A dynamic fuzzy neural network modeling method for intelligent dimming of LED system

Abstract

The present invention discloses a dynamic fuzzy neural network modeling method for intelligent dimming of an LED system, which relates to the technical field of nonlinear optical characteristics of LED systems, and includes the following steps: pre-calibrating the input variable characteristics and expected output value characteristics of the photoelectric thermal model of an LED module; building and training a dynamic fuzzy neural network model, wherein the dynamic fuzzy neural network model includes an input layer, a fuzzification layer, a precondition layer, a postcondition layer, and an output layer, and the training of the dynamic fuzzy neural network model includes network structure optimization and parameter learning. The present invention adopts an online modeling method, and the model structure parameters can be optimized and adjusted in real time, which has stronger adaptability. The combined measurement of heat sink temperature and ambient temperature is used to indirectly reflect the influence of junction temperature on the photometric output of the LED system, thereby avoiding the complexity of junction temperature measurement.

Description

Technical Field

[0001] The invention relates to the technical field of nonlinear optical characteristics of LED systems, and in particular to a dynamic fuzzy neural network modeling method for intelligent dimming of LED systems. Background Art

[0002] High-power LED systems have the advantages of high light efficiency and flexible controllability, and have great potential in providing energy-saving dynamic or interactive lighting, such as lighting rendering of large buildings, crop cultivation, medical surgical lights, etc. However, in some special light sources, especially when used as standard light sources for measurement, such as high-power LED solar simulators, such applications have high requirements for optical quality: usually requiring operation point switching / continuous adjustment within a wide range, and ensuring that the system output reaches the predetermined optical performance target (such as luminosity, chromaticity) within a limited time. Therefore, the study of the optical dynamic quality of high-power LED systems has gradually become a focus of attention. Although the photoelectric conversion frequency of LED itself can reach megahertz, as a power device, especially in a compact design with high power density, the heating effect of system operation is very significant, and the system heat dissipation is usually several times or even dozens of times that of ordinary lighting. This heat is dissipated in the form of heat conduction and heat convection, and its dynamic process has obvious uncertain and complex nonlinear characteristics such as time lag and time variation. In addition, there is a coupling relationship between the optical, electrical, and thermal parameters of LEDs, which makes it difficult to confirm the law of system luminosity and chromaticity changes.

[0003] In order to establish the mathematical expression model of light, electricity and heat of LED system, most scholars start from the mechanism analysis of equivalent resistance and capacitance network. However, the mathematical form of these models often involves a large number of free parameters and high calculation amount. At the same time, there are difficulties in parameter extraction. For example, junction temperature measurement requires expensive test equipment or complex experiments, which is difficult to achieve dynamic measurement in array design. In addition, this type of model is generally an offline model. Once its parameters or structure are established, they cannot be changed. Therefore, when the ambient temperature changes (such as temperature changes throughout the year) or when the device aging and other conditions change, the model has mismatch problems and it is difficult to achieve accurate dimming. For example, in the existing current feedback and temperature feedforward dimming control methods, the steady-state relationship table of luminous flux-current is used, which is sufficient for application in ordinary lighting scenarios, but there is still a large gap in achieving more accurate dimming control scenarios.

[0004] A search of Chinese invention patent CN104482445 A discloses an intelligent dimming LED plant lamp and its intelligent dimming system. The intelligent dimming system includes an LED power module, a microprocessor system, a light sensing system, and a storage system. The storage system stores the required lighting data of different plants at different growth stages. The intelligent dimming LED plant lamp includes a lamp body and an LED module arranged in the lamp body. The LED module includes a substrate, a plurality of LED arrays arranged on the substrate, and a controller using the intelligent dimming system. The microprocessor system can make the best brightness adjustment instructions to the LED power module according to the data of the storage system and the light sensing system, control the output of the LED power module, and realize the dimming of the LED plant lamp, so that the output light intensity and band of the plant lamp reach the optimal light intensity value and spectral distribution required by the plant according to the growth stage of the plant and the natural light intensity on site, thereby promoting plant growth. However, there is still a large gap in realizing more precise dimming control scenarios.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] In view of the problems in the related art, the present invention proposes a dynamic fuzzy neural network modeling method for intelligent dimming of an LED system to overcome the above technical problems existing in the existing related art.

