Parent-child two-way relation evaluation and adjustment system fusing psychological and behavioral data

Abstract

The invention provides a parent-child two-way relationship evaluation and adjustment system fusing psychological and behavioral data, and the system comprises a data collection module which collects parent-child data under the guarantee of a hierarchical authorization mechanism through a parent end and child end dual-terminal architecture, and achieves the data desensitization and synchronous processing through a sensitive information fuzzification and time-space alignment technology; the feature extraction and fusion module generates a unified feature vector through weighted fusion by extracting audio-visual features, cognitive deviation vectors, physiological time sequence features and parent-child relationship vectors; and the evaluation and strategy generation module adopts a double-branch neural network to perform bidirectional relation evaluation on the fusion features, matches a strategy template based on an evaluation result, dynamically optimizes strategy parameters of the parent end and the child end through a reinforcement learning framework, and establishes a strategy parameter iteration mechanism according to implementation feedback. According to the method, objective quantitative evaluation and personalized adjustment strategy generation of the parent-child relationship under multi-modal data fusion are realized, and the intervention accuracy is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent mental health assessment and intervention, and specifically to a parent-child two-way relationship assessment and regulation system that integrates psychological and behavioral data. Background Art

[0002] Current parent-child relationship assessment methods mainly rely on subjective scales or single-dimensional analysis of behavioral observations, which have significant flaws. Traditional scale assessments often use static questionnaires from the perspective of either parents or children, which makes it difficult to capture dynamic psychological changes and non-verbal behavioral signals in interactions, resulting in one-sided assessment results. Although behavioral analysis based on manual observation can record some interaction patterns, it is limited by the observer's subjective judgment and short-term sampling bias, and cannot quantify deep interaction indicators such as physiological synchronization and micro-expression feedback. More importantly, existing technologies lack a quantification mechanism for parent-child two-way cognitive biases - the differences in the perceptions of parents and children on the same interaction event are often overlooked, and this perspective separation phenomenon is precisely the core cause of relationship conflict. The shortcomings of the above methods in data source integration, dynamism and bidirectionality restrict the objectivity of the assessment results and the precise positioning of intervention targets.

[0003] When it comes to parent-child relationship regulation, existing intervention programs are generally static and one-dimensional. Most systems rely on pre-set, universal policy templates, failing to consider the unique and dynamic nature of parent-child interaction patterns, leading to a mismatch between strategies and families' actual needs. Furthermore, traditional policy generation mechanisms lack closed-loop optimization capabilities: They are unable to dynamically adjust policy parameters based on historical intervention effectiveness data, while also ignoring the differentiated role requirements of parents and children in the parent-child relationship regulation process. This "one-size-fits-all" intervention model struggles to adapt to the complex state transitions that occur in parent-child relationships over time, and it is even unable to address the coordination failures caused by asynchronous parent-child regulatory responses. There is an urgent need to develop an intelligent regulatory system that integrates two-way evaluation, dynamic policy optimization, and differentiated role adaptation. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the current background technology, the purpose of the present invention is to provide a parent-child two-way relationship evaluation and adjustment system that integrates psychological and behavioral data.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A parent-child bidirectional relationship assessment and regulation system that integrates psychological and behavioral data. The system is divided into a parent side and a child side, and includes: The data collection module collects first parent-child data in real time based on a preset authorization mechanism; performs desensitization processing on the first parent-child data to generate second parent-child data; and performs preprocessing and time synchronization on the second parent-child data to obtain third parent-child data; A feature extraction and fusion module extracts an audio-visual feature vector, a cognitive bias vector, a time series feature vector, and a parent-child relationship vector based on the third parent-child data and performs a fusion operation to generate a fused feature vector; An evaluation and strategy generation module inputs the fused feature vector into a preset evaluation model to output a bidirectional evaluation result; constructs an adjustment strategy library, where the adjustment strategies in the adjustment strategy library consist of strategy templates and strategy parameters; selects a strategy template from the adjustment strategy library based on the bidirectional evaluation result; and dynamically optimizes the strategy parameters corresponding to the strategy template through a reinforcement learning framework.

