Health data classification method based on Ising machine training TSK fuzzy system

Abstract

The invention discloses a health data classification method based on an Ising machine training TSK fuzzy system, and the method comprises the steps: collecting target health data to construct fuzzy sets, calculating leading parameters of different fuzzy sets through a Gaussian membership function, solving the activation degree of each sample for different rules, carrying out the weighted average of the output of the rules according to the activation degree, and carrying out the classification of the health data. The output of the TSK fuzzy system is expressed as a linear regression form; linear regression output is converted into a QUBO problem through equivalent mathematical changes; and solving the optimal unwinding solution by using a quantum simulated annealing algorithm, and realizing classification of the target health data by analyzing the unwinding solution. According to the method, efficient approximate calculation is performed locally through a quantum simulated annealing algorithm, or the method is deployed on a quantum computer platform to obtain a higher-speed global optimal solution. While the calculation efficiency is ensured, the quantum optimization process can effectively avoid a common local optimal trap in a traditional method, so that the overall performance of the model in a complex nonlinear problem is improved.

Description

Technical Field

[0001] This invention belongs to the field of quantum machine learning, and in particular relates to a health data classification method based on Ising machine training of a TSK fuzzy system. This method is suitable for auxiliary modeling and feature classification of health-related structured data, and is particularly useful for feature recognition and pattern extraction of multidimensional health data such as speech and physiological signals. Background Art

[0002] With the rapid development of artificial intelligence, machine learning (ML) has become an important tool for extracting patterns and making predictions from massive amounts of data. It is widely used in fields such as medical imaging diagnosis, electronic health record analysis, financial risk assessment, autonomous driving decision-making, and natural language understanding. In the field of medical information processing, technologies such as deep learning-based image segmentation, statistical learning-based clinical risk prediction, and feature engineering-based genomic analysis have made significant progress in improving image segmentation performance and feature modeling efficiency. However, modern medical data often exhibits multi-source heterogeneity, high-dimensionality, nonlinearity, and noise. Multiple perspectives, such as imaging data, laboratory test indicators, gene expression profiles, and vital signs, must be integrated into a single model. These high-dimensional features make classical optimization algorithms prone to local optima and significantly increase computational complexity. Furthermore, black-box models struggle to meet clinical interpretability and regulatory compliance requirements. Faced with these challenges, traditional methods such as support vector machines, random forests, and deep neural networks often struggle to generalize and maintain robustness in complex medical scenarios. Consequently, researchers are seeking new approaches to advance machine learning, with quantum computing being one of the technologies attracting considerable attention.

[0003] The potential of quantum computing lies in its ability to leverage the properties of quantum mechanics to break through the bottlenecks of traditional computing systems, bringing about new computing power and speed. Like other classical theories, machine learning can be embedded in quantum mechanical formalisms, a field known as quantum machine learning (QML). As an emerging paradigm, quantum machine learning leverages physical properties such as quantum superposition and entanglement, offering a natural advantage in solving combinatorial optimization and non-convex problems. Leveraging the parallel computing power of quantum bits and the global search mechanism of quantum annealing, it is possible to efficiently avoid local optima in high-dimensional parameter spaces, significantly accelerating model training and improving generalization performance. Furthermore, through quantum feature mapping, it is possible to capture complex data associations that are difficult to express using classical methods, potentially breaking through the bottleneck of classical machine learning in medical big data.

