Vital sign signal regression model construction method based on fuzzy set

Abstract

The invention discloses a vital sign signal regression model construction method based on fuzzy sets, and belongs to the technical field of vital sign signal detection.The construction method comprises the steps that firstly, a plurality of fuzzy sets are defined, then labels corresponding to all samples are mapped to respective fuzzy classes of all the fuzzy sets, and the membership degree of each fuzzy class is obtained; obtaining a membership degree vector of each sample so as to calculate a fuzzy class condition probability distance from a source domain to a target domain; and finally, iteratively updating an initial regression model for vital sign signal regression by taking the minimum fuzzy class condition probability distance from the source domain to the target domain as a target or basis until the iterated initial regression model reaches a preset condition, and finally obtaining a vital sign signal regression model. According to the design, information of known source domain samples and regression fuzzy class information are considered, information of known labels can be better utilized, and a vital sign signal regression model with better generalization is constructed in an auxiliary mode.

Description

Technical Field

[0001] The present invention belongs to the technical field of vital sign signal detection, and more specifically, relates to a method for constructing a vital sign signal regression model based on fuzzy sets. Background Art

[0002] With the rapid development of machine learning technology, many mature classification algorithms have been studied for vital sign signals, such as electromyography, electrooculography, electrocardiography, functional near-infrared spectroscopy, and electroencephalography.

[0003] Taking EEG signals collected by brain-computer interfaces as an example, individual differences can cause different users to produce different EEG signals for the same task, significantly impacting the algorithm's portability. Transfer learning, which uses data or knowledge from a source domain (old users) to assist learning in a target domain (new users), can mitigate the impact of individual differences and significantly reduce the need for labeled data in the target domain.

[0004] Transfer learning algorithms can be broadly categorized into traditional transfer learning based on traditional machine learning and deep transfer algorithms based on neural networks. Traditional transfer learning algorithms primarily achieve transfer by modeling the distribution differences between the source and target domains and constructing optimization problems. For example, they use the maximum mean distance to measure the difference in the marginal probability distribution of features between domains. Furthermore, they introduce category information and use class conditional probabilities to estimate conditional probabilities, approximately modeling the difference in the joint probability distribution between domains to achieve transfer. While maintaining inter-domain transferability, they also consider category discriminability, requiring more pronounced category differences in the transferred distribution. Deep transfer learning algorithms primarily model inter-domain distances through a loss, effectively aligning the features extracted by the feature extractor with the distributions of the source and target domains. For example, they use the joint probability of the outputs of different layers to approximate the joint probability distribution of the original data. The original domain is divided into subdomains by category, and alignment within the subdomains further enhances transfer performance.

[0005] However, most of the above algorithms require discrete category information to model conditional probability distributions, class-conditional probability distributions, or joint probability distributions. This makes the above classification transfer learning algorithms unsuitable for the problem of vital sign signal regression with continuous labels. Summary of the Invention

[0006] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a method for constructing a vital sign signal regression model based on fuzzy sets, the purpose of which is to solve the technical problem that the probability distribution constructed with discrete category information cannot be used to construct a vital sign signal regression model.

[0007] To achieve the above object, according to one aspect of the present invention, a method for constructing a vital sign signal regression model based on fuzzy sets is provided, comprising:

[0008] S1: Determine the target domain samples of vital sign signals, the source domain samples of vital sign signals and their corresponding regression-type source domain labels; concatenate all the source domain labels to obtain a label domain, and convert n s Source domain labels Divide into multiple dividing points, use the dividing points to define C f fuzzy sets;

[0009] S2: For each sample x k and its labels are mapped to C f Each fuzzy class is obtained from the fuzzy classes of the fuzzy sets f Membership Each fuzzy class Class f The membership degree μ f Combined to get the sample x k The membership vector of k is the serial number of the sample; if the sample x k If the sample x is a source domain sample, the label is the true label. i If it is a target domain sample, the label is a pseudo label;

[0010] S3: Using formula Calculate the fuzzy class conditional probability distance from the source domain to the target domain Identify the source domain, Identify the target domain, n s is the number of source domain samples, n t is the number of samples in the target domain, is the i-th source domain sample x s,i The corresponding membership of the cth fuzzy class The normalized value of is the jth target domain sample x t,j The corresponding membership of the cth fuzzy class The normalized value of , dist is the distance representation;

[0011] S4: Fuzzy class conditional probability distance from the source domain to the target domain An optimization function or loss function is designed with the minimum as the objective, and the initial regression model for vital sign signal regression is iteratively updated. During the iterative update process, the initial regression model after iteration is used to generate pseudo labels for target domain samples in the next iteration, and returns to S2. When the initial regression model after iteration meets the preset conditions, the vital sign signal regression model is considered to be obtained.

