A system and method for modeling the psychological and emotional attribution and autonomous intervention of teenagers based on the relationship network between agents and families

Abstract

The application discloses a kind of based on agent and family relationship network modeling adolescent psychological emotion attribution and autonomous intervention system and method, comprising: the interactive data of each object is collected in real time, and dynamic weighted graph is constructed;Topological structure quantitative analysis is executed to dynamic weighted graph;When detecting negative emotion or network anomaly, edge ablation, node ablation and substructure ablation operation are executed to reference graph, emotion prediction change quantity of each ablation operation is calculated using graph neural network emotion prediction model, and the structure root contribution degree of each network structure element is output;Agent is embedded as new node in dynamic weighted graph, selects optimal strategy, and actively intervenes in the interactive process between each object to implement targeted topological repair;And through the comparison of network structure and emotional state before and after intervention, closed-loop optimization is formed.The application promotes emotion attribution from individual event chain to relationship network topological level, and realizes the technical leap from passive response to agent autonomous intervention.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and mental health monitoring, and in particular to a system and method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling. Background Technology

[0002] In recent years, artificial intelligence technology has rapidly penetrated into home settings, and intelligent agents are gradually evolving from passively responding to user commands into intelligent home assistants with autonomous perception, decision-making, and action capabilities. At the same time, the mental health of adolescents has received increasing social attention, and how to leverage intelligent agent technology to continuously monitor and effectively intervene in their mental state has become a research hotspot in the fields of affective computing and human-computer interaction. However, existing technologies still face many bottlenecks in practical applications.

[0003] Current family intelligence systems mostly focus on recognizing individual emotions of single members, analyzing adolescents' psychological states as isolated events, and lack the ability to perceive the overall family relationship structure. However, family systems theory suggests that adolescents' emotional problems are often not isolated events at the individual level, but rather a reflection of the implicit relationship structure formed by long-term interactions among family members—structural imbalances such as the marginalization of a member, the formation of exclusive alliances between subsystems, and the unidirectional accumulation of emotions along specific paths are the deeper causes. Existing technologies simplify family relationships to single indicators such as intimacy scores or conflict frequency, losing the topological information in the relationship structure and failing to answer key questions such as "What does this family's relationship network look like?" and "Are emotions propagating along what paths?" At the same time, existing causal reasoning methods are limited to tracing linear event chains of "preceding events → subsequent emotions," attributing the cause to "anxiety caused by a certain criticism," failing to attribute emotional problems to persistent structural defects in the family network. Furthermore, existing intelligent agents exist outside the family as observers or tools, and have not yet embedded themselves in the family relationship network as endogenous nodes to proactively adjust family dynamics by changing the network topology.

[0004] Based on this, the present invention proposes a system and method for attribution of adolescent psychological emotions and self-intervention based on modeling of intelligent agents and family relationship networks. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention proposes a system and method for attribution and autonomous intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling. It utilizes dynamic weighted directed graphs to model family relationships in multiple dimensions, characterizing the interaction features between members from four dimensions: emotional valence, interaction intensity, physiological synchronicity, and power relations. Combined with a streaming update mechanism, it realizes real-time tracking of the evolution trajectory of the family relationship network, thus optimizing the perception and expression process of relationship structure.

[0006] On the one hand, to achieve the above objectives, this invention provides a system for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling, comprising: A multimodal data acquisition module is deployed in the target scene to collect multimodal interaction data of each object in real time. The multimodal interaction data includes at least visual perception data, voice perception data and physiological perception data. The dynamic relationship network construction module is used to extract the relationship features between objects based on the multimodal interaction data, construct a dynamic weighted graph as the mathematical representation of the relationship network with each object as a node and the relationship features as directed edge weights, and continuously update the dynamic weighted graph. The network topology analysis module is used to perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies. The network ablation attribution module is used to perform edge ablation, node ablation, and substructure ablation operations on the current network snapshot as the reference graph when negative emotions or network structure anomalies are detected. It uses a pre-trained graph neural network emotion prediction model to calculate the change in emotion prediction of each ablation operation relative to the reference graph, and outputs the contribution of each network structure element to the structural root cause of negative emotions based on the change in emotion prediction. The intelligent agent intervention module is used to embed the intelligent agent as a new node into the dynamic weighted graph, determine the intervention target based on the structural root contribution and match candidate intervention strategies, transform each candidate intervention strategy into a predictive topological modification of the dynamic weighted directed graph and input it into the graph neural network emotion prediction model for simulation evaluation, select the candidate intervention strategy with the best expected improvement, and have the intelligent agent actively intervene in the interaction process between objects to implement targeted topological repair. The closed-loop feedback optimization module is used to reconstruct the network snapshot after the intervention and after a preset observation period. It compares the actual changes in network structure indicators and adolescent emotional state before and after the intervention with the expected improvement in the pre-simulated range, determines the effectiveness of the intervention, updates the strategy confidence score in the intervention strategy knowledge base, and adjusts the behavioral parameters of the agent to form a closed-loop optimization.

[0007] On the other hand, to achieve the above objectives, the present invention also provides a method for attribution and self-intervention of adolescent psychological emotions based on modeling of intelligent agents and family relationship networks, including: The system acquires multimodal interaction data of each object collected in real time by a multimodal perception unit deployed in the target scene. The multimodal interaction data includes at least visual perception data, speech perception data, and physiological perception data. Based on the multimodal interaction data, the relationship features between each object are extracted. Using each object as a node and the relationship features as directed edge weights, a dynamic weighted graph is constructed as a mathematical representation of the relationship network, and the dynamic weighted graph is continuously updated. Perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies; When negative emotions or network structure anomalies are detected, the current network snapshot is used as the baseline graph. Edge ablation, node ablation, and substructure ablation operations are performed on the baseline graph to construct corresponding ablation graphs. The baseline graph and each ablation graph are input into a pre-trained graph neural network emotion prediction model to obtain the corresponding emotion prediction value. The change in emotion prediction relative to the baseline graph is calculated for each ablation operation. The contribution of each network structure element to the structural root of negative emotions is output according to the magnitude of the change in emotion prediction. The agent is embedded as a new node in the dynamic weighted graph. The intervention target is determined according to the structural root contribution and the candidate intervention strategy is matched. Each candidate intervention strategy is transformed into a predictive topological modification of the dynamic weighted directed graph and input into the graph neural network emotion prediction model for simulation evaluation. The candidate intervention strategy with the best expected improvement is selected. The agent acts as the execution subject to actively intervene in the interaction process between objects and implement targeted topological repair. After the intervention is implemented, a pre-set observation period is followed by the reconstruction of the network snapshot after the intervention. The actual changes in network structure indicators and adolescent emotional states before and after the intervention are compared with the expected improvement range simulated before the intervention. The effectiveness of the intervention is determined, the strategy confidence score in the intervention strategy knowledge base is updated, and the behavioral parameters of the agent are adjusted to form a closed-loop optimization.

[0008] Preferably, the relationship features between the objects are extracted, including: The multimodal interaction data is sliced ​​using fixed time windows, and within each time window: Facial expression features are extracted from the visual perception data of each object, and mapped to the emotional valence score by the pre-trained emotion recognition model. The mean within the window is taken as the emotional valence feature. Based on the spatial location data and voice activity detection results of each object, the ratio of co-existence time and the ratio of dialogue time are integrated as the interaction intensity feature. Based on the physiological perception data of each subject, the Pearson correlation coefficient of the heart rate sequences of the two subjects was calculated, and normalized after Fisher Z transformation as a physiological synchronicity feature. Based on the speech perception data of each object, the speech duration ratio, interruption frequency ratio and volume average ratio are calculated and normalized after dimensionality reduction as power relationship features. The emotional valence feature, interaction intensity feature, physiological synchronicity feature and power relationship feature constitute a four-dimensional relationship feature vector.

