Noise suppression embedded migration component analysis method

Abstract

A noise suppression embedded migration component analysis method relates to the technical field of intelligent cabin body area networks, and comprises the following steps: firstly, adopting a migration component analysis algorithm to perform feature matching and dimension reduction, so that edge distribution between a source domain and a target domain is as close as possible, and source domain data, labeled target domain data and unlabeled target domain data are obtained; secondly, performing sample screening on source domain data mahalanobis distance according to the aligned data, and removing samples in the source domain data to obtain a final source domain data set; and finally, using the final source domain data set as a training set, and inputting the training set into a basic classifier for training to obtain a final classification model. And performing classification prediction on the target domain label-free data obtained in the first stage by using a classification model, and outputting a prediction label. According to the method, the anti-interference capability and the transfer learning efficiency of the model in a dynamic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent cockpit body area network technology, specifically to a noise suppression embedded transfer component analysis method, used to solve the destructive impact of noise samples on transfer learning efficiency in cross-domain cognitive tasks. Background Technology

[0002] With the development of intelligent vehicle technology, users' demand for personalized services is increasing, and intelligent cockpits have become the core of enhancing user experience. Intelligent cockpit body area networks (WBSNs), as a sensor network covering the user's body, integrate intelligent, low-power body-domain correlation sensors and brakes, combined with wireless communication and multimedia technologies, to provide a data foundation for cognitive services such as driver fatigue monitoring, tracking of physical health parameters, and multi-user emotion recognition. To achieve these complex cognitive services, WBSNs typically employ fusion decision models supported by prior knowledge or experience. In the current mainstream supervised or semi-supervised learning paradigms, a core premise is that the prior database used to train the model and the actual collected multimodal test data must satisfy the independent and identically distributed assumption.

[0003] In real-world applications, due to differences in physiological characteristics, activity habits, and in-cabin sensing environments among users, the target user's body domain data always exhibits varying degrees of domain bias relative to a general prior database. This distributional inconsistency leads to poor generalization performance of existing fusion decision models on new user data, making it difficult to meet the accuracy and effectiveness requirements of local model learning in smart cabins. In WBSN technology, the core challenge of cross-domain cognitive tasks lies in the distributional differences between the source and target domains. To address the failure of traditional machine learning caused by this data distribution inconsistency, transfer learning has been introduced into this field. Transfer learning aims to transfer knowledge from the source domain (such as a prior database) to the target domain (such as specific user data) to reduce domain differences. The classic transfer component analysis (TCA) method improves feature transferability by maximizing edge distribution alignment, but it has the following drawbacks:

[0004] 1. The destructive impact of noise samples. Abnormal noise samples, unavoidable during physiological signal acquisition, severely distort the conditional distribution, leading to systematic bias in the classifier's performance in the target domain. Existing methods lack proactive screening mechanisms for noise samples, causing them to continuously reduce knowledge transfer efficiency.

[0005] 2. Insufficient adaptability to dynamic environments. The intelligent cockpit environment is highly dynamic; changes in user activity habits lead to continuous variations in the distribution of noise samples. Existing methods struggle to adapt to these dynamic changes, resulting in a significant decline in model performance as the environment changes.

[0006] 3. Limitations of Edge Alignment. Traditional TCA methods only focus on edge alignment, neglecting the importance of conditional alignment. This results in poor transfer learning performance when there are significant differences in the conditional distributions of the source and target domains.

[0007] 4. Uniqueness of sample weighting. While existing instance weighting methods can dynamically adjust sample weights, they lack an active screening mechanism for noisy samples. This means that noisy samples can still negatively impact model training during the transfer process. Summary of the Invention

[0008] This invention proposes a noise-suppressed embedded transfer component analysis method to address the aforementioned problems in existing technologies. This method embeds a noise suppression mechanism into the transfer component analysis framework, achieving joint optimization of distribution alignment and noise suppression, thereby improving the model's robustness to interference and transfer learning efficiency in dynamic environments.

