Infant affective interaction robot system based on multi-modal context perception and adaptive decision

Abstract

The application provides a kind of infant emotional interaction robot system based on multi-modal situation awareness and adaptive decision, it relates to artificial intelligence and computer technology field, including first data processing unit, second data processing unit, regulator unit, feedback control unit and decision output unit, the application realizes the profound control effect through closed loop feedback architecture;Using the unified predictive emotional risk index as a dynamic control signal, the model parameters and macro interaction behavior strategy of the system are nonlinearly modulated;Enhanced the system's early capture ability of key emotional triggers in complex environments;Improved the forward-looking and appropriateness of the robot interaction behavior, reduced the reaction delay;Realize the intelligent on-demand allocation of computing resources, improve the overall robustness and safety of the system without increasing the average power consumption.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer technology, specifically to a preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making. Background Technology

[0002] With the deep integration of artificial intelligence and robotics, emotional interaction robot systems are evolving from simple command responses to complex interactions capable of understanding and adapting to human emotional states. Driven by technologies such as computer vision and speech analysis, utilizing computer systems based on specific computational models (neural networks) to process multimodal data streams such as images and sounds to achieve the perception and understanding of human emotions has become a research hotspot in this field, showing significant application prospects, especially in early childhood care where refined and personalized interaction is required. Existing technologies typically face the following challenges in achieving efficient and accurate emotional interaction with young children:

[0003] In many existing solutions, the output of the context awareness and sentiment analysis module and the input of the robot's behavior decision-making module often have a one-way, linear information transmission relationship. This mechanism causes the robot's behavior strategy adjustment to lag behind its perception results, making it difficult to achieve proactive and preventative interaction based on the prediction of sentiment state trends.

[0004] The underlying diagnostic models in current emotional interaction systems (e.g., algorithms for event detection or causal inference) typically have their operating parameters (such as event detection thresholds and analysis time depth) set to fixed values ​​after deployment. This static configuration makes it difficult to adapt to dynamically changing situations. When key emotional precursor signals are weak, the system may miss the best intervention opportunity due to insufficient sensitivity, while under normal conditions, excessive sensitivity may introduce unnecessary computational overhead and noise interference. Summary of the Invention

[0005] The purpose of this invention is to provide a preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The preschool emotional interaction robot system based on multimodal situational perception and adaptive decision-making includes a first data processing unit, a second data processing unit, a regulator unit, a feedback control unit, and a decision output unit.

[0008] The first data processing unit is configured to acquire multimodal data streams in real time and generate first data for immediate emotion attribution based on cross-modal causal inference processing of the multimodal data streams.

[0009] The second data processing unit is configured to receive the first data and, based on the dynamic analysis of the first data, generate second data for predicting trend sentiment states.

[0010] The regulator unit is configured to receive the second data and map the second data to the configuration instructions of the diagnostic model according to a preset diagnostic strategy mapping table.

[0011] The feedback control unit is configured to receive configuration instructions from the diagnostic model and dynamically adjust the operating parameters of the first data processing unit based on the configuration instructions;

[0012] The decision output unit is configured to control the interaction strategy of the emotional interaction robot based on the updated second data used to trigger this decision and the newly generated first data after dynamic adjustment.

[0013] Furthermore, the first data processing unit identifies events in the multimodal data stream as nodes in a time sequence graph, and generates first data by calculating the causal transmission relationship between nodes, wherein the first data is a cognitive focus graph sequence.

[0014] Furthermore, the multimodal data stream includes visual data stream, voice data stream, tactile data stream, and environmental state data stream.

[0015] Furthermore, the second data processing unit performs statistical analysis on the topological structure of the cognitive focus map sequence within a sliding time window to calculate macro-dynamic indicators, and generates second data based on the macro-dynamic indicators, wherein the second data is an emotional breakdown risk index.

[0016] Furthermore, the calculation process for the emotional breakdown risk index is as follows:

[0017] The input is a multimodal data stream provided by a complete set of multi-channel sensors integrated into the emotional interaction robot system, synchronized with time stamps.

[0018] The system performs event recognition on each data stream in parallel, and continuously generates event nodes with modal labels, timestamps and feature vectors based on preset visual event trigger thresholds, voice event trigger thresholds, contact thresholds and environmental fluctuation thresholds.

[0019] Event nodes of all modalities generated in the recent period are placed into a unified graph structure; the graph neural network model is used to calculate the causal transit relationship between any two temporally adjacent nodes, and the output is the weight of the directed edge connecting the two nodes; all nodes and weighted edges together constitute the cognitive focus graph sequence at the current time point.

[0020] Calculate and output the causal relaxation time for the input baseline cognitive focus map sequence;

[0021] The system maintains a buffer that stores the cognitive focus map sequence of the past minute; based on the cognitive focus map sequence of the past minute, it calculates and outputs the full-spectrum modal mutual information gain.

[0022] Causal relaxation time and full-spectrum modal mutual information gain are used as observation variables and input into the pre-trained dynamic Bayesian network inference engine;

[0023] The pre-trained dynamic Bayesian network inference engine infers the posterior probability distribution of state nodes representing the true emotional stability of young children based on the observed causal relaxation time and the value of full-spectrum modal mutual information gain.

[0024] The posterior probability value of the unstable state is normalized and then output as the final emotional breakdown risk index.

[0025] Furthermore, the regulator unit employs an emotional homeostasis adaptive regulator, and the built-in diagnostic strategy mapping table defines a nonlinear mapping relationship between the emotional breakdown risk index and the event detection threshold in the first data processing unit.

[0026] Furthermore, the calculation process for the emotion steady-state adaptive regulator to generate the final configuration instruction after receiving the emotion collapse risk index is as follows:

[0027] Enter the real-time updated risk index of emotional breakdown;

[0028] Call the non-linear mapping function to convert the input emotional breakdown risk index into the diagnostic resource budget for the current period; initialize the set of configuration instructions to be executed, containing the current diagnostic parameter values;

[0029] The process enters a cyclical decision-making process, which terminates when the remaining diagnostic resource budget is insufficient to cover any effective parameter adjustments.

[0030] In each iteration, all adjustable diagnostic parameters are iterated over, and candidate adjustment actions are generated for each parameter; then, the benefit-cost ratio is calculated for each candidate adjustment action.

[0031] From all candidate adjustment actions, select the adjustment action with the highest benefit-cost ratio as the optimal action for this cycle;

[0032] Determine whether the parameter adjustment cost of the optimal action is less than or equal to the current remaining diagnostic resource budget;

[0033] If the determination is yes, execute the optimal action, that is, update the value of the corresponding parameter in the set of configuration instructions to be executed; subtract the cost of the optimal action from the current diagnostic resource budget; the process returns to the loop stage and begins the next optimization loop;

[0034] If the result is negative, it indicates that the remaining budget is insufficient to cover any worthwhile adjustments, and the optimization loop terminates.

[0035] The final output is the set of configuration instructions to be executed when the optimization loop terminates; the set of configuration instructions to be executed contains a set of new diagnostic parameter target values ​​that have been co-optimized and will be sent to the feedback control unit to update the actual operating status of the first data processing unit.

[0036] Furthermore, the feedback control unit is configured to apply the configuration instructions of the diagnostic model to the first data processing unit; and the decision output unit is configured to select the corresponding interaction strategy mode based on the emotional breakdown risk index, and generate specific robot behavior instructions based on the updated first data.

[0037] Furthermore, the complete calculation and execution steps of robot behavior instructions are as follows:

[0038] The input consists of two parallel, real-time updated data streams: a set of configuration instructions from the diagnostic model of the emotion homeostasis adaptive regulator, and an emotion breakdown risk index from the second data processing unit.