[0006] Furthermore, the first parent-child data includes interaction situation data, subjective psychological data, physiological behavior data, and audio-visual data; Interaction context data is collected by parents or children through interaction events, including but not limited to timestamp, duration, location, frequency, and event type; Subjective psychological data is collected directly through the scales, which have the same structure and complementary perspectives, pushed to both the parent and child sides. Physiological and behavioral data are collected through wearable devices. Audio and video data are collected through cameras and microphone arrays deployed in designated interactive areas; The parent end, the child end, the wearable device, and the camera and microphone array constitute a data acquisition terminal; The authorization mechanism is to provide authorization interfaces to the parent and child respectively before the first parent-child data is collected, clearly stating the purpose, scope and usage of each type of data collected; Parents and children are free to choose the type of data to authorize and can withdraw authorization; for minor children, parents are responsible for authorizing and withdrawing data; for adult children, dual authorization from both parents and children is required, and either parent or child can cancel the authorized data.

[0007] Furthermore, the specific steps of the desensitization treatment are as follows: 1) Identify sensitive fields in the interaction context data, subjective psychological data, and physiological behavior data based on a preset sensitive field rule library, and use different technical means to desensitize different sensitive fields; 2) Desensitization processing of the audio and video data includes visual desensitization and audio desensitization; The face area in the video is blurred using a pixelated convolution kernel, and background objects containing sensitive information are identified using the preset YOLOv5 model. Pixel-level random scrambling is applied to the identified area. The Mel-frequency cepstral coefficients in the audio are extracted and retained as voiceprint features, and then Gaussian white noise is superimposed to blur the audio.

[0008] Furthermore, the preprocessing calculates the mean μ and standard deviation σ of the interaction context data, subjective psychological data and physiological behavior data based on the sliding window, and removes the data that deviates from the mean by more than The value of; linear interpolation compensation is used for continuous data; a null value mark is given to discrete data; The specific steps of the time synchronization are as follows: 1) Synchronizing the data acquisition terminal to a preset master clock through the NTP protocol; 2) Setting the timestamp of the interaction context data as the reference time ; The time axis of the physiological behavior data and audio-visual data is based on the reference time Align to the origin.

[0009] Furthermore, the audio-visual feature vector is extracted based on the audio-visual data, including a visual layer, an audio layer, and a fusion layer; The visual layer is used to extract visual feature vectors, and the specific steps are as follows: 1) Calculate the dense optical flow field of the video and construct the motion matrix; 2) Improving the convolution kernel size and the number of feature map channels of the original dilated 3D convolutional neural network; processing the motion matrix based on the improved dilated 3D convolutional neural network to output a visual feature vector; The audio layer extracts audio feature vectors based on the speech emotion recognition model of the Transformer architecture. The specific steps are as follows: 1) performing frame processing on the speech to extract basic acoustic features; and inputting the basic acoustic features into the speech emotion recognition model; 2) The speech emotion recognition model captures short-term acoustic patterns through convolution operations to generate local feature vectors. The multi-head self-attention mechanism then analyzes the contextual dependencies of the local feature vectors. The model then combines dynamic acoustic properties and emotion keyword detection to output an audio feature vector. The fusion layer concatenates the visual feature vector and the audio feature vector to obtain an audio-visual feature vector; The mean μ and standard deviation σ of the audio-visual feature vector are calculated based on the sliding window, and the vectors that deviate from the mean by more than value.

[0010] Furthermore, the cognitive bias vector is extracted based on subjective psychological data, and the specific steps are as follows: 1) Parents’ cognitive bias calculate: Average the absolute differences of all questions in the scale and add the time weight: in, is the preset attenuation coefficient, e is a natural constant, is the absolute difference of the i-th question, Submission time for the scale relative to the base time The time offset of 2) Through the parental cognitive bias value Calculation method for calculating children's cognitive bias value ; 3) Through the parental cognitive bias value and children's cognitive bias Construct a cognitive bias vector.

[0011] Furthermore, the physiological behavior data is input into a double-layer LSTM to extract the time series feature vector.