[0012] Furthermore, the n in the tag field s Source domain labels Dividing multiple dividing points includes: dividing the dividing points according to the percentage of the source domain label value; or dividing the dividing points according to the percentage of the interval length corresponding to the theoretical upper and lower bounds of the label in the vital sign signal problem.

[0013] Furthermore, the fuzzy class conditional probability distance from the source domain to the target domain is The optimization function is designed as the minimum objective, and the initial regression model for vital sign signal regression is iteratively updated, including: using the fuzzy class conditional probability distance from the source domain to the target domain The fuzzy joint probability distance is obtained by summing the class edge probability, and the fuzzy joint probability distance is used as the optimization target. The optimization function is set as To update and iterate the initial regression model;

[0014]

[0015] in, is the joint probability distance matrix between the source domain and the target domain, λ is the degree of weighting coefficient between the intra-class distance and the inter-class distance, is the joint probability distance matrix between source domain and target domain classes, is the jth sample x in the target domain t,j Belong to The membership degree of fuzzy classes, A is the feature transformation, is the squared norm on a reproducible Hilbert space.

[0016] Furthermore, the fuzzy class conditional probability distance from the source domain to the target domain is The optimization function is designed as the minimum objective, and the initial regression model for vital sign signal regression is iteratively updated, including: the fuzzy class conditional probability distance from the source domain to the target domain Minimum is the goal, and the optimization function is set to To update and iterate the initial regression model;

[0017]

[0018] Among them, C' f C f The selected value in the value range; λ is the degree of weighting coefficient between the intra-class distance and the inter-class distance; is the joint probability distance matrix of the fuzzy class conditional probability between the source domain and the target domain, is the fuzzy class conditional probability distance matrix between the source domain and the target domain classes, is the jth sample x in the target domain t,j Belong to The membership degree of fuzzy classes, A is the feature transformation, is the squared norm on a reproducible Hilbert space.

[0019] Furthermore, C' f To utilize The optimal value found; max fs is a hyperparameter used to limit the upper limit of the traversal range; Δ(y i ,C f )=max(μ i )-sec(μ i ); sec(μ i ) is μ i The second largest value in y i is the i-th source domain label.

[0020] Furthermore, the class conditional probability distance from the source domain to the target domain is Minimize as the goal, iteratively update the initial regression model for vital sign signal regression, including: setting the loss function to To update and iterate the initial regression model based on the neural network; wherein, express The output of the layer, Represents the membership matrix corresponding to the source domain label, Represents the membership matrix corresponding to the pseudo-label of the target domain, express The eigenspace of tensor products of rank 1.

[0021] Furthermore, the fuzzy class conditional probability distance from the source domain to the target domain is The optimization function is designed with the minimum as the goal, and the initial regression model for vital sign signal regression is iteratively updated, including: class conditional probability distance from the source domain to the target domain Based on the minimum, the fuzzy subdomain distance loss is constructed as the loss function To update the initial regression model based on the neural network iteratively; where φ is the neural network model, x i is the source domain sample, x j is the target domain sample, is the squared norm on a reproducible Hilbert space.

[0022] According to another aspect of the present invention, a regression method for a vital sign signal is provided, comprising: collecting a vital sign signal of a current user, preprocessing the signal and inputting the signal into a vital sign signal regression model to obtain a regression result corresponding to the vital sign signal of the current user.

[0023] According to another aspect of the present invention, a vital sign signal processing system is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a method for constructing a vital sign signal regression model or a method for regressing a vital sign signal.

[0024] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for constructing a vital sign signal regression model or the steps of a vital sign signal regression method are implemented.