[0039] After receiving the configuration instruction set of the diagnostic model, the feedback control unit immediately applies the configuration instruction set to the first data processing unit; this is an atomic operation, that is, the new parameter values ​​in the configuration instruction set are overwritten to the corresponding configuration register inside the first data processing unit; this ensures that the diagnostic engine can immediately run with higher sensitivity or a deeper level of analysis.

[0040] At the same time, the decision output unit executes the following two parallel sub-steps:

[0041] Based on the input emotional breakdown risk index, the threshold state machine is invoked to determine the current strategic interaction mode;

[0042] Obtain the latest cognitive focus map sequence output by the updated first data processing unit, and calculate the attribution focus vector;

[0043] Entering the core decision-making logic is achieved through a preset strategic and tactical decision matrix. Based on the current strategic interaction mode and attribution focus vector, the corresponding cell is found in the strategic and tactical decision matrix. The cell stores the final action instruction primitive to be executed.

[0044] Before outputting the final instruction, perform a sanity check; check whether the timestamp of the emotional breakdown risk index is within the preset data freshness window.

[0045] If the determination is yes, the process continues normally;

[0046] If the judgment is negative, it indicates that the prediction has failed; the degradation mechanism is triggered; the strategic interaction mode is forced to be set to the predefined standard response mode, but the decision output unit still executes two parallel sub-steps; during decision-making, the most conservative action instruction primitive is selected only in the row corresponding to the standard response mode of the strategic and tactical decision matrix according to the attribution focus vector.

[0047] The final output is to send the selected behavioral instruction primitives to the underlying behavior executor of the emotional interaction robot through a standardized interface.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This invention introduces an emotional homeostasis adaptive regulator as the core control unit of the system; the emotional homeostasis adaptive regulator uses the predicted emotional breakdown risk index as the sole decision-making basis to generate a control effect;

[0050] The inward control effect of the emotional steady-state adaptive regulator of this invention transforms the macroscopic emotional breakdown risk index into precise parameter configuration instructions for the first data processing unit (microscopic diagnostic model) at the bottom layer of the system through a preset diagnostic strategy mapping table. This enables the system's perception ability to intelligently and dynamically adjust the sensitivity and analysis depth of the diagnostic model according to the predicted emotional change trend. When an increased risk is predicted, the system will automatically sharpen its insight to capture subtle triggers that may have been previously overlooked. When the risk is low, it will return to normal sensitivity to save computing resources.

[0051] The outward control effect of the emotional homeostasis adaptive regulator of this invention synchronously provides the emotional breakdown risk index to the decision output unit for selecting the overall robot interaction strategy that matches the current risk level. This means that the robot's behavior is no longer just a passive response to events that have already occurred, but an active adjustment based on the prediction of future states. That is, the emotional breakdown risk index will drive the robot to switch from the regular play mode to the forward-looking preventive soothing mode, and call on more refined attribution data provided by the sharpened perception system to guide specific behaviors.

[0052] This invention features a unique architecture that synchronously couples adaptive adjustment of perception capabilities with forward-looking selection of robot behavior strategies, achieving a fundamental shift from passive response to active adaptation. The system's perception and actions are driven by a unified and forward-looking internal state assessment, forming an organic whole. This enhances the system's understanding of children's emotional changes and the effectiveness of interaction in complex and dynamic scenarios. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the integration of physical scene and information flow in this invention;

[0054] Figure 2 This is a schematic diagram of the overall system structure of the present invention;

[0055] Figure 3 This is a schematic diagram illustrating the calculation process steps of the emotional breakdown risk index of this invention.

[0056] Figure 4 This is a schematic diagram of the calculation process steps for the final configuration instruction of this invention;

[0057] Figure 5 This is a schematic diagram illustrating the complete calculation and execution steps of the robot behavior instructions of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0059] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0060] Example 1:

[0061] Please see Figures 1 to 5 The present invention provides a technical solution: a preschool emotional interaction robot system based on multimodal situation perception and adaptive decision-making, comprising a first data processing unit, a second data processing unit, a regulator unit, a feedback control unit and a decision output unit;

[0062] The first data processing unit is configured to acquire multimodal data streams in real time and generate first data for immediate emotion attribution based on cross-modal causal inference processing of the multimodal data streams.

[0063] The second data processing unit is configured to receive the first data and, based on the dynamic analysis of the first data, generate second data for predicting trend sentiment states.

[0064] The regulator unit is configured to receive the second data and map the second data to the configuration instructions of the diagnostic model according to a preset diagnostic strategy mapping table.

[0065] The feedback control unit is configured to receive configuration instructions from the diagnostic model and dynamically adjust the operating parameters of the first data processing unit based on the configuration instructions;

[0066] The decision output unit is configured to control the interaction strategy of the emotional interaction robot based on the second data used to trigger this decision and the first data newly generated after dynamic adjustment.

[0067] The first data processing unit identifies events in the multimodal data stream as nodes in a time sequence graph, and generates first data by calculating the causal transmission relationship between nodes, wherein the first data is a cognitive focus graph sequence;

[0068] The first data processing unit receives multimodal data streams from a built-in microphone array, CMOS camera, pressure sensor, temperature and humidity sensor, and ambient light sensor. The microphone array is used to collect the child's voice, locate the sound source, and perform environmental noise reduction. The CMOS camera captures the child's facial expressions, body posture, and ambient light. The pressure sensor detects the child's hugging, patting, and other contact behaviors. The temperature and humidity sensor and ambient light sensor help determine the child's comfort level. The event detection module built into the first data processing unit identifies changes in the data stream as event nodes and assigns a timestamp and feature vector to each node, based on graph neural networks. The network model calculates the weights of directed connections between adjacent timestamp nodes, which characterize the strength of causal transmission between nodes. The time-series graph composed of weighted nodes and edges is the cognitive focus graph sequence. The first data processing unit continuously outputs the sequence of cognitive focus graphs as the first data. The abstract attribution of immediate emotions is concretized into a causal inference method based on graph networks. Compared with traditional feature fusion classification methods, by constructing cognitive focus graphs, the dynamic relationship between external stimuli and children's internal responses can be revealed more profoundly. This provides a structured input rich in causal information for subsequent trend analysis, improving the accuracy and interpretability of attribution.

[0069] Multimodal data streams include visual data streams, speech data streams, tactile data streams, and environmental state data streams;

[0070] The visual processing module in the first data processing unit acquires a sequence of video frames from the CMOS camera mounted on the robot's head as a visual data stream. The visual processing module applies a facial key point detection algorithm (in this embodiment, based on, but not limited to, algorithms and deep learning models from the Dlib library in the regression tree ensemble method) to the facial region in the video frame sequence, tracking and outputting the coordinates of multiple key points representing facial expressions in real time. When the visual processing module detects the coordinate displacement velocity of three key points—the left corner of the mouth, the right corner of the mouth, and the center point of the chin tip—the corner of the mouth will rapidly shift laterally or vertically in response to strong emotions such as laughing, crying, and surprise. The chin tip will significantly shift downwards during mouth-opening actions (in this embodiment, surprise and shouting). The triangular region formed by these three key points can very sensitively capture most abrupt changes in facial expression events. Compared to subtle expression areas such as the eyes, the displacement is more macroscopic, the signal-to-noise ratio is higher, and it is more conducive to stable detection. Furthermore, within three consecutive sampling periods (in this embodiment, the period is set to 0.1s), the velocity exceeds the preset visual event trigger threshold. When this happens, an expression mutation event node is generated;

[0071] All visual data streams are processed in real time within the local first data processing unit of the robot terminal; the visual processing module only extracts and outputs anonymous key point coordinate data without biometric information and event nodes generated based on it; the original visual data stream is not stored or uploaded to the cloud after processing, thus physically cutting off the path of leakage of the original facial images; in scenarios where data transmission or storage is necessary (in this embodiment, for algorithm debugging or model optimization), the data will be strictly desensitized; in this embodiment, only the relative displacement vectors of key points are retained, not the absolute coordinates in the image; or the face area is blurred or cartoonized to ensure that the individual's identity cannot be reverse-identified; the CMOS camera is only activated when the presence of a child is detected and emotion analysis is required, following the principles of minimization and necessity of data collection; when in standby or without interaction, the camera is physically turned off or in sleep mode.