[0009] To achieve the above objectives, the technical solution adopted by this invention is: a noise suppression embedded migration component analysis method, the steps of which are as follows:

[0010] Phase 1: Employing the transfer component analysis algorithm for feature matching and dimensionality reduction, the source domain... With the target domain Minimize the edge distribution distance between them to obtain the source domain data. Labeled target domain data and unlabeled target domain data ;

[0011] The specific steps are as follows:

[0012] Step 1: Data Input and Initialization

[0013] Source domain dataset Labeled datasets are represented as ,in It is the first Feature vectors of source domain samples It is its corresponding tag. It is the total number of samples in the source domain;

[0014] Target domain dataset : Contains a small amount of labeled data and a large amount of unlabeled data The dataset is represented as ,in It is a labeled subset of the target domain. It is an unlabeled subset of the target domain. and These are their sample sizes, and ;

[0015] The TCA algorithm is used to reduce the dimensionality of the source and target domains, and the dimensions are set. , It is a positive integer, and and the weighting factors used in the subsequent noise screening stage. , ;

[0016] Step 2: High-dimensional kernel mapping and MMD distance minimization

[0017] Suppose there exists a feature mapping This maps the data from the source and target domains to a reproducing kernel Hilbert space, such that the marginal distributions of the data in the two domains are as close as possible after the mapping. ;

[0018] TCA uses the maximum mean difference to measure the distributional distance between two domains in the mapped space. The mathematical definition of MMD is:

[0019]

[0020] in, and These represent the number of samples in the source domain and the target domain, respectively. Denotes the norm in a Hilbert space;

[0021] By introducing a Gaussian kernel function Convert MMD to matrix Solve the following:

[0022]

[0023] Step 3: Solve for the optimal mapping matrix

[0024] To minimize the MMD distance and preserve the variance of the mapped data, we seek the optimal solution for the linear objective function under semidefinite matrix constraints. The optimized objective function is:

[0025]

[0026]

[0027] in, It is the low-dimensional mapping matrix that needs to be solved. It is a regularization parameter used to balance MMD minimization and data variance preservation; It is a central matrix. It is used for centralized data processing; Identity matrix.

[0028] By solving the Lagrange dual problem of this optimization problem, the matrix is ​​obtained. The solution, ultimately, is the optimal mapping matrix. From the matrix The former The eigenvectors are composed of the eigenvalues ​​corresponding to the largest eigenvalues;

[0029] Step 4: Data Projection and Segmentation

[0030] Using the obtained mapping matrix Projection dimensionality reduction is performed on all samples in the source and target domains to obtain the projected data matrix. ; this matrix Based on the original data's attribution, the data is segmented into projected source domain data. Labeled target domain data and unlabeled target domain data Thus, we obtained data that is distributed and aligned in the new feature space.

[0031] The second stage involves using the aligned data from the first stage to perform sample filtering based on the Mahalanobis distance of the source domain data, removing samples to obtain the final source domain dataset. ;

[0032] The specific steps are as follows:

[0033] Step 5: Calculate the statistics for each category in the target domain.

[0034] With projected labeled target domain data For reference, calculate the internal category. Statistical information, including sample mean Covariance Matrix ;

[0035] For category Its mean The calculation formula is:

[0036]

[0037] in, yes Belongs to the category The total number of samples;

[0038] For category Its covariance matrix The calculation formula is:

[0039]

[0040] The matrix reflects the category The dispersion and correlation of internal sample features;

[0041] Step 6: Calculate the Mahalanobis distance from the source domain samples to each category in the target domain.

[0042] Using Mahalanobis distance to measure the projected source domain samples With each category of the target domain The similarity for each sample in the source domain and its labels Calculate its value to each category in the target domain. Mahalanobis distance The calculation formula is as follows:

[0043]

[0044] Introducing innovative weighting factors This is used to enhance the algorithm's ability to handle intra-class noise, when source domain samples... tags With target domain category When they are the same, the calculation method for Mahalanobis distance is adjusted as follows:

[0045]

[0046] Among them, weighting factors This factor functions to adjust the tolerance for within-class bias:

[0047] when When the distance between samples of the same type is reduced, the algorithm tends to retain those samples that deviate from the mean of the same type, which is suitable for scenarios where the target domain data is sparse.

