Self-adaptive excitation method and system for maintaining healthy behaviors

Abstract

The invention relates to a self-adaptive excitation method for maintaining healthy behaviors. The method comprises the following steps: acquiring environmental data in a target area and physical signs and behavior data of a person; extracting behavior characteristics of the person according to the physiological indexes and the behavior data; constructing a short-term dynamic behavior and a long-term trend behavior according to the behavior characteristics; fusing the short-term dynamic behavior and the long-term trend behavior by using a cross attention mechanism and a gating fusion formula; establishing a health behavior thinking map; generating personalized incentive content; matching the character features with the excitation content through a multi-dimensional depth similarity algorithm, and generating excitation feedback in a text, voice or video form; and iteratively updating the excitation triggering condition and the content generation strategy according to the character behavior corresponding data. Through the technical scheme of the invention, not only can the probability of behavior change be improved, but also the interpretability of the excitation method can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent nursing, and in particular to a self-adaptive motivation method for maintaining healthy behavior, a system thereof, a readable storage medium and a computer device. Background Art

[0002] An individual's health behaviors significantly impact their physical health, and the two are highly correlated. Following healthy behaviors has a positive impact on disease prevention and treatment, playing a key role. However, the willingness and frequency of healthy behaviors decline with age. According to social cognitive theory, both personal and environmental factors can influence the maintenance of healthy behaviors or the development of unhealthy behaviors. Self-management theories posit that individuals should be motivated to achieve goals at a certain level. Nudge theory embraces the concept of liberal, mild authoritarianism. Nudges stimulate behavioral tendencies. Mild nudges involve minimal effort, and the methods used should minimize side effects to avoid excessive coercion. Authoritarian nudges involve managers using nudges to appropriately alter people's behavior, ensuring they behave in a desired manner and maintain healthy behaviors. However, existing intervention techniques face limitations in addressing when individuals should be motivated. Existing incentives can be inappropriate and fail to capture individual interest. With the development of Internet of Things technology, various behaviors of the elderly can be collected in real time. In-depth user modeling and more complex recommendation algorithms can better understand individual emotions and the needs behind the emotions. The recommended incentive methods are more personalized, and the accuracy and effectiveness of the recommendations will be greatly improved.

[0003] Existing incentive-related technical solutions based on large language models may learn information from a large amount of text data during training, which may contain biases and stereotypes. This may lead to model bias problems, in which the output of the model may reflect social or cultural biases and may lead to unfair or harmful suggestions or content. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0005] To this end, the purpose of the present invention is to provide an adaptive incentive method and device for maintaining healthy behavior, which can not only increase the probability of behavioral change, but also effectively improve the interpretability of the incentive method.

[0006] To achieve the above objectives, the technical solution of the first aspect of the present invention provides an adaptive motivation method for maintaining healthy behavior, comprising the following steps:

[0007] Obtain environmental data as well as physiological indicators and behavioral data of people in the target area;

[0008] extracting behavioral characteristics of the person based on the physiological indicators and the behavioral data;

[0009] According to the behavioral characteristics, Bi-LSTM+GRU is used to construct short-term dynamic behavior, and Transformer is used to construct long-term trend behavior through self-attention mechanism.

[0010] The short-term dynamic behavior and the long-term trend behavior are fused using a cross-attention mechanism and a gated fusion formula to obtain a behavior fusion feature;

[0011] Establish a health behavior mind map GoT = (V, E, D), where the vertex set V represents the behavior context elements, the edge set E defines the dependency, influence, or association relationship between elements; the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism;

[0012] Based on the health behavior mind map, personalized incentive content is generated by combining the health behavior incentive knowledge map and the large language model;

[0013] Use a multi-dimensional deep similarity algorithm to match character traits with motivational content and generate motivational feedback in the form of text, voice, or video.

[0014] Iteratively update incentive trigger conditions and content generation strategies based on corresponding data of character behavior.

[0015] In the above technical solution, preferably, the GRU state update formula is:

[0016] z t =σ(W z x t +U z h t-1 +b z );

[0017] r t =σ(W r x t +U r h t-1 +b r );

[0018]

[0019]

[0020] Among them, x t Input feature vector for the current time step t, representing the currently collected behavioral data; h t-1 is the previous time step t -The hidden state of 1 contains the memory of past information and updates the gate z t : To control the current candidate hidden state Compared with the previous hidden state h t-1 Reset gate r t : Determine how much previous hidden state information to forget for calculating candidate states; Candidate hidden state, calculated by combining the current input and partially retained historical state; h t : The final hidden state of the current time step is a weighted combination of the update gate and the candidate state; W z 、W r 、W h :Corresponding to input x t The weight matrix of U z 、U r 、U h : They correspond to the hidden state h t-1 The weight matrix of b z 、b r 、b h : is the bias vector; σ(·) is the Sigmoid activation function, the output value is in the interval [0,1]; tanh(·): is the hyperbolic tangent activation function, the output value is in the interval [-1,1]; ⊙: is the element-by-element multiplication operation;

[0021] Bi-LSTM state update formula: For the input behavior data sequence X = (x1, x2, ..., x T ), bidirectional LSTM calculates the forward hidden state and the backward hidden state

[0022] in, The hidden state from time 1 to T; Hidden state from time T to direction 1; [;]: vector concatenation operation to integrate bidirectional information.

[0023] In the above technical solution, preferably, the calculation formula of the self-attention mechanism is:

[0024] Among them, Q is the query matrix, which comes from the linear transformation of the input features; K is the key matrix, which is also obtained by the linear transformation of the input features; V is the value matrix, which is used for weighted summation; d k is the dimension of the key vector, which is used for scaling to prevent the dot product value from being too large; softmax(·) is a normalization function to ensure that the sum of the output probabilities is 1.

[0025] In the above technical solution, preferably, the calculation formula of the cross attention mechanism is:

[0026]

[0027] Among them, Q B The query matrix of the short-term dynamic behavior; K T 、V T are the key and value matrices of the long-term trend behavior; Q T The query matrix of the long-term trend behavior; K B 、V B are the key and value matrices of the short-term dynamic behavior respectively; d k is a scaling factor; the two formulas respectively realize the information fusion from the short-term dynamic behavior to the long-term trend behavior and from the long-term trend behavior to the short-term dynamic behavior;

[0028] Subsequently, the cascade fusion mechanism is used to perform gated fine-tuning on the above cross-attention fusion results:

[0029]

[0030] The calculation formula of cross attention fusion weight: H fusion =λh B +(1-λ)h T ; Among them, λ is the adaptive fusion weight, which takes the value in [0,1] and is obtained by softmax normalization and is used for global weighting; W Q 、W K is a linear transformation matrix that maps the original features to the same dimension, d k is the scaling factor, H fusion It is the intermediate feature representation after attention fusion;

[0031] The gated fusion calculation formula is:

[0032] g t =σ(W g h B +U g h T +b g );

[0033] H final =g t ⊙H fusion +(1-g t )⊙h B ;

[0034] Among them, H final is the adaptive fusion weight λ, ranging from [0,1]; h B and h T : represent the characteristics obtained from the short-term dynamic behavior and the long-term trend behavior respectively; gt : Gating coefficient, after the Sigmoid function, its value is in [0,1], which is used to dynamically adjust the weights of the two features; W g and U g : Linear transformation matrix, mapping the corresponding features to the same scale; b g : bias term; ⊙: element-by-element multiplication operation; H final : The final fused feature representation is used for subsequent prediction tasks.