[0072] The audio processing module within the first data processing unit acquires multi-channel audio signals as a speech data stream from a microphone array integrated into the robot's torso. The audio processing module performs short-time energy analysis on the speech data stream. When the audio processing module detects that the increment of the calculated energy value within a 100-millisecond time window exceeds a preset speech event trigger threshold... When this occurs, a sound pulse event node is generated, and the peak intensity is recorded;

[0073] The tactile signal processing module in the first data processing unit acquires real-time pressure distribution map data as a tactile data stream based on a matrix pressure sensor array deployed on the robot's arm or abdomen. The tactile signal processing module performs a total pressure integral calculation on the pressure distribution map. During the time period from when the child touches the robot to when they leave, the tactile signal processing module detects that the total pressure value drops from below the contact threshold. The state jumps to the contact threshold When the above conditions are met, an object contact event node is generated, and the contact position and peak pressure are recorded.

[0074] The environmental status monitoring module within the first data unit acquires environmental status data streams from the following sensors: continuous readings of ambient temperature and humidity from temperature and humidity sensors; and continuous readings of ambient light intensity from ambient light sensors. The environmental status monitoring module independently calculates the first derivative for each time series of environmental data streams to monitor the rate of change in real time. If the absolute value of the first derivative of any data stream exceeds a preset environmental fluctuation threshold for a period exceeding 500 milliseconds... When (in this embodiment, the rate of change of light intensity exceeds the light sudden change threshold, or the rate of change of temperature exceeds the temperature sudden change threshold), a corresponding environmental event node, an ambient light sudden change event node, or an ambient temperature sudden change event node is generated, and the direction and magnitude of the change are recorded.

[0075] In this embodiment, the visual event trigger threshold In this embodiment, the value is 0.85; the voice event trigger threshold. In this embodiment, the value is 0.9; contact threshold. Used to distinguish between meaningless light touches and intentional interactive contact; in this embodiment, the value is set to 10% of the sensor's range; it can effectively filter noise such as sensor drift and airflow, while maintaining high sensitivity to behaviors such as active grasping or leaning by young children; environmental fluctuation threshold. Tailored to different environments; in this embodiment, the light change threshold is set to a rate of change of 500 lux per second; corresponding to typical light change events that can affect a child's mood, such as switching lights on or off indoors or opening and closing curtains; the temperature change threshold is set to a rate of change of 0.5 degrees Celsius per minute; exceeding the range that a child's body can comfortably adapt to through its own physiological regulation in a resting state; corresponding to events such as the sudden opening of doors and windows leading to the intrusion of cold / hot airflow, or the start / stop of the air conditioning system, which are common physical causes that can cause physical discomfort and may lead to emotional irritability.

[0076] The first data processing unit transforms the raw multimodal data stream into event nodes with clear semantic labels and quantified attributes, providing structured input for the subsequent construction of cognitive focus map sequences.

[0077] The second data processing unit performs statistical analysis on the topological structure of the cognitive focus map sequence within a sliding time window to calculate macro-dynamic indicators, and generates second data based on these macro-dynamic indicators, whereby the second data is an emotional breakdown risk index.

[0078] The second data processing unit receives the cognitive focus map sequence output by the first data processing unit; within a preset sliding time window (in this embodiment, the sliding time window is set to 5 minutes), it uses graph theory algorithms to calculate two macroscopic dynamic indicators of the cognitive focus map sequence in real time: cognitive coupling entropy, used to quantify the disorder of causal connections in the graph; and perceptual desensitization slope, used to quantify the changing trend of the intensity of the child's response to environmental events; the cognitive coupling entropy and perceptual desensitization slope are input into a pre-trained time series prediction model, and the time series prediction model outputs a continuous value between 0 and 1, which is the emotional breakdown risk index, as the second data; the prediction of trend-based emotional states is concretized into topological dynamic analysis of the causal relationship graph, which can extract macroscopic emotional system stability characteristics from microscopic event associations.

[0079] The computation steps of the graph theory algorithm are as follows:

[0080] The graph theory algorithm traverses all cognitive focus graphs generated within a sliding time window (set to 5 minutes in this embodiment); extracts the weight values ​​of all directed edges in the cognitive focus graph to form a large set of weight values; since the weight values ​​of the edges themselves have been normalized to between 0 and 1;

[0081] The weight value range from 0 to 1 is divided into a fixed number of probability intervals of equal width. In this embodiment, to obtain sufficient statistical resolution, the number of intervals is set to one hundred (i.e., intervals [0, 0.01), [0.01, 0.02), ..., [0.99, 1.0]).

[0082] Graph theory algorithms count how many values ​​in the set of weights fall into each probability interval, thus obtaining the frequency of each interval;

[0083] Divide the frequency of each probability interval by the total number of weight values ​​to obtain the probability value of the probability interval; the sum of the probability values ​​of all intervals should equal 1 to form a complete probability distribution;

[0084] Implement the calculation of Shannon entropy in information theory; initialize the accumulator with a graph theory algorithm, with an initial value of zero;

[0085] The graph theory algorithm traverses each probability interval. For each interval with a non-zero probability value, it performs the following operations: obtain the probability value of the probability interval; calculate the logarithm of the probability value to the base 2; multiply the probability value and the logarithm; and subtract the product from the accumulator.

[0086] After traversing all intervals, the final value in the accumulator is determined as the cognitive coupling entropy of the current time window.

[0087] The desensitization slope is used to quantify the changing trend of a child's response intensity to stimuli from the external environment within a time window. A continuously decreasing trend (negative slope) indicates that the child is becoming desensitized or introverted; conversely, an increasing trend (positive slope) indicates that the reaction is becoming more intense, i.e., hypersensitivity. The calculation steps are as follows:

[0088] Graph theory algorithms can traverse all cognitive focus graphs within a sliding time window;

[0089] In each cognitive focus map, identify all nodes whose node type is environmental state event (in this embodiment, nodes generated by sudden changes in light, sound, and temperature).

[0090] For each identified environmental state event node, the instantaneous response strength is calculated. The instantaneous response strength is defined as the sum of the weights of all directed edges originating from the environmental state event node. The sum quantifies the extent to which the environmental event triggers other subsequent cognitive events (in this embodiment, other cognitive events include facial expressions, actions, sounds, etc.).

[0091] Each calculated instantaneous response intensity is combined with the timestamp of the corresponding environmental state event node to form a data point, namely, timestamp and response intensity.

[0092] Over the entire 5-minute sliding time window, all such data points together constitute the time series of the response strength;

[0093] The graph theory algorithm performs least squares linear regression analysis on the generated time series of response intensity; and fits a straight line that best describes the overall trend of the time series of response intensity.

[0094] The slope of the fitted straight line is the slope of the perception desensitization that is ultimately determined for the current time window.

