CN115116619A - Intelligent analysis method and system for stroke data distribution rule - Google Patents

Abstract

The invention relates to an intelligent analysis method and system of stroke data distribution law. The methods include: obtaining stroke case data; performing ZCA whitening processing on stroke case data to obtain whitening data; using stacked sparse autoencoder to perform dimensionality reduction processing on the whitening results to obtain dimensionality reduction data; using deep reinforcement learning to optimize Learning the vector quantization clustering method to cluster the dimensionality reduction data to obtain the clustering results; based on the clustering results, the stroke data distribution rules are generated. Based on this method, the present invention can improve the data processing efficiency, thereby improving the generation efficiency and accuracy of the stroke data distribution law.

Description

A method and system for intelligent analysis of stroke data distribution law

technical field

The invention relates to the technical field of data processing, in particular to a method and system for intelligent analysis of stroke data distribution law.

Background technique

At present, the data of stroke cases are mainly analyzed by traditional statistical methods. With the in-depth research of artificial intelligence in the field of intelligent medicine, the mining of stroke data distribution rules based on machine learning methods has become a research hotspot in recent years. However, when analyzing the stroke data in the prior art, the method of manually analyzing the distribution law is generally adopted. Therefore, providing a method or system for quickly and efficiently discovering the distribution law of the stroke data has become a technical problem to be solved urgently in this field. .

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an intelligent analysis method and system of stroke data distribution law, which can improve the data processing efficiency, thereby improving the generation efficiency and accuracy of the stroke data distribution law.

For achieving the above object, the present invention provides the following scheme:

An intelligent analysis method for stroke data distribution law, comprising:

Obtain stroke case data;

Performing ZCA whitening processing on the stroke case data to obtain whitening processing data;

The dimensionality reduction processing is performed on the whitening processing result by using a stacked sparse autoencoder to obtain dimensionality reduction data;

The dimensionality reduction data is clustered by adopting the learning vector quantization clustering method optimized by deep reinforcement learning to obtain a clustering result;

A stroke data distribution rule is generated based on the clustering result.

Preferably, performing ZCA whitening processing on the stroke case data to obtain whitening processing data, specifically including:

converting the stroke case data into a numerical matrix;

standardizing the numerical matrix to obtain a first matrix;

determining the sample covariance matrix of the first matrix, and determining the eigenvalues of the sample covariance matrix;

Arrange the eigenvalues in descending order, and extract the eigenvectors of each eigenvalue after the descending arrangement to obtain a second matrix;

determining the rotation matrix of the second matrix;

The whitening process data is determined based on the rotation matrix.

Preferably, a clustering result is obtained by clustering the dimensionality reduction data using a learning vector quantization clustering method optimized by deep reinforcement learning.

Preferably, the learning vector quantization clustering method optimized by deep reinforcement learning is a learning vector quantization clustering method incorporating a deep reinforcement learning state set and a deep reinforcement learning action set.

Preferably, the selection process of the state set is:

The prototype vector obtained in each iteration of the learning vector quantization clustering method is taken as a state, and the number of iterations is taken as the number of states to form the state set.

Preferably, the construction process of the action set includes:

The action set is obtained by exploring and utilizing the mechanism, and a reward function suitable for the clustering method is designed to determine the action corresponding to the maximum reward value, and each iteration selects the action corresponding to the maximum reward to obtain the iterative clustering result;

Among them, the reward function design:

代表执行动作a_i后所得到质心与各类簇样本的平均距离，d_i代表第i次迭代原始LVQ各类簇样本与其质心平均距离。Among them, a _i represents the action selected from the action set A={a ₁ ,a ₂ ,...,a _i ,...,a _L } in the ith iteration,

Represents the average distance between the centroid obtained after performing action a _i and various cluster samples, and d _i represents the average distance between various cluster samples and their centroids in the original LVQ iteration of the i-th iteration.