[0004] Although previous research has attempted to integrate quantum algorithms with machine learning, shortcomings remain. Among existing patents, [Patent No.: CN202411245125.0] describes a method for improving the generalization performance of quantum machine learning models using two-step optimization. This method solves a quadratic programming algorithm to find a set suitable for a small, sampled dataset. The eigenvalues ​​of this set are then used as a measurement operator to optimize gate circuit parameters on the training set, thereby improving the model's generalization and classification performance. However, this method is still prone to falling into local optima when processing high-dimensional medical data, and the resulting measurement operator is difficult to intuitively interpret as a basis for classification. [Patent No.: CN202311515541.3] proposes a machine learning optimization algorithm based on parameterized quantum circuits (PQCs) and quantum eigenmaps. This method uses quantum eigenmaps for data transformation, executes and measures the loss function using a quantum computer, and calculates gradients using the finite difference method on a classical computer. The PQC parameters are dynamically optimized through adaptive gradient descent to improve training stability and optimization efficiency. This method, which uses traditional gradient-stepping algorithms, is still subject to local optimality issues. Furthermore, the finite difference method is computationally complex, especially in high-dimensional parameter spaces. Calculating gradients in multiple directions incurs additional overhead, impacting efficiency and making it less resistant to noise. In summary, a new quantum machine learning solution is urgently needed that leverages the advantages of quantum parallel global search while also ensuring model interpretability, to better meet the needs of multi-perspective, complex medical data classification. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a health data classification method based on the Ising machine training TSK fuzzy system. It is suitable for feature extraction and auxiliary modeling tasks for health-related structured data (such as speech, physiological signals, etc.). The method first converts the TSK fuzzy system into a linear regression form, and further converts it into a quadratic unconstrained binary optimization (QUBO) problem, and then uses the quantum simulated annealing algorithm to solve the QUBO matrix, thereby achieving efficient model parameter optimization. By introducing the global search capability and parallelism of quantum computing, this method can effectively deal with the local optimal dilemma in high-dimensional complex nonlinear data modeling, and improve the system's generalization ability and stability. At the same time, with the advantages of the TSK fuzzy system in rule expression, this method retains good interpretability on the basis of performance improvement, which helps to improve users' understanding and trust in the basis of model decision-making. Overall, the present invention has significant advantages in accuracy, transparency and application flexibility, and is suitable for multi-domain data modeling tasks including health monitoring, auxiliary assessment and complex behavior analysis.

[0006] Technical solution: A health data classification method based on Ising machine training TSK fuzzy system of the present invention includes the following steps:

[0007] Step 1: Collect target health data to construct fuzzy sets, use Gaussian membership functions to calculate the prior parameters of different fuzzy sets, then find the activation degree of each sample for different rules, and take the weighted average of the rule output according to the activation degree. The output of the TSK fuzzy system is expressed as a linear regression form.

[0008] Step 2: Transform the linear regression output of the TSK fuzzy system into a QUBO problem through equivalent mathematical transformations;

[0009] Step 3: Use the quantum simulated annealing algorithm to find the optimal detangle solution of the QUBO problem, and classify the target health data by analyzing the detangle solution.

[0010] Furthermore, step 1 is specifically as follows:

[0011] The TSK fuzzy system is set as follows:

[0012]

[0013] in is a fuzzy set, f k (x) represents the output of the kth rule of TSKFS, 1≤k≤K, k is the number of rules in the rule base, is the output function f k The parameters of (x), represents the i-th sample under the k-th rule, m k represents the number of samples under the kth rule; when multiplication is used as conjunction and implication operators, addition is used as combination operator, and centroid method is used as clarification operator, the output of TSK fuzzy system is expressed as follows:

[0014]

[0015] Among them, μ k (x) and They represent the fuzzy membership and the normalized fuzzy membership, respectively. (k' and k have the same meaning and are both rule numbers, just to avoid duplication of names). They are defined as follows:

[0016]

[0017] is the i-th dimension x in the input vector x i Fuzzy Set The degree of membership;

[0018] The input space of the example is divided by clustering technology, and then the prior parameters are estimated based on the space division results; the Gaussian membership function is used as the fuzzy membership function in the fuzzy rule:

[0019]

[0020] Parameters in membership functions and It is estimated by applying the clustering results of FCM to the input data of the training dataset, that is:

[0021]

[0022]

[0023] where u jk is the input vector x j The membership degree of the kth cluster, h, is an adjustable parameter determined by the learning strategy;

