A deep t-s persistence rule optimization dermatitis classification method
By employing a deep TS persistent rule optimization method, and utilizing variance threshold selection and fuzzy neural network to optimize dermatitis classification, the problem of insufficient rule optimization in TS model dermatitis classification is solved, achieving a more efficient dermatitis classification effect.
Patent Information
- Application Number
- CN202310937276.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-27
- Publication Date
- 2026-01-20
- Estimated Expiration
- 2043-07-27
AI Technical Summary
Existing TS models suffer from insufficient rule optimization in dermatitis classification, low utilization rate, and complex structure, making it difficult to achieve good classification results.
We employ a deep TS persistent rule optimization method, using variance threshold selection for feature reduction, and construct a deep TS fuzzy neural network, which includes an input layer, a membership layer, and a rule layer. We use membership functions and clustering algorithms to generate fuzzy rules, optimize the rule base of the rule layer, reduce random rule generation, and improve model efficiency and structural simplification.
It improves the accuracy and efficiency of dermatitis classification, simplifies the model structure, enhances the expressive power of rules, and approximates the target output layer by layer, thus ensuring the classification performance of the model.
Smart Images
Figure CN117056795B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of deep learning, and particularly relates to a deep T-S persistence rule optimized dermatitis classification method. BACKGROUND
[0002] Takagi-Sugeno model (T-S): T-S fuzzy model is a typical fuzzy dynamic model in fuzzy modeling of complex nonlinear systems, which was proposed by Takagi and Sugeno in 1985. Its main features are: the antecedent is divided according to whether there is a local linear relationship between the system input and output, and the conclusion part is expressed by a polynomial linear equation, so as to form a linear combination between each rule, so that the global output of the nonlinear system has good linear description characteristics.
[0003] The variance threshold selection method is a fast and lightweight method that can quickly eliminate features with small contribution. The principle is to use variance as a feature scoring standard. If the value difference of a certain feature is small, it is generally considered that the feature has little contribution to distinguishing samples, so the features with variance less than the threshold are removed in the feature construction process. Moreover, the variance selection method is suitable for discrete features, and continuous features need to be discretized before use. SUMMARY
[0004] The purpose of the application is to overcome the problems of low utilization and complex structure of the existing T-S model rule optimization, and to propose a deep T-S persistence rule optimized dermatitis classification method to achieve better classification level.
[0005] Technical scheme: In order to achieve the above purpose, the application proposes a deep T-S persistence rule optimized dermatitis classification method, which comprises the following steps:
[0006] S1: Obtain dermatitis data, and reduce features by variance threshold selection method. The data samples are randomly divided into training set S and test set S' according to the ratio of 7:3;
[0007] S2: Construct a deep T-S fuzzy neural network, which includes an input layer, a membership degree layer, a rule layer and an output layer, wherein the number of rule layers is DP, and the number of layers of the remaining layers is 1; generate fuzzy rules by inputting the training set S;
[0008] S3: Evaluate and optimize the dp layer rules, select part of the rules into the dp+1 layer fuzzy rule library, dp=1,2,…,DP, finally fix the DP layer fuzzy rule library and calculate the training classification result, DP is the number of rule layers;
[0009] S4: Input the test set S' into the DP layer fuzzy rule library training to obtain the test classification result.
[0010] Furthermore, the specific method of S1 is as follows: The original dermatitis data is normalized, and then the variance thresholding method is used to filter the features, that is, according to the variance formula... Calculate the mean squared error of the j-th feature in sequence, x ij It is the j-th feature value of the i-th sample. It is the mean of the j-th feature data, j = 1, 2, ..., D, where D is the initial number of features and n is the number of samples. A variance threshold is set. Features with a mean squared error greater than a threshold are removed, and data samples are randomly assigned to the training set in a 7:3 ratio. With test set and It represents the target classification results of training and testing, where n is the number of training samples, n′ is the number of test samples, and d is the number of existing features.