[0009] Preferably, the dynamic weighted graph is continuously updated in a streaming manner, including: The exponentially weighted moving average method is used to update the newly calculated relation feature vector for each time window and the edge weights after merging with the previous time window. A series of network snapshots are continuously output at intervals equal to the width of the time window. Historical snapshots are stored in chronological order, forming the evolutionary trajectory data of the family relationship network.

[0010] Preferably, topology quantitative analysis is performed on the dynamically weighted graph to calculate network structure indicators and detect whether there are network topology anomalies, including: Calculate the weighted node centrality of each object, where the weighted node centrality is the sum of the absolute values ​​of the sentiment valence of all incoming and outgoing edges of the corresponding object; Calculate the betweenness centrality of each object; Calculate network density; Enumerate all triples in the network, detect closed triangles formed by directed edges, extract the sentiment valence symbols on each edge, and identify triangles with the same symbol pattern as the preset pattern as unbalanced triangles and mark them as high-risk structures. Each network structure indicator is compared with a preset threshold. When any indicator exceeds the threshold range or an imbalance triangle is detected, a network structure abnormality signal is generated.

[0011] Preferably, the edge ablation is to reset the weight of each directed edge in the reference graph to zero; the node ablation is to remove each node in the reference graph and all its associated edges; and the substructure ablation is to set the weight of the smaller of the two positive edges in each unbalanced triangle detected in the reference graph to zero.

[0012] Preferably, determining intervention targets and matching candidate intervention strategies based on the structural root cause contribution includes: Extract the structural element with the largest marginal contribution from the structural root contribution; If the structural element is a directed edge, then the intervention target is set to weaken the corresponding directed edge; If the structural element is a node, the intervention target is set to increase or decrease the centrality of the node according to the centrality anomaly type of the corresponding node. If the structural element is an unbalanced triangle, then the intervention target point is set to break the unbalanced triangle; Based on the intervention target, a matching candidate intervention strategy is retrieved from a pre-built intervention strategy knowledge base.

[0013] Preferably, each candidate intervention strategy is transformed into a predictive topological modification of the dynamically weighted directed graph and input into the graph neural network sentiment prediction model for simulation evaluation, selecting the candidate intervention strategy with the optimal expected improvement, including: For each candidate intervention strategy, it is converted into a predictive topology modification of the baseline graph to obtain the corresponding simulated network; Each simulation network is input into the graph neural network emotion prediction model to calculate the post-intervention emotion prediction value corresponding to each candidate intervention strategy. The expected improvement of each candidate intervention strategy is calculated based on the difference between the post-intervention emotion prediction value and the baseline emotion prediction value. A comprehensive score is calculated by combining the historical success rate of each candidate intervention strategy with the expected improvement, and the candidate intervention strategy with the highest comprehensive score is selected as the optimal strategy.

[0014] Preferably, the intelligent agent, as the executing subject, actively intervenes in the interaction process between various objects to implement targeted topology repair, including: When the intervention target is a weakening directed edge, the agent actively intervenes when it detects the target interaction, introducing a third party or guiding the attention to shift. When the intervention target is an enhanced directed edge, the agent initiates a joint activity invitation involving both parties and provides positive feedback; When the intervention target is to break the imbalance triangle, the agent continuously increases its interaction with the rejected object; When the intervention target is to adjust the centrality of nodes, the agent undertakes some emotional support functions or creates interaction opportunities for objects with low centrality.

[0015] Preferably, determining the effectiveness of the intervention and updating the strategy confidence score in the intervention strategy knowledge base includes: Calculate the actual changes in network structure indicators and emotional state across multiple dimensions after the intervention compared to before the intervention; The actual changes in each dimension are compared with the expected changes in the pre-simulation to calculate the intervention effectiveness index. If the intervention effectiveness index is greater than or equal to the first threshold, the intervention is deemed effective, and the confidence score of the corresponding strategy is increased. If the intervention effectiveness index is less than or equal to the second threshold, the intervention is deemed ineffective, and the confidence score of the corresponding strategy is reduced. Otherwise, the condition is deemed partially valid, and the applicable boundary conditions are recorded.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: (1) This invention aims to address the lack of perception of the overall family relationship structure in existing technologies through the construction of dynamic relationship networks. A dynamic weighted directed graph construction mechanism is designed, which can automatically extract relationship features in four dimensions—emotional valence, interaction intensity, physiological synchronicity, and power relations—from the multimodal interaction data of each member in the family environment. Using family members as nodes and relationship features as directed edge weights, family relationships are elevated from binary interaction analysis to network topology expression, while continuously updating network snapshots in a streaming manner. This mechanism effectively compensates for the shortcomings of traditional methods that simplify family relationships to a single score and lose topological structure information, allowing the system to clearly depict the structural form of the family relationship network and the path of emotion propagation.

[0017] (2) This invention aims to address the problem that existing causal reasoning is limited to single event chains through network ablation attribution analysis. A network ablation analysis mechanism is introduced. When negative emotions or abnormal network structures are detected in adolescents, edge ablation, node ablation, and substructure ablation operations are performed on the dynamically weighted directed graph one by one. A graph neural network model is used to predict the adolescent's emotional state after each ablation. By calculating the marginal contribution of each network structural element, the structural root cause of the emotional problem is located. This mechanism effectively overcomes the limitation of traditional methods that simply attribute adolescent emotional problems to a single conflict event, allowing the system to attribute emotions to structural defects such as persistent imbalance triangles, centrality anomalies, and network density imbalances within the family network.

[0018] (3) In this invention, the agent is designed as a new node in the family relationship network, enabling it to establish and continuously maintain multi-dimensional relationship edges with each family member. The candidate intervention strategies are transformed into predictive modifications to the network topology, and the intervention effect is simulated through a graph neural network. After selecting the strategy with the optimal expected improvement, the agent actively intervenes in the interaction process of family members to implement targeted topology repair. This mechanism transforms the agent from "only making suggestions" to "being able to personally participate in changing the family relationship structure." At the same time, by comparing the network structure and emotional state before and after intervention, a closed-loop optimization is formed, fundamentally improving the psychological and emotional state of adolescents. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for attribution of adolescent psychological and emotional states and self-intervention based on modeling of intelligent agents and family relationship networks, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the construction process of the dynamic relationship network according to an embodiment of the present invention. Figure 3This is a flowchart of the network topology quantitative analysis according to an embodiment of the present invention; Figure 4 This is a flowchart of the intelligent agent embedding network and targeted intervention according to an embodiment of the present invention; Figure 5 This is a general framework diagram of an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] This embodiment provides a system for attribution and self-intervention of adolescent psychological emotions based on modeling of intelligent agents and family relationship networks, such as... Figure 5 ,include: A multimodal data acquisition module is deployed in the target scene to collect multimodal interaction data of each object in real time. The multimodal interaction data includes at least visual perception data, voice perception data and physiological perception data. The dynamic relationship network construction module is used to extract the relationship features between objects based on the multimodal interaction data, construct a dynamic weighted graph as the mathematical representation of the relationship network with each object as a node and the relationship features as directed edge weights, and continuously update the dynamic weighted graph. The network topology analysis module is used to perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies. The network ablation attribution module is used to perform edge ablation, node ablation, and substructure ablation operations on the current network snapshot as the reference graph when negative emotions or network structure anomalies are detected. It uses a pre-trained graph neural network emotion prediction model to calculate the change in emotion prediction of each ablation operation relative to the reference graph, and outputs the contribution of each network structure element to the structural root cause of negative emotions based on the change in emotion prediction. The intelligent agent intervention module is used to embed the intelligent agent as a new node into the dynamic weighted graph, determine the intervention target based on the structural root contribution and match candidate intervention strategies, transform each candidate intervention strategy into a predictive topological modification of the dynamic weighted directed graph and input it into the graph neural network emotion prediction model for simulation evaluation, select the candidate intervention strategy with the best expected improvement, and have the intelligent agent actively intervene in the interaction process between objects to implement targeted topological repair. The closed-loop feedback optimization module is used to reconstruct the network snapshot after the intervention and after a preset observation period. It compares the actual changes in network structure indicators and adolescent emotional state before and after the intervention with the expected improvement in the pre-simulated range, determines the effectiveness of the intervention, updates the strategy confidence score in the intervention strategy knowledge base, and adjusts the behavioral parameters of the agent to form a closed-loop optimization.