[0024] The output of the TSK fuzzy system is expressed in the form of linear regression as follows:

[0025]

[0026] in

[0027] x e =[1,x T ] T ∈R (d+1)×1

[0028]

[0029] p g =[(p 1 ) T ,(p 2 ) T ,...,(p K ) T ] T ∈R K(d+1)×1

[0030] x g represents the eigenvector obtained by fuzzy mapping of the antecedent part of the TSK fuzzy system, p g Represents the consequent parameters of the TSK fuzzy system, through the known x g Solve using the least squares method.

[0031] Furthermore, step 2 is specifically as follows:

[0032] In the total sample matrix Each row of is expanded by adding a constant 1 to obtain the expanded regression training data matrix Regression training label Indicates that the regression weights are used For given X and Y, the training linear regression model is expressed as follows:

[0033]

[0034] Where E(w) is the Euclidean error function. The regression problem has an analytical solution given by

[0035] w=(X T X) -1 X T Y

[0036] The formula can be rewritten as follows

[0037]

[0038] Introduce a K-dimensional precision vector P = [p1, p2, ..., p K ] T , each element in P is an integer power of 2 and is positive or negative. Then introduce a vector of the same K dimension Replace with binary coefficients to make the inner product Can get a scalar This scalar represents the i-th element in the weight vector (1≤i≤d+1). The elements in the vector P are ordered; Is a binary decision variable used to select or ignore elements in P. When the value is 1, it is selected and when it is 0, it is ignored. and the precision matrix is defined as follows:

[0039]

[0040]

[0041] Among them I d+1 represents a (d+1)-dimensional identity matrix, represents the Kronecker product, matrix The dimension is (d+1)×K(d+1), and the following formula is used to restore the original weight vector:

[0042]

[0043] Finally, use the precision matrix and binary vectors To represent the weight vector, the original minimization problem is transformed into an equivalent QUBO problem, and the following formula is obtained:

[0044]

[0045] Furthermore, step 3 is specifically as follows:

[0046] The weight vector is expressed as the product of the precision matrix and the binary variable in represents a binary decision variable, The precision matrix, P is the precision vector composed of powers of 2. Here represents the Kronecker product, I d+1 is the identity matrix. This transformation converts the real-valued parameter optimization problem into a QUBO problem on discrete binary variables.

[0047] After constructing the square loss objective function in the linear regression model, substituting it into the above transformation, the QUBO problem can be obtained as follows:

[0048]

[0049] in To transform and expand the sample feature matrix through precision The obtained coefficient matrix and bias term.

[0050] The QUBO problem is equivalently converted to the Ising Hamiltonian form so that it can be solved using quantum simulated annealing. This conversion is done by variable substitution. where s i ∈{-1,1} represents the Ising spin variable, and the following Ising model expression is obtained:

[0051]

[0052] The coupling term and bias term are converted from the QUBO matrix, specifically:

[0053]

[0054] After the model is built, the quantum simulated annealing algorithm is used to solve it. Parameters such as the initial temperature, annealing attenuation coefficient, and number of iterations at the temperature are set to simulate the evolution of the system under quantum fluctuations, and the optimal spin solution s corresponding to the lowest Hamiltonian is obtained. *

[0055] Map the optimal spin solution back to the original binary variable space:

[0056]

[0057] The final restored optimized model parameter weights are:

[0058]

[0059] The obtained weight w *The output function of the TSK fuzzy system constructed in step 1 can be directly substituted into the prediction and classification of the target health data, thus completing the entire modeling and classification process. This solution combines the global search capabilities of quantum optimization with the interpretability of fuzzy systems, significantly improving the model's adaptability and classification accuracy for complex, high-dimensional health data.

[0060] The present invention further discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0061] The present invention further discloses a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the method of the present invention when the computer program / instruction is executed by a processor.