[0035] In the above technical solution, preferably, extracting the behavioral characteristics of the person based on the physiological indicators and the behavioral data includes:

[0036] Construct a spatiotemporal graph G = (V, E, A), where each node v i Represents a sensor, edge weight w ij It represents the spatiotemporal correlation between sensors and is calculated as:

[0037] Among them, p i is the spatial position of sensor i, t i is the timestamp of sensor i; p j is the spatial position of sensor j, t j is the timestamp of sensor j, τ is the time attenuation coefficient, which is used to control the influence of time difference on correlation, and σ is the spatial attenuation coefficient, which is used to control the influence of spatial distance on correlation;

[0038] The value of τ: first calculate the time difference Δt of all sensors ij =|t i -t j The median of | Let the initial value Perform a small grid search in the range of {0.5τ0,τ0,2τ0} and select the optimal τ through cross-validation of downstream prediction performance;

[0039] The value of σ: First calculate the Euclidean distance between all sensors || p i -p j The median of || Let the initial value A small grid search is performed in the range of {0.5σ0, σ0, 2σ0}, and the optimal σ is selected through cross-validation of downstream prediction performance.

[0040] In the above technical solution, preferably, the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism, including:

[0041] The cross-layer dependency matrix D is modeled using a multi-layer attention mechanism, and the calculation formula is as follows:

[0042] Among them, h i 、h j denote the feature vectors of nodes i and j respectively, h is generated by the health behavior analysis framework; ‖ denotes the vector concatenation operation, which is used to enhance the information fusion capability; W d For a trainable weight matrix, the weights between each layer are learned independently to ensure that the dependencies between layers are dynamically adjusted; b d is the bias term, which provides additional flexibility; σ is the activation function, which is used to control the nonlinear transformation of weights;

[0043] The cross-layer dependency matrix D uses a multi-head attention mechanism to integrate cross-layer features. The specific calculation steps are as follows:

[0044] in: They are query matrix, key matrix, and value matrix respectively, h represents different attention heads; d k is a scaling factor used to keep the gradient stable; the final cross-layer dependency matrix D is obtained by weighted summation of the outputs of multiple attention heads: in, is the output transformation matrix and H is the number of attention heads.

[0045] In the above technical solution, preferably, based on the health behavior mind map, combined with the health behavior motivation knowledge map and the large language model, personalized motivation content is generated, including:

[0046] Using multi-dimensional deep similarity algorithm: Among them, the user behavior feature vector is a; the copy image feature vector is b, a i and b i are weighting coefficients, respectively, used to emphasize the importance of different attributes;

[0047] Generate motivational copy based on the multimodal tag system, the tag includes attribute tag L attribute =f attribute (A, D), behavior label L behavior =f behavior (C, T) and environmental label L environment =f environment (E, P); where L attribute is a label generated based on attributes, A is the attribute of the elderly, D is the feature data of the copy image, and f attribute is the attribute matching function; L behavior is the label generated based on the health behavior goal, C is the content feature of the copy image, T is the target behavior type, and f behavior is the behavior recognition function; L environment is a label generated based on environmental features, E is the environmental feature, P is the scene information of the copy image, and fenvironment Identify the function for the environment;

[0048] Tag optimization is performed based on historical data and effect similarity. The formula is:

[0049] Among them, S similarity is the similarity score, α is the adjustment parameter, d current and d history are the feature differences between the current copy image and the historical copy image, and exp is an exponential function;

[0050] Generate label weights: Among them, w i is the weight of label i, similarity i Score the relevance between label i and the text image, ∑ j The sum of the relevance scores for all tags;

[0051] Multi-dimensional label weight fusion: Among them, Final Weight is the final label weight, L i is the label value of label i.

[0052] The technical solution of the second aspect of the present invention provides an adaptive incentive system for maintaining healthy behavior, which is used to implement the steps of the adaptive incentive method for maintaining healthy behavior provided by any technical solution of the first aspect of the present invention, including:

[0053] A data acquisition module is configured to acquire environmental data within a target area and physiological indicators and behavioral data of a person through a non-contact multimodal sensor;

[0054] a feature extraction module configured to extract the behavioral features of the person based on the physiological indicators and the behavioral data;

[0055] A behavior model building module is configured to construct short-term dynamic behavior using Bi-LSTM+GRU and long-term trend behavior using Transformer through a self-attention mechanism according to the behavior characteristics;

[0056] a feature fusion module configured to fuse the short-term dynamic behavior and the long-term trend behavior using a cross-attention mechanism and a gated fusion formula to obtain a behavior fusion feature;

[0057] The behavioral mind map building module is configured to build a healthy behavioral mind map GoT = (V, E, D), where the vertex set V represents the behavioral context elements, the edge set E defines the dependency, influence, or association relationship between the elements; and the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism.

[0058] An incentive content generation module is configured to generate personalized incentive content based on the health behavior mind map, combined with the health behavior incentive knowledge map and the large language model;

[0059] An incentive feedback module is configured to match character features with incentive content using a multi-dimensional deep similarity algorithm to generate incentive feedback in the form of text, voice, or video;

[0060] The incentive content iteration module is configured to iteratively update the incentive trigger conditions and content generation strategies based on the corresponding data of the character's behavior.

[0061] The technical solution of the third aspect of the present invention provides a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the adaptive motivation method for maintaining healthy behavior provided by the technical solution of the first aspect.

[0062] The technical solution of the fourth aspect of the present invention provides a computer device, including a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to implement the steps of the adaptive motivation method for maintaining healthy behavior provided by the technical solution of the first aspect above.

[0063] Compared with the existing technology, the advantages of the adaptive incentive method and system for maintaining healthy behaviors provided by the present invention are: by obtaining indicators of the elderly's health-related behaviors and their environment in real time, and combining the elderly's static characteristics and historical behavioral characteristics, the context of the target behavior is generated, and the corresponding incentive content is matched to generate incentive videos, voices or texts to promote the elderly to change unhealthy behaviors and maintain healthy behaviors. The present invention aims to propose an incentive method for maintaining healthy behaviors based on a multimodal large model, providing appropriate incentives for various health behaviors in the daily lives of the elderly, continuing to monitor changes in the elderly's behavior after the incentives, and iteratively updating the triggering conditions of the incentives. The incentive method proposed by the present invention can not only increase the probability of behavioral changes, but also effectively improve the interpretability of the incentive method. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0065] Figure 1 A flowchart of an adaptive excitation method according to an embodiment of the present invention is shown;

[0066] Figure 2 The figure shows a structural block diagram of the adaptive excitation system involved in the embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0068] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0069] like Figure 1 As shown, according to an embodiment of the present invention, the adaptive motivation method for maintaining healthy behavior includes the following steps:

[0070] S1, obtains environmental data in the target area and physiological indicators and behavioral data of the person;

[0071] In this step, the above data is collected in real time through various sensors, such as high-precision millimeter-wave radar, door magnetic sensors, privacy cameras, and other contactless sensors. These sensors collect real-time behavioral parameters of the elderly (such as sedentary behavior, smoking behavior, eating behavior, social participation behavior, and physical activity). Environmental sensors also collect environmental indicators of the elderly's space, such as temperature and humidity. At the same time, this data can be used to obtain the elderly's human-computer interaction behavior and personal information (such as physiological indicators) through wearable devices, mobile apps, robots, etc.