[0048] when When the distance between samples of the same type is magnified, the algorithm tends to retain only those samples that are very close to the mean of the same type, making it more sensitive to noise and suitable for scenarios with high-quality target domain data.

[0049] Step 7: Sample Decision and Elimination

[0050] For each source domain sample After calculating the Mahalanobis distances to all classes in the target domain, the following decision logic is executed:

[0051] Step 7.1 Find the match The target domain category with minimum Mahalanobis distance ;

[0052] Step 7.2 Comparison tags and Self-label ;

[0053] Step 7.3 If the two are the same, the source domain sample is considered to be consistent with the distribution of a certain category in the target domain and is a high-quality sample, so it is retained; if the two are different, the source domain sample is considered to be an outlier or a noise sample, which does not match any category in the target domain and is removed.

[0054] Through this process, the source domain dataset After filtering, the final source domain dataset is obtained. .

[0055] Phase 3: Using the final source domain dataset obtained after Phase 2 filtering. As the training set, it is input into a base classifier. The model is trained in this way to obtain the final classification model. Using classification models The unlabeled data of the target domain obtained in the first stage Perform classification prediction and output predicted labels.

[0056] The beneficial effects of this invention are as follows: the method effectively solves the problems of domain differences and noise interference, significantly improves robustness to noisy data, and can operate stably in complex environments. It is not only applicable to human activity recognition tasks but also to intelligent cockpit body domain cognition tasks. This invention can improve the performance of intelligent cockpit body domain cognition tasks, such as driver fatigue detection, health parameter tracking, and multi-user emotion recognition, meeting users' needs for personalized and precise services. Attached Figure Description

[0057] Figure 1 This is a flowchart of the algorithm of the present invention.

[0058] Figure 2 This is a schematic diagram of the pseudocode algorithm of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the implementation cases of the present invention, and not all of the implementation cases.

[0060] A noise suppression embedded migration component analysis method is disclosed. This method effectively solves the problems of domain differences and noise interference in intelligent cabin body area network data by combining migration component analysis (TCA) with a Mahalanobis distance-based noise sample suppression (WMD) strategy. The technical solution, operation steps, and parameter settings of the present invention will be described in detail below with reference to specific embodiments.

[0061] Example 1: Migration task from dataset 1 to dataset 2

[0062] 1. Data preparation and preprocessing

[0063] Source Domain Dataset: Dataset 1 is used as the source domain. This dataset contains data on 18 different physical activities collected from 9 subjects wearing 3 inertial measurement units (IMUs) and heart rate detectors. In this embodiment, four behavioral labels are selected: "sitting, lying down, walking, and running". Each sample contains 21 dimensions of raw data, including triaxial acceleration of the chest, arms, and ankles, angular velocities of the arms and ankles, and magnetic field strength.

[0064] Target Domain Dataset: Dataset 2 was used as the target domain. This dataset contains data on 12 physical activities collected by sensors worn by 10 volunteers on their chest, right wrist, and left ankle. Four behavioral labels were also selected: "sitting, lying down, walking, and running." Each sample contained the same 21-dimensional raw data as the source domain.

[0065] Data preprocessing: The raw data in the source and target domains are standardized to have a mean of 0 and a variance of 1. Subsequently, feature extraction is performed on each dimension of the data. The extracted features include time-domain features such as mean, variance, range, skewness, and kurtosis, as well as the top five highest peak values, their corresponding frequencies, and 11 autocorrelation coefficients, ultimately forming a 546-dimensional feature vector.

[0066] 2. Technical Solution and Parameter Settings

[0067] Phase 1: TCA Feature Matching and Dimensionality Reduction

[0068] Input source domain dataset and target domain dataset. Dimensionality reduction. Set to 30; regularization parameter Set to 0.1; the kernel function uses a Gaussian kernel, where the kernel width parameter... The value can be set to a multiple of the average distance between feature vectors through empirical or cross-validation methods. .

[0069] Construct the kernel matrix of the source and target domains. Constructing the MMD matrix Solving the optimization problem The optimal mapping matrix is ​​obtained. . use right and Perform projection dimensionality reduction to obtain the projected data. and and will Divided into and .