[0062] The present invention further discloses a computer program product, comprising a computer program / instruction, which implements the steps of the method of the present invention when executed by a processor.

[0063] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0064] (1) This invention uses an Ising machine to solve model parameters. This can be performed locally using a quantum simulated annealing algorithm for efficient approximate calculations, or deployed on a quantum computer platform to obtain a faster global optimal solution. While ensuring computational efficiency, the quantum optimization process can effectively avoid the local optimality trap common in traditional methods, thereby improving the overall performance of the model in complex nonlinear problems.

[0065] (2) This method is less dependent on the size of the training data and is suitable for feature modeling and classification tasks in scenarios with limited data volume. Experiments have shown that compared with traditional machine learning methods, strategies based on quantum optimization have stronger generalization ability and robustness when processing small samples and high-dimensional feature data. Therefore, it is particularly suitable for practical scenarios with high data acquisition costs, such as individual behavior monitoring and structured health data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is the linear regression model fitting graph;

[0067] Figure 2 This is a flow chart of the Tsk-Ising process of the present invention;

[0068] Figure 3 It is the algorithm Tsk-Ising;

[0069] Figure 4 This is an experimental comparison chart. DETAILED DESCRIPTION

[0070] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0071] The TSK fuzzy system based on Ising machine optimization of the present invention is mainly divided into two components:

[0072] (1) Convert the TSK fuzzy system into linear regression form:

[0073] Fuzzy systems are a type of intelligent model based on fuzzy logic rules. Their notable feature is their interpretability, unlike traditional "black box" models. Furthermore, fuzzy systems demonstrate a strong ability to model uncertainty, resulting in increased robustness in uncertain scenarios. Because fuzzy system modeling methods can be developed in a data-driven manner, optimizing system parameters using available training data becomes much easier. The TSK fuzzy system (TSK-FS) is the most widely used fuzzy system, consisting of a set of "if-then" inference rules:

[0074]

[0075] in is a fuzzy set, f k (x)(1≤k≤K) represents the output of the kth rule of TSKFS, k is the number of rules in the rule base, is the output function f k When multiplication is used as conjunction and implication operators, addition is used as combination operator, and centroid method is used as clarification operator, the output of TSK fuzzy system can be expressed as follows:

[0076]

[0077] Among them, μ k (x) and They represent the fuzzy membership and the normalized fuzzy membership, respectively, and are defined as follows:

[0078]

[0079] In formula (3), is the i-th dimension x in the input vector x i Fuzzy Set The degree of membership.

[0080] There are several ways to estimate the leading parameters of TSKFS. A typical method is to partition the input space of the examples by clustering techniques (such as the classic fuzzy c-means (FCM) algorithm), and then estimate the leading parameters based on the space partitioning results. For example, if the Gaussian membership function in (5) is used as the fuzzy membership function in the fuzzy rule

[0081]

[0082] Parameters in membership functions and It can be estimated by applying the clustering results of FCM to the input data of the training dataset, that is:

[0083]

[0084] where u jk is the input vector x j The degree of membership of the k-th cluster, h, is an adjustable parameter that can be determined manually or through a learning strategy (such as a cross-validation strategy).

[0085] Once the premise parameters are obtained, the output of TSKFS in (2) can be expressed in the form of linear regression as follows:

[0086]

[0087] in

[0088] x e =[1,x T ] T ∈R (d+1)×1 (9a)

[0089]

[0090]

[0091] p g =[(p 1 ) T ,(p 2 ) T ,...,(p K ) T ] T ∈R K(d+1)×1 (9e)

[0092] x g represents the eigenvector obtained by fuzzy mapping of the antecedent part of the TSK fuzzy system, p g represents the consequent parameter of the TSK fuzzy system. Formula (8) is a linear model, p g Through the known x g Solve using the least squares method.