[0007] The technical solution of the present invention is achieved in this way:

[0008] A dynamic fuzzy neural network modeling method for intelligent dimming of an LED system includes the following steps:

[0009] Pre-calibrate the input variable characteristics and expected output value characteristics of the LED module photoelectric thermal model;

[0010] Build and train a dynamic fuzzy neural network model, including;

[0011] The dynamic fuzzy neural network model includes an input layer, a fuzzification layer, a precondition layer, a postcondition layer and an output layer, and the training of the dynamic fuzzy neural network model includes network structure optimization and parameter learning, wherein;

[0012] The construction of the dynamic fuzzy neural network model includes the following steps:

[0013] Step S1, model initialization k = 1, read the first set of observation samples z(1) = [x(1), y(1)], let z 1* (1) = z(1), the symbol * indicates that the point is a cluster, z i* =[x i* ,y i* ], i∈[1,v] represents the i-th cluster center, and v is the number of clusters;

[0014] Define the following divergence function S(z(k)) as the evaluation function of density cluster analysis:

[0015] Where N is a positive integer, n, m are the dimensions of input x and output y; z j (l) represents the jth component of sample z at time l, z i (k) represents the i-th component of sample z at the k-th moment;

[0016] Let S(z(1))=1, c 1 =x 1* (1) represents the neural network fuzzy layer, and initializes the membership width σ 1 and the linear parameters p of the network posterior part r =[p0, p1, p2, ..., p n ] T ;

[0017] Step S2, let k = k + 1, calculate the predicted value of the model photometric output according to the new sampling time input and the existing fuzzy rules

[0018] Step S3, read in the new observation data z(k), and calculate its divergence value using the equivalent recursive form to improve the calculation efficiency, which is expressed as:

[0019]

[0020] in,

[0021]

[0022] Step S4, update the divergence values ​​of all cluster centers, expressed as:

[0023]

[0024] Step S5, calculating the model system error, error rate and generalization factor, where;

[0025] Step S501, obtaining the system error, expressed as: It calculates the minimum Euclidean distance

[0026] Step S502, obtaining the error reduction rate, which is expressed as: from the linear regression equation D = ΓP + E and QR decomposition Γ = MH, where D is the expected output, P is the weight vector, Γ is the regression vector, E is the error vector, M and H are an orthogonal matrix and an upper triangular matrix respectively;

[0027] Get the linear minimum square error solution The error reduction rate is

[0028] Step S503, define the normalization factor based on the error reduction rate

[0029] Step S504, calculate the expected accuracy k e , effectively accommodating the boundary k d ,,k e =max{α k-1 *emin max}, k d =max{α k-1 *dmin max}, where α, e max , e min , d max , d min is a predefined value.

[0030] Furthermore, the input variable characteristics are expressed as:

[0031] X(k)=[I f (k), U f (k), T r (k), T a (k)];

[0032] The expected output value feature is expressed as:

[0033] Y(k)=Φ(k), k=1, 2,...N;

[0034] Among them, I f , U f , T r and T a They represent the LED driving current, forward voltage, heat sink temperature and ambient temperature respectively, Φ is the luminous flux output by the LED system,

[0035] x(k) and y(k) are the vectors of X(k) and Y(k) normalized by the min-max method.

[0036] Furthermore, the input layer includes a plurality of nodes, the number of which is consistent with the dimension of the model input, and each node represents an input language variable;

[0037] The fuzzification layer includes a plurality of nodes, the node function is a fuzzy function, and the fuzzy function adopts a Cauchy-type function, which is a first-order form of a Gaussian function and is expressed as:

[0038]

[0039] Where i = 1, 2, ..., n, r = 1, 2, ..., R. For x i The rth fuzzy function of For x i The rth fuzzy function center, σ r Represents x i The influence radius of the rth fuzzy function, n is the dimension of the input vector, expressed as: x(k) = [x1(k), x2(k), ..., x n (k)],x i (k) is the i-th dimension component of x(k), R is the total number of rules of the fuzzy system, and σ=[σ 1 , σ 2 , ..., σ r ];

[0040] The antecedent layer includes a plurality of nodes, each of which represents the premise part of a fuzzy rule. For the rth rule, its output is:

[0041] in,

[0042] The subsequent layer includes a plurality of nodes, each node representing a linear subsystem y r , that is, the conclusion part of the TSK system, the output of each subsystem is as follows:

[0043] in, Expressed as the linear parameters of the subsequent network;

[0044] The output layer includes at least one node, and each node of the layer corresponds to the dimension of the output variable, which is expressed as:

[0045] Further, step S6, model structure optimization, includes:

[0046] Step S601, rule modification criteria: determine If satisfied, then the fuzzy function center is corrected; let S(z j* (k)) = S (z (k)), c j =x(k),c j Replace the center of the jth fuzzy function, where j is the number corresponding to the minimum distance between the center of the existing fuzzy function and the sample point, and after execution, proceed to step S6;

[0047] Step S602, rule generation criteria: determine If satisfied, generate a new rule, let S(z (v+1)*(k)) = S (z (k)), R = R + 1, c R =x(k),σ R =k0*δ min (k), S(z R* (k)) = S(z(k)), after execution, go to step S6;

[0048] Step S603: if both the rule modification criteria and the rule generation criteria are not satisfied, go to step S6;

[0049] Step S7, parameter learning phase: using the extended Kalman filter algorithm EKF to update the rule radius σ = [σ 1 , σ 2 , ..., σ R ] and the recursive least squares method with forgetting factor to estimate the posterior layer parameters

[0050] Step 701, the extended Kalman filter algorithm EKF algorithm, is expressed as:

[0051]

[0052] in, δ is an appropriate parameter;

[0053] Step 702, the RLS algorithm: is the forgetting factor, Ω is a large positive number, is the covariance matrix of the ith rule at time k, where;

[0054] If a new rule is generated in step S5, the covariance matrix corresponding to the rule is initialized

[0055] Furthermore, the method further comprises the following steps:

[0056] At each sampling moment, steps S2 to S7 are repeated until the set N moments are reached.

[0057] Beneficial effects of the present invention:

[0058] The dynamic fuzzy neural network modeling method for intelligent dimming of an LED system of the present invention pre-calibrates the input variable characteristics and expected output value characteristics of the photoelectric thermal model of the LED module, builds a dynamic fuzzy neural network model and trains it, and adopts an online modeling method. The model structure parameters can be optimized and adjusted in real time, and it has stronger adaptability. The heat sink temperature and the ambient temperature are combined to indirectly reflect the influence of the junction temperature on the luminous output of the LED system, thereby avoiding the complexity of the junction temperature measurement. At the same time, when the ambient temperature is affected by seasonal changes or the performance is degraded, the model does not need to be recalibrated. Not only are the formulas of the online modeling method in a recursive form, which greatly simplifies the calculation complexity, but also because the rules are continuously improved and corrected through incremental learning, the model structure is more streamlined, and its consequent is in a linear form with interpretability. By expanding it to N-step prediction, accurate dimming control based on the prediction model can be achieved to eliminate the influence of the temperature lag factor, which helps to improve the dynamic response quality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 It is a flow chart of a dynamic fuzzy neural network modeling method for intelligent dimming of an LED system according to an embodiment of the present invention;

[0061] Figure 2 It is a dynamic fuzzy neural network structure topology diagram of a dynamic fuzzy neural network modeling method for intelligent dimming of an LED system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0063] According to an embodiment of the present invention, a dynamic fuzzy neural network modeling method for intelligent dimming of an LED system is provided.

[0064] like Figure 1-Figure 2 As shown, the dynamic fuzzy neural network modeling method for intelligent dimming of an LED system according to an embodiment of the present invention includes the following steps:

[0065] Pre-calibrate the input variable characteristics and expected output value characteristics of the LED module photoelectric thermal model;

[0066] Build and train a dynamic fuzzy neural network model, including;

[0067] The dynamic fuzzy neural network model includes an input layer, a fuzzification layer, a precondition layer, a postcondition layer and an output layer, and the training of the dynamic fuzzy neural network model includes network structure optimization and parameter learning, wherein;

[0068] Specifically, it collects observation data of the dimming process of the high-power LED system in real time as sample data. Including input data driving current If, voltage Uf, heat sink temperature Tr, ambient temperature Ta, output data luminous flux φ. After the formation, the collected input and output samples z = (x, y);

[0069] Initialize the fuzzy neural network parameters, the specific parameters can be as follows:

[0070] ρ=0.95, Ω=800, δ=1.0, e min =0.15, e max ε min =0.25,ε max k err =0.0015, k GF =0.99. The data samples are input into the dynamic fuzzy neural network model in the order of sampling, and the dynamic fuzzy neural network model predicts the luminosity value at the next moment.