[0011] Furthermore, the specific method of S2 is as follows: The input layer is responsible for normalizing the input data. The input layer takes the training data S as input, and outputs the normalized data. The membership layer achieves fuzzification of feature values. This layer takes feature values X = [x1, x2, ..., x...] as input. d ] T The membership function is used to convert the feature values into membership values, and the membership values are output. A Gaussian function is used as the membership function, and the membership function of the j-th feature is... Expressed as:
[0012]
[0013] Where i = 1, 2, ..., n, and This represents the cluster center and variance of the m-th cluster, where m = 1, 2, ..., M, and M is the number of clusters. The FCM algorithm is used to cluster the feature values, and the j-th feature x... j =[x 1j ,x 2j ,…,x nj Clustering into M classes generates cluster centers. Clustering variance x represents j The m-th linguistic fuzzy set,
[0014] j = 1, 2, ..., d, where d is the existing characteristic number;
[0015] Each rule layer is input with feature values and their membership values to generate rules and calculate the fitness and error of the rules. High-quality rules are selected and input into the next rule layer. Starting from the second rule layer, additional rules from the previous rule layer are input to enrich the current rule base. The fuzzy rule base of the dp-th layer initially contains K. dpThere are fuzzy rules, dp = 1, 2, ..., DP;
[0016] Generate K dp There are random rules, each rule using the sequence [x1, x2, ..., x]. d ] T As the input vector, each feature is randomly selected from a certain linguistic fuzzy set. The k-th fuzzy rule is represented as:
[0017]
[0018] Among them, K dp It is the number of rules in the dp-th rule layer. It is the fuzzy set of language selected by the k-th rule in the j-th feature. M is the number of cluster categories. x is the weight coefficient of the j-th feature. j It is the j-th eigenvalue. This indicates that the j-th eigenvalue corresponds to The linguistic meaning, f k (X) is the output of the kth fuzzy rule. The rule is divided into an IF part and a Then part, namely the antecedent part and the consequent part.
[0019] Calculate the fitness of each rule and its antecedent parameters, and the fitness of the k-th fuzzy rule. and the antecedent parameters The expression is:
[0020]
[0021] Where k = 1, 2, ..., K dp , It is the membership function of the j-th feature. is the membership value of the i-th sample, n is the number of samples, d is the number of features, 1 is a column vector of all 1s, and X is the input data;
[0022] According to the objective function right Find the partial derivative. It is the target output. Given the antecedent parameter of the k-th rule, calculate the consequent parameter of the k-th rule. for:
[0023]
[0024] Where k = 1, 2, ..., K dp η is a constant, and I is the identity matrix. The target output is calculated by determining the output Y of the dp-th rule layer. dp :
[0025]
[0026] Further, the S3 specific method is to optimize the input space of each rule of the current layer in the dp layer rule layer by using the following formula:
[0027]
[0028] wherein dp=2,...,DP, represents the i-th sample optimization vector of the dp layer, represents the membership value of x i , in the dp layer rule layer, the error E k is used to quantitatively evaluate the rule:
[0029]
[0030] wherein dp=1,2,...,DP, k=1,2,...,K dp , n is the number of samples, is the target output of the i-th sample, y ik is the output of the k-th rule to the i-th sample, μ k (x i ) is the membership value of the i-th sample under the k-th rule, and the error threshold e dp is set, and the high-quality rule is evaluated to enter the next rule layer for training:
[0031]
[0032] wherein Q dp is the serial number set of the high-quality rule of the dp layer, is the error value of the k-th rule of the dp layer, e dp is the set error threshold of the dp layer, K' dp is the number of optimized rules of the dp layer rule library, dp=1,2,...,DP, and the error threshold should decrease with the increase of the number of rule layers, i.e. e DP ≤...≤e2≤e1, and the rule optimization is stopped when the set expectation is reached;
[0033] The training output of the persistent rule optimization T-S classifier is Y dp , the rules of the DP layer are fixed, including the cluster center c j , the cluster variance δ j and the consequent parameter is the consequent parameter of the k-th rule in the DP rule layer, j=1,2,...,d, k=1,2,...,K' DP .
[0034] Further, the specific method of S4 is: input test data S'=[s'1, s'2, …, s'N] and calculate the test classification output Y' of the T-S classifier optimized by the persistence rule according to the rule base of the DP layer in the training phase and according to formulas 1-6. d ] T , according to the rule base of the DP layer in the training phase, the test classification output Y' of the T-S classifier optimized by the persistence rule is calculated according to formulas 1-6.