[0093] (2) Convert the linear model TSK fuzzy system into a QUBO problem:

[0094] The present invention uses adiabatic quantum computing method to solve the linear regression problem, which can find positive and negative real-valued regression weights. In the specific operation, the present invention uses the original Each row of is expanded by adding a constant 1 to increase mathematical convenience, and the expanded regression training data matrix is ​​obtained as Regression training label Indicates that the regression weights are used For given X and Y, the training linear regression model is expressed as follows:

[0095]

[0096] Where E(w) is the Euclidean error function, refer to Figure 1 The green line is represented by the weight w, which is the regression hyperplane that fits the data. The regression problem has an analytical solution, which is given by

[0097] w=(X T X) -1 X T Y (11)

[0098] Formula (10) can be rewritten as follows

[0099]

[0100] Next, we introduce a K-dimensional precision vector P = [p1, p2, ..., p K ] T , each element in P is an integer power of 2 and can be positive or negative. Next, we introduce a vector of the same K dimension Replace with binary coefficients to make the inner product Can get a scalar This scalar represents the i-th element in the weight vector (1≤i≤d+1). The elements in the vector P must be ordered. It can be viewed as a binary decision variable that is used to select or ignore elements in P. When the value is 1, it is selected and when it is 0, it is ignored. and the precision matrix is defined as follows:

[0101]

[0102] Among them I d+1 represents a (d+1)-dimensional identity matrix, represents the Kronecker product, the matrix The dimension of is (d+1)×K(d+1). Now we can use the following formula to recover the original weight vector:

[0103]

[0104] The precision matrix can now be used and binary vectors To represent the weight vector, the minimization problem of formula (12) is transformed into an equivalent QUBO problem, and the following formula is obtained:

[0105]

[0106] In this formula, Y is ignored. T Y, because it is a constant scalar, will not affect the optimal solution of the QUBO problem. At this time, the Ising machine can be used to solve this optimization problem.

[0107] Example

[0108] This embodiment uses the public UCI speech feature dataset for experimental verification. The dataset contains 195 sample records, covering a series of biomedical speech measurement features collected from 31 individuals, including a total of 22 dimensions of structured attribute information. In order to fully retain the distribution characteristics of the original data, only normalization preprocessing is performed, and no imbalanced sample correction strategy is introduced. The TSK-Ising method proposed in the present invention is used in the experiment, and compared with five common traditional machine learning methods (support vector machine SVM, logistic regression LR, random forest RF, ridge regression Ridge, linear discriminant analysis LDA). The evaluation index is the average classification accuracy (ACC), and each method is repeated 10 times to reduce the impact of random fluctuations. The results are as follows Figure 4 As shown in the results, the proposed method performs better in terms of classification accuracy and stability, verifying its generalization ability and robustness in processing complex structured speech data.

Claims

1. A health data classification method based on Ising machine training TSK fuzzy system, characterized by: The steps include: Step 1: Collect target health data to construct fuzzy sets, use Gaussian membership functions to calculate the prior parameters of different fuzzy sets, then find the activation degree of each sample for different rules, and take the weighted average of the rule output according to the activation degree. The output of the TSK fuzzy system is expressed as a linear regression form. Step 2: Transform the linear regression output of the TSK fuzzy system into a QUBO problem through equivalent mathematical transformations; Step 3: Use the quantum simulated annealing algorithm to find the optimal detangle solution of the QUBO problem, and classify the target health data by analyzing the detangle solution.