[0095] The architecture of the time series forecasting model is as follows:

[0096] Input layer: Receives two-dimensional time series data. Each time step of the time series data contains two features: the currently calculated cognitive coupling entropy and the perceptual desensitization slope;

[0097] LSTM layer: contains one or more LSTM layers; in this embodiment, it is set as two stacked LSTM layers, each containing 64 hidden units; the first layer receives the input sequence and outputs the complete sequence to the second layer, the second layer further processes it and only outputs the hidden state of the last time step;

[0098] Fully connected layer: The hidden state of the last time step output of the second LSTM layer is input into a fully connected layer containing 32 neurons, and the ReLU activation function is used for non-linear transformation to integrate the features learned by the LSTM layer.

[0099] Output layer: Contains a single neuron and uses the Sigmoid activation function; the Sigmoid function has the characteristic that the output value range is strictly between 0 and 1; this ensures that the emotional breakdown risk index output by the time series prediction model is a probability value that conforms to physical normalization.

[0100] Training steps for a time series prediction model:

[0101] Collect a large-scale synchronous multimodal dataset containing long-term behaviors of young children in diverse scenarios (in this embodiment, it includes video, audio, environmental data, etc.).

[0102] Data annotation: At least three child development psychology experts were invited to independently annotate the collected synchronous multimodal dataset back-to-back; the annotation task was to provide a continuous risk value between 0 (complete calm) and 1 (breakdown) for the emotional state of the children every second, forming a time series of real risk index; the final annotation result was the average of the three experts.

[0103] Feature engineering: The collected synchronous multimodal dataset is processed by the first data processing unit and graph theory algorithm to generate time series of cognitive coupling entropy and perceptual desensitization slope that are time-aligned with it;

[0104] Sample construction: The feature data and labeled data are sliced; each training sample consists of a fixed-length (in this embodiment, the past 5 minutes) sequence of [cognitive coupling entropy, perceptual desensitization slope] as input, and the corresponding real risk index at the end of the sequence as a label;

[0105] Loss function definition: The mean squared error is used as the loss function; the loss function measures the accuracy of a single prediction by calculating the square of the difference between the risk index predicted by the time series prediction model and the true risk index labeled by experts.

[0106] Optimizer selection: The Adam optimizer is used. The Adam optimizer is a gradient descent algorithm with an adaptive learning rate. It is used to update the weight parameters inside the time series prediction network efficiently and stably based on the gradient calculated from the loss function. The goal is to minimize the value of the loss function over the entire training set.

[0107] Iterative training: The constructed training samples are input into the untrained time series prediction model in batches; the time series prediction model performs forward propagation and gives a predicted risk index for each input sample;

[0108] The predicted values ​​of the time series prediction model are compared with the true labels, and the average loss of the current batch is calculated using the mean squared error loss function. Backpropagation is performed to transmit the loss gradient from the output layer back to the input layer, and the gradient of each weight parameter is calculated.

[0109] The Adam optimizer performs a small update on all weights of the time series prediction model based on the calculated gradient.

[0110] Repeat the above steps until the loss of the time series prediction model on the independent validation set no longer decreases (in this embodiment, an early stopping strategy is used to prevent overfitting), marking the completion of the training process.

[0111] In this embodiment, the key parameters involved are as follows:

[0112] The parameter symbol of the event node is set to Event Node These are discrete spatiotemporal points with specific semantics identified in multimodal data streams, and are the basic units that constitute the cognitive focus map sequence.

[0113] The sign of the parameter for causal relaxation time is set to It quantifies the time required for a cognitive focus map sequence to recover to its pre-disturbance steady state after being subjected to a tiny internal disturbance; it directly reflects the dynamic stability and resilience of the child's cognitive system. The larger the value, the closer the system is to the instability threshold.

[0114] The calculation logic is as follows: Obtain the baseline cognitive focus map sequence at the current time point; apply a virtual perturbation to the baseline cognitive focus map sequence, which is achieved by randomly selecting an edge in the graph and instantaneously increasing the weight by a fixed perturbation intensity factor. To achieve this, at each subsequent time step, the system continuously generates new cognitive focus map sequences and calculates the normalized graph edit distance between the new cognitive focus map sequences and the baseline cognitive focus map sequences. The system starts counting until the normalized graph edit distance is reached. The price first fell back to the preset steady-state recovery threshold. When the count reaches the specified value, stop counting; multiply the count value by the system's time step to obtain the total duration, which is the causal relaxation time.

[0115] Disturbance intensity factor Used to control the magnitude of the virtual disturbance; in this embodiment, the value is set to 0.5; the disturbance needs to be sufficient to produce an observable response in the system, but not so large as to completely destroy the original structure of the system, resulting in irrecoverability; it is determined through simulation experiments using historical steady-state data; steady-state recovery threshold. This is used to define the criteria for determining whether the system has returned to a steady state; in this embodiment, the value is set to 0.05; representing the reasonable range of the cognitive focus map sequence under normal fluctuations, a value lower than this indicates that the impact of the disturbance has basically dissipated.

[0116] The parameter sign of the full-spectrum modal mutual information gain is set to It quantifies the rate of change in the overall information association strength among all different types of event subgraphs (visual, auditory, tactile, and environmental) within a specific time window; it reflects the degree of coupling of the entire sensory system of young children, and a sharp positive increase in the value means that there is a global over-association in sensory processing.

[0117] The calculation logic is as follows:

[0118] The first step involves the system extracting all visual, auditory, tactile, and environmental event nodes and connections from the cognitive focus map sequence within a sliding time window, forming four independent subgraphs. Based on the node degree and edge weight distribution of each subgraph, the graph structure entropy is calculated (i.e., the visual graph structure entropy is calculated separately). auditory map structure entropy tactile image structure entropy Entropy of the environment graph structure Merge all four subgraphs and calculate the joint graph structure entropy of the merged joint graph. By increasing the visual graph structure entropy auditory map structure entropy tactile image structure entropy Entropy of the environment graph structure Add them together, then subtract the joint graph structure entropy. Obtain the instantaneous full-spectrum modal mutual information of the current window. Calculate the difference in instantaneous mutual information between the current window and the previous window, and divide the difference in instantaneous mutual information by the length of the time window to obtain the full-spectrum modal mutual information gain. The time window length in the calculation process is set to 60 seconds in this embodiment, because the window needs to be long enough to smooth out short-term noise, and short enough to capture meaningful trend changes.

[0119] In this embodiment, the calculation process for the emotional breakdown risk index is as follows:

[0120] The input is a multimodal data stream provided by a complete set of multi-channel sensors integrated into the emotional interaction robot system (in this embodiment, the multi-channel sensors include a CMOS image sensor, a microphone array, a pressure sensor, a temperature and humidity sensor, and an ambient light sensor) and synchronized with time stamps.

[0121] Event recognition is performed on each data stream in parallel, based on a preset visual event trigger threshold. Voice event trigger threshold Contact threshold and environmental fluctuation threshold It continuously generates event nodes with modal tags (in this embodiment, modal tags are visual / auditory / tactile / environmental), timestamps, and feature vectors. ;

[0122] All modal event nodes generated in the recent period It is placed in a unified graph structure; the graph neural network model is used to calculate the causal transit relationship between any two temporally adjacent nodes, and the output is the weight of the directed edge connecting the two nodes; all nodes and weighted edges together constitute the cognitive focus graph sequence at the current time point;

[0123] Calculate and output the causal relaxation time from the input baseline cognitive focus map sequence. Regarding causal relaxation time Read the historical maximum causal relaxation time With the historical minimum causal relaxation time From the currently obtained causal relaxation time Subtract the historical minimum causal relaxation time The difference is obtained; then the difference is divided by the historical maximum causal relaxation time. With the historical minimum causal relaxation time The difference; to achieve control over causal relaxation time Perform normalization processing;

[0124] The system maintains a buffer storing the cognitive focus map sequence of the past minute; based on the cognitive focus map sequence of the past minute, it calculates and outputs the full-spectrum modal mutual information gain. For full-spectrum modal mutual information gain Read the historical maximum full-spectrum modal mutual information gain With the historical minimum full-spectrum modal mutual information gain And perform the same computational logic to achieve full-spectrum modal mutual information gain. Perform normalization processing;

[0125] Causal relaxation time and full-spectrum modal mutual information gain They are used as observation variables and input into the pre-trained dynamic Bayesian network inference engine;

[0126] Pre-trained dynamic Bayesian network inference engines are based on observed causal relaxation times. and full-spectrum modal mutual information gain The values ​​are used to infer the posterior probability distribution of state nodes representing the true emotional stability of young children.