[0012] Furthermore, the parent-child relationship vector is extracted based on the interaction scenario data, and the specific steps are as follows: 1) Set a time window and count the total number of interactions between parents and children in the interaction events within the time window, and eliminate the influence of extreme values ​​through logarithmic transformation to obtain the interaction frequency intensity , the extreme value refers to the number of interactions within the time window that deviates from the preset normal range: in, is the total number of interactions between parents and children in the interaction event; 2) Calculate the absolute value of the time difference between the child behavior response and the parent behavior initiation in the interaction event, and take the average value of the time difference in the time window to obtain the response delay : in, The timestamp of the parent initiating the action, Timestamp for child responses; 3) Count the number of conflicts in the interaction events; weight the number of conflicts within the time window to calculate the conflict event density X; and add a time decay factor to give more weight to recent conflicts; 4) Increase the frequency and intensity of interaction , response delay and the conflict event density X are integrated into the parent-child relationship vector; Map the audio-visual feature vector, cognitive bias vector, time series feature vector and parent-child relationship vector to a unified dimension; generate a fused feature vector by weighted feature fusion .

[0013] Furthermore, the evaluation model is constructed by a two-branch neural network, including an input layer, a separation layer, a two-branch layer and an output layer; The input layer inputs the fused feature vector ; The separation layer will Split into parent view vector With children perspective vector , through the preset feature mask matrix accomplish: The double-branch layer is divided into a first branch and a second branch; The first branch is based on Output parents' ratings of the relationship ; The second branch is based on Output the child's rating of the relationship ; The output layer outputs the bidirectional evaluation results in, It is a two-way cognitive bias.

[0014] Furthermore, the steps for generating the adjustment strategy are as follows: 1) Matching strategy template based on the two-way evaluation result L : in, For strategy templates and the semantic relevance of the bidirectional evaluation result L, is a policy template collection, For the nth strategy template, is the preset balance weight; 2) The policy parameters include parent policy parameters With child strategy parameters ; The policy templates are respectively related to the parent policy parameters With child strategy parameters Constructing a parent adjustment strategy and a child adjustment strategy, and pushing them to the parent terminal and the child terminal respectively; The optimization steps of the strategy parameters are as follows: 1) After executing the adjustment strategy, the system re-collects the parent-child data and generates new evaluation results ; 2) Based on new evaluation results Calculate the reward values ​​of the parent adjustment strategy and the child adjustment strategy respectively and ; 3) Optimize parent strategy parameters based on objective function With child strategy parameters ; in, is the preset penalty coefficient, and are the benefits of parental and children's regulation strategies, It means to find the mathematical expectation.

[0015] Compared with the prior art, the advantages of the present invention are: 1. This paper deeply integrates subjective psychological scales, multi-source behavioral signals, and contextualized interaction data to construct a feature vector that integrates cognitive bias, physiological synchronization, interaction quality, and other dimensions. This method, for the first time, enables cross-modal correlation analysis between implicit behavioral patterns and explicit psychological cognition in parent-child relationships, significantly improving the comprehensiveness and timeliness of assessment results. 2. The present invention innovatively adopts a dual-branch neural network architecture to analyze the feature vectors of the parent and child sides respectively and output independent scores. By quantifying the two-way cognitive difference value, it can accurately identify the perception deviation of both parties on the same interactive event. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0017] Figure 1 Schematic diagram of the system workflow of the present invention; Figure 2 Generate a schematic diagram for the bidirectional evaluation results of the present invention; Figure 3 generating a schematic diagram for the regulation strategy of the present invention; Figure 4 It is a schematic diagram of the technical effect of the present invention. DETAILED DESCRIPTION

