CN-122020244-A - Analysis method for noise suppression embedded migration component

Abstract

A method for analyzing an embedded migration component for noise suppression relates to the technical field of intelligent shelter body area networks and comprises the steps of firstly, adopting a migration component analysis algorithm to perform feature matching and dimension reduction to enable a source area to be in a source area With the target domain The edge distribution is as close as possible to obtain source domain data Tagged target domain data And unlabeled target domain data Secondly, carrying out sample screening on the Male distance of the source domain data on the aligned data, removing samples in the Male distance to obtain a final source domain data set And finally, using the final source domain data set As training set, input into a basic classifier Training to obtain final classification model Using classification models For the target domain unlabeled data obtained in the first stage And carrying out classified prediction and outputting a prediction label. According to the method, the anti-interference capability and the migration learning efficiency of the model in a dynamic environment are improved.

Inventors

• SHI HAN
• LIU YANG
• GAO WEI
• SONG BAOYAN
• Zu Wanni
• Cui Ziqian
• WANG BOWEN
• Cui Andi

Assignees

• 辽宁大学

Dates

Publication Date

20260512

Application Date

20260122

Claims (3)

1. 1. The method for analyzing the noise suppression embedded migration component is characterized by comprising the following steps of: the first stage is to adopt migration component analysis algorithm to perform feature matching and dimension reduction to enable the source domain With the target domain Minimizing the edge distribution distance between the two to obtain source domain data Tagged target domain data And unlabeled target domain data ; A second stage of screening samples of the Marshall distance of the source domain data by using the data aligned in the first stage, and removing the samples to obtain a final source domain data set ; Third stage, using the final source domain data set obtained by the second stage screening As training set, input into a basic classifier Training to obtain final classification model Using classification models For the target domain unlabeled data obtained in the first stage And carrying out classified prediction and outputting a prediction label.
2. 2. The method for analyzing a noise suppression embedded migration component of claim 1, wherein the specific steps in the first stage are as follows: step 1, data input and initialization Source domain data set The labeled dataset is represented as Wherein Is the first The eigenvectors of the individual source domain samples, Is the label to which it corresponds, Is the total number of source domain samples; target domain data set Comprising a small amount of tagged data And a large amount of unlabeled data Is expressed as Wherein Is a subset of the target domain that is tagged, Is a subset of the target domain that is unlabeled, And Respectively the number of samples, and ; Reducing dimensions of source domain and target domain by TCA algorithm and setting dimensions , Is a positive integer, and And weighting factors for use in subsequent noise screening stages , ; Step 2: gao Weihe mapping and MMD distance minimization It is provided that there is a feature map Mapping the data of the source domain and the target domain to a regeneration kernel Hilbert space, so that the data edge distribution of the two domains after mapping is as close as possible, namely ; TCA uses the maximum mean difference to measure the distribution distance of two domains in the mapped space, and the mathematical definition of MMD is: Wherein, the And The number of samples of the source domain and the target domain respectively, Representing norms in the hilbert space; by introducing a Gaussian kernel function Converting MMD into matrix And (3) carrying out solving: Step3, solving the optimal mapping matrix In order to minimize MMD distance and maintain variance of mapped data, solving an optimal solution of a linear objective function under the constraint of a semi-definite matrix, wherein the optimal objective function is as follows: Wherein, the Is a low-dimensional mapping matrix of the required solution, Is a regularization parameter for balancing MMD minimization and data variance retention; is a central matrix of the matrix, The data processing device is used for carrying out centering processing on the data; A unit matrix; obtaining a matrix by solving the Lagrange dual problem of the optimization problem And finally, the optimal mapping matrix From a matrix Front of (2) And the feature vector corresponding to the maximum feature value. Step 4, data projection and segmentation Mapping matrix obtained by solving Performing projection dimension reduction on all samples of the source domain and the target domain to obtain a projected data matrix And then the matrix is processed Dividing into projected source domain data according to attribution of original data Tagged target domain data And unlabeled target domain data To this end, data is obtained that is aligned in a distribution in the new feature space.
3. 3. The method for analyzing a noise suppression embedded migration component according to claim 1, wherein the specific steps in the second stage are as follows: Step 5, calculating statistics of each category of the target domain With projected tagged target domain data For reference, calculate each category inside Includes statistical information of the sample mean Sum covariance matrix ; For categories Its average value The calculation formula of (2) is as follows: Wherein, the Is that Belongs to category Is a sample count of (1); For categories Covariance matrix of it The calculation formula of (2) is as follows: The matrix reflects the category The degree of dispersion and correlation of internal sample features; Step 6, calculating the mahalanobis distance from the source domain sample to each category of the target domain Measuring post-projection source domain samples using mahalanobis distance Each category with the target domain For each sample in the source domain And labels therefor Calculate each category to the target domain Mahalanobis distance of (v) The calculation formula is as follows: innovative weighting factors are introduced For enhancing the processing power of the algorithm on noise in class, when the source field sample Is a label of (2) Class with target domain In the same time, the calculation mode of the mahalanobis distance is adjusted as follows: Wherein the weight factor The effect of this factor is to regulate the tolerance to intra-class deviations: When (when) When the distance between similar samples is reduced, the algorithm tends to keep samples with certain deviation from the similar mean value, and the method is suitable for scenes with sparse target domain data; When (when) When the distance between the similar samples is amplified, the algorithm tends to only keep samples which are very close to the similar mean value, is more sensitive to noise, and is suitable for scenes with higher target domain data quality; Step 7, sample decision and rejection For each source domain sample After computing its mahalanobis distance to all classes of the target domain, the following decision logic is executed: Step 7.1 find Target domain class with minimum mahalanobis distance ; Step 7.2 comparison Is a label of (2) Self-label ; Step 7.3, if the two are the same, the source domain sample is considered to be consistent with a certain category distribution of the target domain and is a high-quality sample to be reserved, if the two are different, the source domain sample is considered to be an abnormal point or noise sample and is not matched with any category of the target domain to be removed; Through this process, the source domain dataset Is filtered to obtain a final source domain data set 。