Preferably, the selection process of the action set is:

Use the "exploration" mechanism and the "utilization" mechanism to selectively select data points and prototype vectors to perform "closer or farther" operations;

Among them, the realization process of the "utilization" mechanism is: introduce the parameter m<z, randomly select m samples from the input data set to form a data subset X ^m , and "close" the m data in the data set with the prototype vector or The "far away" operation gets an action; z is the number of samples in the stroke dataset;

The implementation process of the "exploration" mechanism is: introduce the parameter v<z, the exploration coefficient ε is set to 0.1, randomly select v samples from the stroke data set to form a data subset X ^v , and compare all the data in the data subset X ^v with the prototype vector Do a "closer" or "away" operation to get an action;

If the sample x _i is the same as the label of the prototype vector p _j , the "zoom in" operation is performed, and the "zoom in" operation is: use the formula p' _j = x _j +η(x _i -x _j ) to update the prototype vector In the formula, η is the learning rate, p′ _j is the updated prototype vector, and x _j is the attribute value corresponding to the prototype vector p _j ;

The distance between the sample x _i and the updated prototype vector p′ _j is:

||p′ _j -x _i || ₂ =||x _j +η(x _i -x _j )-x _i || ₂ =(1-η)*|p _j -x _i ||

If the sample x _i and the prototype vector p _j have different labels, the "far away" operation is performed, and the "far away" operation is: using the formula p' _j =x _j -η(x _i -x _j ) to update the prototype vector p _j ;

At this time, the distance between x _i and the updated prototype vector p′ _j is:

||p′ _j -xi || ₂ =||x _j -η( _xi -x _j )-x _i || _{2 =} (1+η)*||p _{j -xi} || ₂ .

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

The intelligent analysis method for the distribution law of stroke data provided by the present invention uses a feature dimension reduction method to remove redundant noise, removes redundant features of the stroke data, and uses a clustering method to analyze the redundant feature data to obtain the data distribution law. , to realize the intelligent analysis of stroke case data, which can improve the data processing efficiency, and then improve the generation efficiency and accuracy of stroke data distribution rules.

Corresponding to the above-mentioned intelligent analysis method for stroke data distribution law, the present invention also provides an intelligent analysis system for stroke data distribution law, which includes:

A data acquisition module for acquiring stroke case data;

The whitening processing module is used to perform ZCA whitening processing on the stroke case data to obtain whitening processing data;

a dimensionality reduction processing module, configured to perform dimensionality reduction processing on the whitening processing result by using a stacked sparse autoencoder to obtain dimensionality reduction data;

a clustering processing module, configured to perform clustering on the dimensionality reduction data using a clustering method to obtain a clustering result;

A rule generation module, configured to generate a stroke data distribution rule based on the clustering result.

Preferably, the whitening processing module includes:

a matrix conversion unit, for converting the stroke case data into a numerical matrix;

a normalization processing unit, configured to perform normalization processing on the numerical matrix to obtain a first matrix;

an eigenvalue determining unit, configured to determine the sample covariance matrix of the first matrix, and determine the eigenvalues of the sample covariance matrix;

a feature extraction unit, configured to arrange the feature values in descending order, and extract the feature vector of each feature value after the descending arrangement to obtain a second matrix;

a matrix determination unit for determining the rotation matrix of the second matrix;

A whitening processing unit, configured to determine the whitening processing data based on the rotation matrix.

Preferably, the clustering processing module includes:

A clustering processing unit for clustering the dimensionality reduction data by adopting a learning vector quantization clustering method optimized by deep reinforcement learning to obtain a clustering result; A learning vector quantization clustering method for state sets for reinforcement learning and action sets for deep reinforcement learning.