[0018] To achieve the above objectives, the present invention is implemented through the following technical solutions: the present invention provides a parent-child two-way relationship assessment and adjustment system that integrates psychological and behavioral data. Figures 1 to 3 , the system comprises: The data collection module and system are divided into a parent side and a child side; based on a preset authorization mechanism, first parent-child data is collected in real time; the first parent-child data is desensitized to generate second parent-child data; the second parent-child data is pre-processed and time-synchronized to obtain third parent-child data; Interaction context data collection: When parents or children plan to start a meaningful interaction (e.g., starting homework tutoring, starting family game time, or starting dinner conversation), they can actively click the corresponding preset activity type button to initiate an interaction event. Contextual information related to the current interaction event is recorded, including but not limited to timestamp, duration, location (e.g., family living room), frequency, and event type (e.g., dinner together, homework tutoring, game time); Subjective psychological data collection: The system pushes standardized scales with the same structure and complementary perspectives (such as the "Parent-Child Interaction Satisfaction Scale") to parents and children respectively. The scale includes core dimensions such as emotional connection, communication quality, and behavioral understanding, and subjective psychological data is directly collected through the scale.

[0019] Physiological and behavioral data collection: collected through wearable devices (such as smart watches, bracelets, heart rate monitors) or environmental sensors. Physiological and behavioral data include but are not limited to heart rate (HR), heart rate variability (HRV), electrodermal activity (EDA / GSC), and activity level (number of steps, exercise status).

[0020] Audio and visual data collection: This is done through cameras and microphone arrays deployed in designated interaction areas. Audio and visual data includes but is not limited to facial expressions, body language, voice intonation, and conversation content during parent-child interactions. Before the first parent-child data is collected, the system provides authorization interfaces to both the parent and child sides (for parents and children of appropriate age), clearly stating the purpose, scope, and usage of each type of data collected; Parents and children can freely choose which data to authorize, and users can withdraw authorization at any time; For minor children, the data authorization and withdrawal shall be made by the parents; For adult children, data collection requires dual authorization from both parents and children. Either parent or child can cancel the authorized data at any time. The first parent-child data is desensitized on the local device through edge computing. The specific steps are as follows: 1) Based on a preset sensitive field rule library, the system identifies sensitive fields in interaction context data, physiological behavior data, and subjective psychological data (such as the user's name, geographic location, and identifiable background objects). The system uses different technical means to desensitize different sensitive fields (such as differential privacy mechanisms and geographic gridding). 2) Desensitization of audio and video data includes visual desensitization and audio desensitization; The face area in the video is blurred using a pixelated convolution kernel (kernel size k=16×16). The background objects containing sensitive information are identified using the preset YOLOv5 model, and pixel-level random scrambling is applied to the identified area. Extract the Mel-Frequency Cepstral Coefficients (MFCC) from the audio as voiceprint features. When sampling, retain the 12-dimensional coefficients and discard the fundamental frequency features to reduce identifiability. Then add Gaussian white noise for blurring. Preprocessing includes outlier removal and missing value compensation; The mean μ and standard deviation σ of the interaction situation data, subjective psychological data, and physiological behavior data were calculated based on a sliding window (window length W = 10s), and the data of the second parent-child group with deviations from the mean of more than The value of Linear interpolation compensation is used for continuous data; empty value marks are assigned to discrete data (such as questionnaire options).

[0021] Time synchronization includes hardware clock alignment, synchronization anchoring, and discrete event alignment; Hardware clock alignment synchronizes data collection terminals (parent and child devices, wearable sensors, cameras) to a preset master clock (error <10ms) based on the NTP protocol when an interaction event is triggered.

[0022] Triggering synchronization anchoring based on interaction events. The specific steps are as follows: 1) Through the timestamp of the interaction event in the interaction context data (such as the user clicks "Start parent-child chat") as a benchmark; 2) In this embodiment, physiological behavior data is timestamped Cut the 5-minute clips before and after and resample them to a uniform frequency of 1 Hz; The audio and video data is based on the time of the first frame of the video Offset, calculate the actual start time of the interaction event The moment video recording starts The time deviation between and compensate the audio stream; The steps for compensating the audio stream are as follows: 1) Sequence the original timestamp of the audio stream ... Unified translation , so that the start time of the audio stream is consistent with Alignment: in, is the new timestamp after translation; 2) Audio segment cropping and padding: when (video recording lags), intercept the audio stream Data for the period ( is the duration of the interaction event), discard Duration of the head silence segment; when (Video recording in advance): insert in the audio stream header Zero-amplitude silence data (simulating a period of no sound) to ensure that the audio stream is start; 3) Sampling rate consistency guarantee: After compensation, the audio is resampled according to the device's native sampling rate (such as 48kHz) to ensure that the time axis strictly matches the video frame rate (such as 30fps). The length of the audio block associated with each frame of video is Adjusted to: C Where a is , for , C is the duration of each frame.

