Network evaluation method for neuro-fuzzy system having learning ability

Abstract

To provide a network evaluation method for a neuro-fuzzy system having learning ability.SOLUTION: In a method, a neuro-fuzzy system model adopting a fuzzy logic rule is constructed under a framework of a neural network, the neuro-fuzzy system is trained using an improved particle intelligent optimization algorithm to obtain optimal parameters and a trained neuro-fuzzy system model, and system input is inputted to the trained model according to an actual problem(s) to acquire the ability boundary evaluation results of the system. An alternative model of a real system is constructed by combining the reasoning ability of the fuzzy logic with the infinite approximation function ability of the neural network, and the alternative model is made similar to a real model to the extent possible using the intelligent optimization algorithm.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to a method for evaluating a network of a neural-fuzzy system having a learning ability.

Background Art

[0002] In recent years, an online ability boundary evaluation model based on real-time characteristic parameters has been rapidly developed in the intelligent technology. The research and application fields of this model cover many fields such as industrial production, energy, environmental monitoring, and medical and health. Currently, various online ability boundary evaluation models for specific fields have been proposed and applied. With the development of the Internet of Things (IoT) technology and sensor technology, the real-time data collection technology has been continuously improved, and the speed and accuracy of data collection have also been continuously increased, providing reliable data support for the application of the online ability boundary evaluation model based on real-time characteristic parameters. With the continuous development and improvement of artificial intelligence and machine learning technologies, the algorithm of the online ability boundary evaluation model based on real-time characteristic parameters has also been continuously optimized, and the boundary of the target system can be predicted and controlled more accurately. While gradually promoting practical application, this model continues to promote applications in many fields such as fault detection and early warning of equipment during industrial automated production, fault diagnosis and control in the power system, and health monitoring and diagnosis in the medical and health field, and has been widely applied and verified in any of them. In Patent Document 1, focusing on the evaluation of the temperature control ability of an energy storage container, collecting multi-dimensional information and environmental parameters of the energy storage container, and using an evaluation model to predict the temperature control results of the energy storage container under different environmental conditions, the purpose is to evaluate the temperature control ability, but high requirements are imposed on the complexity of the model, the quality, and the completeness of the data. In Patent Document 2, in order to determine the index weight coefficient for evaluating the emergency response ability of the power grid, a combination of a subjective pre-determination analysis method and an objective pre-determination analysis method is adopted. However, in this method, since the evaluation is performed using quantitative indexes and weight coefficients, there are limitations in dealing with very complex and dynamically changing situations.

[0003] The fuzzy comprehensive evaluation method is an evaluation method based on fuzzy mathematics. By quantitatively describing fuzzy information through the membership function, it solves fuzzy problems. However, in the fuzzy comprehensive evaluation method, the determination of the membership function has no consistent acceptable criteria and is greatly affected by subjectivity. In recent years, type-2 fuzzy theory has been gradually applied and achieved success in the field of fuzzy control, and research on type-2 fuzzy clustering is also underway. However, due to the limitation of the computational complexity of type-2 fuzzy theory, further research is needed for the application of type-2 fuzzy clustering algorithms in evaluation models. The method based on neural networks is mainly suitable for state estimation research. Generally, this method has low interpretability and requires a large amount of state monitoring data and health status display parameter data, so it is difficult to apply online.

[0004] In short, a fuzzy comprehensive evaluation method that can ensure the reliability and accuracy of data in a changing environment, improve the real-time performance and accuracy of the algorithm, and reduce subjectivity by reducing excessive reliance on experts' subjective judgments and experiences has not yet been invented.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] Therefore, an object of the present invention is to provide a network evaluation method for a neuro-fuzzy system with learning ability in order to solve the technical problem that existing evaluation methods are not suitable for complex systems containing multiple unknown parameters.

Means for Solving the Problem

[0007] The technical means adopted by the present invention are as follows.

[0008] A network evaluation method for a neural - fuzzy system with learning ability, which is an aspect of the present invention, S1. Construct a neuro - fuzzy system model that adopts fuzzy logic rules under the framework of a neural network; S2. Use an improved particle swarm intelligence optimization algorithm to train the neuro - fuzzy system to obtain optimal parameters and a trained neuro - fuzzy system model; S3. Define a system input value according to an actual problem, input it into the trained neuro - fuzzy system model, and obtain an evaluation result of the system's ability boundary.

