gradually increase until the maximum number of rules is reached. Through multiple generations of iterations, the individuals with the best fitness in the population are finally obtained, and the optimized interval type-2 fuzzy neural network structure and parameters are obtained.

[0021] Furthermore, the scaling factor and the crossover probability in the adaptive differential evolution algorithm are both adaptively adjusted.

[0022] Furthermore, the inference module performs soft measurement prediction of unmeasurable process variables based on the optimized interval type-II fuzzy neural network structure and parameters in combination with the real-time data collected by the input module.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] The present invention adopts an adaptive differential evolution algorithm to optimize the interval type-2 fuzzy neural network model, which can accelerate the convergence process of the model, reduce calculation time and improve calculation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to illustrate the embodiments of the invention. Where appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0026] Figure 1 It is the structure diagram of interval type-2 fuzzy neural network;

[0027] Figure 2 Flowchart of optimizing IT2FNN structure and initial parameters for MSADE;

[0028] Figure 3 is the fitness value under different rules;

[0029] Figure 4 The training results of the three models;

[0030] Figure 5 The test results of three models are shown in Figure 2.

[0031] Figure 6 is the training error of the three models;

[0032] Figure 7 is the test error of the three models. DETAILED DESCRIPTION

[0033] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0034] The structure of interval type-2 fuzzy neural network is as follows Figure 1 As shown in the figure, it is a network with five layers, namely input layer, membership function layer, rule layer, reduction layer and output layer. In this paper, the network fuzzy rule adopts the Mamdani form, and its consequent parameter is a centroid interval. The jth fuzzy rule in the interval type-2 fuzzy neural network is:

[0035]

[0036] where x i =[x1,x2,…,x I ] represents the input variable, y represents the output variable, represents the interval type-2 fuzzy membership function, j is the number of fuzzy rules (j=1,2,…,J).

[0037] The functions of each layer of IT2FNN are as follows:

[0038] The first layer is the input layer. Each node represents an input. This layer does not perform any calculations and directly transmits it to the next layer.

[0039] The second layer is the fuzzification layer, also known as the membership function layer, where each node represents a membership function. The Gaussian interval type II membership function has two forms: uncertain center or uncertain standard deviation. This paper uses the Gaussian interval type II membership function with uncertain center, and its expression is as follows:

[0040]

[0041] in,

[0042]

[0043]

[0044] The third layer is the rule layer, also known as the incentive layer. In this layer, each node represents a fuzzy rule. The meet operation under the product-norm rule is used to calculate the incentive strength of the rule antecedent. The incentive interval F of the jth rule is j (x) is represented as follows:

[0045]

[0046] The fourth layer is the reduction layer, which mainly completes the reduction operation of the interval type-2 fuzzy system. Its output is a traditional fuzzy set, which can be expressed as [y l ,y r ], the output of this layer is as follows:

[0047]

[0048] in,

[0049]

[0050] To find the left and right turning points, The KM algorithm can be used to calculate y l(r) The calculation steps of the turning point L(R) are as follows:

[0051] Step 1: Sort in ascending order At the same time, re-sort the J rules by index, and re-sort the Substitute into formula (6) to calculate in

[0052] Step 2: Find contentment The turning point L1(R1) under the condition, when j≤L1(R1), When j>L1(R1), Substitute into formula (6) to calculate y l '(y r ').

[0053] Step 3: If Then continue to step 4. Otherwise, let y l(r) =y l(r) 'And L=L1, R=R1.

[0054] Step 4: Order Return to step 2.

[0055] The fifth layer is the output layer, whose nodes represent the output of the entire network. In order to improve the performance of the network, a weighting factor q is introduced, and the network output is as follows:

[0056] y=qy l +(1-q)y r (9)

[0057] The weighting factor satisfies 0≤q≤1, and y is changed by adaptively adjusting the q value. l and r ratio to improve the accuracy of the network.

[0058] Structural parameter learning algorithm

[0059] The scaling factor F and crossover probability factor CR of the traditional differential evolution algorithm remain fixed during the entire iteration process and cannot change with the iteration process, which makes the algorithm prone to fall into local optimality. Compared with the traditional DE algorithm, the MSADE algorithm proposed in this paper has made three major improvements: first, it has improved the initialization of the population; second, it has improved the mutation strategy; third, it has introduced adaptive scaling factors and crossover probabilities. The following is a detailed description of the MSADE algorithm.