Because the technical effect achieved by the intelligent analysis system for stroke data distribution law provided by the present invention is the same as the technical effect achieved by the intelligent analysis method for stroke data distribution law provided above, it will not be repeated here.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is the flow chart of the intelligent analysis method of stroke data distribution law provided by the present invention;

2 is a schematic structural diagram of a stacked sparse autoencoder provided by an embodiment of the present invention;

3 is a process flow diagram of a deep reinforcement learning-based learning vector quantization clustering method provided by an embodiment of the present invention;

4 is a data processing flowchart of a deep Q network provided by an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of an intelligent analysis system for stroke data distribution law provided by the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The purpose of the present invention is to provide an intelligent analysis method and system of stroke data distribution law, which can improve the data processing efficiency, thereby improving the generation efficiency and accuracy of the stroke data distribution law.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

As shown in Figure 1, the intelligent analysis method for stroke data distribution law provided by the present invention includes:

Step 100: Obtain stroke case data.

Step 101: Perform ZCA whitening processing on the stroke case data to obtain whitening processing data. The purpose of performing ZCA whitening processing on the original stroke case data in step 101 is to make all attributes have the same variance, and each attribute has low correlation or no correlation, thereby reducing the redundancy of the original data . For example, the dimension of the original stroke case data is 40, including 1000 stroke data samples, and the data is divided into 8 categories, namely cerebral infarction, cerebral hemorrhage, TIA, unruptured intracranial aneurysm, spontaneous subarachnoid space Bleeding, arteriovenous malformation (AVM), carotid artery stenosis or occlusion, and moyamoya disease.

ZCA whitening is performed on the data of stroke cases, so that the variances of 40-dimensional attributes such as age, hospital admission, and discharge department are the same, and each attribute has a low correlation or irrelevance, thereby reducing the redundancy of the original data. . Assuming that the original data set has z samples, and the dimension of each sample is d, the specific processing process is as follows:

(1) All original stroke data sets are converted into a d×z numerical matrix X, and after normalization, the mean value of each attribute is made zero, and the final matrix is denoted as A (ie, the first matrix).

(2) First calculate the sample covariance matrix Σ corresponding to the matrix A, find the eigenvalues of the sample covariance matrix Σ, and mark them as λ ₁ , λ ₂ , ..., λ _d in the order of size, corresponding to them _{The eigenvectors of} u ₁ , _{u 2} , .

(3) Calculate the rotated matrix as:

(i＝1，2，...，d)去乘以矩阵X_rot对应第g(g＝1，2，...，d)行的各元素，以使旋转后矩阵对应的每个属性为单位方差。Among them, x _z is the input data, the dimension of the input data is d, and z is the number of data samples.

( _i =1, 2, . is the unit variance.

All features are made to have the same variance based on the above step (3).

(4) Multiply the matrix X _rot to the left by the matrix U, and the obtained matrix X ₁ =UX _rot is the result of ZCA whitening of the original data set. Each column of the matrix corresponds to the sample data after ZCA whitening. Based on the processing of this step, after whitening The correlation between the features of the data is low.

Step 102: Use the stacked sparse autoencoder to perform dimension reduction processing on the whitening processing result to obtain dimension reduction data. For example, dimensionality reduction of ZCA-whitened stroke data by stacking sparse autoencoders (SAE). The structure of the stacked sparse autoencoder is shown in Figure 2, where x represents the input layer data, corresponding to X ₁ in the whitening process, h represents the first layer of the SAE network structure in the hidden layer, and o and k represent the second and first layers, respectively. Three-layer SAE network structure, y represents the output layer, and h, o, and k together form the hidden layer. In the process of training the three-layer SAE network, the output layer neurons are initialized first, and the weight of each neuron is assigned a random value close to zero and the learning rate η is initialized (decreases as time increases).

Based on the structure shown in Figure 2, calculate the distance between the input data and the weights of all output layers, select the neuron with the smallest weight vector distance as the winning neuron and adjust the weight of the winning neuron to reduce the learning rate, Determine whether the stopping condition is met (the learning rate is zero or the maximum number of iterations is reached). If the stopping condition is not met, the distance between the input data and the weight of the output layer is calculated to find the winning neuron. Finally, the dataset obtained by SAE dimensionality reduction is sent to the Deep Reinforcement Learning Optimized Learning Vector Quantization Clustering (DQN-LVQ) algorithm for clustering.