[0009] Furthermore, in S1, the neuro - fuzzy system model is NFS = {P, T, I, O, M, W, f} represented by where P = {p1, p2, ···, p n} means an initial library containing finite elements representing avionics system facilities, subsystems, and system destruction events, where

Number

Number

[0010] Furthermore, S2 initializes to generate random particles and performs iterations, the particles continuously update themselves by tracking two extreme values in each iterative algorithm, updates the velocity and position of the particles, and performs iterations on the random particles to obtain the optimal solution.

[0011] Furthermore, among the two extreme values mentioned above, one is the optimal solution obtained by the particle itself, called the personal extreme value p best , and the other is the optimal solution obtained by the entire population so far, called the global extreme value g best .

[0012] Furthermore, the velocity and position of the particles are updated according to the following update formulas

Number

Number

Number

[0013] Furthermore, in the update formula for the velocity of the particles, the inertial weight is updated according to the following update formula: m k =m min +(m max -m min )(k max -k) / k max Among them, m k , m min and m max represent the current inertial weight, the minimum inertial weight, and the maximum inertial weight corresponding to the maximum number of iterations, respectively, and k represents the current number of iterations.

[0014] Furthermore, the system input in S3 includes a certainty input and an uncertainty input.

[0015] The present invention further provides a storage medium storing a program for executing a network evaluation method of the neuro-fuzzy system having the learning ability described above.

[0016] The present invention further provides an electronic device including a memory, a processor, and a computer program stored in the memory and executed by the processor, wherein the processor operates the computer program to execute the network evaluation method of the neuro-fuzzy system having the learning ability described above.

Advantages of the Invention

[0017] The present invention has the following beneficial effects compared with the prior art.

[0018] The present invention combines the inference ability of fuzzy logic with the infinite approximation function ability of a neural network to construct an alternative model of a real system, and then uses an intelligent optimization algorithm to make the alternative model resemble the real model as much as possible. The ability boundary evaluation model constructed from a neuro-fuzzy system introduces fuzzy rules into the framework of a neural network. This method has universality by combining the infinite approximation function ability of a neural network with the learning ability of a fuzzy system. According to experimental results and analysis, the evaluation inclusiveness and rationality model can be applied to the ability evaluation of complex systems. In particular, when a complex system contains many unknown parameters, this method has the advantage of uniqueness and provides a certain reference for improving system performance.

Brief Description of the Drawings

[0019] To more clearly explain the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the attached drawings required for the description of the embodiments or the prior art. The following attached drawings are some embodiments of the present invention. It goes without saying that those skilled in the art can obtain other attached drawings based on these attached drawings without creative labor.

[0020]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Modes for Carrying Out the Invention

[0021] To make the objectives, technical means, and advantages of the embodiments of the present invention clearer, hereinafter, while referring to the drawings in the embodiments of the present invention, the technical means in the embodiments of the present invention will be clearly and completely described. It goes without saying that the described embodiments are only some embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art without creative labor based on the embodiments in the present invention shall all be included in the scope protected by the present invention.

[0022] It should be noted that terms such as "first" and "second" in the specification, claims, and the above drawings of the present invention are used to distinguish similar objects and are not used to describe a specific order or priority. Since the embodiments of the present invention described herein can be implemented in an order other than that shown herein, it is possible to understand that the data thus adopted may be exchanged appropriately when appropriate. Moreover, the terms "comprising" and "having", and any variations thereof, refer to the meaning of non-exclusive inclusion. For example, a process, method, system, product, or device comprising a series of steps or units need not be limited to those steps or units clearly listed, and may further include other steps or units not clearly listed or specific to those processes, methods, products, or devices.

[0023] As shown in FIG. 1, the network evaluation method of the neural-fuzzy system with learning ability according to the present invention includes the following steps.

[0024] S1. Model the structured neural-fuzzy system.