[0023] This embodiment also proposes a method for attribution and self-intervention of adolescent psychological emotions based on modeling of intelligent agents and family relationship networks, including: The system acquires multimodal interaction data of each object collected in real time by a multimodal perception unit deployed in the target scene. The multimodal interaction data includes at least visual perception data, speech perception data, and physiological perception data. Based on the multimodal interaction data, the relationship features between each object are extracted. Using each object as a node and the relationship features as directed edge weights, a dynamic weighted graph is constructed as a mathematical representation of the relationship network, and the dynamic weighted graph is continuously updated. Perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies; When negative emotions or network structure anomalies are detected, the current network snapshot is used as the baseline graph. Edge ablation, node ablation, and substructure ablation operations are performed on the baseline graph to construct corresponding ablation graphs. The baseline graph and each ablation graph are input into a pre-trained graph neural network emotion prediction model to obtain the corresponding emotion prediction value. The change in emotion prediction relative to the baseline graph is calculated for each ablation operation. The contribution of each network structure element to the structural root of negative emotions is output according to the magnitude of the change in emotion prediction. The agent is embedded as a new node in the dynamic weighted graph. The intervention target is determined according to the structural root contribution and the candidate intervention strategy is matched. Each candidate intervention strategy is transformed into a predictive topological modification of the dynamic weighted directed graph and input into the graph neural network emotion prediction model for simulation evaluation. The candidate intervention strategy with the best expected improvement is selected. The agent acts as the execution subject to actively intervene in the interaction process between objects and implement targeted topological repair. After the intervention is implemented, a pre-set observation period is followed by the reconstruction of the network snapshot after the intervention. The actual changes in network structure indicators and adolescent emotional states before and after the intervention are compared with the expected improvement range simulated before the intervention. The effectiveness of the intervention is determined, the strategy confidence score in the intervention strategy knowledge base is updated, and the behavioral parameters of the agent are adjusted to form a closed-loop optimization.

[0024] Specifically, such as Figure 1This embodiment first constructs a dynamically weighted directed graph with family members as nodes and multi-dimensional relationship features as edges using multimodal perception data, elevating family relationships from binary interactions to a network topology level. Secondly, it performs quantitative analysis of the network topology, automatically detecting high-risk structures such as unbalanced triangles, centrality anomalies, and network density imbalances. When emotional problems in adolescents are detected, network ablation analysis is used to locate the structural root causes of the emotional problems, outputting a ranking of the contribution of each structural element. Subsequently, an intelligent agent is embedded into the network as a new node, actively intervening in the interaction process of family members to perform targeted topology repair. Finally, a closed-loop optimization is formed by comparing the network structure and emotional state before and after intervention, achieving a technological leap from individual monitoring to structural attribution, and from passive response to autonomous intervention.

[0025] Step 1: Deployment covers key activity areas in the home, including the living room, bedroom, and study. Data acquisition devices include a depth camera, microphone array, and wearable physiological signal acquisition devices. The depth camera captures facial images, body posture, and spatial position at 30 frames per second, with a resolution of 1920×1080 pixels. The microphone array captures speech signals at a 16kHz sampling rate, providing sound source localization and multi-channel speech separation capabilities. The wearable device simultaneously captures physiological signals such as heart rate, heart rate variability, and skin conductance at a 100Hz sampling rate. Each family member takes turns standing in front of the depth camera to capture multi-angle facial images, record voiceprint samples, wear the wearable device, and record the binding relationship between the device number and the identity tag. After all family members' registration information is entered, the system generates an identity tag for each member, along with corresponding facial templates, voiceprint templates, and physiological device numbers.

[0026] Step 2: Synchronously collect multimodal interaction data from each family member and perform preprocessing and identity labeling. The depth camera, microphone array, and wearable device maintain millisecond-level time synchronization based on the system clock, continuously collecting data. For the video stream captured by the depth camera, face alignment and pose normalization are performed. The face region in each frame is cropped and scaled to a fixed size, and facial action unit intensity vectors are extracted using a facial action unit detection model. For the speech streams of each member separated by the microphone array, pre-emphasis, framing, and windowing are performed. Acoustic features such as fundamental frequency and Mel-frequency cepstral coefficients are extracted, and speech activity detection is performed to distinguish between effective speech segments and silent segments. For the physiological signals collected by the wearable device, denoising and baseline drift correction are performed, and heart rate and skin conductance signals are smoothed.

[0027] Identity labeling is achieved as follows: Facial images of each member captured by a depth camera are compared with the facial template library entered in Step 1, and member identity labels are output for each frame of data; the voice stream is compared with the voiceprint template library entered in Step 1 using voiceprint features, and when the audio and video identity labels are inconsistent, the facial recognition result takes priority and is labeled. The labeled modal data are aligned with millisecond-level timestamps and assigned to each member's dedicated data channel according to their identity label.

[0028] Furthermore, extract the relationship features between the objects, including: The multimodal interaction data is sliced ​​using fixed time windows, and within each time window: Facial expression features are extracted from the visual perception data of each object, and mapped to the emotional valence score by the pre-trained emotion recognition model. The mean within the window is taken as the emotional valence feature. Based on the spatial location data and voice activity detection results of each object, the ratio of co-existence time and the ratio of dialogue time are integrated as the interaction intensity feature. Based on the physiological perception data of each subject, the Pearson correlation coefficient of the heart rate sequences of the two subjects was calculated, and normalized after Fisher Z transformation as a physiological synchronicity feature. Based on the speech perception data of each object, the speech duration ratio, interruption frequency ratio and volume average ratio are calculated and normalized after dimensionality reduction as power relationship features. The emotional valence feature, interaction intensity feature, physiological synchronicity feature and power relationship feature constitute a four-dimensional relationship feature vector.

[0029] Specifically, Step 3: Extract the four-dimensional relationship feature vector between every two family members. For any two members in the family... and The standardized multimodal time series data output from Step 2 is sliced ​​using a fixed time window.

[0030] Time window width set to seconds, adjacent window overlap rate set to That is, it slides once every 30 seconds. The time interval covered by each time window is ,in Within each time window, the following four-dimensional relational feature vectors are extracted: (1) Characteristics of affective valence Let the members be... There are M valid facial images within the time window. Based on the facial action unit intensity vector output from Step 2, the pre-trained facial expression emotion recognition model is mapped to a single-frame emotion valence score. .member For members The emotional valence score is defined as: (1); Positive values ​​indicate a positive emotional tendency within that time window, while negative values ​​indicate a negative one; the absolute value represents the intensity. Emotional valence characteristics reflect the member's... The emotional tone expressed during the interaction.