Step 103: Clustering the dimensionality reduction data by using a clustering method to obtain a clustering result. In the present invention, it is a learning vector quantization clustering method (DQN-LVQ) based on deep reinforcement learning, which integrates the state set and action set in reinforcement learning into the clustering process, and establishes the Markov decision corresponding to the clustering problem. In the process, the action set is optimized by exploration and utilization mechanism to explore more possibilities, and a reward function suitable for this clustering method is designed to determine the action corresponding to the maximum reward, and each iteration selects the best action to get the iteration the best clustering result.

Among them, the reinforcement learning adopted in the present invention is a deep Q network. Deep Q network (DQN) belongs to the deep reinforcement learning algorithm based on value function. Through a deep neural network, the current target state action Q function value (hereinafter referred to as the Q value) is estimated, and another structure similar to, However, networks with different parameters are used to estimate the Q value of the current target state, that is, Q(s, a θ _i ), where s represents the state, a represents the action to be performed, and the current state Q value is measured. The principle of data processing is shown in Figure 4. The environment is initialized, the state s _t is obtained, the action a is executed, the state s _t+1 and the return r of a new round are obtained, and the playback unit stores (s _t , a, r, s _t+1 ), through the samples collected from the playback unit, the gradient descent method is used to solve the problem, and the parameters are updated.

Learning vector quantization clustering (LVQ) belongs to the prototype clustering method. First, a set of prototype vectors is randomly selected in the data set as the cluster center, and the clustering space is divided into multiple clusters. For each input sample, it is classified into the nearest cluster, requiring the data to be labeled with a class. The core idea of LVQ is to iteratively optimize the prototype vectors corresponding to various clusters. Each time, for all the labeled training samples, find the prototype vector closest to its position, and check whether the category identification values of the two are the same. vector to make appropriate updates. Based on this, the state set selection process of the present invention is:

The prototype vector obtained by each iteration of the LVQ algorithm is regarded as a state, the number of iterations is regarded as the number of states, the initialized prototype vector is the state s ₀ , and the state s ₁ is entered after an action is performed, that is, an LVQ operation is performed. For example, if the number of iterations is set to f, the state set contains f states s _f , the prototype vector corresponding to the final state s _f is the center of each type of cluster, and the prototype vector with the closest distance to each sample is the cluster to which it belongs.

The action set is selected as:

The traditional LVQ algorithm performs operations on all data points and prototype vectors, and the resulting clustering result is not necessarily the optimal solution, because some data points may not need to be zoomed in or away from the operation, resulting in the clustering performance of the algorithm. Lifting, and unnecessarily close or far away operations increase the computational complexity of the algorithm. In order to reduce these unnecessary operations, the present invention further adopts the "exploration and utilization" mechanism to selectively select data points and prototype vectors to perform "closer or farther" operations, so as to reduce the algorithm complexity and improve the clustering effect. The present invention takes the "closer or farther" operation of the prototype vector and the selected point in the data set as an action. More actions mean more possibilities, and better clustering effects may be obtained, but too many actions will increase the amount of calculation, and a balance needs to be achieved. The set action set contains L actions.

Exploration and utilization: different data points and prototype vectors will obtain different clustering results, so it cannot be guaranteed that the best clustering performance can be obtained by operating on all samples and prototype vectors in the original data set. Further, the "exploration" and "utilization" mechanisms are adopted in the selection of data sets. Using the "exploration" mechanism can increase the chance of selecting unmined actions, and the "utilization" mechanism is to make all data as possible to be selected for "closer" or "farther away" operations.

"Using" mechanism to achieve: introduce the parameter m<z (where z is the number of samples in the stroke data set), randomly select m samples from the input data set to form a data subset X ^m , and match the m data in the data set with the prototype The vector does "closer" or "away" operation as an action, and this process embodies the "exploitation" mechanism to perform "closer" or "away" operation in all possible m input samples, so that the resulting prototype vector Represents the class center information of the input samples as much as possible.

Implementation of the "exploration" mechanism: introduce the parameter v<z, the exploration coefficient ε is set to 0.1, randomly select v samples in the input stroke data set to form a data subset X ^v , and do "pull" with all the data in the data set and the prototype vector The "closer" or "farther away" operation is used as an action.