[0025] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0026] (1) The present invention provides a method for constructing a vital sign signal regression model based on fuzzy sets, which first defines C f fuzzy sets, and then for each sample x k and its labels are mapped to C f Each fuzzy class is obtained from the fuzzy classes of the fuzzy sets f Membership Then we get each sample x k The membership vector of Secondly, use Calculate the fuzzy class conditional probability distance from the source domain to the target domain Finally, the fuzzy class conditional probability distance from the source domain to the target domain is The minimum is the target or basis, and the initial regression model for vital sign signal regression is iteratively updated until the initial regression model after iteration meets the preset conditions, ultimately obtaining a vital sign signal regression model. This design considers information about known source domain samples and regression fuzzy class information, making better use of known labels and assisting in the construction of a more generalizable vital sign signal regression model.

[0027] (2) In this scheme, the percentage of label values is used to divide the data based on the known source domain data, which can ensure that the number of samples in each fuzzy class is close. The fuzzy set will change according to the distribution density of the known source domain samples, and more fuzzy sets will be divided in the area with higher density. When the number of samples is large and the label distribution is relatively dispersed, it can more naturally reflect the data distribution. In addition, the percentage of interval length is used to divide the label interval based on prior knowledge. It ensures that each fuzzy set covers labels of a fixed length interval by dividing the label interval of the problem, better adapts to out-of-distribution information, and improves generalization. At the same time, when the label distribution is a clustered unimodal distribution, the percentage of interval length can avoid the situation where similar labels are mistakenly divided into different fuzzy classes in the percentage of label value division.

[0028] (3) In this scheme, the optimization function is set as To update and iterate the initial regression model; Among them, through the above method, the joint probability distance in the classification algorithm can be applied to the regression problem, and a better generalized model can be constructed in the regression scenario where the label distribution of the source domain and the target domain is similar.

[0029] (4) This solution sets the optimization function as To update and iterate the initial regression model; Through the above method, we can avoid the instability caused by the large difference in label distribution between the source domain and the target domain, and achieve better generalization model construction in real scenarios with large differences in label distribution.

[0030] (5) C' in this scheme f To utilize The optimal value is found; through the above parameter selection method, the parameters that maximize the discrimination of fuzzy classes can be automatically selected according to the data, and the regression model with better generalization can be constructed in conjunction with the fuzzy class conditional probability alignment.

[0031] (6) This plan sets The initial regression model based on the neural network is updated and iterated; through the above method, the joint maximum mean distance loss in the classification problem can be applied to the regression problem, expanding the scope of application of the existing algorithm and constructing a regression joint adaptation model.

[0032] (7) This scheme sets the loss function as The initial regression model based on the neural network is updated and iterated; through the above method, the sub-domain distance loss in the classification task can be applied to the regression problem to construct a regression sub-domain adaptation network. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the framework of the method for constructing a vital sign signal regression model based on fuzzy sets provided in Example 1 of the present invention;

[0034] Figure 2 This is an example diagram of fuzzy set and membership calculation provided by Example 1 of the present invention;

[0035] Figure 3 This is a structural diagram of a regression model provided in Example 1 of the present invention;

[0036] Figure 4 is a structural diagram of another regression model provided in Example 1 of the present invention;

[0037] Figure 5 This is a framework diagram of the method for constructing a vital sign signal regression model based on fuzzy sets provided in Example 2 of the present invention. The diagram takes the brain-computer interface regression application scenario as an example. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0039] Example 1

[0040] This embodiment provides a method for constructing a vital sign signal regression model based on fuzzy sets, including: S1-S4, such as Figure 1 shown.

[0041] S1: Determine the target domain samples of vital sign signals, the source domain samples of vital sign signals and their corresponding regression source domain labels; concatenate all source domain labels to obtain the label domain, and convert n s Source domain labels Divide into multiple cutoff points and use the cutoff points to define C f A fuzzy set.

[0042] The demarcation point can be determined based on the percentage of the source domain label value; or based on the percentage of the interval length corresponding to the upper and lower bounds of the label in the brain-computer interface problem. Specifically, assuming that there are multiple known subjects, the source domain is constructed together. and a target domain consisting of a target subject This method first concatenates the source domain subject labels to obtain the label domain. s regression labels Divide the cutoff point according to its percentage, and use the cutoff point to define C f Fuzzy sets, get linguistic variables C f =3 as an example, use n s The 5th, 50th and 95th percentile values p5, p 50 and p 95 Define fuzzy sets. When C f When p5 and p 95 is kept constant and the middle percentile is adjusted. An example of a linguistic variable is Figure 2 shown.