[0071] Specifically, the construction of the dynamic fuzzy neural network model includes the following steps:

[0072] Step S1, model initialization k = 1, read the first set of observation samples z(1) = [x(1), y(1)], let z 1* (1) = z(1), the symbol * indicates that the point is a cluster, z i* =[x i* ,y i* ], i∈[1,v] represents the i-th cluster center, and v is the number of clusters;

[0073] Define the following divergence function S(z(k)) as the evaluation function of density cluster analysis:

[0074] Where N is a positive integer, n, m are the dimensions of input x and output y; z i (l) represents the jth component of sample z at time l, z i (k) represents the i-th component of sample z at the k-th moment;

[0075] Let S(z(1))=1, c 1=x 1* (1) represents the neural network fuzzy layer, and initializes the membership width σ 1 and the linear parameters p of the network posterior part r =[p0, p1, p2, ..., p n ] T ;

[0076] Step S2, let k = k + 1, calculate the predicted value of the model photometric output according to the new sampling time input and the existing fuzzy rules

[0077] Step S3, read in the new observation data z(k), and calculate its divergence value using the equivalent recursive form to improve the calculation efficiency, which is expressed as:

[0078]

[0079] in,

[0080]

[0081] Step S4, update the divergence values ​​of all cluster centers, expressed as:

[0082]

[0083] Step S5, calculating the model system error, error rate and generalization factor, where;

[0084] Step S501, obtaining the system error, expressed as: It calculates the minimum Euclidean distance

[0085] Step S502, obtaining the error reduction rate, which is expressed as: from the linear regression equation D = ΓP + E and QR decomposition Γ = MH, where D is the expected output, P is the weight vector, Γ is the regression vector, E is the error vector, M and H are an orthogonal matrix and an upper triangular matrix respectively;

[0086] Get the linear minimum square error solution The error reduction rate is

[0087] Step S503, define the normalization factor based on the error reduction rate

[0088] Step S504, calculate the expected accuracy k e , effectively accommodating the boundary k d ,,k e =max{α k-1 *emin max}, k d =max{α k-1*dmin max}, where α, e max , e min , d max , d min is a predefined value.

[0089] Among them, the input variable characteristics are expressed as:

[0090] X(k)=[I f (k), U f (k), T r (k), T a (k)];

[0091] The expected output value feature is expressed as:

[0092] Y(k)=Φ(k), k=1, 2,...N;

[0093] Among them, I f , U f , T r and T a They represent the driving current, forward voltage, heat sink temperature and ambient temperature of the LED respectively, Φ is the luminous flux output by the LED system, x(k) and y(k) are the vectors of X(k) and Y(k) normalized by the min-max method respectively.

[0094] The input layer includes a plurality of nodes, the number of which is consistent with the dimension of the model input, and each node represents an input language variable;

[0095] The fuzzification layer includes a plurality of nodes, the node function is a fuzzy function, and the fuzzy function adopts a Cauchy-type function, which is a first-order form of a Gaussian function and is expressed as:

[0096]

[0097] Where i = 1, 2, ..., n, r = 1, 2, ..., R. For x i The rth fuzzy function of For x i The rth fuzzy function center, σ r Represents x i The influence radius of the rth fuzzy function, n is the dimension of the input vector, expressed as: x(k) = [x1(k), x2(k), ..., x n (k)],x i (k) is the i-th dimension component of x(k), R is the total number of rules of the fuzzy system, and σ=[σ 1 , σ 2, ..., σ r ];

[0098] The antecedent layer includes a plurality of nodes, each of which represents the premise part of a fuzzy rule. For the rth rule, its output is:

[0099] in,

[0100] The subsequent layer includes a plurality of nodes, each node representing a linear subsystem y r , that is, the conclusion part of the TSK system, the output of each subsystem is as follows:

[0101] in, Expressed as the linear parameters of the subsequent network;

[0102] The output layer includes at least one node, and each node of the layer corresponds to the dimension of the output variable, which is expressed as:

[0103] Wherein, step S6, model structure optimization, includes:

[0104] Step S601, rule modification criteria: determine If satisfied, then the fuzzy function center is corrected; let S(z j* (k)) = S (z (k)), c j =x(k),c j Replace the center of the jth fuzzy function, where j is the number corresponding to the minimum distance between the center of the existing fuzzy function and the sample point, and after execution, proceed to step S6;

[0105] Step S602, rule generation criteria: determine If satisfied, generate a new rule, let S(z (v+1) *(k))=S(z(k)), R=R+1, c R =x(k),σ R =k0*δ min (k), S(z R* (k)) = S(z(k)), after execution, go to step S6;

[0106] Step S603: if both the rule modification criteria and the rule generation criteria are not satisfied, go to step S6;