2. A health data classification method based on Ising machine training TSK fuzzy system according to claim 1, characterized in that: Step 1 is as follows: The TSK fuzzy system is set as follows: in is a fuzzy set, f k (x) represents the output of the kth rule of TSKFS, 1≤k≤K, k is the number of rules in the rule base, is the output function f k The parameters of (x), represents the i-th sample under the k-th rule, m k represents the number of samples under the kth rule; when multiplication is used as conjunction and implication operators, addition is used as combination operator, and the centroid method is used as clarification operator, the output of the TSK fuzzy system is expressed as follows: Among them, μ k (x) and They represent the fuzzy membership and the normalized fuzzy membership respectively, k' and k are two rule numbers, which are defined as follows: is the i-th dimension x in the input vector x i Fuzzy Set The degree of membership; The input space of the example is divided by clustering technology, and then the prior parameters are estimated based on the space division results; the Gaussian membership function is used as the fuzzy membership function in the fuzzy rule: Parameters in membership functions and It is estimated by applying the clustering results of FCM to the input data of the training dataset, that is: where u jk is the input vector x j The membership degree of the kth cluster, h, is an adjustable parameter determined by the learning strategy; The output of the TSK fuzzy system is expressed in the form of linear regression as follows: in x e =[1,x T ] T ∈R (d+1)×1 p g =[(p 1 ) T ,(p 2 ) T ,...,(p K ) T ] T ∈R K(d+1)×1 x g represents the eigenvector obtained by fuzzy mapping of the antecedent part of the TSK fuzzy system, p g Represents the consequent parameters of the TSK fuzzy system, through the known x g Solve using the least squares method.

3. The health data classification method based on Ising machine training TSK fuzzy system according to claim 1 is characterized in that: Step 2 is as follows: In the total sample matrix Each row of is expanded by adding a constant 1 to obtain the expanded regression training data matrix Regression training label Indicates that the regression weights are used For given X and Y, the training linear regression model is expressed as follows: Where E(w) is the Euclidean error function. The regression problem has an analytical solution given by w=(X T X) -1 X T Y The formula can be rewritten as follows Introduce a K-dimensional precision vector P = [p1, p2, ..., p K ] T , each element in P is an integer power of 2 and is positive or negative. Then introduce a vector of the same K dimension Replace with binary coefficients to make the inner product Can get a scalar This scalar represents the i-th element in the weight vector (1≤i≤d+1). The elements in the vector P are ordered; Is a binary decision variable used to select or ignore elements in P. When the value is 1, it is selected, and when it is 0, it is ignored. Binary variables and the precision matrix is defined as follows: Among them I d+1 represents a (d+1)-dimensional identity matrix, represents the Kronecker product, matrix The dimension is (d+1)×K(d+1), and the following formula is used to restore the original weight vector: Finally, use the precision matrix and binary vectors To represent the weight vector, the original minimization problem is transformed into an equivalent QUBO problem, and the following formula is obtained:

4. The health data classification method based on Ising machine training TSK fuzzy system according to claim 1 is characterized in that: Step 3 is as follows: The weight vector is expressed as the product of the precision matrix and the binary variable in represents a binary decision variable, Precision matrix, P is the precision vector composed of powers of 2, where represents the Kronecker product, I d+1 is the identity matrix. This transformation transforms the real-valued parameter optimization problem into a QUBO problem on discrete binary variables. After constructing the square loss objective function in the linear regression model, substituting it into the transformation, the QUBO problem can be obtained as follows in To transform and expand the sample feature matrix through precision The obtained coefficient matrix and bias term; The QUBO problem is equivalently converted to the Ising Hamiltonian form so that it can be solved using quantum simulated annealing. This conversion is done by variable substitution. where s i ∈{-1,1} represents the Ising spin variable, and the following Ising model expression is obtained: The coupling term and bias term are converted from the QUBO matrix, specifically: After the model is built, the quantum simulated annealing algorithm is used to solve it. The initial temperature, annealing attenuation coefficient, and number of iterations at the temperature are set to simulate the evolution of the system under quantum fluctuations to obtain the optimal spin solution s corresponding to the lowest Hamiltonian. * ; Map the optimal spin solution back to the original binary variable space: The final restored optimized model parameter weights are: The obtained weight w * It can be directly substituted into the TSK fuzzy system output function constructed in step 1 to predict and classify the target health data, thus completing the entire modeling and classification process.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 1.

6. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

7. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images