[0072] S2, extracting behavioral characteristics of the person based on the physiological indicators and the behavioral data;

[0073] In this step, the feature extractor is used to extract six types of features from the three modalities, including text-audio common features, text-video common features, audio-video common features, text characteristic features, audio characteristic features, and video characteristic features, and construct cross-layer dependency relationships to analyze the dynamic impact of different levels of health behaviors.

[0074] Feature Classification

[0075]

[0076]

[0077] S3, based on the behavioral characteristics, uses Bi-LSTM+GRU to construct short-term dynamic behavior and uses Transformer to construct long-term trend behavior through self-attention mechanism;

[0078] In this step, a two-stream network combining LSTM and Transformer is used, divided into a behavior stream (BS) and a trend stream (TS), to model short-term dynamic behaviors and long-term trend behaviors, respectively. The bidirectional long short-term memory network (Bi-LSTM) is responsible for extracting and storing the time series features of behavioral data, and the gated recurrent unit (GRU) further optimizes the memory capacity of short-term behavioral patterns, capturing the short-term dynamic behavior patterns of the elderly, such as falls, smoking, and eating.

[0079] Short-term dynamic behavior modeling: Bi-LSTM+GRU is used to extract and store time series features to optimize the memory capacity of short-term behavior patterns; Transformer enhances the modeling ability of long-term behavior trends through self-attention mechanism and position encoding, enabling it to accurately identify sudden behaviors (such as falls) and long-term activity changes (such as decreased activity).

[0080] S4, using a cross-attention mechanism and a gated fusion formula to fuse the short-term dynamic behavior and the long-term trend behavior to obtain a behavior fusion feature;

[0081] In this step, the cross-attention mechanism is used to realize the information interaction between the behavior stream (BS) and the trend stream (TS); while the gated fusion mechanism is used to adaptively adjust the contribution of short-term behavior and long-term trends.

[0082] S5. Build a health behavior mind map GoT = (V, E, D), where the vertex set V represents the behavior context elements, the edge set E defines the dependency, influence, or association relationship between elements, and the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism.

[0083] In this step, the health behavior thought graph can be modeled as a tuple (G, T, E, R, D), where G represents the model reasoning process (i.e., all thought relationships in the health behavior context), T represents possible thought transformations, E is the evaluator function used to obtain thought scores, ε is an evaluator function used to obtain a large number of thoughts, and R is the ranking function used to select the most relevant thoughts. D represents the cross-layer dependency matrix, which models the dynamic interactions between health behaviors at different levels (individual, environmental, and social) to improve the ability to predict personalized health behaviors.

[0084] The reasoning process of health behavior is modeled as a multi-layer directed graph G = (V, E, D), where the vertex set V represents various health behavior context elements identified by the health behavior model, covering health behavior types, behavior subjects, and behavior situations (such as situational clues, emotional states, behavioral intentions, behavioral plans, etc.).

[0085] Edge set E: defines the dependency, influence, or association relationship between elements, reflecting the interaction between various contextual elements of health behavior, such as the matching between behavior type and behavior subject, the stimulation of situational cues on behavioral intention, and the influence of emotional state on behavior planning.

[0086] S6, based on the health behavior mind map, combining the health behavior motivation knowledge map and the large language model to generate personalized motivation content;

[0087] In this step, the output results of the mind map generation model and the six types of features about health behaviors extracted by the feature extractor are input into the incentive content learning module. Combined with the health behavior incentive knowledge graph + large language model, the incentive content corresponding to the current behavior characteristics is obtained.

[0088] S7, uses a multi-dimensional deep similarity algorithm to match character features with motivational content and generates motivational feedback in the form of text, voice, or video;

[0089] In this step, in the construction of the knowledge graph of healthy behavior incentives:

[0090] Obtain text, audio, and video materials on healthy behavior motivation and process them to obtain a knowledge graph of healthy behavior motivation. The specific steps are as follows:

[0091] Step 1: Preprocessing the health behavior motivation text material

[0092] Text cleaning and word segmentation

[0093] Text cleaning: Clean the original text to remove noise and irrelevant content (such as HTML tags, special characters, spaces, etc.).

[0094] Word segmentation: Divide the text into meaningful lexical units (such as nouns, verbs, etc.), which helps further analyze entities and relationships.

[0095] Part-of-speech tagging: Perform part-of-speech tagging (such as verbs, nouns, etc.) on the content after word segmentation to help extract behavioral entities and actions.

[0096] Extracting health behavior motivation knowledge entities and relationships: Use natural language processing (NLP) methods to extract health behavior-related entities and relationships. For example, named entity recognition (NER) is used to extract health behavior-related entities (such as "walking," "fitness," and "elderly"). The initial generated health behavior motivation knowledge graph contains entities and their relationships.

[0097] Step 2: Knowledge graph completion and optimization

[0098] Multimodal entity and relationship extraction: With the help of a large multimodal model, entities and relationships in the knowledge graph of healthy behavior incentives are automatically extracted. For each new text, audio, or video input, the model dynamically updates the knowledge graph. For example, if information related to "elderly people taking a morning walk" is extracted from a video, the system will add the entities "elderly" and "morning walk," as well as the relationship between them (such as "behavior").

[0099] Completing the initial knowledge graph: Use the entities and relationships output by the multimodal model to complete the initial behavior-motivated knowledge graph.

[0100] Step 3: Expert review and graph optimization

[0101] Experts in the health behavior field reviewed the completed health behavior incentive knowledge graph based on their expertise to ensure its accuracy, reliability, and practicality. They optimized the graph by annotating and modifying entities, relationships, and their weights. Based on expert feedback, a reward function was designed to evaluate and optimize the quality of the health behavior incentive knowledge graph. This feedback was further used to adjust relationship weights, entity connections, and other aspects of the graph to ensure its accuracy and rationality.

[0102] Healthy behavior incentive copywriting and personalized recommendations

[0103] 1. Incentive copywriting generation

[0104] (1) Text processing and sentiment analysis

[0105] Word segmentation and feature extraction: Segment the text content and extract text features.

[0106] Sentiment classification: Use large language models to classify text sentiment labels.