[0107] Step S7, parameter learning phase: using the extended Kalman filter algorithm EKF to update the rule radius σ = [σ 1 , σ 2 , .., σ R] and the recursive least squares method with forgetting factor to estimate the posterior layer parameters

[0108] Step 701, the extended Kalman filter algorithm EKF algorithm, is expressed as:

[0109]

[0110] in, δ is an appropriate parameter;

[0111] Step 702, the RLS algorithm: is the forgetting factor, Ω is a large positive number, is the covariance matrix of the ith rule at time k, where;

[0112] If a new rule is generated in step S5, the covariance matrix corresponding to the rule is initialized

[0113] The following steps are also included:

[0114] At each sampling moment, steps S2 to S7 are repeated until the set N moments are reached.

[0115] In addition, specifically, at different operating points, the linear combination form of each subsystem can be recorded and derived, including the center position and membership width of each rule, so as to select a better network threshold parameter.

[0116] With the help of the above technical solution, the following can be achieved: 1) In the establishment of LED dimming models, most of them are offline models. The online modeling method adopted by the present invention can optimize and adjust the model structure parameters in real time, which has stronger adaptability. 2) The present invention avoids the complexity of junction temperature measurement and adopts the joint measurement of heat sink temperature and ambient temperature to indirectly reflect the influence of junction temperature on the luminous output of LED system. At the same time, when the ambient temperature changes with the seasons or the performance degrades, the model does not need to be recalibrated.

[0117] In addition, specifically, 1) the formulas of the online modeling method are all in recursive form, which greatly simplifies the computational complexity and can design a low-cost embedded controller; 2) because the rules are continuously improved and corrected through incremental learning, the model structure is more streamlined, and its consequent is in linear form and is interpretable; 3) by expanding to N-step prediction, accurate dimming control based on the prediction model can be achieved to eliminate the influence of temperature lag factors, which helps to improve the dynamic response quality of the system. 4) This method can be further expanded or customized, such as introducing more prediction indicators such as chromaticity and spectral functions, to facilitate different application purposes.

[0118] In summary, with the help of the above-mentioned technical scheme of the present invention, by pre-calibrating the input variable characteristics and expected output value characteristics of the LED module photoelectric thermal model, a dynamic fuzzy neural network model is built and trained, and an online modeling method is adopted. The model structure parameters can be optimized and adjusted in real time, and it has stronger adaptability. The two joint measurements of heat sink temperature and ambient temperature are used to indirectly reflect the influence of junction temperature on the luminous output of the LED system, avoiding the complexity of junction temperature measurement. At the same time, when the ambient temperature is affected by seasonal changes or the performance is degraded, the model does not need to be recalibrated. Not only are the formulas of the online modeling method in recursive form, which greatly simplifies the calculation complexity, but also because the rules are continuously improved and corrected through incremental learning, the model structure is more streamlined, and its consequent is in linear form, which is interpretable. By expanding it to N-step prediction, precise dimming control based on the prediction model can be achieved to eliminate the influence of temperature lag factors, which helps to improve the dynamic response quality of the system.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A dynamic fuzzy neural network modeling method for intelligent dimming of LED system, characterized in that: The following steps are involved: Pre-calibrate the input variable characteristics and expected output value characteristics of the LED module photoelectric thermal model; Build and train a dynamic fuzzy neural network model, including; The dynamic fuzzy neural network model includes an input layer, a fuzzification layer, a precondition layer, a postcondition layer and an output layer, and the training of the dynamic fuzzy neural network model includes network structure optimization and parameter learning, wherein; The construction of the dynamic fuzzy neural network model includes the following steps: Step S1, model initialization k = 1, read the first set of observation samples z(1) = [x(1), y(1)], let z 1* (1) = z(1), the symbol * indicates that the point is a cluster, z i* =[x i* ,y i* ], i∈[1,v] represents the i-th cluster center, and v is the number of clusters; Define the following divergence function S(z(k)) as the evaluation function of density cluster analysis: Where N is a positive integer, n, m are the dimensions of input x and output y; z j (l) represents the jth component of sample z at time l, z i (k) represents the i-th component of sample z at the k-th moment; Let S(z(1))=1, c 1 =x 1* (1) represents the neural network fuzzy layer, and initializes the membership width σ 1 and the linear parameters p of the network posterior part r =[p0,p1,p2,...,p n ] T ; Step S2, let k = k + 1, calculate the predicted value of the model photometric output according to the new sampling time input and the existing fuzzy rules Step S3, read in the new observation data z(k), and calculate its divergence value using the equivalent recursive form to improve the calculation efficiency, which is expressed as: in, Step S4, update the divergence values ​​of all cluster centers, expressed as: Step S5, calculating the model system error, error rate and generalization factor, where; Step S501, obtaining the system error, expressed as: It calculates the minimum Euclidean distance Step S502, obtaining the error reduction rate, which is expressed as: from the linear regression equation D = ΓP + E and QR decomposition Γ = MH, where D is the expected output, P is the weight vector, Γ is the regression vector, E is the error vector, M and H are an orthogonal matrix and an upper triangular matrix respectively; Get the linear minimum square error solution The error reduction rate is Step S503, define the normalization factor based on the error reduction rate Step S504, calculate the expected accuracy k e , effectively accommodating the boundary k d , k e =max{α k-1 *emin max }, k d =max{α k-1 *dmin max }, where α, e max , e min , d max , d min is a predefined value.