[0023] A feature extraction and fusion module extracts an audio-visual feature vector, a cognitive bias vector, a time series feature vector, and a parent-child relationship vector based on the third parent-child data and performs a fusion operation to generate a fused feature vector; The extraction of audio-visual feature vectors includes visual feature extraction, audio feature extraction and feature fusion; The visual feature extraction process is as follows: 1) Calculate the dense optical flow field for the desensitized video frame sequence and construct the motion matrix: in, is the motion matrix at time t, The coordinates of the i-th pixel in the face area, and They are Instantaneous displacement in horizontal and vertical directions, for Total exercise intensity; 2) Improve the expanded 3D convolutional neural network by reducing the convolution kernel size from Change to ; Change the number of feature map channels from Changed to 256; Based on the time window, the motion matrix is ​​reorganized into a three-channel tensor (horizontal displacement, vertical displacement and motion intensity); the improved expanded three-dimensional convolutional neural network is input and the output visual feature vector is : in, is the time window length (default 3 seconds); Audio features are extracted through a preset speech emotion recognition model; The speech emotion recognition model is built using the Transformer architecture (Wav2Vec 2.0 type), which includes an input layer, a feature processing layer, an intent parsing layer, and an output layer. By dividing the continuous speech segment into frames (the default frame length is 25ms and the frame shift is 10ms), the time domain features such as energy and zero-crossing rate of each frame are calculated to form a preliminary acoustic feature sequence; Input the desensitized basic acoustic feature sequence (including MFCC, energy, zero-crossing rate and other temporal features); The specific steps of the feature processing layer are as follows: 1) Use convolution operations to extract short-term acoustic patterns (such as phoneme boundaries and formant characteristics) and generate local feature vectors; 2) Analyze the global dependencies of local feature vectors through a multi-head self-attention mechanism to capture long-range speech context features; Output hidden state sequence containing temporal semantics ... .

[0024] The intent parsing layer steps are as follows: 1) Based on Statistical dynamic acoustic properties, including: Pitch fluctuation intensity: Calculate the variance of the hidden state in the spectral dimension; Speech rate quantification: output the number of phonemes per unit time through syllable boundary detection; Silent interval ratio: the ratio of time steps used to identify silent segments; 2) Detecting the activation intensity of predefined emotional keywords (such as "happy" and "angry") in the hidden state; Analyze the state transition pattern between adjacent time steps, mark the interruption events (state mutation) and response delays (state retention time).

[0025] The output layer aggregates the temporal hidden state and intent parsing results to generate 256-dimensional audio features ; Concatenate visual features and audio features to obtain audio-visual feature vectors .

[0026] The mean μ and standard deviation σ of the audio-visual feature vector are calculated based on the sliding window, and the audio-visual feature vectors that deviate from the mean by more than The value of The specific steps for calculating the bidirectional deviation value are as follows: 1) Parents’ cognitive bias calculate: Based on subjective psychological data, calculate the absolute difference in ratings: in, is the absolute difference of the i-th question, is the parent's rating of their own behavior for question i, For question i, the children's rating of their parents' behavior; Average the absolute differences across all questions and add the time weight: in, is the time weight coefficient, Submission time for the scale relative to the base time The time offset of 2) Calculate the children’s cognitive bias value in the same way ; 3) The calculation formula of the time weight coefficient is as follows: in, is the preset attenuation coefficient, and e is a natural constant; 4) Generate a two-dimensional cognitive bias vector : Input physiological behavior data into the double-layer LSTM to extract the time series feature vector; The two-layer LSTM (also known as stacked LSTM) adopts a hierarchical architecture. Its core principle is to enable the network to learn more complex and deep feature representations and abstract patterns in the input sequence data by cascading two LSTM layers. The first layer of LSTM performs hidden state calculation on the original input sequence at each time step and outputs the corresponding hidden state sequence; The second layer LSTM performs hidden state calculations time-step by time-step based on the hidden state sequence, and finally generates a high-level hidden state output.