次,即对随机构成的

个

数据子集，依次对每个

中所有样本进行“拉近”或“远离”运算，共计执行

次。探索过程则为

次，即对随机构成的

个

数据子集，依次对每个

中所有的数据进行“拉近”或“远离”运算，共计执行

次。If the setting requires L actions, use the process to perform

times, that is, for randomly composed

indivual

subsets of data, in turn for each

All samples in the "zoom in" or "far away" operation, a total of

Second-rate. The exploration process is

times, that is, for the randomly formed

indivual

subsets of data, in turn for each

All the data in the "closer" or "farther" operation, a total of

Second-rate.

The "closer" operation is: if the sample x _i and the prototype vector _pj have the same label, the "closer" operation is performed, that is, the prototype vector is updated according to the following formula to make the two closer:

p′ _j =x _j +η(x _i -x _j )

The distance between x _i and the updated prototype vector is:

||p′ _j -x _i || ₂ =||x _j +η(x _i -x _j )-x _i || ₂ =(1-η)*||p _j -x _i ||

The updated prototype vector is closer to _xi ; where the learning rate η is set to η=0.1,

The "far away" operation is: if the sample x _i and the prototype vector p _j have different labels, the "far away" operation is performed, that is, the prototype vector is updated according to the following formula:

p′ _j =x _j -η(x _i -x _j )

The distance between x _i and the updated prototype vector is:

||p′ _j -x _i || ₂ =||x _j -η(x _i -x _j )-x _i || ₂ =(1+η)*||p _j -x _i || ₂

After the update, the prototype vector is further away from _xi ;

Reward function design:

代表执行动作a_i后所得到质心与各类簇样本的平均距离。d_i代表第i次迭代原始LVQ各类簇样本与其质心平均距离。α代表其比值。质心与簇内样本点欧氏距离越小，聚类效果越好，所以α>1时，代表本次迭代聚类效果优于LVQ。α<1时，表明本次迭代聚类性能劣于LVQ算法。Among them, a _i represents the action selected from the action set A={a ₁ ,a ₂ ,...,a _i ,...,a _L } in the ith iteration,

Represents the average distance between the centroid obtained after performing action a _i and the samples of various clusters. d _i represents the average distance between the original LVQ cluster samples of the i-th iteration and their centroids. α represents its ratio. The smaller the Euclidean distance between the centroid and the sample points in the cluster, the better the clustering effect, so when α>1, it means that the iterative clustering effect is better than LVQ. When α<1, it indicates that the clustering performance of this iteration is inferior to that of the LVQ algorithm.

Among them, it is used to find the shortest distance from one point to another point in a coordinate system, taking the distance between two points as an example, assuming that the coordinates of A _i and B _j are (x _i , y _i ) and (x i , respectively) _j , y _j ), then the Euclidean distance d _ij between the two is:

Among them, the data processing process of DQN-LVQ is shown in Figure 3 and Table 1.

Table 1 Data processing flow chart of DQN-LVQ

Step 104: Generate a stroke data distribution rule based on the clustering result.

Based on the above description, the present invention also has the following advantages:

1. In view of the defect of redundant features in the stroke data, the present invention first performs ZCA whitening preprocessing on the original stroke data, so as to reduce the redundancy of the data.

2. For the learning vector quantization clustering method, as the data dimension increases, the calculation amount increases, but the data analysis effect shows a downward trend. The encoding performs dimensionality reduction on the stroke dataset, and then clusters the dimensionality-reduced data.

3. The present invention combines deep reinforcement learning and learning vector quantization clustering method (LVQ) to propose a new deep reinforcement learning LVQ clustering method, and then uses the proposed new learning vector quantization clustering method to perform a stroke data set. Analyzed and obtained the distribution law of stroke data.

Corresponding to the above-mentioned intelligent analysis method for stroke data distribution law, the present invention also provides an intelligent analysis system for stroke data distribution law, as shown in FIG. 5 , the system includes:

The data acquisition module 1 is used for acquiring stroke case data.