[0060] The traditional differential evolution algorithm randomly generates the initial population individuals, which may result in a certain probability that the solution space cannot be well covered. In this paper, IT2FNN uses an interval type II Gaussian membership function with an uncertain center. If the deviation between the model training data and the center of the membership function is too large, it will lead to no solution in the later degradation process. Therefore, the center of the membership function should be set within a reasonable interval. Therefore, the left and right centers of the Gaussian membership function need to be and width and the centroid of the subsequent piece Encode them separately and set different interval ranges.

[0061] The center vector of the rule antecedent membership function of the initialized population is as follows:

[0062]

[0063] The width vector of the membership function is as follows:

[0064]

[0065] The vector of the centroid of the consequent of the fuzzy rule is as follows:

[0066]

[0067] In summary p The solution vector can be expressed as:

[0068]

[0069] The initial population is represented as

[0070]

[0071] Where k = 1, 2, ..., N P , j is the number of rules.

[0072] In the process of performing mutation operations, it is difficult to balance the development and exploration capabilities of the algorithm by selecting a single strategy. Therefore, this paper adopts multi-strategy mutation operations to improve the optimization efficiency of the algorithm. The two strategies are shown as follows:

[0073] Strategy 1:

[0074]

[0075] Strategy 2:

[0076] V k,E =X k,E +F k,E ×(X best -X pbest,E ) (twenty one)

[0077] where r1, r2, r3, r4, r5 are intervals [1, N p ] are random integers that are not equal to k and are mutually unequal. best Is the best individual in the population. pbest,E and X pworst They are The individuals with the best fitness value and the individuals with the worst fitness in the population evolve in the direction of the difference between the two excellent vectors, thereby producing better individuals and improving the local search ability. The triggering conditions of the above two strategies are as follows:

[0078]

[0079] If the fitness value of an individual in the population is less than or equal to the average fitness value of the individuals in the population, that is, If the population is not optimal, strategy 2 is adopted to continue development based on the better parent generation. Otherwise, strategy 1 is adopted to enhance the exploration ability of individuals in the population, jump out of the local optimum, and seek the global optimum.

[0080] The values ​​of scaling factor and crossover probability in traditional DE are fixed, which has poor adaptability, is prone to fall into local optimal solution, and is prone to premature phenomenon. In order to solve this problem, this paper adopts an adaptive scaling factor F k,E and the crossover probability CR k,E .

[0081]

[0082] Among them, F u =0.9, F l =0.1, F k,E The value of changes adaptively according to the fitness value of the generated new individual.

[0083]

[0084] The crossover probability in the above formula makes different individuals in each generation have different crossover probabilities, CR u and CR l are the upper and lower limits of the crossover probability, and their values ​​are 0.7 and 0.1 respectively.

[0085] MSADE is used to optimize the IT2FNN structure and initial parameters to improve the accuracy of the model. The specific steps of the algorithm are as follows:

[0086] 1: Set the initial number of rules j = 2 and the population size N p =50, initial algebra E=1, maximum number of iterations G1=30, maximum number of rules J=30.

[0087] 2: Generate the initial population under j rules according to the initialization method mentioned above.

[0088] 3: Use sample data to train IT2FNN and calculate the error between IT2FNN output and sample output.

[0089] 4: Calculate the individual fitness function value. The fitness function is used to compare the quality of individuals in the population. In this paper, the root mean square error (RMSE) is selected as the fitness function, that is, the smaller the value of RMSE, the higher the accuracy of IT2FNN. The expression of the fitness function is as follows:

[0090]

[0091] where N is the number of training samples; y(t) and are the output of IT2FNN and the sample output value respectively.

[0092] 5: Perform mutation, crossover, and selection operations, let E = E + 1. If E < G1, return the updated population individuals to step 3. Otherwise, go to step 6.

[0093] 6: Let E = 1, j = j + 1, and return to step 2. When j > 30, stop the iteration. The number of fuzzy rules corresponding to the minimum value of the fitness function RMSE is the optimal number of rules, that is, the network structure and the initial parameters of the network of IT2FNN are determined. Figure 2 It is the flowchart of MSADE optimizing the structure and initial parameters of IT2FNN.