[0027] The steps for generating the parent-child relationship vector are as follows: 1) Set a time window and count the total number of interactions between parents and children in the interaction events within the time window (such as 24 hours), and eliminate the influence of extreme values ​​through logarithmic transformation to obtain the interaction frequency intensity , extreme values ​​refer to data where the number of interactions within the time window deviates from the preset normal range: in, is the total number of interaction events; 2) Calculate the absolute value of the time difference between the child behavior response and the parent behavior initiation in the interaction event, and take the average of the absolute value of the time difference in the time window to obtain the response delay : in, The timestamp of the parent initiating the action, Timestamp for child responses; 3) Count the number of conflicts in interactive events; perform a weighted calculation on the number of conflicts within the time window to obtain the conflict event density X; and add a time decay factor to give more weight to recent conflicts; 4) Increase the frequency and intensity of interaction , response delay and conflict event density X are integrated into the parent-child relationship vector : Interaction frequency and intensity The larger the value, the higher the interaction frequency (positive indicator), and the response delay The larger the value, the more serious the response delay (negative indicator), and the larger the conflict event density X, the higher the conflict frequency (negative indicator).

[0028] The generation process of the fused feature vector is as follows: 1) Map the audio-visual feature vector, cognitive bias vector, temporal feature vector, and parent-child relationship vector to a unified dimension; 2) Generate a fused feature vector by fusing the audio-visual feature vector, cognitive bias vector, time series feature vector, and parent-child relationship vector through weighted features : in, 、 、 and is the preset feature weight ( ), through the experience of experts and based on different family types, for example, for families with high conflict, the weight of cognitive bias is increased; for families with communication alienation, the weight of audio-visual features is increased.

[0029] An evaluation and strategy generation module inputs the fused feature vector into a preset evaluation model to output a bidirectional evaluation result; constructs an adjustment strategy library, where the adjustment strategies in the adjustment strategy library consist of strategy templates and strategy parameters; selects a strategy template from the adjustment strategy library based on the bidirectional evaluation result; and dynamically optimizes the strategy parameters corresponding to the strategy template through a reinforcement learning framework.

[0030] The bidirectional evaluation model is constructed through a two-branch neural network, including an input layer, a separation layer, a two-branch layer, and an output layer; The input layer inputs the fused feature vector; The separation layer will Split into parent view vector With children perspective vector , through the preset feature mask matrix accomplish: in, and is the preset binary mask matrix, is element-wise multiplication; The double-branch layer is divided into a first branch and a second branch; The first branch is based on Output parents' ratings of the relationship ; The second branch is based on Output the child's rating of the relationship ; In some embodiments, the first branch and the second branch are fully connected layers of the same dimension.

[0031] The output layer outputs the bidirectional evaluation results : in, Bidirectional cognitive bias ( The smaller the value, the more consistent the perspectives of both parties are).

[0032] The steps to generate a reconciliation strategy are as follows: 1) Matching strategy template based on the two-way evaluation result L : in, For strategy templates and the semantic relevance of the bidirectional evaluation result L, is a policy template collection, For the nth strategy template, is the preset balance weight; 2) Policy parameters include parent policy parameters With child strategy parameters ; Policy templates and parent policy parameters With child strategy parameters A parent adjustment strategy and a child adjustment strategy are formed and pushed to the parent end and the child end respectively.

[0033] The optimization steps of the strategy parameters are as follows: 1) After executing the adjustment strategy, the system re-collects parent-child data and generates new evaluation results ; 2) Based on new evaluation results Calculate the reward values ​​of parent regulation strategy and child regulation strategy respectively and : in, and g is the preset weight coefficient; 4) Optimize parental strategy parameters based on the objective function by maximizing the long-term parent-child relationship improvement effect and minimizing the difference in regulatory response between parents and children With child strategy parameters ; in, is the preset penalty coefficient, and are the benefits of parental and children's regulation strategies, It means to find the mathematical expectation.

