CN-121996934-A - Cross-user wearable activity recognition method based on group specific concept perception characterization learning

CN121996934ACN 121996934 ACN121996934 ACN 121996934ACN-121996934-A

Abstract

The invention discloses a cross-user wearable activity recognition method based on group specific concept perception characterization learning, which comprises the steps of collecting multi-user sensor data and preprocessing, measuring concept offset degrees from a time sequence view angle and a semantic view angle respectively, fusing multi-view angle measurement results to perform user clustering to generate a group specific concept label, constructing a perception model comprising an activity encoder, a user encoder and a plurality of classifiers, performing supervised learning by minimizing joint loss of activities, user and group specific concept classification, introducing a condition discriminator to construct a characterization pair of joint distribution and edge distribution, obtaining a trained model by minimizing condition mutual information of the activity characterization and the user characterization under the group specific concept through countertraining, and inputting test data into the trained model to perform activity recognition when the method is applied. The method eliminates the cross-user concept offset by explicitly modeling the group specific concepts and decoupling the activities and the user features, and remarkably improves the activity identification generalization capability of the model on new users.

Inventors

CHEN LING
HU RONG

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A method of identifying wearable activity across users based on group-specific concept-aware token learning, comprising the steps of: collecting wearable sensor time sequence data and activity labels thereof from a plurality of users, and preprocessing to obtain a data set with the user labels and activity categories; Based on the data set, measuring the concept offset degree among users from a time sequence view angle and a semantic view angle for each pair of users and each activity category respectively, and fusing multi-view measurement results to perform user clustering to generate a group specific concept label of each data sample; The method comprises the steps of constructing a group specific concept perception model, wherein the group specific concept perception model is used for taking a data set as input, and firstly extracting activity characterization and user characterization through an activity encoder and a user encoder respectively; the method comprises the steps of training, updating group specific concept perception model parameters by minimizing joint loss comprising activity classification, user classification and group specific concept detection, introducing a condition discriminator, and alternately updating the condition discriminator parameters and various encoder parameters by countermeasure training to obtain a trained group specific concept perception model; when the method is applied, wearable sensor data of a test user is taken as input, and activity category prediction is obtained through the trained group specific concept perception model, so that activity result recognition is realized.
2. The method for identifying wearable activities across users based on learning of group-specific concept perception characterizations according to claim 1, wherein the preprocessing results in a data set with user tags, comprising: Performing outlier processing and normalization on the collected wearable sensor time sequence data of a plurality of users to obtain initial time sequence data; Dividing the initial time sequence data through a sliding window, and labeling the corresponding activity label to obtain a data set with a user label and an activity category, wherein the time sequence data of each user corresponds to a domain.
3. The method for identifying wearable activities across users based on group-specific concept-aware token learning of claim 2, wherein for each pair of users and each activity category, the degree of concept offset between users is measured from a temporal perspective and a semantic perspective, respectively, comprising: Constructing an activity center sample based on the user and the activity category; Based on the activity center sample, measuring the concept offset degree between users in each activity category from a time sequence view angle, and calculating the similar inconsistent distance and the heterogeneous separable distance between the users by adopting a dynamic time alignment method to respectively obtain a similar distance matrix and a heterogeneous distance matrix; And simultaneously, measuring the concept offset degree between users under each activity category from a semantic view angle, generating a text prompt by using the physiological attribute of each user, obtaining activity characterization through a text encoder, and calculating semantic distances between every two users to obtain a semantic distance matrix.
4. The method for identifying wearable activities across users based on learning of group-specific conceptual perceptions characterization of claim 3, wherein the method for averaging center of gravity based on dynamic time warping is used to construct activity center samples based on each pair of users and each activity, comprising: taking all sample sets of a specific user and a specific activity class as input, and calculating a point-by-point average value of all samples on an original time axis to be used as an initial center sequence; Iteratively executing the following steps until convergence, namely aligning each sample with the current center sequence by using a dynamic time alignment method to obtain the corresponding relation between each time point in the sample and each time point of the center sequence; Taking the converged center sequence as an activity center sample.
5. The method for identifying wearable activities across users based on group-specific concept perception token learning according to claim 3, wherein the step of fusing multi-perspective measurement results to perform user clustering, generating a group-specific concept tag for each data sample, comprises: Under each activity category, converting the similar distance matrix, the heterogeneous distance matrix and the semantic distance matrix into corresponding similarity matrices; Carrying out similarity network fusion on three similarity matrixes under each activity category to obtain a fused similarity matrix, carrying out spectral clustering on the fused similarity matrix, automatically determining the number of user groups and obtaining user groups under the activity category; the user groups for all activity categories are collectively numbered, and a group-specific concept tag is generated for each data sample.
6. The method for identifying the wearable activity across users based on the group-specific concept perception characterization learning according to claim 1, wherein the activity encoder and the user encoder are composed of a plurality of layers of convolutional neural networks, each layer of convolutional neural network comprises a convolutional layer, a batch normalization layer and an activation function, the output layer of the encoder is an average pooling layer, and the convolutional neural networks do not share parameters; the activity encoder specifically takes activity sensor data in a data set as input, and carries out one-dimensional convolution on a time dimension through a convolution layer, a batch normalization layer and an activation function, so as to extract dynamic information changing along with time in an activity category; The user encoder specifically takes sensor data in a data set as input, performs one-dimensional convolution on a time dimension through a convolution layer, a batch normalization layer and an activation function, is used for extracting dynamic information changing along with time in a user, and obtains user characterization through an average pooling layer.
7. The group-specific concept-aware token learning-based cross-user wearable activity recognition method of claim 1, wherein the activity classifier, the group-specific concept detector, and the user classifier each comprise a full connectivity layer and an activation function; The activity classifier specifically takes activity characterization as input, outputs an activity class prediction value through a full-connection layer, and maps the activity class prediction value into activity class prediction probability through an activation function; the group specific concept detector specifically takes activity characterization as input, outputs a group specific concept prediction value through a full connection layer, and maps the group specific concept prediction value into group specific concept prediction probability through an activation function; the user classifier specifically takes user characterization as input, outputs a user prediction value through a full connection layer, and maps the user prediction value into user prediction probability through an activation function.
8. The method for identifying wearable activities across users based on group-specific concept perception characterization learning according to claim 1, wherein cross entropy loss is calculated based on activity class prediction probability output by an activity classifier and real activity labels to obtain activity class loss; Calculating cross entropy loss based on the user prediction probability output by the user classifier and the user label to obtain user classification loss; The cross entropy loss is calculated based on the group-specific concept prediction probability output by the group-specific concept detector and the generated group-specific concept tag, and the group-specific concept loss is obtained.
9. The method for identifying wearable activity across users based on group-specific concept-aware token learning of claim 1, wherein introducing a condition discriminant, updating condition discriminant parameters by challenge training, comprises: Constructing a characterization pair of joint distribution and edge distribution, which specifically comprises splicing the activity characterization of the same data sample with the user characterization under the same set of specific concepts to obtain a characterization pair from joint distribution, splicing the activity characterization with the randomly disturbed user characterization under the same set of specific concepts to obtain a characterization pair from edge distribution, and labeling a classification label; Inputting the characterization pairs and the corresponding unique heat codes of the group specific concepts into a condition discriminator to obtain discrimination probability; And calculating the discrimination loss according to the discrimination probability and the classification labels, and updating parameters of the condition discriminator with the minimum discrimination loss as a target.
10. The method for identifying wearable activity across users based on group-specific concept-aware token learning of claim 9, wherein introducing a condition arbiter updates various types of encoder parameters through countermeasure training, comprising: And fixing the condition discriminator parameters, and calculating the discrimination loss again to maximize the recalculated discrimination loss as the parameters of the target update active encoder and the user encoder.