[0043] S2: For each sample xk and its labels are mapped to C f Each fuzzy class is obtained from the fuzzy classes of the fuzzy sets f Membership Each fuzzy class Class f The membership degree μ f Combine them to get sample x k The membership vector of k is the serial number of the sample; if the sample x k If the sample x is a source domain sample, the label is the true label. i If it is a target domain sample, the label is a pseudo label.

[0044] Specifically, the language variables obtained by S1 Continuous labels can be fuzzified into discrete categories. f =3 as an example, define three fuzzy classes corresponding to Class1, Class2, and Class3 fuzzy sets. For a label y of a subject in the source domain or target domain, i , we can get the vectors corresponding to the three fuzzy classes: Each element represents y i Belongs to the corresponding fuzzy class. After obtaining the membership vector, a continuous value label can be converted into a discrete fuzzy class. Different from the traditional classification category, here a label can belong to multiple fuzzy classes at the same time with different degrees. For each supervised EEG sample {x i ,y i}, are converted into membership form {x i ,μ i}, for the unsupervised samples of the target subjects, the membership vector can be generated by iteration or pseudo labels generated by the neural network.

[0045] S3: Exploitation Calculate the fuzzy class conditional probability distance from the source domain to the target domain Identify the source domain, Identify the target domain, n s is the number of source domain samples, n t is the number of samples in the target domain, is the i-th source domain sample x s,i The corresponding membership of the cth fuzzy class The normalized value of is the jth target domain sample x t,j The corresponding membership of the cth fuzzy class is the normalized value of , and |·| is the cardinality of the set.

[0046] Specifically, the EEG samples {xi ,μ i}, we can define the inter-domain fuzzy class conditional probability distance. Here, we start from the existing classification transfer learning algorithm class conditional probability distance formula to show the construction process of the regression algorithm. Assume that in a C-class brain-computer interface classification problem, and The subdomain composed of the c-th category in and A certain classification transfer learning algorithm defines the class conditional probability distance between domains as:

[0047]

[0048] dist represents the distance metric based on discrete category information in the classification transfer learning method, |·| is the set cardinality, is a subdomain consisting of c class samples in the source domain. After fuzzification, each dimension of the fixed-length μ vector represents a fuzzy class, corresponding to the discrete categories in the above formula. Since each EEG sample belongs to a virtual fuzzy class to varying degrees, when calculating the class conditional probability, a membership weight needs to be added for each EEG sample: The fuzzy class distance measure in regression is:

[0049]

[0050] The target domain membership is obtained by fuzzifying the pseudo labels generated by the regressor. Thus, the class-based distance metric in regression can be calculated, and the classification transfer learning method can be restructured for regression problems.

[0051] S4: Class conditional probability distance from source domain to target domain The minimum is the target or basis, and the initial regression model for vital sign signal regression is iteratively updated; during the iterative update process, the initial regression model after the iteration generates pseudo labels for the target domain samples in the next iteration and returns to S2; until the initial regression model after the iteration meets the preset conditions, it is considered that the vital sign signal regression model is obtained.

[0052] Specifically, by reducing S3 The class conditional probabilities of the source domain subjects and the target subjects can be aligned. An optimization function or loss term L can be designed transThis helps align EEG samples. After data alignment, the distribution differences between different subjects are reduced, and the regressor trained on the source domain subjects can be directly applied to the target subjects. In traditional transfer learning algorithms, it is necessary to build a regressor on the aligned source domain samples to generate target domain pseudo-labels for the next iteration. The next iteration then repeats steps S2-S4 to gradually achieve alignment. For example, in deep learning algorithms, it is necessary to use the target domain pseudo-labels output by the neural network to calculate the membership vector to achieve alignment.

[0053] As an optional implementation, the class conditional probability distance from the source domain to the target domain is Minimize as the goal, iteratively update the initial regression model for vital sign signal regression, including: setting the optimization function to To update and iterate the initial regression model;

[0054]

[0055] in, is the intra-class joint probability distance matrix between the source domain and the target domain, is the joint probability distance matrix between source domain and target domain classes, is the membership degree of the i-th sample in the source domain to the c-th fuzzy class, The jth sample in the target domain belongs to The membership of a fuzzy class, A is a certain feature transformation (changes with different algorithms) is the squared norm on a reproducible Hilbert space.