[0107] Emoji matching: Match corresponding emojis based on emotion labels.

[0108] (2) Copy image generation

[0109] Combination of pictures and text: Combine the sorted emoticons with text to generate copy.

[0110] Background template application: Apply the preset background image template to generate a preliminary copy image.

[0111] Multimodal annotation: Input a large multimodal model and automatically label each copy image.

[0112] S8, iteratively update the incentive trigger conditions and content generation strategies based on the corresponding data of character behavior.

[0113] In the above embodiment, preferably, the GRU state update formula is:

[0114] z t =σ(W z x t +U z h t-1 +b z );

[0115] r t =σ(W r x t +U r h t-1 +b r );

[0116]

[0117] Among them, x t Input feature vector for the current time step t, representing the currently collected behavioral data; h t-1 is the hidden state of the previous time step t-1, which contains the memory of past information and updates the gate z t : To control the current candidate hidden state Compared with the previous hidden state h t-1 Reset gate r t : Determine how much previous hidden state information to forget for calculating candidate states; Candidate hidden state, calculated by combining the current input and partially retained historical state; h t : The final hidden state of the current time step is a weighted combination of the update gate and the candidate state; W z 、W r 、W h :Corresponding to input x t The weight matrix of U z 、U r 、U h : They correspond to the hidden state h t-1 The weight matrix of b z 、b r 、b h : is the bias vector; σ(·) is the Sigmoid activation function, the output value is in the interval [0,1]; tanh(·): is the hyperbolic tangent activation function, the output value is in the interval [-1,1]; ⊙: is the element-by-element multiplication operation;

[0118] Bi-LSTM state update formula: For the input behavior data sequence X = (x1, x2, ..., x T ), bidirectional LSTM calculates the forward hidden state and the backward hidden state

[0119] in, The hidden state from time 1 to T; Hidden state from time T to direction 1; [;]: vector concatenation operation to integrate bidirectional information.

[0120] In the above embodiment, preferably, the calculation formula of the self-attention mechanism is:

[0121] Among them, Q is the query matrix, which comes from the linear transformation of the input features; K is the key matrix, which is also obtained by the linear transformation of the input features; V is the value matrix, which is used for weighted summation; d k is the dimension of the key vector, which is used for scaling to prevent the dot product value from being too large; softmax(·) is a normalization function to ensure that the sum of the output probabilities is 1.

[0122] In the above embodiment, preferably, the calculation formula of the cross attention mechanism is:

[0123]

[0124] Among them, Q B The query matrix of the short-term dynamic behavior; K T 、V T The key and value matrix of the long-term trend behavior; Q T The query matrix of the long-term trend behavior; K B 、V B The key and value matrix of the short-term dynamic behavior; d k is a scaling factor; the two formulas respectively realize the information fusion from the short-term dynamic behavior to the long-term trend behavior and from the long-term trend behavior to the short-term dynamic behavior;

[0125] Subsequently, the cascade fusion mechanism is used to perform gated fine-tuning on the above cross-attention fusion results:

[0126]

[0127] The calculation formula of cross attention fusion weight: H fusion =λh B +(1-λ)h T ; Among them, λ is the adaptive fusion weight, which takes the value in [0,1] and is obtained by softmax normalization and is used for global weighting; W Q 、W K is a linear transformation matrix that maps the original features to the same dimension, d k is the scaling factor, H fusion It is the intermediate feature representation after attention fusion;

[0128] The gated fusion calculation formula is:

[0129] g t =σ(W g h B +U g h T +b g );

[0130] H final =g t ⊙H fusion +(1-g t )⊙h B ;

[0131] Among them, H final is the adaptive fusion weight λ, ranging from [0,1]; h B and h T : represent the characteristics obtained from the short-term dynamic behavior and the long-term trend behavior respectively; g t : Gating coefficient, after the Sigmoid function, its value is in [0,1], which is used to dynamically adjust the weights of the two features; W g and U g : Linear transformation matrix, mapping the corresponding features to the same scale; b g : bias term; ⊙: element-by-element multiplication operation; H final : The final fused feature representation is used for subsequent prediction tasks.

[0132] In the above embodiment, preferably, extracting the behavioral characteristics of the person based on the physiological indicators and the behavioral data includes:

[0133] Construct a spatiotemporal graph G = (V, E, A), where each node represents a sensor and the edge weight is w ij It represents the spatiotemporal correlation between sensors and is calculated as:

[0134] Among them, p i is the spatial position of sensor i, t i is the timestamp of sensor i; p j is the spatial position of sensor j, t j is the timestamp of sensor j, τ is the time attenuation coefficient, which is used to control the influence of time difference on correlation, and σ is the spatial attenuation coefficient, which is used to control the influence of spatial distance on correlation;

[0135] The value of τ: first calculate the time difference Δt of all sensors ij =|t i -t j The median of | Let the initial value Perform a small grid search in the range of {0.5τ0,τ0,2τ0} and select the optimal τ through cross-validation of downstream prediction performance;

[0136] The value of σ: First calculate the Euclidean distance between all sensors || p i -p j The median of || Let the initial value A small grid search is performed in the range of {0.5σ0,σ0,2σ0], and the optimal σ is selected through cross-validation of downstream prediction performance.

[0137] During the calculation process, the mechanism first organizes data through a time window slicing strategy, splitting sensor data into multiple time windows at fixed time intervals and using positional encoding to preserve time information.

[0138] Example of edge weight calculation: Assume that the timestamps of sensors i and j are t i =10:00 and t j =10:05, its spatial position is p i =(1,1) and p j =(1,2), then the edge weight w ij The calculation is as follows:

[0139]

[0140] Assuming τ = 5 seconds and σ = 1 meter, we can substitute the above values ​​for the calculation:

[0141]

[0142] This shows that timestamps and spatial distances have a significant impact on the correlation between sensors. If the sensors are close in distance and close in time, their weight is greater, otherwise the weight is smaller.

[0143] Subsequently, a dynamic attention allocation mechanism is used to calculate the contribution of each sensor data to the target behavior (such as fall detection) and assign different weights during the fusion process. For detecting sudden events (such as falls), GAT uses a neighborhood information aggregation mechanism to improve the relevance of local signals and reduce false positives.

[0144] Furthermore, this mechanism supports adaptive modal weighting, dynamically adjusting the contribution of different modal data to the decision-making process based on sensor signal-to-noise ratio, data integrity, and historical performance. For example, at night when lighting is poor, the system automatically reduces the weight of privacy-destroying camera data and increases the weight of millimeter-wave radar to enhance detection reliability. This mechanism optimizes the model's adaptability to different behavioral scenarios and improves the accuracy and robustness of behavior recognition.

[0145] The specific calculation process is as follows:

[0146] Dynamic Attention Calculation:

[0147] 1. Derivation of GAT Attention Coefficient

[0148] Assume that the feature vectors of sensors i and j are X i =[1.0,0.5] and X j =[0.8,0.6], set training weight And the attention vector a=[0.5,0.7].