Description

Cross-user wearable activity recognition method based on group specific concept perception characterization learning Technical Field The invention belongs to the field of wearable human activity recognition, and particularly relates to a cross-user wearable activity recognition method based on group specific concept perception and characterization learning. Background With the rapid development of wearable sensing technology and the increasing popularity of smart devices, wearable activity recognition (WHAR) has become one of the key technologies in the pervasive computing field. WHAR aims at identifying the activities of users through time sequence data acquired by the wearable sensor, and shows wide application prospects in the fields of health medical monitoring, sports fitness management, man-machine interaction, intelligent home and the like. Wearable activity recognition faces the key challenge of generalizing across users in that individual differences of users can lead to distribution shifts between activity data of different users, resulting in significant degradation of activity recognition performance of trained models on new users. A Domain Generalization (DG) method may be used to perform cross-user generalization of the activity recognition model. DG regards each user's data as a domain, aiming to learn generalizable models from data from multiple training users, thus achieving good activity recognition performance on test data from new users. The existing domain generalization method can be summarized into three major categories, wherein the generalization capability is improved by a data manipulation-based method through data enhancement and other modes, the generalization capability is improved by applying domain-invariant constraint to the activity labels by a characteristic learning-based method, and model generalization is promoted by a learning strategy-based method through strategies such as integrated learning, meta learning, self-supervision learning and the like. Methods based on token learning are the most widely used type of methods at present, and the methods eliminate individual differences by learning tokens which remain unchanged on all users, so that generalization across users is realized. However, the distribution offset between the activity data of different users can be divided into two types, including a change in the distribution of the input sensor data, i.e., a covariate offset, and a change in the potential mapping function from the sensor data to the activity category, i.e., a conceptual offset. The method based on the characterization learning is only applicable to the situation of covariate offset, but is not applicable to the situation of concept offset. Under the concept deviation, the discrimination rule of the activity category is changed, namely, the definition of various activities is changed, and the conventional method forcedly learns the uniform unchanged characterization for all users, which can prevent the model from capturing the changed definition of the category, thereby limiting the activity recognition capability of the model for different users. In the WHAR scenario, the conceptual offset may appear as a systematic variation between different user groups, rather than as a difference between individual users. For example, young people generally exhibit greater activity intensity than elderly people, so that the definition of "walking" and "jogging" activities of young users themselves and their classification boundaries both exhibit greater activity intensity than elderly users, forming similarities in discriminant rules between different users within the two groups, and differences in discriminant rules between the two groups. Furthermore, there may be differences in the user group structure under different activity categories. For example, young people may differ significantly from elderly people in dynamic activity "walking", but may behave similarly in static activity "lying down". In order to detect group-level concept offset, the degree of concept offset between users can be measured, and then clustering is performed to obtain a user group structure. However, the existing concept offset index can only measure the overall concept offset degree between different domains, and cannot measure the concept offset degree of each class, so that the existing concept offset index cannot be used for detecting the user group structure with differences under different activities. To capture class definitions that change under concept bias, some recent work is generalized by learning to distinguish data from different domains, and performing domain-aware token learning. However, capturing class definitions that vary under group-level conceptual offsets may allow the model to learn the differences between different user groups, but may not remove the personalized features of users within the group. The individuation features are only applicable to a single u