[0031] (2) Characteristics of interaction intensity Based on the spatial coordinates of each member output by the depth camera, the members are defined. With members The threshold for determining coexistence in the same spatial region is Meters. Assume a total of S frames are captured within the window, with the number of frames occurring in a given location being... Proportion of time spent together for: (2); Based on the voice activity detection results output by the microphone array, the number of members within the window is counted. With members Number of times participating in the same dialogue round and total conversation duration (seconds). The weighted average of the time spent together and the time spent talking is then combined: (3); in, For preset weighting coefficients, The width of the time window. The larger the value, the higher the degree of intimacy between the two people in terms of physical space and language interaction.

[0032] (3) Physiological synchronicity characteristics Based on the heart rate time-series signals of each member output from Step 2, the heart rate of each member is acquired within the time window. and members Heart rate sequence and ,in This represents the number of heart rate sampling points within the window.

[0033] Calculate the Pearson correlation coefficient between the two sequences: (4); In the formula, For members With members Pearson correlation coefficient of heart rate sequences For members Average heart rate within the time window. For members Average heart rate within the time window; The correlation coefficients were normalized to the [0,1] interval after Fisher's Z-transform to obtain the physiological synchronicity score: (5); A value close to 1 indicates a high degree of synchronization between the autonomic nervous systems of both parties, reflecting a strong emotional resonance.

[0034] (4) Characteristics of power relations Based on the voice streams and voice activity detection results of each member output from Step 2, three sub-indicators are calculated within the time window: member Relative to members Percentage of voice recording time : ,in, For members Effective voice duration, For members The effective voice duration.

[0035] Interruption frequency ratio: , For members interrupt members Number of times, For members interrupt members The number of times; Volume Averaging Ratio: , For members The effective average volume of the voice signal. For members The average effective volume of the voice signal.

[0036] The above three-dimensional feature vectors Principal component analysis was used to reduce the dimensionality to one dimension, and then the scores were normalized to the [0,1] interval to obtain the power relationship scores: (6); in, is the load vector of the first principal component. The higher the score, the more members... relative members The more dominant a position is in the interaction.

[0037] Using the above four equations, a four-dimensional relational feature vector is output for each time window. : (7); This vector serves as a member in a dynamically weighted directed graph. Pointer to member The directed edge at time The weight.

[0038] Furthermore, the dynamically weighted graph is continuously updated in a streaming manner, including: The exponentially weighted moving average method is used to update the newly calculated relation feature vector for each time window and the edge weights after merging with the previous time window. A series of network snapshots are continuously output at intervals equal to the width of the time window. Historical snapshots are stored in chronological order, forming the evolutionary trajectory data of the family relationship network.

[0039] Specifically, Step 4: Construct a dynamically weighted directed graph. Let the total number of members in the family be N. Using the four-dimensional relationship feature vectors between members output in Step 3 as edge weights, construct the time-varying directed graph. Dynamically weighted directed graph : (8); in, Let N be a set of nodes, containing all N family members, with each node representing one member. It is a set of directed edges, consisting of directed relations between all its members. Indicates from member Pointer to member The directed edges; Let be the weight matrix, where For members Pointer to member The four-dimensional relational feature vector: (9); In the formula, As an affective valence feature, As a feature of interaction intensity, It is a physiological synchronicity characteristic. This is a characteristic of power relations.

[0040] For any directed edge, the relation feature vector output from Step 3 is updated smoothly across time windows using an exponentially weighted moving average method.

[0041] Let the first The newly calculated relational feature vectors for each time window are: The edge weights after merging the previous time window are The edge weights after merging the current time window are updated as follows: (10); in, The update rate is used to control the balance between historical cumulative weights and new observations; in this embodiment, it is set to [value missing]. For the first time window k=1 after system startup, the edge weights after merging from the previous time window... All four dimensions are initialized to 0.5.

[0042] Using the above method, a dynamically weighted directed graph is output for each time window. This graph, as a mathematical representation of a family relationship network, not only preserves multi-dimensional quantitative information about the relationships between members, but also achieves a smooth evolution of edge weights through an exponentially weighted moving average.

[0043] Step 5: Continuously update the dynamically weighted directed graph in a streaming manner and output a network snapshot sequence. The exponentially weighted moving average update in Step 4 is run in streaming computation mode. Each time new data arrives, the system only updates the affected data. Instead of reconstructing the entire graph, the weights of directed edges are calculated, ensuring computational efficiency. The output period of the network snapshot is consistent with the time window width of Step 3, i.e., based on the weights of all possible directed edges. Output a network snapshot frame at second intervals: 1,2,3,…(11) Snapshot per frame Includes: a complete set of nodes ; set of directed edges Four-dimensional weight matrix .

[0044] Historical snapshots are stored chronologically, forming the evolutionary trajectory data of family relationship networks: .

[0045] Through Steps 3 to 5 above, the continuous dynamic process of family interaction is discretized into a series of equally spaced network snapshots, making the family relationship network indexable, traceable, and analyzable in the time dimension, providing a data foundation for subsequent quantitative analysis of topology and anomaly detection.

[0046] like Figure 2 Build a flowchart for dynamic relationship networks.

[0047] Furthermore, a topology quantitative analysis is performed on the dynamically weighted graph to calculate network structure indicators and detect whether there are network topology anomalies, including: Calculate the weighted node centrality of each object, where the weighted node centrality is the sum of the absolute values ​​of the sentiment valence of all incoming and outgoing edges of the corresponding object; Calculate the betweenness centrality of each object; Calculate network density; Enumerate all triples in the network, detect closed triangles formed by directed edges, extract the sentiment valence symbols on each edge, and identify triangles with the same symbol pattern as the preset pattern as unbalanced triangles and mark them as high-risk structures. Each network structure indicator is compared with a preset threshold. When any indicator exceeds the threshold range or an imbalance triangle is detected, a network structure abnormality signal is generated.

[0048] Specifically, Step 6: Perform topological quantization analysis on the dynamically weighted directed graph, such as... Figure 3 .

[0049] For time Network snapshots Calculate the following five network structure metrics: (1) Weighted node centrality : Degree centrality based on absolute value weighting of affective valence features.

[0050] member The weighted node centrality is defined as the sum of the absolute values ​​of the sentiment valence of all its incoming and outgoing edges: (12); in, For members Pointer to member Emotional valence weights (inbound edges). For members Pointer to member The emotional valence weight (outbound).

[0051] This indicator represents the member The overall activity level and emotional involvement level in the family relationship network range from [0, 2(N-1)]. High centrality indicates that the member is carrying an excessive emotional burden, which is a potential risk signal; low centrality indicates that the member is on the edge of the network and is at risk of being isolated.

[0052] (2) Middle Centrality : set up For members To members The number of all shortest paths (based on the reciprocal of the absolute value of sentiment valence as the edge weight, i.e., the stronger the relationship, the smaller the edge weight). For members The number of shortest paths.

[0053] member The definition of betweenness centrality is: (13); This indicator represents the member The degree of criticality in the family's information and emotion transmission channels ranges from [0, 1]. Members with excessively high mediator centrality pose a single point of failure risk in the network—if this member experiences an emotional breakdown, it will disrupt the entire family's information transmission channels.

[0054] (3) Network density : Network density is defined as the proportion of the number of actual relational edges (ignoring weights and only considering the existence of edges) to the total number of edges in the complete graph. Since this graph is directed, the maximum possible number of edges is... : (14); in, The number of relation edges existing at the current moment (i.e. The number of edges), with a value range of [0,1].

[0055] Excessive network density indicates blurred boundaries and entanglement among family members; insufficient network density indicates emotional alienation and a lack of support.

[0056] (4) Triangle structure detection and balance determination: Enumerate all triples in the network The algorithm detects whether three directed edges form a closed triangle. For each detected triangle, it extracts the sentiment valence symbols on the three edges. .