[0094] The parameters of the MSADE-IT2FNN model proposed in this invention are learned using the gradient descent method. Here, the error function is selected as the performance function as shown in the following formula:

[0095]

[0096] where y(t) and are the output of the MSADE-IT2FNN model and the expected output respectively.

[0097] The parameters that need to be learned using the gradient descent method in this paper include the uncertainty center and width of the interval type-2 Gaussian membership function, the consequent parameters and and the weighting factor q. The parameter update formulas are as follows:

[0098]

[0099]

[0100] where ε m ,ε σ ,ε w and ε qThey represent the learning rate of the uncertain center, width, rule consequent parameter and weighting factor respectively. Since there are left and right turning point problems in the KM algorithm, when the function is derived, the parameters should be learned and adjusted according to the turning point partition. The specific derivation process is as follows:

[0101]

[0102]

[0103] When j≤L,

[0104]

[0105] When j>L,

[0106]

[0107] When j≤R,

[0108]

[0109] When j>R,

[0110]

[0111]

[0112]

[0113] Simulation examples and analysis.

[0114] In order to demonstrate the performance of the proposed MSADE-IT2FNN model, it was applied to the soft measurement of starch utilization during wine fermentation and compared with the performance of the FNN and DE-IT2FNN models. In order to fairly compare the performance of the three models, we introduced the root mean square error RMSE as a performance indicator for comparison.

[0115] The liquor brewing process is a complex biochemical reaction process. The various indicators that determine the quality and output of liquor, such as temperature, moisture, starch content, acidity, etc., influence and restrict each other, which makes the design and implementation of online measurement, automatic control and real-time optimization of key parameters in the liquor brewing process very difficult. In the actual production process, operations are often based on the experience of operators, which makes the production cost and labor intensity of operators high, seriously restricting the benefits and development of enterprises. To establish an automatic control and real-time optimization system for the liquor brewing process, it is necessary to first realize the online measurement of key parameters.

[0116] Starch utilization is a key parameter that reflects the yield of liquor in the process of liquor brewing, but its measurement is generally carried out by laboratory chemical analysis, which is an offline measurement and has problems such as measurement lag and large errors. The application of soft measurement technology provides a practical solution to the online measurement of starch utilization. We select five parameters directly related to starch utilization, namely, the amount of reducing sugar, moisture when entering the pool, moisture when leaving the pool, acidity when entering the pool, and acidity when leaving the pool as the input variables of the starch utilization soft measurement model, and starch utilization is the output variable of the soft measurement model. The MSADE-IT2FNN proposed in this paper is used as a soft measurement model, and a starch utilization soft measurement prediction model based on MSADE-IT2FNN is established.

[0117] In a large brewery, 150 sets of sample data were collected. Abnormal data will cause large errors in the trained model. Therefore, before model training, the abnormal data collected in the data should be removed and then normalized. The data normalization formula is shown in formula (54). The normalized data is used as the input of the MSADE-IT2FNN soft sensor model. The first 80 sets of data in the 124 sets of sample data are used for model training, and the last 44 sets of sample data are used to test the model effect.

[0118]

[0119] The fitness values ​​of the DE-IT2FNN model and the MSADE-IT2FNN model at different numbers of fuzzy rules are obtained from the structural learning algorithm above. Figure 3 As shown in the figure, it can be seen from the curve that the DE-IT2FNN model has the smallest fitness value when it has 9 fuzzy rules, while the MSADE-IT2FNN model has the smallest fitness value when it has 4 fuzzy rules, thus determining that the number of fuzzy rules of the two models are 9 and 4 respectively.

[0120] Figure 4 and Figure 5 The training results and test results of the FNN model, DE-IT2FNN model and the MSADE-IT2FNN model proposed in the present invention are shown in the figure. As can be seen from the curves in the figure, the MSADE-IT2FNN model not only has a good fitting effect in the training process, but also has a relatively high prediction accuracy in the test process, and has a strong generalization ability. Figure 6 and Figure 7 They are the training errors and test errors of the three models respectively. From the curves in the figure, it can be seen that compared with the other two models, the errors of the MSADE-IT2FNN model are generally stable and smaller in both the training and testing processes.