The whitening processing module 2 is used to perform ZCA whitening processing on the stroke case data to obtain whitening processing data.

The dimensionality reduction processing module 3 is used to perform dimensionality reduction processing on the whitening processing result by using the stacked sparse autoencoder to obtain dimensionality reduction data.

The clustering processing module 4 is used for clustering the dimensionality reduction data by using a clustering method to obtain a clustering result.

The rule generation module 5 is used for generating the stroke data distribution rule based on the clustering result.

As a preferred embodiment of the present invention, in order to improve the real-time performance and accuracy of stroke data processing, the above-mentioned whitening processing module 1 includes:

The matrix transformation unit is used to transform the stroke case data into a numerical matrix.

The normalization processing unit is configured to perform normalization processing on the numerical matrix to obtain the first matrix.

The eigenvalue determining unit is used for determining the sample covariance matrix of the first matrix and determining the eigenvalues of the sample covariance matrix.

The feature extraction unit is used for arranging the eigenvalues in descending order, and extracting the eigenvectors of each eigenvalue after the descending ordering to obtain the second matrix.

The matrix determination unit is used for determining the rotation matrix of the second matrix.

The whitening processing unit is used for determining the whitening processing data based on the rotation matrix.

As another preferred embodiment of the present invention, the clustering processing module 4 adopted above includes:

The clustering processing unit is used for clustering the dimensionality reduction data by adopting the learning vector quantization DQN-LVQ clustering method optimized by deep reinforcement learning to obtain the clustering result. The DQN-LVQ clustering method is a learning vector quantization clustering method that integrates the state set of reinforcement learning and the action set of reinforcement learning.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

In the present invention, specific examples are used to illustrate the principles and implementations of the present invention, and the descriptions of the above embodiments are only used to help understand the method and the core idea of the present invention; There will be changes in the specific implementation manner and application scope of the idea of the invention. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (10)

1. a stroke data distribution law intelligent analysis method, is characterized in that, comprises: Obtain stroke case data; Performing ZCA whitening processing on the stroke case data to obtain whitening processing data; The dimensionality reduction processing is performed on the whitening processing result by using a stacked sparse autoencoder to obtain dimensionality reduction data; The dimensionality reduction data is clustered by adopting the learning vector quantization clustering method optimized by deep reinforcement learning to obtain a clustering result; A stroke data distribution rule is generated based on the clustering result. 2. The method for intelligent analysis of stroke data distribution law according to claim 1, wherein the described stroke case data is subjected to ZCA whitening processing to obtain whitening processing data, specifically comprising: converting the stroke case data into a numerical matrix; standardizing the numerical matrix to obtain a first matrix; determining the sample covariance matrix of the first matrix, and determining the eigenvalues of the sample covariance matrix; Arrange the eigenvalues in descending order, and extract the eigenvectors of each eigenvalue after the descending arrangement to obtain a second matrix; determining the rotation matrix of the second matrix; The whitening process data is determined based on the rotation matrix. 3 . The method for intelligent analysis of stroke data distribution law according to claim 1 , wherein a clustering result is obtained by clustering the dimensionality reduction data using a learning vector quantization clustering method optimized by deep reinforcement learning. 4 . 4. The method for intelligent analysis of stroke data distribution law according to claim 3, wherein the learning vector quantization clustering method optimized by deep reinforcement learning is to incorporate the state set of deep reinforcement learning and the action of deep reinforcement learning A set of learning vector quantization clustering methods. 5. intelligent analysis method of stroke data distribution law according to claim 3, is characterized in that, the selection process of described state set is: The prototype vector obtained in each iteration of the learning vector quantization clustering method is taken as a state, and the number of iterations is taken as the number of states to form the state set. 6. The method for intelligent analysis of stroke data distribution law according to claim 3, wherein the construction process of the action set comprises: The action set is obtained by exploring and utilizing the mechanism, and a reward function suitable for the clustering method is designed to determine the action corresponding to the maximum reward value, and each iteration selects the action corresponding to the maximum reward to obtain the iterative clustering result; Among them, the reward function design:

Among them, a _i represents the action selected from the action set A={a ₁ , a ₂ , ..., a _i , ..., a _L } in the ith iteration, and d _ai represents the result obtained after executing the action a _i The average distance between the centroid and various cluster samples, d _i represents the average distance between the original LVQ cluster samples of the i-th iteration and their centroids. 7. the stroke data distribution law intelligent analysis method according to claim 6, is characterized in that, the selection process of described action set is: Use the "exploration" mechanism and the "utilization" mechanism to selectively select data points and prototype vectors to perform "closer or farther" operations; Among them, the realization process of the "utilization" mechanism is: introduce the parameter m<z, randomly select m samples from the input data set to form a data subset X ^m , and "close" the m data in the data set with the prototype vector or The "far away" operation gets an action; z is the number of samples in the stroke dataset; The implementation process of the "exploration" mechanism is: introduce the parameter v<z, the exploration coefficient ε is set to 0.1, randomly select v samples from the stroke data set to form a data subset X ^v , and compare all the data in the data subset X ^v with the prototype vector Do a "closer" or "away" operation to get an action; If the sample x _i is the same as the label of the prototype vector p _j , the "zoom in" operation is performed, and the "zoom in" operation is: use the formula p' _j = x _j +η(x _i -x _j ) to update the prototype vector In the formula, η is the learning rate, p′ _j is the updated prototype vector, and x _j is the attribute value corresponding to the prototype vector p _j ; The distance between the sample x _i and the updated prototype vector p′ _j is: ||p′ _j -x _i || ₂ =||x _j +η(x _i -x _j )-x _i || ₂ =(1-η)*||p _j -x _i || If the sample x _i and the prototype vector p _j have different labels, the "far away" operation is performed, and the "far away" operation is: using the formula p' _j =x _j -η(x _i -x _j ) to update the prototype vector p _j ; At this time, the distance between x _i and the updated prototype vector p′ _j is: ||p′ _j -xi || ₂ =||x _j -η( _xi -x _j )-x _i || _{2 =} (1+η)*||p _{j -xi} || ₂ . 8. An intelligent analysis system for stroke data distribution law, characterized in that, comprising: A data acquisition module for acquiring stroke case data; The whitening processing module is used to perform ZCA whitening processing on the stroke case data to obtain whitening processing data; a dimensionality reduction processing module, configured to perform dimensionality reduction processing on the whitening processing result by using a stacked sparse autoencoder to obtain dimensionality reduction data; a clustering processing module, configured to perform clustering on the dimensionality reduction data by adopting a learning vector quantization clustering method optimized by deep reinforcement learning to obtain a clustering result; A rule generation module, configured to generate a stroke data distribution rule based on the clustering result. 9. The intelligent analysis system for stroke data distribution law according to claim 8, wherein the whitening processing module comprises: a matrix conversion unit, for converting the stroke case data into a numerical matrix; a normalization processing unit, configured to perform normalization processing on the numerical matrix to obtain a first matrix; an eigenvalue determining unit, configured to determine the sample covariance matrix of the first matrix, and determine the eigenvalues of the sample covariance matrix; a feature extraction unit, configured to arrange the feature values in descending order, and extract the feature vector of each feature value after the descending arrangement to obtain a second matrix; a matrix determination unit for determining the rotation matrix of the second matrix; A whitening processing unit, configured to determine the whitening processing data based on the rotation matrix. 10. The intelligent analysis system for stroke data distribution law according to claim 8, wherein the clustering processing module comprises: A clustering processing unit, used for clustering the dimensionality reduction data by adopting the learning vector quantization DQN-LVQ clustering method optimized by deep reinforcement learning to obtain a clustering result; the learning vector quantization clustering method optimized by deep reinforcement learning is: A learning vector quantization clustering method incorporating a deep reinforcement learning state set and a deep reinforcement learning action set.

Images