[0057] According to the social equilibrium theory, they can be classified as follows (see Table 1): Table 1 In an unbalanced triangle, if the members With members For a positive relationship (+), members With members For a positive relationship (+), members With members If the relationship is negative (-), then the member At the apex of the triangle, members and There is hostility between them. The system records the member triples, symbol configuration, and top node of the unbalanced triangle.

[0058] (5) Tracking the path of emotion transmission: Based on the product of physiological synchronicity characteristics and the absolute value of emotional valence, the definition is derived from the member... To members The weight of emotional contagion on the directed edge : (15); In the formula, It is a physiological synchronicity characteristic. As an affective valence feature, For members To members The emotional contagion weight on the directed edge.

[0059] Larger indicates a larger number of members The emotions spread along that side to the members The stronger the tendency, the value range is [0,1]. Dijkstra's algorithm is used to search for members on the dynamically weighted directed graph. To members The maximum emotional contagion path is the path that maximizes the cumulative contagion weight along the path.

[0060] Step 7: Compare each network structure index with the preset threshold to generate a network structure anomaly signal.

[0061] The preset thresholds for each indicator are determined based on the statistical distribution of the family's historical baseline data for the first 30 days after the system is started.

[0062] make As an indicator The average of the previous 30 days, If the standard deviation is used, then the anomaly detection threshold is set as follows: (16); Anything outside this range is considered abnormal.

[0063] The comparison rules are as follows: Table 2 Table 2 When any of the above conditions are triggered, the system generates a network structure anomaly signal. An anomaly signal includes the following fields: anomaly type; identifier of the member or structure involved; and the deviation between the current indicator value and the threshold.

[0064] Further, Step 8: Detect negative emotional states in adolescents and trigger an attribution process. The adolescents' emotional states are analyzed in real-time by a multimodal fusion classifier based on facial expressions, voice tone, and physiological signals collected in Step 2. For a moment The emotion category labels represent negative, neutral, and positive, respectively. Confidence level for negative emotions.

[0065] The system automatically triggers the attribution procedure when any of the following conditions are met: Triggering condition one: ,in This is a preset threshold.

[0066] Triggering condition two: Step 7 outputs a network structure anomaly signal. .

[0067] When any of the triggering conditions are met, the system will take a network snapshot of the current moment. Marked as the baseline map.

[0068] Furthermore, the edge ablation is to reset the weight of each directed edge in the reference graph to zero; the node ablation is to remove each node in the reference graph and all its associated edges; and the substructure ablation is to set the weight of the smaller of the two positive edges in each unbalanced triangle detected in the reference graph to zero.

[0069] Specifically, Step 9: Perform network ablation analysis on the dynamically weighted directed graph to locate the structural root causes of emotional problems.

[0070] Based on the baseline diagram As the initial state, perform the following three types of ablation operations: (1) Edge ablation: For each directed edge in the base graph One by one, set all the four-dimensional weight vectors of this side to zero, while keeping the rest of the structure unchanged, to construct the ablation map. 。; (2) Node ablation: For each node in the baseline graph Remove the node and all its associated edges one by one, leaving the rest of the structure unchanged, and construct the ablation graph. .

[0071] (3) Substructure ablation: For each unbalanced triangle detected in Step 6 To break the triangle structure, the weight of the two positive edges with the smaller weight is set to zero, thus constructing an ablation graph. .

[0072] Let the pre-trained graph neural network sentiment prediction model be... , base map Input the data into the model to obtain the baseline sentiment prediction value. : (17); Each ablation map Input the data into the model to obtain the ablation-predicted sentiment values. .

[0073] The marginal contribution of this ablation operation is defined as: (18); This indicates that removing this structural element increases (improves) the predictive value of adolescent emotions. The more negative the value, the greater the contribution of that element to negative emotions. Sort the marginal contributions of all ablation operations from smallest to largest, and output an attribution report, including the ranking of the contribution of each structural element and the location of the main structural root causes.

[0074] Further, Step 10: The agent is embedded into the dynamically weighted directed graph as a newly added node. (Agent) It exists in the home environment in either a physical or virtual form. (In a dynamically weighted directed graph) Create its own node The updated node set is .

[0075] Initial relationship edges are established by performing pre-defined interaction procedures with each family member. For family members... Initial edge weight Set as: (19); In the formula, For agents and members The initial edge weight vector between them.

[0076] In subsequent operation, the intelligent agent continuously senses and interacts with its members through its own sensors. The interaction is processed in the same way as in Step 3 to extract relation features. And update the edge weights using an exponentially weighted moving average: (20); in, , For the current time window, the agent and its members The edge weight vector after merging between them This is the edge weight vector after merging from the previous time window.

[0077] After the agent becomes a formal node in the network, the total number of nodes is changed from Increase to As a result, global structural metrics such as network density and shortest path change accordingly.

[0078] Further, determining intervention targets and matching candidate intervention strategies based on the structural root cause contribution includes: Extract the structural element with the largest marginal contribution from the structural root contribution; If the structural element is a directed edge, then the intervention target is set to weaken the corresponding directed edge; If the structural element is a node, the intervention target is set to increase or decrease the centrality of the node according to the centrality anomaly type of the corresponding node. If the structural element is an unbalanced triangle, then the intervention target point is set to break the unbalanced triangle; Based on the intervention target, a matching candidate intervention strategy is retrieved from a pre-built intervention strategy knowledge base.

[0079] Specifically, Step 11: Identify intervention targets and match candidate strategies based on the attribution report. Extract the structural elements with the largest marginal contribution from the attribution report and convert them into intervention targets according to the following rules. .

[0080] If it is a directed edge If the target of intervention is to weaken that edge, the propagation channel of the negative impact will be blocked; if it is a node If the centrality anomaly type is determined, the centrality of the node is adjusted by either increasing or decreasing it to change its role in the network; if it is an unbalanced triangle... The intervention target is set to break the triangular structure, causing the triangle to return to its balanced state.

[0081] After identifying the intervention target, the system uses a pre-built intervention strategy knowledge base. The system retrieves matching candidate intervention strategies. The knowledge base is organized by target type: weakening edge correspondences reduces the time spent alone between the two parties and introduces a third party; increasing centrality correspondences increases the frequency of interaction between the member and other members and provides positive feedback; reducing centrality correspondences allows the agent to assume part of its emotional support function; breaking the imbalance triangle correspondence strengthens the positive connection between the excluded member and the top node.

[0082] Furthermore, each candidate intervention strategy is transformed into a predictive topological modification of the dynamically weighted directed graph and input into the graph neural network sentiment prediction model for simulation evaluation. The candidate intervention strategy with the optimal expected improvement is selected, including: For each candidate intervention strategy, it is converted into a predictive topology modification of the baseline graph to obtain the corresponding simulated network; Each simulation network is input into the graph neural network emotion prediction model to calculate the post-intervention emotion prediction value corresponding to each candidate intervention strategy. The expected improvement of each candidate intervention strategy is calculated based on the difference between the post-intervention emotion prediction value and the baseline emotion prediction value. A comprehensive score is calculated by combining the historical success rate of each candidate intervention strategy with the expected improvement, and the candidate intervention strategy with the highest comprehensive score is selected as the optimal strategy.

[0083] Specifically, Step 12: Perform a pre-simulation evaluation of the candidate strategies and execute the optimal strategy.

[0084] For each candidate strategy This is transformed into a predictive modification Δ to the network topology. To obtain the simulated network : (twenty one); The baseline graph (i.e., a snapshot of the network at the current moment saved when the attribution process is triggered); Will Input a graph neural network emotion prediction model Φ, and calculate the emotion prediction value after intervention. : ( )(twenty two); Expected improvement for: (twenty three); Combining the strategy's historical success rate Calculate the overall score : (twenty four); choose The largest solution implement.