[0025] The neural-fuzzy system combines the characteristics of the neural network and the fuzzy system, and has the infinite approximation ability of the neural network and the fuzzy inference ability of the fuzzy logic system. The neural-fuzzy system (NFS) model is constructed and defined as follows. The NFS consists of seven elements defined as follows.

[0026] NFS = {P, T, I, O, M, W, f}

[0027] Among them, P = {p1, p2, ···, p n} means an initial library containing finite elements representing avionics system facilities, subsystems, and system destruction events, where

Number

Number

[0028] The NFS model adopts fuzzy logic rules in the framework of the neural network to make the inference of the logical relationship of the system clear and smooth. The introduction of the parameters W, M, and the non - linear function f clearly describes the complex weight relationship of the system. The advantages of avoiding the complexity of modeling and reducing the modeling cost by this method are not available in the conventional methods.

[0029] The terminal library T is the set of outputs from all subsystems and the entire system. The elements of the terminal library are the elements of the initial libraries of other subsystems. The input library I is a subset of the initial library and represents the elements of the initial library that have been transformed by the elements of all non-terminal libraries. The output library O is a subset of the terminal library. In the case of a large and complex system composed of multiple subsystems, some of the outputs of the subsystems can be measured, but some of the outputs of the other subsystems cannot be obtained due to restrictions or high costs. The output library includes all the measurable outputs of the subsystems. Due to uncertain factors such as measurement errors, external interference, and signal transmission losses, a parameter M(p i ) is assigned to each element in the initial library to represent the truth degree of the value. This operation improves the robustness of the model and enhances its experimental value. The non-linear function f is introduced based on the fuzzy logic system.

[0030] Output continuous function [Number] In the case of, the input state is defined as follows.

[0031] [Number]

[0032] The neuro-fuzzy system usually includes a fuzzy inference unit and a neural network layer, introduces fuzzy inference into the neural network structure, and makes the modeling of complex systems more flexible. This system is constructed based on a series of "if-then" rules. The "if-then" rules establish the relationship between the output and the input. The "if-then" rules are concretized as follows.

[0033] A j1 ,A j2 ,···,A jn and B jDefine it as a fuzzy set. Rule j: Set x1 to A j1 and set x2 to A j2 and so on, and set x n to A jn . When this is done, y will be B j . Then, the inference error ε of the system is described as follows.

[0034] y = W T Φ(x) + ε

[0035] Here, the basis function vector is expressed as follows,

Number

Number

[0036]

Number

[0037] Among them, W represents a weight-adjustable matrix, and ε0 represents the maximum inference error.

[0038] S2 performs training and learning based on data-driven.

[0039] Learning is performed on the unknown weight parameters in the fuzzy neural system constructed in the first stage using an improved particle swarm intelligence optimization algorithm. This algorithm is a global optimization method based on swarm intelligence and performs effective search in complex spaces. The present application performs adaptive adjustment on a large number of weight parameters in the vulnerability assessment model of the above-mentioned aviation protection system using an improved particle swarm optimization algorithm. This method is simple and requires few parameter adjustments, so it can effectively solve a large number of complex problems such as non-linearity, non-differentiability, and multiple extrema. This algorithm is first initialized to generate random particles, and then iterations are performed to find the optimal solution. The optimal solution is the minimum 2-norm of the error between the predicted value and the true value. In each iterative algorithm, the particle always updates itself by tracking two extreme values, one of which is the optimal solution obtained by the particle itself, called the personal extreme value p best and the other is the optimal solution obtained so far by the entire personal swarm, called the global extreme value g best After that, the velocity and position are updated.

[0040]

Number

[0041] Among them,

Number

Number

[0042] m k = m min +(m max - m min )(k max - k) / k max

[0043] Among them, m k , m min and m max represent the inertia weight at the current time, the minimum inertia weight, and the maximum inertia weight corresponding to the maximum number of iterations at the current time respectively. k represents the number of iterations at the current time. Therefore, the inertia weight has the self - adapting ability to ensure the convergence of the search process.

[0044] S3. Perform online system capacity boundary evaluation.