[0149] Calculate the attention score e between two nodes ij :

[0150] First, perform a weighted sum:

[0151] WX i =[0.2,0.3]·[1.0,0.5] T =[0.2·1.0+0.3·0.5]=[0.35];

[0152] WX j =[0.2,0.3]·[0.8,0.6] T =[0.2 0.8 + 0.3 0.6] = [0.34];

[0153] Then concatenate these two feature vectors: WX i ||WX j =[0.35,0.34]

[0154] Compute the dot product:

[0155]

[0156] Finally, apply LeakyReLU:

[0157] eij=LeakyReLU(0.427)=0.427;

[0158] Normalized attention coefficient:

[0159] Assume that the neighbor set of node i is N(i) = {j}, then

[0160]

[0161] This shows that j has a greater influence on i.

[0162] The aggregation derivation process is as follows:

[0163] Assume that the feature representation of sensor i at layer l is The feature representation of neighbor node j at layer l is And a ij =1, transformation matrix

[0164] Perform feature aggregation:

[0165]

[0166] If there is only one neighbor node j, substitute the data:

[0167]

[0168]

[0169] Finally, the activation function is applied:

[0170]

[0171] The calculation process of modal weight adaptive adjustment is as follows:

[0172] Weight calculation example: Assume we have three modal data: camera, millimeter-wave radar, and accelerometer. Calculate the confidence score Sm for each modality.

[0173] Sensor signal-to-noise ratio (SNR)

[0174] Camera: SNRcam = 0.85

[0175] Millimeter-wave radar: SNRradar = 0.95

[0176] Accelerometer: SNRacc = 0.75

[0177] Data Integrity:

[0178] Camera: integritycam=0.90

[0179] Millimeter-wave radar: integrity radar = 0.95

[0180] Accelerometer: integrity acc = 0.80

[0181] Historical performance:

[0182] Camera: Historical performance cam = 0.90

[0183] Millimeter-wave radar: Historical performance radar = 0.95

[0184] Accelerometer: Historical performance acc = 0.85

[0185] According to the above data: calculate the confidence Sm of each mode

[0186] S cam =λ1·0.85+λ2·0.90+λ3·0.92

[0187] S radar =λ1·0.95+λ2·0.95+λ3·0.88

[0188] S acc =λ1·0.75+λ2·0.80+λ3·0.85

[0189] Assume λ1 = 0.5, λ2 = 0.3, λ3 = 0.2, then:

[0190] S cam =0.5·0.85+0.3·0.90+0.2·0.92=0.879

[0191] S radar =0.5·0.95+0.3·0.95+0.2·0.88=0.936

[0192] S acc =0.5·0.75+0.3·0.80+0.2·0.85=0.785

[0193] Calculate the modal weight:

[0194] w cam =0.879 / (0.879+0.936+0.785)=0.338

[0195] w radar =0.936 / (0.879+0.936+0.785)=0.360

[0196] w acc =0.785 / (0.879+0.936+0.785)=0.302

[0197] The above weights are used to adjust the contribution of each modality in the final decision.

[0198] In the above embodiment, preferably, the cross-layer dependency matrix D is dynamically calculated by a multi-head attention mechanism, including:

[0199] Cross-layer dependency matrix D: Models the dynamic interactions of different levels of health behavior and optimizes personalized health behavior prediction. The cross-layer dependency matrix D is modeled using a multi-layer attention mechanism and is calculated as follows:

[0200] Among them, h i 、h jdenote the feature vectors of nodes i and j respectively, h is generated by the health behavior analysis framework; || denotes the vector concatenation operation, which is used to enhance the information fusion capability; W d For a trainable weight matrix, the weights between each layer are learned independently to ensure that the dependencies between layers are dynamically adjusted; b d is a bias term that provides additional flexibility; σ is an activation function that controls the nonlinear transformation of weights; through adaptive learning, the weight of D changes over time to adapt to the dynamic adjustment needs of users' health behaviors.

[0201] The cross-layer dependency matrix D uses a multi-head attention mechanism to integrate cross-layer features, which can dynamically adjust the information propagation between health behavior feature layers. The specific calculation steps are as follows:

[0202] in: They are query matrix, key matrix, and value matrix respectively, h represents different attention heads; d k is a scaling factor used to maintain gradient stability; the calculation results of multiple attention heads are finally integrated through the splicing operation to ensure effective information propagation between different health behavior feature layers. The final cross-layer dependency matrix D is obtained by the weighted sum of the outputs of multiple attention heads: in, is the output transformation matrix, and H is the number of attention heads. This computational approach ensures that D can dynamically adapt to changes in health behaviors and improves personalized prediction capabilities through cross-layer information propagation. Furthermore, compared to traditional single-layer association modeling methods, this approach can more accurately characterize the multi-level interactions of health behaviors, enhancing the effectiveness of personalized health interventions.

[0203] Through the connection of these edges, the mind transformation graph G forms a dynamic network that can reflect the decision-making and evolution process of health behavior. Each step or strategy of behavior transformation can be modeled as a function T(G,p θ ), where G = (V, E, D) is a behavioral network diagram reflecting the current health behavior status, p θ is the LLM used. T simulates the behavior transformation by modifying G. Therefore, the behavior network graph after transformation is G′=T(G,p θ )=(V′,E′), where V′=(V∪{V +})\{V -} and E′=(E∪{E +})\{E -}. V + and E + Represents the newly added health behavior state vertices and the relationships between them. These newly added elements together depict the characteristics of the new behavior and its evolution path;

[0204] To maximize the expressive power of GoT, health behavior experts can delete thoughts by specifying the corresponding vertices and edges to be deleted (V - and E - ). V - and E - It represents the old states and their relationships that may be eliminated or adjusted during the transformation process.

[0205] The purpose of scoring thinking is to gain a deeper understanding of whether the current solution has reached a sufficient standard. This scoring process is modeled as a universal function E(v,G,p θ ), where v represents the specific thinking content to be evaluated.

[0206] In order to determine which health behavior suggestions (thinking) are most important and effective, a function R(G,p θ ,h) to sort and select them. Here, G represents a network diagram consisting of various health behavior recommendations and their relationships. θ is a predictive model trained on data to evaluate the effectiveness and applicability of actions, while h specifies the top number of action suggestions we want to select from all recommended actions. Based on the score of each action suggestion (derived by the E function described above), the top h recommended actions with the highest scores are selected.

[0207] In the above technical solution, preferably, S6, based on the health behavior mind map, generates personalized incentive content in combination with the health behavior incentive knowledge map and the large language model, including:

[0208] Using multi-dimensional deep similarity algorithm: Among them, the user behavior feature vector is a; the copy image feature vector is b, a i and b i They are weighted coefficients used to emphasize the importance of different attributes. Based on the results of multi-dimensional similarity calculation, the copy image that best matches the user's characteristics is output;

[0209] User feedback optimization: To ensure continuous optimization of image and text recommendations, we've designed a feedback mechanism. Every time a user views or interacts with a piece of text, the system adjusts the recommendation strategy in real time based on the user's behavior (such as likes, comments, or shares), and updates the weighting coefficient a_i, gradually improving the accuracy of personalized recommendations.