[0056] For example, the Joint Probability Distribution Adaptation (JPDA) algorithm is reconstructed. JPDA simultaneously enhances domain transferability and category discrimination by reducing the Discriminative Joint Probability Maximum Mean Discrepancy (DJP-MMD) between the source and target domains. Assuming there are C categories, DJP-MMD is defined as: λ is a hyperparameter that balances transferability and discrimination, where:

[0057] is a reproducible Hilbert space, and A is the transformation matrix to be learned. Due to the use of discrete category information, the original JPDA algorithm cannot be directly used in regression problems.

[0058] In this solution, JPDA can be reconstructed: for a given source domain label y s,i and target domain pseudo labels (iterative generation), first in y s,i C f Fuzzy classes, after fuzzification, the source domain membership vector is obtained and target domain After the generalization algorithm, the regression and can be rewritten as:

[0059]

[0060] Therefore, DJP-MMD can be calculated in regression problems, and JPDA can be successfully applied to regression.

[0061] As an optional implementation, the class conditional probability distance from the source domain to the target domain is Minimize as the goal, iteratively update the initial regression model for vital sign signal regression, including: setting the optimization function to To update and iterate the initial regression model;

[0062]

[0063] C' f C f The selected value of is the joint probability distance matrix of the fuzzy class conditional probability between the source domain and the target domain, is the fuzzy class conditional probability distance matrix between the source domain and the target domain. f To use the formula The optimal value found; max fs is a hyperparameter used to limit the upper limit of the traversal range; Δ(y i ,C f )=max(μ i )-sec(μ i ); sec(μ i ) is μ i The second largest value in y i is the i-th source domain label.

[0064] It should be noted that in addition to the reconstruction of the existing classification transfer learning, a new transfer algorithm can be constructed based on the method of the present invention: Fuzzy Set Selection and Subdomain Distribution Adapation for Regression (FSS-SDAR). The details of the algorithm are as follows: In the generalization of the JPDAR algorithm, C fThe parameters are manually selected. FSS-SDAR uses the concept of fuzzy class confusion to enable the algorithm to automatically select the number of fuzzy sets with the minimum class confusion. The fuzzy set selection mainly includes the following steps:

[0065] First, interval division. When defining fuzzy sets, FSS-SDAR uses the length of the interval to define fuzzy sets, which is different from the previous percentage definition of fuzzy sets. FSS-SDAR divides the label range into several intervals and uses the endpoints of these intervals to define fuzzy sets. In addition, when the label has theoretical upper and lower limits (for example, in the brain-computer interface problem, the fatigue level is between [0,1]), FSS uses the theoretical interval of the label to define the range of the fuzzy set. Now, p5, p 50 and p 95 The 5%, 50%, and 95% intervals are set to the length of the interval, rather than the percentile values of the labels. Using the above interval endpoints, you can define the linguistic variables Because fuzzy sets are defined based on the theoretical range of labels, this approach can capture information outside the source domain distribution, thereby improving the model's generalization ability. Furthermore, in non-percentage settings, the number of fuzzy class samples can be inconsistent, effectively alleviating the problem of the number of samples in each fuzzy set gradually decreasing as the number of samples increases.

[0066] Then, using S1 to get FSS-SDAR selects the appropriate number of fuzzy sets C' f , to minimize fuzzy category confusion and reduce the probability of samples falling between two fuzzy sets. FSS-SDAR uses Δ(y i ,C f ) metric to select C' f : For sample label y i and a specific C' f , let its membership vector be Note sec(μ i ) is μ i The second largest value in . Then Δ(y i ,C f )=max(μ i )-sec(μ i ), where Δ represents the f The difference between the maximum and second-largest membership values of a label under the setting. Δ can be used to measure the distance between the sample and the fuzzy category peak. The larger the Δ value, the closer the label is to the fuzzy category peak corresponding to the maximum membership, indicating that the fuzzy category confusion is lower. Before alignment, FSS-SDAR f Traverse: C f ∈3,4,…,max fs , where max fsIt is a hyperparameter used to limit the upper limit of the traversal range. FSS-SDAR calculates the samples in each source domain at different C f Select the average Δ under , and then select C' that maximizes the average Δ f ,Right now

[0067]

[0068] Therefore, the selected C' f The maximum membership of most samples is increased, thereby alleviating the impact of category confusion on transfer learning performance. It is worth noting that C' f It is calculated based on one source domain. When there are multiple source domains, the fuzzy category may change with different source domains.