[0127] The system normalizes the posterior probability values ​​of the unstable state and outputs them as the final emotional breakdown risk index.

[0128] In this embodiment, the graph neural network model adopts a graph attention network architecture; the specific hierarchical structure of the graph neural network model is as follows:

[0129] Input layer and node feature vector construction: The input to the graph neural network model is a series of event nodes ordered in time; before feeding the nodes into the network, it is necessary to construct a comprehensive initial feature vector.

[0130] The initial feature vector is composed of the following three parts:

[0131] Modal features: One-hot encoded vectors representing the modalities (visual / auditory / tactile / environmental) of a node;

[0132] Temporal characteristics: the timestamp of the node occurrence, and the normalized value;

[0133] Content features: When generating event nodes, the original high-dimensional feature embedding vectors output by the underlying recognition model (in this embodiment, including the facial expression recognition model) are used; through concatenation, each event node is fused to form a rich feature vector;

[0134] The graph attention layer consists of two stacked graph attention layers;

[0135] The working principle of each layer can be described as follows: For any event node in the graph, calculate the attention score of the influence exerted by all neighboring nodes (nodes that are temporally adjacent) on the event node; the score is calculated by linearly transforming the feature vectors of the target node and the neighboring nodes through a learnable weight matrix, and then passing it through a non-linear activation function (in this embodiment, the non-linear activation function is LeakyReLU); the attention score is then normalized by the Softmax function, and its physical meaning is transformed into attention weight, which represents the contribution of each neighboring node to the event node updating its own state;

[0136] The new feature vector of the target node is obtained by weighting and summing the feature vectors of all neighboring nodes according to the attention weights mentioned above;

[0137] By stacking two layers, graph neural network models are able to capture more complex dependencies between nodes over longer distances;

[0138] Output layer: To calculate the causal propagation weights between any two temporally adjacent nodes, the graph neural network model obtains the updated final feature vector from the second graph attention layer; these two feature vectors are concatenated; the concatenated long vector is input into a small feedforward neural network containing two fully connected layers; the last output layer of the feedforward neural network contains only one neuron and uses the sigmoid activation function to ensure that the final output weight values ​​are strictly between 0 and 1, which conforms to the probabilistic meaning of the strength of causal relationship.

[0139] Training steps for a graph neural network model:

[0140] Preparation and annotation of the training dataset; collection of a large-scale synchronous multimodal dataset containing long-term behaviors of young children in diverse scenarios; processing the synchronous multimodal dataset into a sequence of event nodes with complete feature vectors; inviting at least three child development psychology experts to annotate the extracted event node sequences; the annotation task is to examine each pair of temporally adjacent event nodes (in this embodiment, event A occurs at t1, and event B occurs at t2) and give a causal relationship score between 0 (completely unrelated) and 1 (strong causal relationship); in this embodiment, a sudden loud noise in the environment (event A) is followed immediately by a child's face showing a look of fear (event B), and the final annotation result uses the average of the three experts; each training sample consists of an initial feature vector of a pair of adjacent event nodes as input and the causal relationship score annotated by the experts as a label;

[0141] Mean squared error is used as the loss function. The loss function measures the accuracy of prediction by calculating the square of the difference between the causal weights predicted by the model and the causal scores labeled by experts. The Adam optimizer is used to efficiently and stably update all learnable parameters inside the GAT network based on the gradient calculated by the loss function. A standard supervised learning process is adopted, in which the constructed training samples are input into the graph neural network model in batches, and forward propagation, loss calculation, backpropagation and parameter update are performed in a loop. The loss is monitored on an independent validation set and an early stopping strategy is used to prevent the graph neural network model from overfitting until the model converges.

[0142] The pre-trained dynamic Bayesian network inference engine is a probabilistic graphical model. Its core function is to act as a fusion module, receiving two time series indicators representing macroscopic dynamic states and inferring the hidden states that cannot be directly observed and represent the true emotional stability of young children.

[0143] In this embodiment, the pre-trained dynamic Bayesian network inference engine is specifically a hidden Markov model; the hidden Markov model is a special but very effective form of the pre-trained dynamic Bayesian network inference engine, with a clear structure, which is very suitable for modeling hidden states with temporal dependencies.

[0144] Hidden state nodes: In the Hidden Markov Model, there are discrete hidden state variables that represent the emotional stability of the child at any point in time; the emotional stability is divided into three levels: stable, irritable, and unstable.

[0145] Observation Node: The Hidden Markov Model receives two observation variables at each time point: the normalized causal relaxation time and the full-spectrum modal mutual information gain. To be compatible with discrete hidden states, the continuous observations of the causal relaxation time and the full-spectrum modal mutual information gain need to be discretized. In this embodiment, each index is divided into three intervals: low, medium, and high, based on its numerical range. Therefore, at each time point, the state of the observation node is a combination of these two index intervals (in this embodiment, causal relaxation time is high, and full-spectrum modal mutual information gain is medium). In a specific preferred embodiment, the method and results based on the numerical range are as follows:

[0146] Collect at least 1000 hours of expert-annotated multimodal data of young children; calculate two continuous time series data points, causal relaxation time and full-spectrum modal mutual information gain, aligned with the time of the annotated data.

[0147] Causal relaxation time reflects the time required for a cognitive focus map to recover to a stable state from the shock of an event; the shorter the time, the more stable the system's cognition; the longer the time, the more confused the system's cognition, and the higher the risk of emotional instability. Statistical analysis of all calculated causal relaxation times revealed that the values ​​are mainly distributed in the range of 0-500ms. Low range: causal relaxation time < 50ms; corresponding to the top 30% quantile; indicating that the system can resolve causal relationships in a very short time, a typical characteristic of high emotional stability; Middle range: 50ms ≤ causal relaxation time ≤ 200ms; corresponding to the 30% to 80% quantile; the system needs some time to process, and may be in an agitated or transitional state; High range: 200ms < causal relaxation time ≤ 500ms; this range corresponds to the bottom 20% quantile; the system's inability to establish stable causal cognition for a long time is a strong signal of high emotional instability or impending collapse.

[0148] Full-spectrum modal mutual information gain measures the consistency and information coupling between different modal data (in this embodiment, voice, facial expression, and action). A higher full-spectrum modal mutual information gain indicates greater synchronization of physiological and behavioral indicators and a clearer emotional state. A lower full-spectrum modal mutual information gain may indicate contradictory signals, a sign of emotional instability. Full-spectrum modal mutual information gain is mainly concentrated between 0 and 1.0. High range: Full-spectrum modal mutual information gain > 0.7; corresponding to the top 25% quantiles of the data distribution; high synergy among modalities, a typical characteristic of highly stable emotions. Middle range: 0.3 ≤ Full-spectrum modal mutual information gain ≤ 0.7; corresponding to the 25% to 75% quantiles; a certain degree of coupling exists between modalities, corresponding to irritability or a normal state. Low range: 0 ≤ Full-spectrum modal mutual information gain < 0.3; corresponding to the bottom 25% quantiles; severe information imbalance between modalities, a strong signal of highly unstable emotions.