[0034] In summary, after matching the strategy template with the results of a two-way evaluation, the present invention uses a reinforcement learning framework to optimize the strategy parameters of parents and children respectively. By continuously tracking the relationship state migration after the strategy is executed, the targeted iteration of the adjustment strategy and the differentiated adaptation of the roles are achieved, which completely solves the problem of mismatch between universal intervention plans and dynamic family needs.

[0035] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A parent-child bidirectional relationship assessment and adjustment system that integrates psychological and behavioral data. The system is divided into parent and child sides, and is characterized by: The system comprises: The data collection module collects first parent-child data in real time based on a preset authorization mechanism; performs desensitization processing on the first parent-child data to generate second parent-child data; and performs preprocessing and time synchronization on the second parent-child data to obtain third parent-child data; A feature extraction and fusion module extracts an audio-visual feature vector, a cognitive bias vector, a time series feature vector, and a parent-child relationship vector based on the third parent-child data and performs a fusion operation to generate a fused feature vector; An evaluation and strategy generation module inputs the fused feature vector into a preset evaluation model to output a bidirectional evaluation result; constructs an adjustment strategy library, where the adjustment strategies in the adjustment strategy library consist of strategy templates and strategy parameters; selects a strategy template from the adjustment strategy library based on the bidirectional evaluation result; and dynamically optimizes the strategy parameters corresponding to the strategy template through a reinforcement learning framework.

2. The system according to claim 1, wherein: The first parent-child data includes interaction situation data, subjective psychological data, physiological behavior data, and audio-visual data; Interaction context data is collected by parents or children through interaction events, including but not limited to timestamp, duration, location, frequency, and event type; Subjective psychological data is collected directly through the scales, which have the same structure and complementary perspectives, pushed to both the parent and child sides. Physiological and behavioral data are collected through wearable devices. Audio and video data are collected through cameras and microphone arrays deployed in designated interactive areas; The parent end, the child end, the wearable device, and the camera and microphone array constitute a data acquisition terminal; The authorization mechanism is to provide authorization interfaces to the parent and child respectively before the first parent-child data is collected, clearly stating the purpose, scope and usage of each type of data collected; Parents and children are free to choose the type of data to authorize and can withdraw authorization; for minor children, parents are responsible for authorizing and withdrawing data; for adult children, dual authorization from both parents and children is required, and either parent or child can cancel the authorized data.

3. The system according to claim 2, characterized in that The specific steps of the desensitization treatment are as follows: 1) Identify sensitive fields in the interaction context data, subjective psychological data, and physiological behavior data based on a preset sensitive field rule library, and use different technical means to desensitize different sensitive fields; 2) Desensitization processing of the audio and video data includes visual desensitization and audio desensitization; The face area in the video is blurred using a pixelated convolution kernel, and background objects containing sensitive information are identified using the preset YOLOv5 model. Pixel-level random scrambling is applied to the identified area. The Mel-frequency cepstral coefficients in the audio are extracted and retained as voiceprint features, and then Gaussian white noise is superimposed to blur the audio.

4. The system according to claim 2, wherein: The preprocessing calculates the mean μ and standard deviation σ of the interactive context data, subjective psychological data and physiological behavior data based on the sliding window, and removes the data that deviates from the mean by more than The value of; linear interpolation compensation is used for continuous data; a null value mark is given to discrete data; The specific steps of the time synchronization are as follows: 1) Synchronizing the data acquisition terminal to a preset master clock through the NTP protocol; 2) Setting the timestamp of the interaction context data as the reference time ; The time axis of the physiological behavior data and audio-visual data is based on the reference time Align to the origin.