[0070] Phase 2: Suppression of WMD Noise Samples

[0071] Input the projected source domain data and labeled target domain data Weighting factors Setting it to 0.0 means that the distance between similar samples will not be adjusted.

[0072] Calculate the target domain with labeled data Each category sample mean Covariance Matrix .for Each sample Calculate its to Mahalanobis distance for all categories Find with The category with the smallest distance ,Compare tags and tags If the tags are the same, keep them. Otherwise, remove. .

[0073] Output the filtered, high-quality source domain dataset .

[0074] Phase 3: Model Training and Prediction

[0075] enter and .

[0076] use As the training set, a K-Nearest Neighbors (KNN) classifier is trained, where Using a trained KNN classifier to... Perform classification prediction to obtain the predicted labels for the unlabeled data in the target domain.

[0077] 3. Experimental Results and Analysis

[0078] In the transfer learning task from dataset 1 to dataset 2, the accuracy of the method of this invention reached 76.04%, significantly higher than the comparison algorithms TCA (65.21%), TJM (52.71%), JDA (68.33%), and BDA (74.58%), and also higher than KNN (72.92%) without transfer learning. This indicates that by embedding a noisy sample suppression strategy, the method of this invention can effectively improve the performance of the intelligent cabin body domain cognitive task.

[0079] Example 2: Migration task from dataset 2 to dataset 1

[0080] 1. Data preparation and preprocessing

[0081] Dataset 2 was used as the source domain. Dataset 1 was used as the target domain. Similar to Example 1, standardization and feature extraction were performed to obtain a 546-dimensional feature vector.

[0082] 2. Technical Solution and Parameter Settings

[0083] Phase 1: TCA Feature Matching and Dimensionality Reduction

[0084] Same as in Example 1, , Use the Gaussian kernel function.

[0085] Phase 2: Suppression of WMD Noise Samples

[0086] Weighting factors Setting it to 0.2 slightly increases the distance between similar samples, enhancing sensitivity to noise. The operation steps are the same as in Example 1.

[0087] Phase 3: Model Training and Prediction

[0088] The operating steps are the same as in Example 1.

[0089] 3. Experimental Results and Analysis

[0090] In the transfer learning task from dataset 2 to dataset 1, the accuracy of the method of this invention reached 67.64%, which is significantly higher than the comparison algorithms TCA (65.15%), TJM (62.71%), JDA (65.15%), and BDA (65.15%), and also higher than KNN (62.93%) without transfer learning. This further verifies the effectiveness and robustness of the method of this invention in different transfer directions.

[0091] Example 3: Parameter Sensitivity Analysis (Taking Dataset 1 to Dataset 2 as an example)

[0092] To verify the weighting factors A parameter sensitivity analysis was performed to assess the impact on the performance of the method of this invention. While keeping other parameters constant ( If the condition remains unchanged, Variations within a certain range. Experimental results show that:

[0093] when At that time, the number of noise samples removed decreased, and the presence of noise samples led to a decrease in accuracy of approximately 74%.

[0094] when At that time, the number of deleted noisy samples increased, but excessively large ones... It is possible to accidentally delete normal samples, which may lead to a decrease in accuracy of about 75%.

[0095] when exist Within the specified range, the method of the present invention achieves the highest accuracy, approximately 76.04%, and Variations within this range have little impact on accuracy.

Claims

1. A method for analyzing embedded migration components with noise suppression capabilities, characterized in that, The steps are as follows: Phase 1: Employing the transfer component analysis algorithm for feature matching and dimensionality reduction, the source domain... With the target domain Minimize the edge distribution distance between them to obtain the source domain data. Labeled target domain data and unlabeled target domain data ; The second stage involves using the aligned data from the first stage to perform sample filtering based on the Mahalanobis distance of the source domain data, removing samples to obtain the final source domain dataset. ; Phase 3: Using the final source domain dataset obtained after Phase 2 filtering. As the training set, it is input into a base classifier. The model is trained in this way to obtain the final classification model. Using classification models The unlabeled data of the target domain obtained in the first stage Perform classification prediction and output predicted labels.