[0045] Define system inputs including certainty inputs and uncertainty inputs according to the actual problem. Define several hyperparameters in the model and input them into the trained NFS model to obtain the system capacity boundary evaluation results.

[0046] (Example) The improved particle swarm optimization algorithm is integrated into the neuro - fuzzy system network model to learn from sample data to predict the reference weight value and evaluate the capacity boundary online. Its steps are executed as follows.

[0047] 1) Prepare a sample dataset. The dataset contains 250 samples. Among them, 200 samples are used for training, and 50 samples are used as a test set to test the training accuracy. That is, Ω = {(I1, O1), (I2, O2), ···, (Im, OM)}, where m = 250. Since there is not enough experimental data, these samples are obtained by simulation to test the method proposed in this application.

[0048] 2) Determine the weight numbers and the search space. As shown in Figure 3, assume there are four fuzzy logic systems, and if each fuzzy logic system has five weights, a total of 20 weight parameters are required for optimization. All weights are searched in the interval [0, 1].

[0049] 3) Initialize the parameters in the optimization algorithm. Treat the weight parameters to be learned as search space particles, and determine the maximum speed of the particles. Initialize the position and speed of each particle, and the particle swarm is randomly generated. The number of particles is set to 100, and the maximum number of iterations is 50 (c1 = 0.5, c2 = 0.5).

[0050] 4) Update the speed and position of the particles in each dimension.

[0051] 5) Decode each particle according to the sample data, obtain the marked values of the terminal library of the NFS model, and use them for the system's performance evaluation. Obtain the difference between the predicted value and the true value of the marked values in each terminal library. Next, update the individual optimum and the group optimum according to the fitness.

[0052] 6) Execute Step 4 in an iterative manner until the end rule is reached and the parameters are learned.

[0053] 7) Obtain the optimized weights and end the program. Based on the NFS model inference algorithm, by introducing an improved particle swarm optimization algorithm, the ability of parameter learning is provided, and many unknown parameters in the system can be freed from relying on experts and experience.

[0054] Emulation is executed according to three situations, each representing the fitness value of the entire subsystem. According to this, the algorithm can be achieved under three situations where the improved particle swarm optimization is then used for convergence for 20 iterations. Comparing the curves in Figure 2 under the three situations, it can be concluded that the optimization accuracy can be improved by increasing the knowledge information of the system. Regarding the weights in Example 3, compare the optimal weights and the optimized weights in Figure 3. Furthermore, the test results are shown in Figure 4, indicating the excellent performance of the proposed NFS model. In this way, the trained fuzzy neural network model can be used for the evaluation of the ability system. The input library obtains values online and in real time when changing over time to obtain the results. Figure 5 shows the real-time evaluation results of the system ability when there is one variable in the input library. Moreover, it can be seen that variable x1 has a small impact on the system ability, while variable x10 has a greater impact. Therefore, the sensitivity of the elements in the input library can also be obtained through the network model of the fuzzy neurosystem for the system ability evaluation results. When there are two variables in the input library of the system, the system ability evaluation results form a surface, and according to the surface in Figure 6, the ability boundary of the system can also be called an interval. For example, when the change range of variable x1 is [0.3, 0.6] and the change range of variable x9 is [0.05, 0.35], the ability range of the system is [0.1101, 0.1766], that is, the upper and lower ability boundaries are 0.1101 and 0.1766 respectively. When the change range of variable x9 is [0.05, 0.35] and the change range of variable x10 is [0.3, 0.6], the ability range of the system is [0.0111, 0.3029], that is, the upper and lower ability boundaries are 0.0111 and 0.3029 respectively. Similarly, when there are multiple system input libraries for variables, the system evaluation results can be easily obtained through the NFS model. The application of the fuzzy neural network greatly simplifies the description of the system ability. The system ability to construct an evaluation model with learning ability is verified by calculating examples in a system with four subsystems as the research object for its learning efficiency and accuracy.According to the experimental results and analysis, the model of evaluation inclusiveness and rationality can be applied to the ability evaluation of complex systems. In particular, when a complex system contains many unknown parameters, this method has the advantage of uniqueness and provides certain reference for improving system performance.