[0210] In this step, personalized copy recommendations are made by calculating the similarity between the user's health behavior feature vector and the copy image feature vector. Traditional similarity calculation methods are mainly based on cosine similarity. However, we introduced Deep Multidimensional Similarity (DMS) to comprehensively compare the user's health behavior feature vector with multiple attributes of the copy image (such as emotional tone, health theme, background elements, etc.), thereby obtaining more accurate personalized matching results.

[0211] Generate motivational copy based on the multimodal tag system, the tag includes attribute tag L attribute =f attribute (A, D), behavior label L behavior =f behavior (C, T) and environmental label L environment =f environment (E, P); where L attribute is a label generated based on attributes, A is the attribute of the elderly, D is the feature data of the copy image, and f attribute is the attribute matching function; L behavior is the label generated based on the health behavior goal, C is the content feature of the copy image, T is the target behavior type, and f behavior is the behavior recognition function; L environment is a label generated based on environmental features, E is the environmental feature, P is the scene information of the copy image, and f environment Identify the function for the environment;

[0212] In this step, the image is labeled based on the personalized attributes of the elderly user. Based on each elderly user's personalized information, such as gender and medical history, the model automatically identifies the match between the image and the user's attributes and assigns labels such as "female" and "urinary incontinence" to the image. These labels help the multimodal model more accurately determine the applicability of the image content.

[0213] Each copy image corresponds to a specific health behavior goal, such as diet, sleep, physical activity, social participation, etc. The model derives the behavior goal by analyzing the details in the copy image (such as food location, menu design, etc.) and labels the copy image according to the target health behavior. For example, if the copy image involves eating behavior, the model may identify the layout difference between healthy and unhealthy foods and add labels to them, such as "eating behavior", "healthy diet" or "unhealthy diet choice".

[0214] The spatiotemporal characteristics of the environment (such as time of day, seasonal changes, weather conditions, temperature, and holidays) have a significant impact on the behavioral patterns and psychological states of the elderly. Multimodal models acquire and analyze environmental data in real time, generating corresponding environmental labels such as "morning," "spring," and "cloudy." This information provides context for subsequent analysis of users' behavioral and emotional states.

[0215] Tag optimization is performed based on historical data and effect similarity. The formula is:

[0216] Among them, S similarity is the similarity score, α is the adjustment parameter, d current and d history are the feature differences between the current copy image and the historical copy image, and exp is an exponential function;

[0217] In this step, to enhance the relevance and effectiveness of the image, the model compares the current image with historically effective images, identifies similarities between the two, and labels them based on their similarity levels. These similarity levels are divided into several tiers, with labels assigned to each image, from high to low, to further optimize the image's effectiveness in guiding users.

[0218] Generate label weights: Among them, w i is the weight of label i, similarity i Score the relevance between label i and the text image, ∑ j The sum of the relevance scores for all tags;

[0219] In this step, building on the label generation described above, the model calculates label weights based on the relevance of each label. The generation of label weights relies not only on the inherent connections between labels but also combines expert judgment with the self-learning optimization process of the multimodal model to ensure that the sum of the weights of each label is 1. Label weights are adjusted through multimodal model calculations, combining multidimensional data streams such as natural language processing, computer vision, and user behavior data.

[0220] Multi-dimensional label weight fusion: Among them, Final Weight is the final label weight, L i is the label value of label i;

[0221] In this step, the weight of each tag in multiple label dimensions is dynamically adjusted based on its relevance to the image and visuals, combining multimodal data streams (such as visual information, user history, and environmental conditions). Ultimately, through weight fusion, the model achieves a comprehensive score for the image and visuals, further improving personalized recommendations and enhancing the user experience.

[0222] Choice of incentive content format:

[0223] ① Obtaining incentive content and generating category labels: The system first obtains the incentive content to be pushed and analyzes and categorizes it using a pre-trained multimodal classification model. Each incentive content is assigned one or more category labels (such as "emotional support," "health tips," "daily reminders," etc.) based on its application scenario, form, and function. Through this process, the system determines the application scenario category label for each incentive content.

[0224] Calculation formula:

[0225] L categorv =f categorv (C, S);

[0226] Among them, L categorv is the category label of the incentive content, C is the feature of the incentive content, S is the application scenario data, and f categorv is the category identification function.

[0227] ②Generation of basic labels for incentive content (X1): Based on the generated category labels, the system conducts statistical analysis of historical data on the incentive content and calculates the basic labels X1 for each category label. These basic labels reflect the frequency of use and effect evaluation of a certain category of incentive forms in history.

[0228] Calculation formula:

[0229] X1=f history (L category , D history ); where X1 is the basic label, L category is the category label, D history is the data in the form of historical incentives, f history It is a historical data analysis function.

[0230] ③ Incentive content model label generation (X2): Based on the basic label X1, the system uses a machine learning model to perform modeling analysis to generate a model label X2 for the incentive content. This label is based on multi-dimensional features such as user historical behavior, environmental data, and emotional state, and represents the potential utility and adaptability of a certain incentive content.

[0231] Calculation formula:

[0232] X2=f model (X1, U, E); where X2 is the model label, X1 is the basic label, U is the user feature data (such as health status, emotional state, etc.), E is the environmental feature data (such as time, weather, etc.), f model Modeling functions for incentive forms.

[0233] ④ Incentive Content Preference Label Generation (X3): Based on the basic label X1 and model label X2, the system further leverages the elderly user's incentive content preference history and combines it with a personalized recommendation algorithm to predict the user's preference for a specific incentive form and generate a predicted label X3. This label reflects the user's current interest in and acceptance of a particular type of incentive content.

[0234] Calculation formula:

[0235] X3=f prediction (X1, X2, U history ); where X3 is the predicted label, X2 is the model label, X1 is the basic label, and U history is the user's historical preference data, f prediction is the preference prediction function.

[0236] ⑤ Calculation of Incentive Content Preference Rating: Based on the basic label X1, the model label X2, and the predicted label X3, the system calculates the user preference rating for each incentive form. This rating comprehensively considers the contribution of the three labels in a weighted manner to ultimately determine the appeal of each incentive content.

[0237] Calculation formula:

[0238] Preference c =b+a1X1+a2X2+a3X3; where b is a parameter, a1, a2, and a3 are the weights of the three tags, and the incentive form with the highest score is the one that is ultimately pushed to the user.

[0239] An example of the reasoning process for health behavior: Morning exercise behavior among the elderly

[0240] (1) Initial state modeling:

[0241] Assume that the elderly take morning walking as the initial behavior.

[0242] The vertex set of the initial graph G0 is: V0 = {behavior subject: elderly, situational clue: morning, behavior type: walking};

[0243] The edge set E0 is: E0 = {(morning, walking)},

[0244] The cross-layer dependency matrix D0 calculates the influence of walking behavior based on individual, environmental, and social factors.