[0085] Furthermore, the intelligent agent, acting as the execution subject, actively intervenes in the interaction process between various objects to implement targeted topology repair, including: When the intervention target is a weakening directed edge, the agent actively intervenes when it detects the target interaction, introducing a third party or guiding the attention to shift. When the intervention target is an enhanced directed edge, the agent initiates a joint activity invitation involving both parties and provides positive feedback; When the intervention target is to break the imbalance triangle, the agent continuously increases its interaction with the rejected object; When the intervention target is to adjust the centrality of nodes, the agent undertakes some emotional support functions or creates interaction opportunities for objects with low centrality.

[0086] Specifically, intelligent agents The intervention methods of the implementing entity are as follows, such as... Figure 4 : Weakening edge: When a target interaction is detected, proactively intervene by introducing a third party or guiding attention away from the target. Strengthen the edge: Initiate activity invitations involving both parties and provide positive feedback; Breaking the triangle: Continuously increase interaction with excluded members to enhance their sense of emotional fulfillment; Adjusting centrality: Taking on some emotional support functions, or creating opportunities for interaction for members with low centrality.

[0087] Further, determining the effectiveness of the intervention and updating the strategy confidence score in the intervention strategy knowledge base includes: Calculate the actual changes in network structure indicators and emotional state across multiple dimensions after the intervention compared to before the intervention; The actual changes in each dimension are compared with the expected changes in the pre-simulation to calculate the intervention effectiveness index. If the intervention effectiveness index is greater than or equal to the first threshold, the intervention is deemed effective, and the confidence score of the corresponding strategy is increased. If the intervention effectiveness index is less than or equal to the second threshold, the intervention is deemed ineffective, and the confidence score of the corresponding strategy is reduced. Otherwise, the condition is deemed partially valid, and the applicable boundary conditions are recorded.

[0088] Specifically, Step 13: Reconstruct the network snapshot after intervention, compare the data before and after intervention, and update the knowledge base and agent parameters.

[0089] After the intervention is implemented, the system cannot immediately determine its effectiveness—structural changes in family relationship networks require a period of daily interaction to fully manifest. Therefore, the system sets a waiting time window. The waiting period is 7 days, which is the observation period for structural changes. During this period, the real-time monitoring and abnormal triggering mechanisms of Step 1-Step 5 will continue to operate normally and will not affect emergency intervention. The 7-day snapshot is only used for long-term macro-level assessment of policy confidence.

[0090] After this window closes, repeat the network construction process from Step 2 to Step 5 to obtain a snapshot of the network after intervention. Following Step 6, various network structure indicators were calculated after the intervention, and the emotional state data of adolescents within that time window were also read. .

[0091] To quantify the intervention effect, the system calculates the changes in the following four indicators: (25); (26); (27); (28); in, This refers to the number of unbalanced triangles. , , , These represent the changes in weighted node centrality, network density, number of unbalanced triangles, and adolescent emotional state, respectively.

[0092] Actual change Compared with the expected change in the pre-simulation in Step 12 To make comparisons, define indicators of intervention effectiveness: (29); in, The closer the value is to 1, the better the actual effect matches the expectation.

[0093] The judgment rules are as follows: If Validity is determined by the policy confidence level. Increase by 0.1; if If the determination is partially valid, record the applicable condition boundaries; if Invalid decision, strategy confidence. Decrease by 0.15.

[0094] Policy confidence The updated constraint is set to the interval [0,1]. The agent is then adjusted synchronously. Three behavioral parameters: relation edge maintenance frequency (In case of an anomaly, encryption is reduced from 60 seconds to 30 seconds, and then restored after stabilization), Intervention trigger threshold (If the effective intervention response is timely, adjust downwards by 0.05; if the ineffective intervention is frequent, adjust upwards by 0.05). Minimum acceptable improvement. (Dynamically adjusted based on the actual improvement level of the high-confidence strategy). After the update is complete, return to Step 3 to begin the next round of continuous monitoring and optimization.

[0095] To more clearly illustrate the technical solution of the present invention, specific embodiments are provided below for description: Step 1: Deploy multimodal data acquisition devices and register family member identity profiles. Deployment covers key activity areas in the home, including the living room, bedroom, and study. Acquisition devices include a depth camera (1920×1080 resolution, 30fps), a microphone array (16kHz sampling rate), and a wearable device (100Hz sampling rate, collecting heart rate, heart rate variability, and skin conductance). Each family member sequentially acquires multi-angle facial images, records a 10-second voiceprint sample, wears the wearable device, and is associated with an identity tag. The system generates identity tags, a 512-dimensional facial feature template, and a 256-dimensional voiceprint feature template for each member.

[0096] Step 2: Synchronously collect multimodal interaction data from each family member and perform preprocessing and identity labeling. The depth camera, microphone array, and wearable device maintain millisecond-level time synchronization based on the system clock, continuously collecting data. The video stream undergoes face alignment and pose normalization, is cropped and scaled to 224×224 pixels, and 20 facial motion unit intensity vectors (values ​​0-1) are extracted. The audio stream undergoes pre-emphasis (coefficient 0.97), framing (25ms frame length, 10ms frame shift), and Hamming windowing to extract fundamental frequency and 12-dimensional MFCC features, with a VAD energy threshold of 0.3. Physiological signals undergo wavelet denoising (db4, threshold 0.03) and baseline drift correction (0.5Hz high-pass filtering). Identity labeling is jointly determined by facial recognition (cosine similarity threshold 0.7) and voiceprint recognition (threshold 0.65), with facial recognition results given priority. The labeled modal data are aligned with millisecond-level timestamps and categorized into each member's dedicated data channel according to their identity label.

[0097] Step 3: Slice the standardized multimodal time series data with a time window of 60 seconds and an overlap rate of 50% (slide once every 30 seconds) to extract the four-dimensional relational feature vector: (a) Emotional valence features: Based on the facial action unit intensity vector, the pre-trained facial expression emotion recognition model maps it to a single frame valence score (-1 to +1), and the mean within the window is taken.

[0098] (b) Interaction intensity characteristics: The proportion of time spent together is calculated based on spatial location (distance < 2 meters), and the number of dialogue rounds and duration are statistically analyzed based on VAD. Each is weighted at 50% and the values ​​range from 0 to 1.

[0099] (c) Physiological synchronicity characteristics: The Pearson correlation coefficient of the heart rate sequences of the two members was calculated and normalized to the 0-1 interval after Fisher Z transformation.

[0100] (d) Power relationship characteristics: Calculate the proportion of speech duration, interruption frequency ratio, and average volume ratio, and normalize them to the 0-1 range after PCA dimensionality reduction.

[0101] Step 4: Construct a dynamically weighted directed graph. Using each family member as a node and the four-dimensional relationship feature vector output from Step 3 as the directed edge weights, construct a dynamically weighted directed graph. Update the edge weights using an exponentially weighted moving average: new features in the current window account for 30%, and historical accumulation accounts for 70%. Initialize each dimension to 0.5 in the first time window.

[0102] Step 5: Continuously update the dynamically weighted directed graph in a streaming manner and output a network snapshot sequence. Output one network snapshot every 60 seconds, containing the complete set of nodes, the set of directed edges, and a four-dimensional weight matrix. Historical snapshots are stored in chronological order, forming the evolutionary trajectory of the family relationship network.

[0103] Step 6: Perform topology quantization analysis on the dynamically weighted directed graph. For the network snapshot at time t, calculate five network structure metrics: (a) Weighted node centrality: the sum of the absolute values ​​of the sentiment valence of all incoming and outgoing edges, with a value range of [0, 2(N-1)].