[0121] Table 1 lists the three performance indicators of the number of rules, training RMSE and test RMSE of the FNN model, DE-IT2FNN model and the MSADE-IT2FNN model proposed in this paper. It can be clearly seen from Table 1 that the MSADE-IT2FNN model proposed in this paper has the smallest training RMSE value and test RMSE value compared with the other two models, and it has only 4 fuzzy rules and the model structure is relatively simple. The results show that the MSADE-IT2FNN model is suitable for modeling nonlinear systems and complex industrial processes.

[0122] Table 1 Performance comparison of three models

[0123]

[0124] In view of the fact that the key parameters of complex industrial processes cannot be detected online, an online soft measurement modeling method of interval type-2 fuzzy neural network (MSADE-IT2FNN) based on multi-strategy and adaptive differential evolution algorithm optimization is proposed. Compared with the DE algorithm, the MSADE algorithm proposed in the present invention has a fitness value of less than that of the DE-IT2FNN model under almost all rule numbers when the initial parameter settings are all the same, indicating that the MSADE algorithm has good global search capabilities. The parameter learning process is based on the KM algorithm. The left and right turning points are divided into regions and the gradient descent method is used for learning. Since the initial parameters have been optimized in the structural learning process, the network can quickly optimize the model when learning parameters, and the accuracy is high. Finally, the proposed MSADE-IT2FNN model is applied to the soft measurement modeling problem of starch utilization in the wine fermentation process. The experimental results show that the MSADE-IT2FNN model has higher prediction accuracy than the DE-IT2FNN model and the IT2FNN model.

[0125] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization, characterized in that: The system comprises: Input module, used to collect data of measurable variables in the system; Optimization module, including multi-strategy, adaptive differential evolution algorithm, used to optimize the structure and parameters of interval type-2 fuzzy neural network; The estimation module performs soft measurement of unmeasurable process variables based on the optimized interval type-II fuzzy neural network; The output module is used to output the prediction results and pass them to the industrial control system for feedback and adjustment.

2. The interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization according to claim 1 is characterized in that: The interval type-2 fuzzy neural network includes an input layer, a membership function layer, a rule layer, a reduction layer and an output layer; Wherein, the input layer is used to receive external input data, and each node represents an input variable; Each node of the membership function layer represents a fuzzy membership function, which is responsible for converting the input signal into a fuzzy membership value; The rule layer uses a plurality of fuzzy rules to perform reasoning according to the membership function value output by the membership function layer; The reduction layer completes the conversion from the interval type II fuzzy result to the traditional fuzzy set, and reduces the interval fuzzy result to the traditional fuzzy output. This layer calculates the left and right boundaries of the output interval according to the reasoning result, and obtains the final fuzzy output through the turning point calculation method. The output layer calculates the final output result using weighting factors according to the output of the degradation layer.

3. The interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization according to claim 1 is characterized in that: The multi-strategy, adaptive differential evolution algorithm is specifically: First, the initial parameters are set, including the number of rules, the number of populations, the initial generation, the maximum number of iterations and the maximum number of rules. Based on these initial parameters, the initial population is generated. Each population individual contains multiple parameter vectors, involving the center, width and consequent centroid of the antecedent membership function of the fuzzy rule. For each generation of individuals in the population, the fitness of the individuals is evaluated by calculating the error between the output of the interval type-2 fuzzy neural network and the actual sample output; In each generation, mutation, crossover and selection operations are performed to gradually update the individuals in the population. When the number of generations reaches the maximum number of iterations, the iteration process is stopped. In each generation, the number of rules will gradually increase until the maximum number of rules is reached. Through multiple generations of iterations, the individuals with the best fitness in the population are finally obtained, and the optimized interval type-2 fuzzy neural network structure and parameters are obtained.

4. The interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization according to claim 3 is characterized in that: The scaling factor and crossover probability in the adaptive differential evolution algorithm are both adaptively adjusted.

5. The interval type-2 fuzzy neural network soft measurement system based on multi-strategy and adaptive differential evolution algorithm optimization according to claim 1 is characterized in that: The inference module performs soft measurement prediction of unmeasurable process variables based on the optimized interval type-II fuzzy neural network structure and parameters in combination with the real-time data collected by the input module.

Images