[0149] Transition dependency: The hidden state at the current time point depends only on the hidden state at the previous time point; the dependency relationship is quantified by the state transition probability matrix;

[0150] Emission dependency: The observation at the current time point depends only on the hidden state at the current time point; the dependency is quantified by the observation emission probability matrix;

[0151] When the Hidden Markov Model receives a continuous sequence of observations, the function of the pre-trained dynamic Bayesian network inference engine is to calculate the posterior probability of being in each possible hidden state (stable / irritable / unstable) at each time point.

[0152] In this embodiment, the pre-trained dynamic Bayesian network inference machine adopts the classic forward algorithm; the forward algorithm uses dynamic programming to efficiently and recursively calculate the probability of being in each hidden state given all observations up to the current time; the final output emotional collapse risk index is the posterior probability value of being in an unstable state.

[0153] The training of a pre-trained dynamic Bayesian network inference engine is essentially a process of learning three core probability parameter sets; it requires a large-scale multimodal dataset annotated by experts; experts need to annotate the discrete categories (stable, irritable, or unstable) of the emotional states of young children second by second; the original multimodal dataset is processed into time series of causal relaxation time and full-spectrum modal mutual information gain that are time-aligned with it; and the time series of causal relaxation time and full-spectrum modal mutual information gain are discretized; the training data consists of two sets of parallel sequences: one set is the discretized observation sequence, and the other set is the hidden state sequence annotated by experts;

[0154] The goal of training is to learn the following three parameter matrices:

[0155] Initial state probability vector: the probability of being in each hidden state at the beginning of the sequence;

[0156] State transition probability matrix: the probability of transitioning from one state to another (in this embodiment, the probability of transitioning from an agitated state to an unstable state).

[0157] Observation emission probability matrix: the probability of observing a specific combination of observations (in this embodiment, causal relaxation time - high, mutual information gain - high) when in a certain hidden state (unstable).

[0158] Since there are hidden state sequences annotated by experts, maximum likelihood estimation is directly used for training.

[0159] The algorithm estimates the above probabilities using a direct counting method: it calculates the initial state probability by counting the frequency of each state in the training data as the starting state of the sequence; and it calculates the state transition probability by counting the number of times a state transitions to another state and dividing the result by the total number of times the initial state appears.

[0160] The observation emission probability is calculated by counting the number of times a specific observation value occurs in a specific state and dividing by the total number of times the state occurs. A state counter, which can be a one-dimensional array or a hash table, is created to record the total number of times each hidden state (stable, agitated, unstable) occurs in the training data. An emission counter, which can be a two-dimensional array or a nested hash table, is created to record the number of times a specific combination of observation values ​​is observed in a certain hidden state. There are a total of 3×3=9 possible combinations. The aligned hidden state sequence and the discretized observation sequence are synchronously traversed from time point t=1 to T (in this embodiment, T is the total length of the hidden state sequence and the discretized observation sequence). At each time point t, the hidden state is obtained; the combination of observation values ​​at the current time is obtained; the counter is updated; the emission probability matrix is ​​calculated; after the traversal is completed, the observation emission probability matrix is ​​calculated based on the final result of the counter.

[0161] The regulator unit uses an emotional homeostasis adaptive regulator. The built-in diagnostic strategy mapping table defines the nonlinear mapping relationship between the emotional breakdown risk index and the event detection threshold or causal graph search depth parameter in the first data processing unit.

[0162] The core of the emotional homeostasis adaptive regulator is a lightweight decision logic unit; the received emotional breakdown risk index is used as an independent variable, and the system outputs the configuration instructions of the diagnostic model by consulting or calculating the internal diagnostic strategy mapping table (in this embodiment, the diagnostic strategy mapping table is implemented by a small neural network or fuzzy logic system).

[0163] An intelligent feedback loop was constructed, which transforms the macro-level emotional breakdown risk index into a precise calibration of the micro-level diagnostic model. This makes the system's perception capability no longer static, but intelligently allocate computing resources based on predicted emotional changes, automatically sharpening insight at critical moments. Thus, without increasing average power consumption, it significantly improves the ability to capture key emotional triggers.

[0164] The diagnostic model is a fuzzy logic reasoning system;

[0165] The construction of a fuzzy logic reasoning system includes the following core components:

[0166] Input variable: The only input variable is the emotional breakdown risk index, which takes continuous values ​​in the range [0,1]. The variable is fuzzified into three linguistic terms: low risk, medium risk, and high risk. Each term is defined by a membership function, which describes the extent to which any specific risk index value belongs to that term. In this embodiment, the membership function is defined as follows:

[0167] Low risk: A trapezoidal membership function is used, with a membership degree of 1 in the interval [0, 0.2] and linearly decreasing to 0 in the interval [0.2, 0.4]. Medium risk: A triangular membership function is used, with a membership degree of 0 at 0.3, reaching a peak of 1 at 0.5, and linearly decreasing to 0 at 0.7. High risk: A trapezoidal membership function is used, with a membership degree linearly increasing to 1 in the interval [0.6, 0.8] and remaining at 1 in the interval [0.8, 1.0].

[0168] Output variables: Contains two output variables, corresponding to the two dimensions of the adjustment command:

[0169] Output variable 1: The event threshold adjustment factor is a multiplicative factor used to adjust the detection thresholds for all events (visual, speech, tactile, and environmental); its value range is [0.5, 1.0], where 1.0 represents no adjustment and 0.5 represents doubling the sensitivity; it is fuzzified into three linguistic terms: no adjustment, slight reduction, and significant reduction.

[0170] Output variable 2: The causal graph search depth adjustment value is an additive value used to adjust the temporal depth of the backtracking events when the graph neural network constructs the causal graph; the value range is [0,10], where 0 represents no increase in depth and 10 represents the maximum increase; it is fuzzified into three linguistic terms: no increase, slight increase, and large increase;

[0171] Inference engine and rule base: The inference engine processes the fuzzy input according to the preset rule base to produce a fuzzy output; the rule base is the essence of the diagnostic policy mapping table;

[0172] Deblurring of the output: The fuzzy output obtained from the inference is converted back into a precise and executable value. In this embodiment, the standard centroid method is used for deblurring to calculate the final event threshold adjustment factor and causal graph search depth adjustment value.

[0173] In this embodiment, the key parameters involved in the emotion homeostasis adaptive regulator are as follows:

[0174] The parameters for the diagnostic resource budget are set as follows: ; is a dimensionless numerical value used to quantify the total amount of resources authorized to improve diagnostic precision under the current risk index; it is derived from the emotional breakdown risk index through a nonlinear mapping function, the original source of which is the logistic function.

[0175] The calculation logic is as follows: Obtain the input emotional breakdown risk index, and set the parameter sign of the emotional breakdown risk index to... Risk index for emotional breakdown Subtracting risks activates the inflection point Multiply the difference by the budget growth rate. The calculation results are then processed using the natural exponential function; the maximum budget amount is then calculated. Divide by the sum of the calculated result and 1; map the emotional breakdown risk index, ranging from zero to one, to a value ranging from zero to the maximum budget. Diagnostic resource budget between [the parties].

[0176] Maximum total budget Its function is to limit the upper limit of resource adjustment and prevent system instability due to over-adjustment; in this embodiment, the value is set to one hundred units; the technical balance achieved between ensuring sufficient adjustment space and controlling overall computational overhead is determined through stress testing; risk activation inflection point. A budget allocation of exactly half its maximum value is defined as the level at which the emotional breakdown risk index reaches a certain level; in this embodiment, the value is set to 0.5; the budget growth slope... The sensitivity of the budget to changes in the risk index is controlled; in this embodiment, the value is set to 10, so that the budget grows rapidly when the risk index is close to the inflection point, while the growth is slow in the two ends.