[0069] Furthermore, unlike JPDAR, FSS-SDAR does not explicitly align the joint probability distribution and label distribution of the two domains, but only aligns the marginal probability distribution (fuzzy class conditional probability distribution) of each fuzzy category in the two domains, thereby alleviating the negative transfer caused by the large difference in label distribution. f , the distance between the source and target domains in FSS-SDAR can be defined in a form similar to DJP-MMD, but and was modified to:

[0070]

[0071] Substituting the sample into the empirical estimate is:

[0072]

[0073] Finally, similar to JPDAR, it is possible to minimize The optimization objective of FSS-SDAR is obtained.

[0074] As an optional implementation, the class conditional probability distance from the source domain to the target domain is Minimize as the goal, iteratively update the initial regression model for vital sign signal regression, including: setting the loss function to To update and iterate the initial regression model based on the neural network; express The output of the layer, Represents the membership matrix corresponding to the source domain label, Represents the membership matrix corresponding to the pseudo-label of the target domain, express The eigenspace of tensor products of rank 1.

[0075] Among them, such as Figure 3As shown in the figure, the Joint Adaptation Network (JAN) aligns the outputs of different network layers to approximately align the joint probability distribution of the source domain and the target domain. JAN assumes that the label differences between the source domain and the target domain mainly exist in the neural network layers close to the output, while the feature differences mainly exist in the neural network layers close to the input. Assuming that the network has layer, the output of the i-th layer is Z i ,but It can serve as an effective replacement for the joint probability distribution. Based on this assumption, JAN defines the Joint Maximum Mean Discrepancy Loss (JMMD), and achieves transfer by minimizing JMMD. In practical applications of JAN, it is necessary to obtain a one-hot vector of the class label. However, in regression problems, the labels are continuous, so JAN cannot be directly applied to regression problems. By establishing fuzzy sets and fuzzification, JMMD can be modeled in regression problems as follows:

[0076]

[0077] As an optional implementation, the class conditional probability distance from the source domain to the target domain is Minimize as the goal, iteratively update the initial regression model for vital sign signal regression, including: setting the loss function to To update and iterate the initial regression model based on the neural network; where φ is the neural network model.

[0078] Among them, such as Figure 4 As shown in Figure 1, the deep subdomain adaptation network achieves local alignment instead of global alignment by dividing the domain into several subdomains and calculating the loss on the subdomains. DSAN defines the Local Maximum Mean Discrepancy (LMMD) to measure the difference between subdomains. When the subdomains are divided by categories, minimizing LMMD can be regarded as the alignment of class conditional probabilities. Since class conditional probabilities are difficult to calculate in regression problems, DSANR cannot be used directly in regression problems. Define φ as a mapping to a reproducible Hilbert space. After fuzzification, LMMD can be defined in regression as:

[0079]

[0080]

[0081] For a The network of the layer l is recorded, and the deep network activations generated by the layer l are the source domain and the target domain respectively. and Then, the l-th layer LMMD can be reformulated as:

[0082]

[0083] Therefore, DSAN can be applied to regression problems.

[0084] Example 2

[0085] This embodiment provides a regression method for vital sign signals, such as Figure 5 As shown, it includes: collecting the vital sign signal of the current user, preprocessing it and inputting it into the vital sign signal regression model to obtain the regression result corresponding to the vital sign signal of the current user.

[0086] Example 3

[0087] This embodiment provides a vital sign signal processing system, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method for constructing a vital sign signal regression model or the method for regressing vital sign signals.

[0088] Example 4

[0089] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the steps of the method for constructing a vital sign signal regression model or the method for regressing vital sign signals.