[0104] (b) Betweenness centrality: the proportion of the number of shortest paths through this node to the total number of shortest paths among all members, with a value range of [0,1].

[0105] (c) Network density: The proportion of the actual number of directed edges to N(N-1), with a value range of [0,1].

[0106] (d) Triangle structure detection and balance determination: Enumerate all triples, and those with the sign pattern (+,+,-) are determined to be unbalanced triangles and marked as high-risk structures.

[0107] (e) Emotional contagion path tracking: Using physiological synchronicity × absolute value of emotional valence as the edge weight, Dijkstra's algorithm is used to search for the maximum emotional contagion path.

[0108] Step 7: Compare each network structure indicator with a preset threshold to generate a network structure anomaly signal. The threshold is determined based on the historical baseline of the first 30 days after system startup (mean ± 2 standard deviations). When any indicator exceeds the threshold, or an imbalance triangle is detected, a network structure anomaly signal is generated, including the anomaly type, involved members, and deviation magnitude.

[0109] Step 8: Detect negative emotional states in adolescents and trigger the attribution process. The adolescents' emotional states are output in real time by a multimodal fusion classifier. When the confidence score of negative emotions is greater than 0.65, or when Step 7 outputs a network anomalous signal, the attribution process is automatically triggered.

[0110] Step 9: Perform network ablation analysis on the dynamically weighted directed graph to locate the structural root causes of emotional problems. Using the network snapshot at the trigger point as the baseline graph, perform three types of ablation operations: (a) Edge elimination: Reset the weight of each directed edge to zero one by one; (b) Node ablation: Remove each node and its associated edges one by one; (c) Substructure ablation: Set the weaker positive edges in the unbalanced triangles to zero one by one. Input the baseline map and each ablation map into the graph neural network sentiment prediction model to obtain the baseline prediction value and each ablation prediction value.

[0111] Calculate the change in sentiment prediction for each ablation operation (ablation prediction minus baseline prediction). The structural element that contributes the most after ranking is the structural root of the sentiment problem, and an attribution report is output.

[0112] Step 10: The agent embeds itself into the dynamically weighted directed graph as a newly added node. The agent creates its own node in the graph, with the initial edge weights set to... Subsequently, the relationship features between the node and each member are extracted in the same way as in Step 3, and the edge weights are updated using an exponentially weighted moving average (new features account for 30%, and historical features account for 70%). The total number of nodes increases from N to N+1.

[0113] Step 11: Determine intervention targets and match candidate strategies based on the attribution report. If an edge contributes the most, the target is to weaken that edge; if a node contributes the most, the target is to adjust its centrality; if an unbalanced triangle contributes the most, the target is to break the triangle. After determining the target, retrieve matching candidate strategies from the intervention strategy knowledge base.

[0114] Step 12: Conduct pre-intervention simulation evaluation of candidate strategies and execute the optimal strategy. Convert each candidate strategy into a network topology modification to obtain a simulated network. Input this network into a graph neural network to predict the post-intervention sentiment value and calculate the expected improvement. Overall score = 0.7 × expected improvement + 0.3 × historical success rate. Select the strategy with the highest score for execution by the agent. Execution methods: Weakening edges involves introducing a third party into the dialogue; strengthening edges involves initiating joint activities and providing positive feedback; breaking the triangle involves increasing daily interaction with excluded members; adjusting centrality involves assuming emotional support functions or creating interaction opportunities.

[0115] Step 13: Reconstruct the network snapshot after intervention, compare data before and after intervention, and update the knowledge base and agent parameters. Reconstruct the network snapshot after 7 days, calculating the changes in weighted node centrality, network density, number of unbalanced triangles, and adolescent emotional states. Compare the actual changes with the expected changes simulated beforehand. If η ≥ 0.8, the strategy is considered effective, and the confidence level is increased by 0.1; if 0.4 < η < 0.8, it is considered partially effective, and the applicable boundary is recorded; if η ≤ 0.4, it is considered invalid, and the confidence level is decreased by 0.15. Simultaneously adjust agent behavior parameters: when the relationship edge maintenance frequency is abnormal, encrypt it to 30 seconds, and restore it to 60 seconds after stabilization; lower the intervention trigger threshold by 0.05 when effective, and raise it by 0.05 when ineffective; dynamically adjust the minimum acceptable improvement based on the actual improvement level of the high-confidence strategy. After completion, return to Step 3 for continuous monitoring.

[0116] This technical solution ingeniously integrates a dynamic relationship network construction module with a network ablation attribution analysis mechanism. It also introduces intelligent agents as endogenous nodes in the network to execute targeted interventions, significantly improving the structural attribution ability and self-intervention level of adolescents' psychological and emotional problems in family settings. The system utilizes a dynamically weighted directed graph to model family relationships in multiple dimensions, characterizing the interaction features between members from four dimensions: emotional valence, interaction intensity, physiological synchronicity, and power relations. Combined with a streaming update mechanism, it achieves real-time tracking of the evolution trajectory of the family relationship network, optimizing the perception and expression process of relationship structure.

[0117] Network topology quantitative analysis enables the system to automatically detect high-risk structures such as imbalanced triangles, centrality anomalies, and network density imbalances. Meanwhile, network ablation attribution, through a counterfactual analysis framework of "removal-prediction-comparison," precisely pinpoints emotional issues to specific network structural defects. The two work together to overcome the limitations of traditional methods that simplify family relationships to a single scoring indicator and restrict emotional attribution to a single event chain. The agent embeds itself into the network as a new node, selecting the optimal intervention strategy through pre-simulation evaluation and executing it autonomously. This not only achieves a leap from "making suggestions" to "personally participating in changing relationship structures," but also forms a closed-loop optimization by comparing network structure and emotional states before and after intervention, allowing the system's intervention strategy to continuously improve over time.

[0118] In summary, this technical solution provides a systematic solution for adolescent mental health monitoring and family relationship intervention through a complete technical chain including multimodal data acquisition, dynamic relationship network construction, topology quantitative analysis, network ablation attribution, agent embedding and targeted intervention, and closed-loop feedback optimization. It can address the complex challenges of difficult identification of latent causes in family relationship structures, inability to attribute emotional problems, and passive agent responses that cannot be actively intervened. It provides a brand-new technical path for adolescent mental health monitoring and family relationship intervention and demonstrates broad practical application potential.

[0119] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A system for attributing adolescent psychological and emotional issues and for self-intervention based on agent-family relationship network modeling, comprising: A multimodal data acquisition module is deployed in the target scene to collect multimodal interaction data of each object in real time. The multimodal interaction data includes at least visual perception data, voice perception data and physiological perception data. The dynamic relationship network construction module is used to extract the relationship features between objects based on the multimodal interaction data, construct a dynamic weighted graph as the mathematical representation of the relationship network with each object as a node and the relationship features as directed edge weights, and continuously update the dynamic weighted graph. The network topology analysis module is used to perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies. The network ablation attribution module is used to perform edge ablation, node ablation, and substructure ablation operations on the current network snapshot as the reference graph when negative emotions or network structure anomalies are detected. It uses a pre-trained graph neural network emotion prediction model to calculate the change in emotion prediction of each ablation operation relative to the reference graph, and outputs the contribution of each network structure element to the structural root cause of negative emotions based on the change in emotion prediction. The intelligent agent intervention module is used to embed the intelligent agent as a new node into the dynamic weighted graph, determine the intervention target based on the structural root contribution and match candidate intervention strategies, transform each candidate intervention strategy into a predictive topological modification of the dynamic weighted directed graph and input it into the graph neural network emotion prediction model for simulation evaluation, select the candidate intervention strategy with the best expected improvement, and have the intelligent agent actively intervene in the interaction process between objects to implement targeted topological repair. The closed-loop feedback optimization module is used to reconstruct the network snapshot after the intervention and after a preset observation period. It compares the actual changes in network structure indicators and adolescent emotional state before and after the intervention with the expected improvement in the pre-simulated range, determines the effectiveness of the intervention, updates the strategy confidence score in the intervention strategy knowledge base, and adjusts the behavioral parameters of the agent to form a closed-loop optimization.