[0035] Beneficial effects: Compared with the prior art, the technical scheme of the present application has the following beneficial technical effects:
[0036] The method of the present application can selectively retain rules with excellent data expression ability, reduce the generation of random rules, improve the efficiency of the model, and simplify the structure. The method of the present application uses the description information of the membership value to the features to influence the importance of the features of the rules, thereby improving the expression ability of the rules. The method of the present application repeatedly judges the error between the target output and the rule output, and layer by layer approaches the target output, thereby ensuring the classification performance of the model. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 Flow chart of the deep T-S fuzzy classifier optimized by the persistence rule. DETAILED DESCRIPTION
[0038] The above description is only the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the present application.
[0039] As shown in Figure 1 , the present application proposes a deep T-S dermatitis classification method optimized by the persistence rule, which comprises the following steps:
[0040] S1: Obtain dermatitis data and perform feature reduction by the variance threshold selection method, and randomly divide the data samples into a training set S and a test set S' according to a ratio of 7:3;
[0041] S2: Construct a deep T-S fuzzy neural network comprising an input layer, a membership degree layer, a rule layer, and an output layer, wherein the number of rule layers is DP, and the number of layers of the remaining layers is 1; generate fuzzy rules by inputting the training set S;
[0042] S3: Evaluate and optimize the rules of the DP layer, select part of the rules to enter the DP+1 layer fuzzy rule base, DP=1, 2, …, DP, finally fix the DP layer fuzzy rule base and calculate the training classification result, and DP is the number of rule layers;
[0043] S4: input the test set S' into the DP layer fuzzy rule base to obtain the test classification result.
[0044] Furthermore, the specific method of S1 is as follows: The original dermatitis data is normalized, and then the variance thresholding method is used to filter the features, that is, according to the variance formula... Calculate the mean squared error of the j-th feature in sequence, x ij It is the j-th feature value of the i-th sample. It is the mean of the j-th feature data, j = 1, 2, ..., D, where D is the initial number of features and n is the number of samples. A variance threshold is set. Features with a mean squared error greater than a threshold are removed, and data samples are randomly assigned to the training set in a 7:3 ratio. With test set and It represents the target classification results of training and testing, where n is the number of training samples, n′ is the number of test samples, and d is the number of existing features.
[0045] Furthermore, the specific method of S2 is as follows: The input layer is responsible for normalizing the input data. The input layer takes the training data S as input, and outputs the normalized data. The membership layer achieves fuzzification of feature values. This layer takes feature values X = [x1, x2, ..., x...] as input. d ] T The membership function is used to convert the feature values into membership values, and the membership values are output. A Gaussian function is used as the membership function, and the membership function of the j-th feature is... Expressed as:
[0046]
[0047] Where i = 1, 2, ..., n, and This represents the cluster center and variance of the m-th cluster, where m = 1, 2, ..., M, and M is the number of clusters. The FCM algorithm is used to cluster the feature values, and the j-th feature x... j =[x 1j ,x 2j ,…,x nj Clustering into M classes generates cluster centers. Clustering variance x represents j The m-th linguistic fuzzy set,
[0048] j = 1, 2, ..., d, where d is the existing characteristic number;
[0049] Each rule layer is input with feature values and their membership values to generate rules and calculate the fitness and error of the rules. High-quality rules are selected and input into the next rule layer. Starting from the second rule layer, additional rules from the previous rule layer are input to enrich the current rule base. The fuzzy rule base of the dp-th layer initially contains K. dpK fuzzy rules, dp = 1, 2, …, DP;
[0050] K fuzzy rules, dp = 1, 2, …, DP; dp K fuzzy rules, dp = 1, 2, …, DP; d ] T K fuzzy rules, dp = 1, 2, …, DP; The kth fuzzy rule is expressed as:
[0051]
[0052] K fuzzy rules, dp = 1, 2, …, DP; dp is the number of rules in the dpth rule layer, is the language fuzzy set selected by the kth rule in the jth feature, M is the number of clustering categories, is the weight coefficient of the jth feature, x j is the jth feature value, indicates that the jth feature value is the language meaning corresponding to f k (X) is the output of the kth fuzzy rule, which is divided into IF part and Then part, i.e. antecedent part and consequent part;
[0053] The fitness of each rule and the antecedent parameters are calculated, and the fitness of the kth fuzzy rule is and the expression of the antecedent parameter is:
[0054]
[0055] K fuzzy rules, dp = 1, 2, …, DP; dp , is the membership function of the jth feature, is the membership value of the ith sample, n is the number of samples, d is the number of features, 1 is the all-1 column vector, and X is the input data;
[0056] According to the objective function , the partial derivative of is calculated, is the target output, is the antecedent parameter of the kth rule, and the consequent parameter of the kth rule is obtained as:
[0057]
[0058] K fuzzy rules, dp = 1, 2, …, DP; dp , η is a constant, I is the unit matrix, is the target output, and the output Y dp of the dpth rule layer is calculated as:
[0059]
[0060] Further, the S3 specific method is to optimize the input space of each rule of the current layer in the dp layer rule layer by using the following formula:
[0061]
[0062] wherein dp = 2, …, DP, represents the i-th sample optimization vector of the dp layer, represents the membership value of x i , in the dp layer rule layer, the error E k is used to quantitatively evaluate the rule:
[0063]
[0064] wherein dp = 1, 2, …, DP, k = 1, 2, …, K dp , n is the number of samples, is the target output of the i-th sample, y ik is the output of the k-th rule to the i-th sample, μ k (x i ) is the membership value of the i-th sample under the k-th rule, and the error threshold e dp is set, and the high-quality rule is evaluated to enter the next rule layer for training:
[0065]
[0066] wherein Q dp is the serial number set of the high-quality rule of the dp layer, is the error value of the k-th rule of the dp layer, e dp is the set error threshold of the dp layer, K' dp is the number of optimized rules of the dp layer rule library, dp = 1, 2, …, DP, and the error threshold should decrease with the increase of the number of rule layers, i.e. e DP ≤…≤e2≤e1, and the rule optimization is stopped when the set expectation is reached;
[0067] The training output of the persistent rule optimization T-S classifier is Y dp , and the rules of the DP layer are fixed, including the cluster center c j , the cluster variance δ j and the consequent parameter is the consequent parameter of the k-th rule in the DP rule layer, j = 1, 2, …, d, k = 1, 2, …, K' DP .
[0068] Further, the S4 specific method is: input test data S' = [s'1, s'2,..., s'N], according to the rule base of the DP layer in the training phase, calculate the test classification output Y' of the persistent rule optimized T-S classifier according to formula 1-6. d ] T , according to the rule base of the DP layer in the training phase, calculate the test classification output Y' of the persistent rule optimized T-S classifier according to formula 1-6.
Claims
1. A dermatitis classification method based on deep TS persistence rule optimization, characterized in that, The method includes the following steps: S1: Obtain dermatitis data and perform feature reduction using the variance threshold selection method. Randomly divide the data samples into the training set S and the test set S′ in a 7:3 ratio. S2: Construct a deep TS fuzzy neural network, which includes an input layer, a membership layer, a rule layer, and an output layer. The rule layer has a depth of DP, and the other layers have a depth of 1. Input the training set S to generate fuzzy rules. S3: Evaluate and optimize the rules of the dp layer, select some rules to enter the fuzzy rule library of the dp+1 layer, dp = 1, 2, ..., DP, and finally fix the fuzzy rule library of the DP layer and calculate the training classification result, where DP is the rule layer number; S4: Input the test set S′ into the fuzzy rule base of the DP layer for training to obtain the test classification results; The specific method of S3 is to optimize the input space of each rule in the current layer of the dp layer using the following formula: Where dp = 2, ..., DP, This represents the optimization vector of the i-th sample in the dp-th layer. Represents x i The membership value, in the dp-th layer of the rule layer, uses the error E k Quantitative evaluation of the rules: Where dp = 1, 2, ..., DP, k = 1, 2, ..., K dp n is the number of samples. y is the target output of the i-th sample. ik It is the output of the k-th rule on the i-th sample, μ k (x i ) is the membership value of the i-th sample under the k-th rule, and an error threshold e is set. dp High-quality rules are evaluated and then moved to the next rule layer for training. Among them, Q dp It is the set of indices of the high-quality rules at the dp-th layer. It is the error value of the k-th rule in the dp layer, e dp It is the set error threshold for the dp-th layer, K′ dp dp is the number of rules in the optimized rule base of the dp-th layer, where dp = 1, 2, ..., DP. This error threshold should decrease as the number of rule layers increases, i.e., e DP ≤…≤e2≤e1, rule optimization stops when the set expectation is reached; The training output of the TS classifier is continuously optimized by rules to achieve Y. dp The rules of the fixed DP layer include the cluster centers c. j Cluster variance δ j With consequent parameters It is the consequent parameter of the k-th rule in the DP rule layer, j = 1, 2, ..., d, k = 1, 2, ..., K′ DP .