[0055] Finally, the following should be explained. Each of the above embodiments is only for explaining the technical means of the present invention and does not limit it. Although the present invention has been described in detail with reference to each of the above embodiments, it is also possible to modify the technical means described in each of the above embodiments or perform equivalent replacements for some or all of its technical features. It is obvious to those skilled in the art that the essence of the corresponding technical means does not deviate from the scope of the technical means of each embodiment of the present invention by these modifications and replacements.

[0056] (Appendix) (Appendix 1) A network evaluation method for a neural-fuzzy system with learning ability, comprising: S1, constructing a neural-fuzzy system model that adopts fuzzy logic rules under the framework of a neural network; S2, training the neural-fuzzy system using an improved particle swarm intelligence optimization algorithm to obtain optimal parameters and a trained neural-fuzzy system model; S3, defining system input values according to actual problems, inputting them into the trained neural-fuzzy system model, and obtaining the evaluation result of the system's ability boundary. A network evaluation method for a neural-fuzzy system with learning ability, characterized by the above.

[0057] (Appendix 2) In S1, the neural-fuzzy system model is represented by NFS = {P, T, I, O, M, W, f}, where P = {p1, p2, ···, p n} represents an initial library containing finite elements representing avionics system equipment, subsystems, and system failure events. [Number] and T = {t1, t2, ···, t m} represents a terminal library with finite elements that quantifies the degree of influence exerted by the elements in the initial library on the subsystem, [Number] and I(O) represents an input or output library that reflects the mapping of variations to the system. M is such that for each library node p i there is a marked value M(p i ), which reflects the truth degree of the proposition represented by the library node and is a mapping that represents the uncertainty of the device and the subsystem. W = {w1, w2, ···, w r} represents a set of weights of rules that reflects the degree of association between the elements of the initial library and the terminal library, and f is a non-linear function that maps the unknown complex relationships within the system. A network evaluation method for a neuro-fuzzy system having the learning ability described in Appendix 1, characterized in that.

[0058] (Appendix 3) S2 is initialized to generate random particles and perform iterations, the particles continuously update themselves by tracking two extreme values in each iteration algorithm, update the velocity and position of the particles, perform iterations on the random particles to obtain the optimal solution. A network evaluation method for a neuro-fuzzy system having the learning ability described in Appendix 1, characterized in that.

[0059] (Appendix 4) Of the two extreme values mentioned above, one is the optimal solution obtained by the particle itself and is called the individual extreme value p best , and the other is the optimal solution obtained by the entire individual group so far and is called the global extreme value g bestA method for evaluating the network of a neuro-fuzzy system having a learning ability as described in Appendix 3, characterized by the following.

[0060] (Appendix 5) The velocity and position of the said particles are updated according to the following update formula: [Number] Here, [Number] and [Number] represent the velocity and position of the d-dimensional component of particle i in the k-th generation respectively, ν max represents the maximum velocity of the particle, x min and x max represent the minimum position and the maximum position of the particle respectively, r1 and r2 represent random numbers between (0, 1), usually c1 = c2 = 0.5 represents the learning factor, and the inertia weight is defined as m. A method for evaluating the network of a neuro-fuzzy system having a learning ability as described in Appendix 4, characterized by the following.

[0061] (Appendix 6) In the update formula of the velocity of the said particles, the inertia weight is updated according to the following update formula: m k = m min +(m max - m min )(k max - k) / k max Here, m k , m min and m max represent the inertia weight at the current time, the minimum inertia weight and the maximum inertia weight corresponding to the maximum number of iterations respectively, and k represents the number of iterations at the current time. A method for evaluating the network of a neuro-fuzzy system having a learning ability as described in Appendix 5, characterized by the following.

[0062] (Appendix 7) In S3, the system input includes a certainty input and an uncertainty input, and the method for evaluating the network of the neuro-fuzzy system having the learning ability according to Supplementary Note 1 is characterized by this.

[0063] (Supplementary Note 8) A storage medium that stores a program for executing the method for evaluating the network of the neuro-fuzzy system having the learning ability according to any one of Supplementary Notes 1 to 7.