[0245] (2) Behavior evolution (T transformation): Assume that the elderly gradually change their morning walk to morning Tai Chi exercise. At this time, the model performs a thinking transformation operation: V + ={Behavior Type: Tai Chi}, E + ={(Morning, Tai Chi)},V- ={Behavior Type: Walking},E - ={(morning, walking)}, forming a new behavior graph G', which reflects the changes in the health behaviors of the elderly.

[0246] (3) Behavioral scoring (E calculation): Score walking and Tai Chi behaviors: E(walking, G, p θ )=0.75,E(Tai Chi,G,p θ )=0.85

[0247] The scoring function is based on personalized data such as health status, exercise habits and other factors.

[0248] (4) Behavior ranking and intervention optimization (R calculation): Behaviors were ranked according to their scores, and Tai Chi with higher scores was selected as the final recommended behavior. If multiple behaviors had similar scores, a mixed approach was recommended. The final health intervention strategy was "Morning Tai Chi to improve physical fitness and mental state."

[0249] By inputting the output of the mind map generation model and the six types of features of the health behavior feature extractor into the incentive content learning module, combined with the health behavior incentive knowledge graph and the large language model (LLM), personalized health behavior incentive content is generated. For example:

[0250] For elderly people who are reluctant to exercise in the morning, the incentive content generated by the model is:

[0251] "Tai Chi in the morning light relaxes the body and mind, improves balance, and boosts energy. Join us and enjoy the morning tranquility!"

[0252] like Figure 2 As shown, according to another embodiment of the present invention, an adaptive motivation system 100 for maintaining healthy behavior is used to implement the steps of the adaptive motivation method for maintaining healthy behavior provided in any of the above embodiments of the present invention, including:

[0253] The data acquisition module 10 is configured to acquire environmental data in a target area and physiological indicators and behavioral data of a person through a non-contact multimodal sensor;

[0254] A feature extraction module 20 is configured to extract the behavioral features of the person based on the physiological indicators and the behavioral data;

[0255] A behavior model construction module 30 is configured to construct short-term dynamic behavior using Bi-LSTM+GRU and long-term trend behavior using Transformer through a self-attention mechanism according to the behavior characteristics;

[0256] A feature fusion module 40 is configured to fuse the short-term dynamic behavior and the long-term trend behavior using a cross-attention mechanism and a gated fusion formula to obtain a behavior fusion feature;

[0257] The behavior mind map construction module 50 is configured to establish a health behavior mind map GoT = (V, E, D), where the vertex set V represents the behavior context elements, the edge set E defines the dependency, influence or association relationship between the elements; and the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism.

[0258] The motivational content generation module 60 is configured to generate personalized motivational content based on the health behavior mind map, combined with the health behavior motivation knowledge map and the large language model;

[0259] The motivational feedback module 70 is configured to match character features with motivational content using a multi-dimensional deep similarity algorithm to generate motivational feedback in the form of text, voice, or video;

[0260] The incentive content iteration module 80 is configured to iteratively update incentive triggering conditions and content generation strategies using data corresponding to character behaviors.

[0261] Based on the above Figure 1 The method shown, accordingly, an embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the adaptive motivation method for maintaining healthy behavior in any of the above embodiments are implemented.

[0262] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, USB flash drive, mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of the present application.

[0263] Based on the above Figure 1 The method shown, and Figure 2 In the virtual device embodiment shown, in order to achieve the above-mentioned purpose, the embodiment of the present application also provides a computer device, including a storage medium and a processor; the storage medium is used to store computer programs; the processor is used to execute the computer program to implement the steps of the adaptive motivation method for maintaining healthy behavior in any of the above-mentioned embodiments.

[0264] Optionally, the computer device may further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a Wi-Fi module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a Wi-Fi interface), etc.

[0265] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0266] The storage medium may also include an operating system and a network communication module. An operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the execution of information processing programs and other software and / or programs. The network communication module facilitates communication between components within the storage medium, as well as with other hardware and software within the physical device.

[0267] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An adaptive motivational method for maintaining healthy behavior, characterized in that The following steps are involved: Obtain environmental data as well as physiological indicators and behavioral data of people in the target area; extracting behavioral characteristics of the person based on the physiological indicators and the behavioral data; According to the behavioral characteristics, Bi-LSTM+GRU is used to construct short-term dynamic behavior, and Transformer is used to construct long-term trend behavior through self-attention mechanism. The short-term dynamic behavior and the long-term trend behavior are fused using a cross-attention mechanism and a gated fusion formula to obtain a behavior fusion feature; Establish a health behavior mind map GoT = (V, E, D), where the vertex set V represents the behavior context elements, the edge set E defines the dependency, influence, or association relationship between elements; the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism; Based on the health behavior mind map, personalized incentive content is generated by combining the health behavior incentive knowledge map and the large language model; Use a multi-dimensional deep similarity algorithm to match character traits with motivational content and generate motivational feedback in the form of text, voice, or video. Iteratively update incentive trigger conditions and content generation strategies based on corresponding data of character behavior.

2. The adaptive motivation method for maintaining healthy behavior according to claim 1, characterized in that: GRU state update formula: z t =σ(W z x t +U z h t-1 +b z ); r t =σ(W r x t +U r h t-1 +b r ); Among them, x t Input feature vector for the current time step t, representing the currently collected behavioral data; h t-1 is the hidden state of the previous time step t-1, which contains the memory of past information and updates the gate z t : To control the current candidate hidden state Compared with the previous hidden state h t-1 Reset gate r t : Determine how much previous hidden state information to forget for calculating candidate states; Candidate hidden state, calculated by combining the current input and partially retained historical state; h t : The final hidden state of the current time step is a weighted combination of the update gate and the candidate state; W z 、W r 、W h :Corresponding to input x t The weight matrix of U z 、U r 、U h : They correspond to the hidden state h t-1 The weight matrix of b z 、b r 、b h : is the bias vector; σ(·) is the Sigmoid activation function, the output value is in the interval [0,1]; tanh(·): is the hyperbolic tangent activation function, the output value is in the interval [-1,1]; ⊙: is the element-by-element multiplication operation; Bi-LSTM state update formula: For the input behavior data sequence X = (x1, x2, ..., x T ), bidirectional LSTM calculates the forward hidden state and the backward hidden state in, The hidden state from time 1 to T; Hidden state from time T to direction 1; [;]: vector concatenation operation to integrate bidirectional information.

3. The adaptive motivation method for maintaining healthy behavior according to claim 2, characterized in that: The calculation formula of the self-attention mechanism is: Among them, Q is the query matrix, which comes from the linear transformation of the input features; K is the key matrix, which is also obtained by the linear transformation of the input features; V is the value matrix, which is used for weighted summation; d k is the dimension of the key vector, which is used for scaling to prevent the dot product value from being too large; softmax(·) is a normalization function to ensure that the sum of the output probabilities is 1.