2. The noise suppression embedded migration component analysis method according to claim 1, characterized in that, The specific steps in the first stage are as follows: Step 1: Data Input and Initialization Source domain dataset Labeled datasets are represented as ,in It is the first Feature vectors of source domain samples It is its corresponding tag. It is the total number of samples in the source domain; Target domain dataset : Contains a small amount of labeled data and a large amount of unlabeled data The dataset is represented as ,in It is a labeled subset of the target domain. It is an unlabeled subset of the target domain. and These are their sample sizes, and ; The TCA algorithm is used to reduce the dimensionality of the source and target domains, and the dimensions are set. , It is a positive integer, and and the weighting factors used in the subsequent noise screening stage. , ; Step 2: High-dimensional kernel mapping and MMD distance minimization Suppose there exists a feature mapping This maps the data from the source and target domains to a reproducing kernel Hilbert space, such that the marginal distributions of the data in the two domains are as close as possible after the mapping. ; TCA uses the maximum mean difference to measure the distributional distance between two domains in the mapped space. The mathematical definition of MMD is: in, and These represent the number of samples in the source domain and the target domain, respectively. Denotes the norm in a Hilbert space; By introducing a Gaussian kernel function Convert MMD to matrix Solve the following: Step 3: Solve for the optimal mapping matrix To minimize the MMD distance and preserve the variance of the mapped data, we seek the optimal solution for the linear objective function under semidefinite matrix constraints. The optimized objective function is: in, It is the low-dimensional mapping matrix that needs to be solved. It is a regularization parameter used to balance MMD minimization and data variance preservation; It is a central matrix. It is used for centralized data processing; identity matrix; By solving the Lagrange dual problem of this optimization problem, the matrix is ​​obtained. The solution, ultimately, is the optimal mapping matrix. From the matrix The former It consists of the eigenvectors corresponding to the largest eigenvalues. Step 4: Data Projection and Segmentation Using the obtained mapping matrix Projection dimensionality reduction is performed on all samples in the source and target domains to obtain the projected data matrix. ; this matrix Based on the original data's attribution, the data is segmented into projected source domain data. Labeled target domain data and unlabeled target domain data Thus, we obtained data that is distributed and aligned in the new feature space.

3. The noise suppression embedded migration component analysis method according to claim 1, characterized in that, The specific steps in the second stage are as follows: Step 5: Calculate the statistics for each category in the target domain. With projected labeled target domain data For reference, calculate the internal category. Statistical information, including sample mean Covariance Matrix ; For category Its mean The calculation formula is: in, yes Belongs to the category The total number of samples; For category Its covariance matrix The calculation formula is: The matrix reflects the category The dispersion and correlation of internal sample features; Step 6: Calculate the Mahalanobis distance from the source domain samples to each category in the target domain. Using Mahalanobis distance to measure the projected source domain samples With each category of the target domain The similarity for each sample in the source domain and its labels Calculate its value to each category in the target domain. Mahalanobis distance The calculation formula is as follows: Introducing innovative weighting factors This is used to enhance the algorithm's ability to handle intra-class noise, when source domain samples... tags With target domain category When they are the same, the calculation method for Mahalanobis distance is adjusted as follows: Among them, weighting factors This factor functions to adjust the tolerance for within-class bias: when When the distance between samples of the same type is reduced, the algorithm tends to retain those samples that deviate from the mean of the same type, which is suitable for scenarios where the target domain data is sparse. when When the distance between samples of the same type is magnified, the algorithm tends to retain only those samples that are very close to the mean of the same type, making it more sensitive to noise and suitable for scenarios with high-quality target domain data. Step 7: Sample Decision and Elimination For each source domain sample After calculating the Mahalanobis distances to all classes in the target domain, the following decision logic is executed: Step 7.1 Find the match The target domain category with minimum Mahalanobis distance ; Step 7.2 Comparison tags and Self-label ; Step 7.3 If the two are the same, the source domain sample is considered to be consistent with the distribution of a certain category in the target domain and is a high-quality sample, so it is retained; if the two are different, the source domain sample is considered to be an outlier or a noise sample, which does not match any category in the target domain and is removed. Through this process, the source domain dataset After filtering, the final source domain dataset is obtained. .

Images