2. The dermatitis classification method based on deep TS persistence rule optimization according to claim 1, characterized in that, The specific method for S1 is as follows: Normalize the original dermatitis data, and then use the variance thresholding method to filter the features, i.e., according to the variance formula... Calculate the mean squared error of the j-th feature in sequence, x ij It is the j-th feature value of the i-th sample. It is the mean of the j-th feature data, j = 1, 2, ..., D, where D is the initial number of features and n is the number of samples. A variance threshold is set. Features with a mean squared error greater than a threshold are removed, and data samples are randomly assigned to the training set in a 7:3 ratio. With test set and It represents the target classification results of training and testing, where n is the number of training samples, n′ is the number of test samples, and d is the number of existing features.
3. The dermatitis classification method based on deep TS persistence rule optimization according to claim 2, characterized in that, The specific method of S2 is as follows: The input layer is responsible for normalizing the input data. The input layer takes the training data S as input, and outputs the normalized data. The membership layer achieves fuzzification of feature values. This layer takes feature values X = [x1, x2, ..., x...] as input. d ] T The membership function is used to convert the feature values into membership values, and the membership values are output. A Gaussian function is used as the membership function, and the membership function of the j-th feature is... Expressed as: Where i = 1, 2, ..., n, and This represents the cluster center and variance of the m-th cluster, where m = 1, 2, ..., M, and M is the number of clusters. The FCM algorithm is used to cluster the feature values, and the j-th feature x... j =[x 1j ,x 2j ,…,x nj Clustering into M classes generates cluster centers. Clustering variance x represents j The m-th linguistic fuzzy set, j = 1, 2, ..., d, where d is the existing characteristic number; Each rule layer is input with feature values and their membership values to generate rules and calculate the fitness and error of the rules. High-quality rules are selected and input into the next rule layer. Starting from the second rule layer, additional rules from the previous rule layer are input to enrich the current rule base. The fuzzy rule base of the dp-th layer initially contains K. dp There are three fuzzy rules, dp = 1, 2, ..., DP; Generate K dp There are random rules, each rule using the sequence [x1, x2, ..., x]. d ] T As the input vector, each feature is randomly selected from a certain linguistic fuzzy set. The k-th fuzzy rule is represented as: Among them, K dp It is the number of rules in the dp-th rule layer. It is the fuzzy set of language selected by the k-th rule in the j-th feature. M is the number of cluster categories. x is the weight coefficient of the j-th feature. j It is the j-th eigenvalue. This indicates that the j-th eigenvalue corresponds to The linguistic meaning, f k (X) is the output of the kth fuzzy rule. The rule is divided into an IF part and a Then part, namely the antecedent part and the consequent part. Calculate the fitness of each rule and its antecedent parameters, and the fitness of the k-th fuzzy rule. and antecedent parameters The expression is: Where k = 1, 2, ..., K dp , It is the membership function of the j-th feature. is the membership value of the i-th sample, n is the number of samples, d is the number of features, 1 is a column vector of all 1s, and X is the input data; According to the objective function right Find the partial derivative. It is the target output. Given the antecedent parameter of the k-th rule, calculate the consequent parameter of the k-th rule. for: Where k = 1, 2, ..., K dp η is a constant, and I is the identity matrix. The target output is calculated by determining the output Y of the dp-th rule layer. dp :
4. The dermatitis classification method based on deep TS persistence rule optimization according to claim 1, characterized in that, The specific method for S4 is as follows: Input test data S′=[s′1,s′2,…,s′] d ] T Based on the rule base of the DP layer during the training phase, the test classification output Y′ of the TS classifier is calculated according to formulas 1-6 to optimize the continuous rules.