5. The system according to claim 2, wherein: The audio-visual feature vector is extracted based on the audio-visual data, including a visual layer, an audio layer and a fusion layer; The visual layer is used to extract visual feature vectors, and the specific steps are as follows: 1) Calculate the dense optical flow field of the video and construct the motion matrix; 2) Improving the convolution kernel size and the number of feature map channels of the original dilated 3D convolutional neural network; processing the motion matrix based on the improved dilated 3D convolutional neural network to output a visual feature vector; The audio layer extracts audio feature vectors based on the speech emotion recognition model of the Transformer architecture. The specific steps are as follows: 1) performing frame processing on the speech to extract basic acoustic features; and inputting the basic acoustic features into the speech emotion recognition model; 2) The speech emotion recognition model captures short-term acoustic patterns through convolution operations to generate local feature vectors. The multi-head self-attention mechanism then analyzes the contextual dependencies of the local feature vectors. The model then combines dynamic acoustic properties and emotion keyword detection to output an audio feature vector. The fusion layer concatenates the visual feature vector and the audio feature vector to obtain an audio-visual feature vector; The mean μ and standard deviation σ of the audio-visual feature vector are calculated based on the sliding window, and the vectors that deviate from the mean by more than The value of .

6. The system according to claim 2, wherein: The cognitive bias vector is extracted based on subjective psychological data, and the specific steps are as follows: 1) Parents’ cognitive bias calculate: Average the absolute differences of all questions in the scale and add the time weight: in, is the preset attenuation coefficient, e is a natural constant, is the absolute difference of the i-th question, Submission time for the scale relative to the base time The time offset of 2) Through the parental cognitive bias value Calculation method for calculating children's cognitive bias value ; 3) Through the parental cognitive bias value and children's cognitive bias Construct a cognitive bias vector.

7. The system according to claim 2, wherein: The physiological behavior data is input into a double-layer LSTM to extract the time series feature vector.

8. The system according to claim 2, wherein: The parent-child relationship vector is extracted based on the interaction scenario data, and the specific steps are as follows: 1) Set a time window and count the total number of interactions between parents and children in the interaction events within the time window, and eliminate the influence of extreme values ​​through logarithmic transformation to obtain the interaction frequency intensity , the extreme value refers to the number of interactions within the time window that deviates from the preset normal range: in, is the total number of interactions between parents and children in the interaction event; 2) Calculate the absolute value of the time difference between the child behavior response and the parent behavior initiation in the interaction event, and take the average value of the time difference in the time window to obtain the response delay : in, The timestamp of the parent initiating the action, Timestamp for child responses; 3) Count the number of conflicts in the interaction events; weight the number of conflicts within the time window to calculate the conflict event density X; and add a time decay factor to give more weight to recent conflicts; 4) Increase the frequency and intensity of interaction , response delay and the conflict event density X are integrated into the parent-child relationship vector; Map the audio-visual feature vector, cognitive bias vector, time series feature vector and parent-child relationship vector to a unified dimension; generate a fused feature vector by weighted feature fusion .

9. The system according to claim 1, wherein: The evaluation model is constructed by a dual-branch neural network, including an input layer, a separation layer, a dual-branch layer and an output layer; The input layer inputs the fused feature vector ; The separation layer will Split into parent view vector With children perspective vector , through the preset feature mask matrix accomplish: The double-branch layer is divided into a first branch and a second branch; The first branch is based on Output parents' ratings of the relationship ; The second branch is based on Output the child's rating of the relationship ; The output layer outputs the bidirectional evaluation results in, It is a two-way cognitive bias.

10. The system according to claim 9, characterized in that The steps for generating the adjustment strategy are as follows: 1) Matching strategy template based on the two-way evaluation result L : in, For strategy templates and the semantic relevance of the bidirectional evaluation result L, is a policy template collection, For the nth strategy template, is the preset balance weight; 2) The policy parameters include parent policy parameters With child strategy parameters ; The policy templates are respectively related to the parent policy parameters With child strategy parameters Constructing a parent adjustment strategy and a child adjustment strategy, and pushing them to the parent terminal and the child terminal respectively; The optimization steps of the strategy parameters are as follows: 1) After executing the adjustment strategy, the system re-collects the parent-child data and generates new evaluation results ; 2) Based on new evaluation results Calculate the reward values ​​of the parent adjustment strategy and the child adjustment strategy respectively and ; 3) Optimize parent strategy parameters based on objective function With child strategy parameters ; in, is the preset penalty coefficient, and are the benefits of parental and children's regulation strategies, It means to find the mathematical expectation.

Images