[0177] The parameter symbol for parameter adjustment cost is set to For each diagnostic parameter that can be adjusted by the emotional homeostasis adaptive regulator, a set value is provided to quantify the cost of parameter adjustment. The diagnostic resource budget required to perform a minimum unit adjustment (in this embodiment, the threshold is reduced by one percent).

[0178] The calculation logic is as follows: For each diagnostic parameter to be calibrated, under the baseline configuration, run the first data processing unit and record the average single-cycle calculation time; adjust the diagnostic parameter in the smallest unit; run again and record the adjusted average single-cycle calculation time; subtract the baseline time from the adjusted time to obtain the time increment, and then multiply it by the time cost conversion factor. That is, to obtain the parameter adjustment cost. Time cost conversion factor This is used to convert physical time units into dimensionless budget units; in this embodiment, the value is set to 100; the computation time increment of approximately one millisecond is mapped to a cost of 10 budget units, providing a reasonable granularity for subsequent optimization allocation.

[0179] The parameter sign of the priority weight is set to For each adjustable diagnostic parameter, a value is set by a domain expert or through a data-driven approach to quantify the parameter's priority weight. The effectiveness or importance of adjusting parameters for improving the ability to detect premonitions of emotions; parameters with higher weights will be given priority in budget allocation.

[0180] The calculation logic is as follows: At least three child development psychology experts are invited to independently score the importance of all adjustable parameters using a Likert scale based on clinical experience; the scores are then averaged after a consistency check to obtain an initial set of expert weights; a gradient boosting decision tree model (in this embodiment, XGBoost is used) is trained using a labeled large-scale dataset to predict emotional states; feature importance scores corresponding to all adjustable parameters are extracted from the trained model; and a weighted average is calculated between the expert weights and the model's feature importance scores, where the expert weights account for a certain percentage of the score. The contribution of expert weights to the feature importance scores of the model is used to adjust the priority weights of each parameter.

[0181] Expert weighting factor The purpose of this is to strike a balance between the objectivity of data-driven approaches and the prior knowledge of expert experience; in this embodiment, the value is set to 4.0, which gives higher weight to the results of data-driven approaches while retaining expert experience as an effective supplement and correction.

[0182] In this embodiment, the calculation process for the emotion steady-state adaptive regulator to generate the final configuration instruction after receiving the emotion collapse risk index is as follows:

[0183] Enter the real-time updated emotional breakdown risk index ;

[0184] Calling a non-linear mapping function to input the emotional breakdown risk index Diagnostic resource budget converted to the current period Initialize the set of configuration instructions to be executed, containing the current diagnostic parameter values;

[0185] The process enters a cyclical decision-making process, which terminates when the remaining diagnostic resource budget is insufficient to cover any effective parameter adjustments.

[0186] In each iteration, all adjustable diagnostic parameters are iterated, and candidate adjustment actions are generated for each parameter (in this embodiment, for threshold parameters, this means decreasing by a minimum step size; for depth parameters, it means increasing by 1). Then, for each candidate adjustment action, the benefit-cost ratio is calculated. The benefit-cost ratio is calculated by obtaining the parameter priority weights corresponding to the action. And divide by the adjustment cost of the action parameters. ;

[0187] From all candidate adjustment actions, select the adjustment action with the highest benefit-cost ratio as the optimal action for this cycle;

[0188] The steps for selecting the optimal action are as follows: calculate the benefit-cost ratio of each candidate adjustment action, and select the action with the highest benefit-cost ratio to form the optimal candidate set; determine whether the optimal candidate set contains more than one action.

[0189] If there is only one action, then that action is the optimal action for this loop.

[0190] If multiple actions are involved, indicating a draw, the preset draw-deciding rule is activated; that is, from the optimal candidate set, further parameter adjustments are made to control the cost. The action with the lowest cost is selected as the final optimal action; when multiple adjustment options have the same expected benefit, the action with the lowest computational resource consumption is selected first to achieve the best overall energy efficiency.

[0191] Determine whether the parameter adjustment cost of the optimal action is less than or equal to the current remaining diagnostic resource budget;

[0192] If the determination is yes, execute the optimal action, that is, update the value of the corresponding parameter in the set of configuration instructions to be executed; subtract the cost of the optimal action from the current diagnostic resource budget; the process returns to the loop stage and begins the next optimization loop;

[0193] If the result is negative, it indicates that the remaining budget is insufficient to cover any worthwhile adjustments, and the optimization loop terminates.

[0194] The final output is the set of configuration instructions to be executed when the optimization loop terminates; the set of configuration instructions to be executed contains a set of new diagnostic parameter target values ​​that have been co-optimized and will be sent to the feedback control unit to update the actual operating status of the first data processing unit.

[0195] The feedback control unit is configured to apply the configuration instructions of the diagnostic model to the first data processing unit; and the decision output unit is configured to select the corresponding interaction strategy mode based on the emotional breakdown risk index, and generate specific robot behavior instructions based on the updated first data.

[0196] When the emotional steady-state adaptive regulator outputs a configuration command: the feedback control unit immediately writes the configuration command into the configuration register of the first data processing unit, making the diagnostic logic take effect immediately; at the same time, after receiving the high-order emotional breakdown risk index, the decision output unit will immediately switch the robot's overall interaction strategy from the normal play mode to the preventive soothing mode; the decision logic will prioritize calling the updated and more sensitive attribution vector to guide the behavior; in this embodiment, if the attribution vector points to a weak noise source, the robot's behavior command will be to actively play a piece of white noise that can mask the noise, rather than performing dance movements that may aggravate the stimulation.

[0197] If the decision output unit does not receive a valid emotional breakdown risk index within a preset time window, it indicates a malfunction in the prediction unit. The decision output unit then locks the interaction strategy in the standard response mode and makes decisions solely based on the output of the first data processing unit under default parameters. Simultaneously, it sends a system status alarm to the user terminal. In this embodiment, the default parameters include a visual event trigger threshold. Voice event trigger threshold Contact threshold and environmental fluctuation threshold and causal graph search depth;

[0198] The causal graph search depth seeks a balance between computational efficiency and the completeness of causal relationship capture; its main task is to understand immediate, short-range causal chains; after analyzing a large amount of behavioral data, the vast majority of direct causal relationships occur within a 5-second time window; setting the default depth to 5 seconds can cover the core causal relationships while avoiding the huge computational burden caused by excessive backtracking time, ensuring that the first data processing unit can maintain real-time responsiveness under any circumstances.

[0199] The internal decision-making of the emotion homeostasis adaptive regulator is transformed into synchronous control of the diagnostic model and robot behavior; this ensures that the improvement of perception ability and the adjustment of behavior strategy are coupled, while the degradation strategy ensures that even in the extreme case of failure of the prediction function, the core diagnostic and response functions of the system are still available, thereby enhancing the reliability and safety of the system.

[0200] In this embodiment, the key parameters involved in the feedback control unit are as follows:

[0201] The parameter sign of the attribution focus vector is set to It is structured data used to characterize the core driving force that contributes the most to the cognitive load or emotional fluctuations of young children in the current cognitive focus map sequence; it not only identifies which event node it is, but also includes the key attributes of the event node, providing precise guidance for subsequent tactical choices;

[0202] The calculation logic is as follows: Obtain the cognitive focus map sequence updated by the latest configuration instructions from the first data processing unit; perform iterative calculations on the cognitive focus map sequence, the iterative calculation process is as follows: Assign an initial influence score to each node in the cognitive focus map sequence; in each iteration, the updated influence score of an event node is equal to the influence scores of all other event nodes pointing to the event node, and the scores are weighted and summed according to the weights of the edges connecting them; this process is repeated until the influence scores of all nodes converge to the influence convergence threshold. The following steps are taken: The node with the highest influence score after convergence is selected as the focal node; the modality type, semantic label, and feature vector of the focal node are extracted to form the attribution focal vector; and the influence convergence threshold is determined. The value is used to control the accuracy and number of iterations in the influence calculation; in this embodiment, the value is set to 0.001, which is a technical balance achieved between ensuring the stability of the calculation results and controlling the calculation time.