[0090] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a vital sign signal regression model based on fuzzy sets, characterized in that: include: S1: Determine the target domain samples of vital sign signals, the source domain samples of vital sign signals and their corresponding regression-type source domain labels; All the source domain labels are spliced together to obtain the label domain, and n s Source domain labels Divide into multiple dividing points, use the dividing points to define C f fuzzy sets; S2: For each sample x k and its labels are mapped to C f Each fuzzy class is obtained from the fuzzy classes of the fuzzy sets f Membership Each fuzzy class Class f The membership degree μ f Combined to get the sample x k The membership vector of k is the serial number of the sample; if the sample x k If the sample x is a source domain sample, the label is the true label. k If it is a target domain sample, the label is a pseudo label; S3: Using formula Calculate the fuzzy class conditional probability distance from the source domain to the target domain Identify the source domain, Identify the target domain, n s is the number of source domain samples, n t is the number of samples in the target domain, is the i-th source domain sample x s,i The corresponding membership of the cth fuzzy class The normalized value of is the jth target domain sample x t,j The corresponding membership of the cth fuzzy class The normalized value of , dist is the distance representation; S4: Fuzzy class conditional probability distance from the source domain to the target domain Designing an optimization function or loss function with the minimum objective, and iteratively updating the initial regression model for vital sign signal regression; During the iterative update process, the initial regression model after iteration is used to generate pseudo labels for target domain samples in the next iteration and return to S2; until the initial regression model after iteration meets the preset conditions, it is considered that the vital sign signal regression model is obtained.

2. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 1, characterized in that: The n in the tag field s Source domain labels Dividing multiple dividing points includes: dividing the dividing points according to the percentage of the source domain label value; or dividing the dividing points according to the percentage of the interval length corresponding to the theoretical upper and lower bounds of the label in the vital sign signal problem.

3. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 1, characterized in that: The fuzzy class conditional probability distance from the source domain to the target domain The optimization function is designed as the minimum objective, and the initial regression model for vital sign signal regression is iteratively updated, including: using the fuzzy class conditional probability distance from the source domain to the target domain The fuzzy joint probability distance is obtained by summing the class edge probability, and the fuzzy joint probability distance is used as the optimization target. The optimization function is set as To update and iterate the initial regression model; in, is the joint probability distance matrix between the source domain and the target domain, λ is the degree of weighting coefficient between the intra-class distance and the inter-class distance, is the joint probability distance matrix between source domain and target domain classes, is the jth sample x in the target domain t,j Belong to The membership degree of fuzzy classes, A is the feature transformation, is the squared norm on a reproducible Hilbert space.

4. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 1, wherein: The fuzzy class conditional probability distance from the source domain to the target domain The optimization function is designed with the minimum as the goal, and the initial regression model for vital sign signal regression is iteratively updated, including: the fuzzy class conditional probability distance from the source domain to the target domain Minimum is the goal, and the optimization function is set to To update and iterate the initial regression model; Among them, C' f C f The selected value in the value range; λ is the degree of weighting coefficient between the intra-class distance and the inter-class distance; is the joint probability distance matrix of the fuzzy class conditional probability between the source domain and the target domain, is the fuzzy class conditional probability distance matrix between the source domain and the target domain classes, is the jth sample x in the target domain t,j Belong to The membership degree of fuzzy classes, A is the feature transformation, is the squared norm on a reproducible Hilbert space.

5. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 4, characterized in that: C' f To utilize The optimal value found; max fs is a hyperparameter used to limit the upper limit of the traversal range; Δ(y i ,C f )=max(μ i )-sec(μ i ); sec(μ i ) is μ i The second largest value in y i is the i-th source domain label.

6. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 1, characterized in that: The class conditional probability distance from the source domain to the target domain Based on this, we construct the regression joint maximum mean distance loss With the goal of minimizing the loss, the initial regression model for vital sign signal regression is iteratively updated, including: setting To update and iterate the initial regression model based on the neural network; wherein, Meaning 1: The output of the layer, Represents the membership matrix corresponding to the source domain label, Represents the membership matrix corresponding to the pseudo-label of the target domain, Represents an integer The eigenspace of tensor products of rank 1.

7. The method for constructing a vital sign signal regression model based on fuzzy sets according to claim 1, wherein: The fuzzy class conditional probability distance from the source domain to the target domain The optimization function is designed with the minimum as the goal, and the initial regression model for vital sign signal regression is iteratively updated, including: class conditional probability distance from the source domain to the target domain Based on the minimum, the fuzzy subdomain distance loss is constructed as the loss function To update the initial regression model based on the neural network iteratively; where φ is the neural network model, x i is the source domain sample, x j is the target domain sample, is the squared norm on a reproducible Hilbert space.

8. A regression method for a vital sign signal, comprising: The vital sign signal of the current user is collected, preprocessed and input into the vital sign signal regression model according to any one of claims 1 to 7 to obtain a regression result corresponding to the vital sign signal of the current user.

9. A vital sign signal processing system, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images