2. A method for attributing adolescent psychological emotions and autonomous intervention based on agent-family relationship network modeling, used to implement the system described in claim 1, characterized in that, include: The system acquires multimodal interaction data of each object collected in real time by a multimodal perception unit deployed in the target scene. The multimodal interaction data includes at least visual perception data, speech perception data, and physiological perception data. Based on the multimodal interaction data, the relationship features between each object are extracted. Using each object as a node and the relationship features as directed edge weights, a dynamic weighted graph is constructed as a mathematical representation of the relationship network, and the dynamic weighted graph is continuously updated. Perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies; When negative emotions or network structure anomalies are detected, the current network snapshot is used as the baseline graph. Edge ablation, node ablation, and substructure ablation operations are performed on the baseline graph to construct corresponding ablation graphs. The baseline graph and each ablation graph are input into a pre-trained graph neural network emotion prediction model to obtain the corresponding emotion prediction value. The change in emotion prediction relative to the baseline graph is calculated for each ablation operation. The contribution of each network structure element to the structural root of negative emotions is output according to the magnitude of the change in emotion prediction. The agent is embedded as a new node in the dynamic weighted graph. The intervention target is determined according to the structural root contribution and the candidate intervention strategy is matched. Each candidate intervention strategy is transformed into a predictive topological modification of the dynamic weighted directed graph and input into the graph neural network emotion prediction model for simulation evaluation. The candidate intervention strategy with the best expected improvement is selected. The agent acts as the execution subject to actively intervene in the interaction process between objects and implement targeted topological repair. After the intervention is implemented, a pre-set observation period is followed by the reconstruction of the network snapshot after the intervention. The actual changes in network structure indicators and adolescent emotional states before and after the intervention are compared with the expected improvement range simulated before the intervention. The effectiveness of the intervention is determined, the strategy confidence score in the intervention strategy knowledge base is updated, and the behavioral parameters of the agent are adjusted to form a closed-loop optimization.

3. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling as described in claim 2, is characterized in that, Extract the relationship features between objects, including: The multimodal interaction data is sliced ​​using fixed time windows, and within each time window: Facial expression features are extracted from the visual perception data of each object, and mapped to the emotional valence score by the pre-trained emotion recognition model. The mean within the window is taken as the emotional valence feature. Based on the spatial location data and voice activity detection results of each object, the ratio of co-existence time and the ratio of dialogue time are integrated as the interaction intensity feature. Based on the physiological perception data of each subject, the Pearson correlation coefficient of the heart rate sequences of the two subjects was calculated and normalized after Fisher Z transform as a physiological synchronicity feature. Based on the speech perception data of each object, the speech duration ratio, interruption frequency ratio and volume average ratio are calculated and normalized after dimensionality reduction as power relationship features. The emotional valence feature, interaction intensity feature, physiological synchronicity feature and power relationship feature constitute a four-dimensional relationship feature vector.

4. The method for attribution and self-intervention of adolescent psychological emotions based on agent-family relationship network modeling according to claim 2, characterized in that, The dynamically weighted graph is continuously updated in a streaming manner, including: The exponentially weighted moving average method is used to update the newly calculated relation feature vector for each time window and the edge weights after merging with the previous time window. A series of network snapshots are continuously output at intervals equal to the width of the time window. Historical snapshots are stored in chronological order, forming the evolutionary trajectory data of the family relationship network.

5. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 2, characterized in that, Perform topology quantitative analysis on the dynamic weighted graph, calculate network structure indicators, and detect whether there are network topology anomalies, including: Calculate the weighted node centrality of each object, where the weighted node centrality is the sum of the absolute values ​​of the sentiment valence of all incoming and outgoing edges of the corresponding object; Calculate the betweenness centrality of each object; Calculate network density; Enumerate all triples in the network, detect closed triangles formed by directed edges, extract the sentiment valence symbols on each edge, and identify triangles with the same symbol pattern as the preset pattern as unbalanced triangles and mark them as high-risk structures. Each network structure indicator is compared with a preset threshold. When any indicator exceeds the threshold range or an imbalance triangle is detected, a network structure abnormality signal is generated.

6. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 5, characterized in that, The edge ablation is to reset the weight of each directed edge in the baseline graph to zero; the node ablation is to remove each node in the baseline graph and all its associated edges; the substructure ablation is to set the weight of the smaller of the two positive edges in each unbalanced triangle detected in the baseline graph to zero.

7. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 2, characterized in that, Based on the structural root cause contribution, intervention targets are determined and candidate intervention strategies are matched, including: Extract the structural element with the largest marginal contribution from the structural root contribution; If the structural element is a directed edge, then the intervention target is set to weaken the corresponding directed edge; If the structural element is a node, the intervention target is set to increase or decrease the centrality of the node according to the centrality anomaly type of the corresponding node. If the structural element is an unbalanced triangle, then the intervention target point is set to break the unbalanced triangle; Based on the intervention target, a matching candidate intervention strategy is retrieved from a pre-built intervention strategy knowledge base.

8. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 2, characterized in that, Each candidate intervention strategy is transformed into a predictive topological modification of the dynamically weighted directed graph and input into the graph neural network sentiment prediction model for simulation evaluation. The candidate intervention strategy with the optimal expected improvement is selected, including: For each candidate intervention strategy, it is converted into a predictive topology modification of the baseline graph to obtain the corresponding simulated network; Each simulation network is input into the graph neural network emotion prediction model to calculate the post-intervention emotion prediction value corresponding to each candidate intervention strategy. The expected improvement of each candidate intervention strategy is calculated based on the difference between the post-intervention emotion prediction value and the baseline emotion prediction value. A comprehensive score is calculated by combining the historical success rate of each candidate intervention strategy with the expected improvement, and the candidate intervention strategy with the highest comprehensive score is selected as the optimal strategy.

9. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 8, characterized in that, The intelligent agent, acting as the executor, actively intervenes in the interaction process between various objects to implement targeted topology repair, including: When the intervention target is a weakening directed edge, the agent actively intervenes when it detects the target interaction, introducing a third party or guiding the attention to shift. When the intervention target is an enhanced directed edge, the agent initiates a joint activity invitation involving both parties and provides positive feedback; When the intervention target is to break the imbalance triangle, the agent continuously increases its interaction with the rejected object; When the intervention target is to adjust the centrality of nodes, the agent undertakes some emotional support functions or creates interaction opportunities for objects with low centrality.

10. The method for attribution and self-intervention of adolescent psychological emotions based on intelligent agent and family relationship network modeling according to claim 2, characterized in that, Determine the effectiveness of the intervention and update the strategy confidence score in the intervention strategy knowledge base, including: Calculate the actual changes in network structure indicators and emotional state across multiple dimensions after the intervention compared to before the intervention; The actual changes in each dimension are compared with the expected changes in the pre-simulation to calculate the intervention effectiveness index. If the intervention effectiveness index is greater than or equal to the first threshold, the intervention is deemed effective, and the confidence score of the corresponding strategy is increased. If the intervention effectiveness index is less than or equal to the second threshold, the intervention is deemed ineffective, and the confidence score of the corresponding strategy is reduced. Otherwise, the condition is deemed partially valid, and the applicable boundary conditions are recorded.

Images