2. The dynamic fuzzy neural network modeling method for intelligent dimming of LED system according to claim 1 is characterized in that: The input variable characteristics are expressed as: X(k)=[I f (k),U f (k),T r (k),T a (k)]; The expected output value feature is expressed as: Y(k)=Φ(k), k=1, 2,...N; Among them, I f , U f , T r and T a They represent the LED driving current, forward voltage, heat sink temperature and ambient temperature respectively, Φ is the luminous flux output by the LED system, x(k) and y(k) are the vectors of X(k) and Y(k) normalized by the min-max method.

3. The dynamic fuzzy neural network modeling method for intelligent dimming of LED system according to claim 1, characterized in that: The input layer includes a plurality of nodes, the number of which is consistent with the dimension of the model input, and each node represents an input language variable; The fuzzification layer includes a plurality of nodes, the node function is a fuzzy function, and the fuzzy function adopts a Cauchy-type function, which is a first-order form of a Gaussian function and is expressed as: Where i = 1, 2, ..., n, r = 1, 2, ..., R. For x i The rth fuzzy function of For x i The rth fuzzy function center, σ r Represents x i The influence radius of the rth fuzzy function, n is the dimension of the input vector, expressed as: x(k)=[x1(k),x2(k),...,x n (k)],x i (k) is the i-th dimension component of x(k), R is the total number of rules of the fuzzy system, and σ=[σ 1 ,s 2 ,...,s r ]; The antecedent layer includes a plurality of nodes, each of which represents the premise part of a fuzzy rule. For the rth rule, its output is: in, The subsequent layer includes a plurality of nodes, each node representing a linear subsystem y r , that is, the conclusion part of the TSK system, the output of each subsystem is as follows: in, Expressed as the linear parameters of the subsequent network; The output layer includes at least one node, and each node of the layer corresponds to the dimension of the output variable, which is expressed as:

4. The dynamic fuzzy neural network modeling method for intelligent dimming of LED system according to claim 3 is characterized in that: Step S6, model structure optimization, includes: Step S601, rule modification criteria: determine If satisfied, then the fuzzy function center is corrected; let S(z j* (k)) = S (z (k)), c j =x(k),c j Replace the center of the jth fuzzy function, where j is the number corresponding to the minimum distance between the center of the existing fuzzy function and the sample point, and after execution, proceed to step S6; Step S602, rule generation criteria: determine If satisfied, generate a new rule, let S(z (v+1)* (k)) = S (z (k)), R = R + 1, c R =x(k),σ R =k0*δ min (k), S(z R* (k)) = S(z(k)), after execution, go to step S6; Step S603: if both the rule modification criteria and the rule generation criteria are not satisfied, go to step S6; Step S7, parameter learning phase: using the extended Kalman filter algorithm EKF to update the rule radius σ = [σ 1 ,σ 2 ,...,σ R ] and the recursive least squares method with forgetting factor to estimate the posterior layer parameters Step 701, the extended Kalman filter algorithm EKF algorithm, is expressed as: in, δ is an appropriate parameter; Step 702, the RLS algorithm: ρ is the forgetting factor, Ω is a large positive number, is the covariance matrix of the ith rule at time k, where; If a new rule is generated in step S5, the covariance matrix corresponding to the rule is initialized 5. The dynamic fuzzy neural network modeling method for intelligent dimming of LED system according to claim 4, characterized in that: The following steps are also included: At each sampling moment, steps S2 to S7 are repeated until the set N moments are reached.

Images