[0064] (Supplementary Note 9) An electronic device including a memory, a processor, and a computer program stored in the memory and executed by the processor, wherein the processor operates the computer program to execute the method for evaluating the network of the neuro-fuzzy system having the learning ability according to any one of Supplementary Notes 1 to 7, and the electronic device is characterized by this.

Claims

1. A method for evaluating a network of a neuro-fuzzy system having learning capability, comprising the steps of: S1, constructing a neuro-fuzzy system model that employs fuzzy logic rules under the framework of neural network; S2. Using an improved particle swarm intelligence optimization algorithm to train the neuro-fuzzy system, and obtain optimal parameters and a trained neuro-fuzzy system model; S3: A method for evaluating a network of a neuro-fuzzy system with learning capability, comprising: defining system input values ​​according to an actual problem, inputting the system input values ​​into the trained neuro-fuzzy system model, and obtaining a system capability boundary evaluation result.

2. In S1, the neuro-fuzzy system model is represented as NFS={P, T, I, O, M, W, f}; Here, P = {p １ , p ２ , ..., p ｎ } means an initial library containing finite elements representing avionics system equipment, subsystems, and system disruptive events; [0010] and T={t １ , t ２ , ..., t ｍ } represents a terminal library with finite elements that quantify the degree to which elements in the initial library affect the subsystem; [0025] where I(O) represents the input or output library that reflects the mapping of variations to the system, and M represents the input or output library for each library node p ｉ The value M(p ｉ ), which reflects the degree of truth of the proposition represented by the library node and represents the uncertainty of the device and subsystem; W = {w １ , w ２ , ..., w ｒ } represents a weight set of rules reflecting the degree of association between elements of the initial library and the terminal library, and f is a nonlinear function that maps unknown complex relationships within the system. The method for evaluating a network of a neuro-fuzzy system with learning capability according to claim 1, characterized in that

3. S2, Initialize the random particles and perform iterations. The particle keeps updating itself by tracking the two extremes in each iteration of the algorithm; Updating the particle's velocity and position; 2. The method for evaluating a network of a neuro-fuzzy system having learning capability according to claim 1, further comprising: performing iteration on the random particles to find an optimal solution.

4. Of the two extreme values, one is the optimal solution obtained by the particle itself, and the individual extreme value p ｂｅｓｔ The other is the optimal solution found so far by the entire population of individuals, and is the global extremum g ｂｅｓｔ 4. The method for evaluating a network of a neuro-fuzzy system having learning capability according to claim 3, characterized in that said method is called a "network evaluation method for a neuro-fuzzy system having learning capability."

5. The velocity and position of the particle are updated according to the update formula: [0030] Where: [0045] and [0050] respectively represent the d-dimensional component velocity and position of the kth generation particle i, and v ｍａｘ represents the maximum velocity of the particle, and x ｍｉｎ and x ｍａｘ represent the minimum and maximum positions of the particle, respectively, and r １ To ２ represents a random number between (0, 1), and usually, c １ = c ２ 5. The method for evaluating a network of a neuro-fuzzy system having learning capability according to claim 4, wherein: =0.5 represents a learning factor, and an inertia weight is defined as m.

6. In the above update formula for the particle velocity, the inertia weight is updated according to the following update formula: m ｋ =m ｍｉｎ +(m ｍａｘ -m ｍｉｎ )(k ｍａｘ -k) / k ｍａｘ Here, m ｋ , m ｍｉｎ and m ｍａｘ The method for evaluating a network of a neuro-fuzzy system having learning capability according to claim 5, characterized in that, respectively, x, y, and y represent the current inertia weight, the minimum inertia weight, and the maximum inertia weight corresponding to the maximum number of iterations, and k represents the number of iterations at the current time.

7. 2. The method for evaluating a network of a neuro-fuzzy system having learning capability according to claim 1, wherein in step S3, the system input includes a certainty input and an uncertainty input.

8. A storage medium for storing a program for executing the method for evaluating a network of a neuro-fuzzy system having a learning capability according to any one of claims 1 to 7.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executed on the processor, An electronic device, characterized in that said processor executes the method for evaluating a network of a neuro-fuzzy system having learning capability according to any one of claims 1 to 7 by running said computer program.

Images