Description

Analysis method for noise suppression embedded migration component Technical Field The invention relates to the technical field of intelligent shelter body area networks, in particular to a noise suppression embedded migration component analysis method which is used for solving the destructive influence of noise samples on migration learning efficiency in a cross-domain cognitive task. Background With the development of intelligent automobile technology, the demand of users for personalized services is increasing, and intelligent cabins become the core for improving user experience. The intelligent cabin body area network (WBSN) is used as a sensor network for covering the body of a user, and a data basis is provided for cognitive services such as fatigue monitoring, body health parameter tracking, multi-user emotion recognition and the like of a driver by integrating an intelligent and low-power-consumption body area association sensor and a brake and combining a wireless communication technology and a multimedia technology. To achieve these complex cognitive services WBSN typically employ a fused decision model supported by a priori knowledge or experience. Under the current mainstream supervision or semi-supervision learning paradigm, one core premise is that independent co-distribution assumptions need to be satisfied between a priori database for training a model and actually acquired multi-modal test data. In a real application scenario, due to differences in physiological characteristics, activity habits and in-cabin sensing environments of different users, the body domain data of the target user always have field deviations of different degrees relative to a general prior database. The distribution inconsistency causes poor generalization effect of the existing fusion decision model on new user data, and the accuracy and effectiveness requirements of intelligent cabin local model learning are difficult to meet. In WBSN technology, a core challenge faced by cross-domain cognitive tasks is the distribution difference between the source domain and the target domain. In order to solve the problem that the data distribution inconsistency causes the failure of the traditional machine learning, the transfer learning is introduced into the field. Transfer learning aims to transfer knowledge of a source domain (e.g., a priori database) to a target domain (e.g., specific user data) to reduce domain differences in conventional transfer learning methods. Classical migration component analysis (TCA) for migration learning promotes feature migratability by maximizing edge distribution alignment, but suffers from the following drawbacks: 1. Destructive effects of noise samples. Unavoidable abnormal noise samples during physiological signal acquisition can severely distort the conditional distribution, resulting in systematic deviations of the classifier in the target domain. The existing method lacks an active screening mechanism for noise samples, so that the noise samples continuously reduce the knowledge migration efficiency. 2. The adaptability of the dynamic environment is insufficient. The intelligent cabin environment is highly dynamic and changes in user activity habits can lead to continuous changes in noise sample distribution. The existing method is difficult to adapt to the dynamic change, so that the performance of the model is obviously reduced along with the environmental change. 3. The limitation of edge distribution alignment. The conventional TCA method only focuses on edge distribution alignment, ignoring the importance of conditional distribution alignment. This makes the migration learning effect poor in the case where the difference in the condition distribution of the source domain and the target domain is large. 4. Sample weighted unity. The existing example weighting method can dynamically adjust the sample weight, but lacks an active screening mechanism for noise samples. This allows noise samples to still negatively impact model training during migration. Disclosure of Invention The invention provides a noise suppression embedded migration component analysis method, which aims to solve the problems in the prior art. According to the method, a noise suppression mechanism is embedded into a migration component analysis framework, so that the joint optimization of distribution alignment and noise suppression is realized, and the anti-interference capability and migration learning efficiency of the model in a dynamic environment are improved. In order to achieve the purpose, the technical scheme adopted by the invention is that the method for analyzing the noise suppression embedded migration assembly comprises the following steps: the first stage is to adopt migration component analysis algorithm to perform feature matching and dimension reduction to enable the source domain With the target domainMinimizing the edge distribution distance between the two to obtain source domain dataTagged ta