4. The adaptive motivation method for maintaining healthy behavior according to claim 3, characterized in that: The calculation formula of the cross attention mechanism is: Among them, Q B is the query matrix of the short-term dynamic behavior; K T 、V T are the key and value matrices of the long-term trend behavior; Q T K is the query matrix of the long-term trend behavior; B 、V B are the key and value matrices of the short-term dynamic behavior respectively; d k is a scaling factor; the two formulas respectively realize the information fusion from the short-term dynamic behavior to the long-term trend behavior and from the long-term trend behavior to the short-term dynamic behavior; The cascade fusion mechanism is used to perform gated fine-tuning on the above cross-attention fusion results: The calculation formula of cross attention fusion weight: H fusion =λh B +(1-λ)h T ; Among them, λ is the adaptive fusion weight, which takes the value in [0,1] and is obtained by softmax normalization and is used for global weighting; W Q 、W K is a linear transformation matrix that maps the original features to the same dimension, d k is the scaling factor, H fusion It is the intermediate feature representation after attention fusion; The gated fusion calculation formula is: g t =σ(W g h B +U g h T +b g ); H final =g t ⊙H fusion +(1-g t )⊙h B ; Among them, H final is the adaptive fusion weight λ, ranging from [0,1]; h B and h T : represent the characteristics obtained from the short-term dynamic behavior and the long-term trend behavior respectively; g t : Gating coefficient, after the Sigmoid function, its value is in [0,1], which is used to dynamically adjust the weights of the two features; W g and U g : Linear transformation matrix, mapping the corresponding features to the same scale; b g : bias term; ⊙: element-by-element multiplication operation; H final : The final fused feature representation is used for subsequent prediction tasks.

5. The adaptive motivation method for maintaining healthy behavior according to claim 4, characterized in that: Extracting the behavioral characteristics of the person according to the physiological indicators and the behavioral data includes: Construct a spatiotemporal graph G = (V, E, A), where each node v i Represents a sensor, edge weight w ij It represents the spatiotemporal correlation between sensors and is calculated as: Among them, p i is the spatial position of sensor i, t i is the timestamp of sensor i; p j is the spatial position of sensor j, t j is the timestamp of sensor j, τ is the time attenuation coefficient, which is used to control the influence of time difference on correlation, and σ is the spatial attenuation coefficient, which is used to control the influence of spatial distance on correlation; The value of τ: first calculate the time difference Δt of all sensors ij =|t i -t j The median of | Let the initial value Perform a small grid search in the range of {0.5τ0,τ0,2τ0} and select the optimal τ through cross-validation of downstream prediction performance; The value of σ: First calculate the Euclidean distance between all sensors || p i -p j The median of || Let the initial value A small grid search is performed in the range of {0.5σ0, σ0, 2σ0}, and the optimal σ is selected through cross-validation of downstream prediction performance.

6. The adaptive motivation method for maintaining healthy behavior according to claim 5, characterized in that: The cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism, including: The cross-layer dependency matrix D is modeled using a multi-layer attention mechanism, and the calculation formula is as follows: Among them, h i , h j denote the feature vectors of nodes i and j respectively, h is generated by the health behavior analysis framework; ‖ denotes the vector concatenation operation, which is used to enhance the information fusion capability; W d For a trainable weight matrix, the weights between each layer are learned independently to ensure that the dependencies between layers are dynamically adjusted; b d is the bias term, which provides additional flexibility; σ is the activation function, which is used to control the nonlinear transformation of weights; The cross-layer dependency matrix D uses a multi-head attention mechanism to integrate cross-layer features. The specific calculation steps are as follows: in: They are query matrix, key matrix, and value matrix respectively, h represents different attention heads; d k is a scaling factor used to keep the gradient stable; the final cross-layer dependency matrix D is obtained by weighted summation of the outputs of multiple attention heads: in, is the output transformation matrix and H is the number of attention heads.

7. The adaptive motivation method for maintaining healthy behavior according to claim 6, characterized in that: Based on the health behavior mind map, personalized incentive content is generated by combining the health behavior incentive knowledge map and the large language model, including: Using multi-dimensional deep similarity algorithm: Among them, the user behavior feature vector is a; the copy image feature vector is b, a i and b i are weighting coefficients, respectively, used to emphasize the importance of different attributes; Generate motivational copy based on the multimodal tag system, the tag includes attribute tag L attribute =f attribute (A, D), behavior label L behavior =f behavior (C, T) and environmental label L environment =f environment (E, P); where L attribute is a label generated based on attributes, A is the attribute of the elderly, D is the feature data of the copy image, and f attribute is the attribute matching function; L behavior is the label generated based on the health behavior goal, C is the content feature of the copy image, T is the target behavior type, and f behavior is the behavior recognition function; L environment is a label generated based on environmental features, E is the environmental feature, P is the scene information of the copy image, and f environment Identify the function for the environment; Tag optimization is performed based on historical data and effect similarity. The formula is: Among them, S similarity is the similarity score, α is the adjustment parameter, d current and d history are the feature differences between the current copy image and the historical copy image, and exp is an exponential function; Generate label weights: Among them, w i is the weight of label i, similarity i Score the relevance between label i and the text image, ∑ j The sum of the relevance scores for all tags; Multi-dimensional label weight fusion: Among them, Final Weight is the final label weight, L i is the label value of label i.

8. An adaptive motivation system for maintaining healthy behavior, used to implement the steps of the adaptive motivation method for maintaining healthy behavior as claimed in any one of claims 1 to 7, characterized in that: include: A data acquisition module is configured to acquire environmental data within a target area and physiological indicators and behavioral data of a person through a non-contact multimodal sensor; a feature extraction module configured to extract the behavioral features of the person based on the physiological indicators and the behavioral data; A behavior model building module is configured to construct short-term dynamic behavior using Bi-LSTM+GRU and long-term trend behavior using Transformer through a self-attention mechanism according to the behavior characteristics; a feature fusion module configured to fuse the short-term dynamic behavior and the long-term trend behavior using a cross-attention mechanism and a gated fusion formula to obtain a behavior fusion feature; The behavioral mind map building module is configured to build a healthy behavioral mind map GoT = (V, E, D), where the vertex set V represents the behavioral context elements, the edge set E defines the dependency, influence, or association relationship between the elements; and the cross-layer dependency matrix D is dynamically calculated through a multi-head attention mechanism. An incentive content generation module is configured to generate personalized incentive content based on the health behavior mind map, combined with the health behavior incentive knowledge map and the large language model; An incentive feedback module is configured to match character features with incentive content using a multi-dimensional deep similarity algorithm to generate incentive feedback in the form of text, voice, or video; The incentive content iteration module is configured to iteratively update the incentive trigger conditions and content generation strategies based on the corresponding data of the character's behavior.

9. A readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps of the adaptive motivation method for maintaining healthy behavior as claimed in any one of claims 1 to 7 are implemented.

10. A computer device, characterized in that: The method comprises a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to implement the steps of the adaptive motivation method for maintaining healthy behavior as described in any one of claims 1 to 7.

Images