[0203] The parameter symbols for the strategic interaction mode are set to ; is a discrete state variable, representing the overall interaction strategy goal that the robot should adopt under the current macro-emotional risk; it does not specify specific actions, but sets the tone and intention of the behavior;

[0204] The calculation logic is as follows: Four strategic interaction modes are preset: normal interaction mode, proactive guidance mode, preventative reassurance mode, and emergency de-escalation mode; an emotional breakdown risk index is obtained. The risk index of emotional breakdown Compare with the preset mode switching threshold; if the emotional breakdown risk index... Below the first threshold If the emotional breakdown risk index is high, then the mode is set to normal interaction; Between the first threshold Second threshold If it falls between these thresholds, it is set to active guidance; if it falls between the second threshold... and the third threshold If the threshold is between 1 and 2, it is set as preventative reassurance; if it exceeds the third threshold... If so, it will be set as an emergency downgrade;

[0205] First threshold Second threshold and the third threshold The activation boundaries of different strategic modes are defined; in this embodiment, the first threshold... The value range is (0.20, 0.40), the first threshold. The value is set to 0.3, the second threshold. The value range is (0.55, 0.70), the second threshold. The value is set to 0.6, the third threshold. The value range is (0.80, 0.90), the third threshold. The value is set to 0.85; it is derived from statistical analysis of a large amount of data on children's emotional development and is calibrated by child psychology experts, corresponding to the completely stable zone, the slightly fluctuating zone, the significantly unstable zone, and the high-risk critical zone of emotional state respectively.

[0206] The parameter notation of the behavior instruction primitive is set to ; is a standardized instruction code, which is the interface between the decision output unit and the robot's underlying actuator; it represents the smallest executable action unit, such as playing soothing music with a specified number, performing a preset head-shaking motion, or adjusting the screen brightness to a specified value.

[0207] The parameter symbol for the data freshness window is set to ; represents a time duration value, defining the decision output unit's risk index for receiving emotional breakdown. At that time, it was used to determine the risk index of emotional breakdown. Whether the maximum acceptable latency is still effective is the key basis for triggering the degradation mechanism, ensuring that decision-making is always based on sufficiently real-time predictive information; it is based on experimental calibration of the latency of the end-to-end data processing pipeline;

[0208] The calculation logic is as follows: In a typical network environment, multiple end-to-end latency tests are conducted, recording the time from the sensor capturing raw data (i.e., the first data) to the second data processing unit outputting the final emotional breakdown risk index. The average time consumed is denoted as the average processing delay. Calculate the difference between the maximum and average latency in the above tests, and multiply it by the jitter safety factor. Obtain network jitter margin The average processing latency is added to the network jitter margin, and the sum is the final value of the data freshness window; in this embodiment, the data freshness window... The preferred value is set to 2 seconds; a reasonable range is 1 to 3 seconds; average processing delay Set to 1.2 seconds; set the jitter safety factor. The obtained network jitter margin is 2. The average processing latency is 0.8 seconds. With network jitter margin The sum is 2 seconds.

[0209] In this embodiment, the complete calculation and execution steps of the robot's behavior instructions are as follows:

[0210] The input consists of two parallel, real-time updated data streams: a configuration instruction set from the diagnostic model of the emotion homeostasis adaptive regulator, and an emotion breakdown risk index from the second data processing unit. ;

[0211] After receiving the configuration instruction set of the diagnostic model, the feedback control unit immediately applies the configuration instruction set to the first data processing unit; this is an atomic operation, that is, the new parameter values ​​in the configuration instruction set are overwritten to the corresponding configuration register inside the first data processing unit; this ensures that the diagnostic engine can immediately run with higher sensitivity or a deeper level of analysis.

[0212] At the same time, the decision output unit executes the following two parallel sub-steps:

[0213] Emotional breakdown risk index based on input Invoke the threshold state machine to determine the current strategic interaction mode;

[0214] Obtain the latest cognitive focus map sequence output by the updated first data processing unit, and calculate the attribution focus vector. ;

[0215] The core decision-making logic is implemented through a pre-defined strategic and tactical decision matrix. The rows of this matrix are indexed by strategic interaction patterns, and the columns are indexed by attribution focus vectors. Semantic tag index in; based on current strategic interaction patterns and attribution focus vectors The corresponding cell is located in the strategic and tactical decision matrix, and the cell stores the primitive of the action instruction to be executed.

[0216] Before issuing the final instruction, perform a sanity check; check the risk index of emotional breakdown. Is the timestamp within the preset data freshness window?

[0217] If the determination is yes, the process continues normally;

[0218] If the judgment is negative, it indicates a prediction failure; a degradation mechanism is triggered; the strategic interaction mode is forcibly set to a predefined standard response mode, but the decision output unit still executes two parallel sub-steps; during decision-making, only in the row corresponding to the standard response mode of the strategic and tactical decision matrix, based on the attribution focus vector... Choose the most conservative action instruction primitive;

[0219] The final output is to send the selected behavioral instruction primitives to the underlying behavior executor of the emotional interaction robot through a standardized interface.

[0220] In this embodiment, the standard response pattern of the strategic and tactical decision matrix is ​​logically implemented as a two-level nested key-value pair structure; the keys of the outer structure are enumerated values ​​of the strategic interaction patterns, and the values ​​are the values ​​of the inner structure; the keys of the inner structure are the attribution focus vectors. The semantic tags in the text are the codes of the final behavioral instruction primitives to be executed.

[0221] When the strategic interaction mode is a preventative appeasement mode, the corresponding inner mapping relationship is: if the attribution focus vector If the semantic tag is sharp noise, then the corresponding behavioral instruction primitive is "AP AUDIO PLAY WHITE NOISE01" (meaning: play white noise audio number one).

[0222] If the attribution focus vector If the semantic tag is "screen flicker", then the corresponding behavior instruction primitive is "APVISUAL SET SCREEN SOFT LIGHT" (meaning: set the screen to soft eye protection mode).

[0223] If the attribution focus vector If the semantic label is repetitive collision with object A, then the corresponding behavioral instruction primitive is "AP MOTION GENTLE REDIRECT B" (meaning: perform the gentle turning action B to divert the child's attention elsewhere).

[0224] When the strategic interaction mode is the proactive guidance mode, its corresponding inner mapping relationship is:

[0225] If the attribution focus vector If the semantic label is repetitive collision with object A, then the corresponding action instruction primitive is "AP AUDIO PLAY GUIDE A NEW WAY" (meaning: play the guiding voice, prompting a new way to play with object A).

[0226] If the attribution focus vector The semantic tag for the character is "gazing out the window for an extended period of time". The corresponding action command primitive is "AP MOTION MOVE TO WINDOW AND_TALK" (meaning: move to the window and start a conversation about the view outside).

[0227] By pre-storing the structured and well-defined decision matrix shown above, efficient, reliable, and interpretable dual-axis collaborative decision-making can be achieved.

[0228] Figure 1 It is a view that integrates the physical scene and information flow. The isometric hidden line removal diagram on the left shows the application scenario of the system. The perception range around the emotional interaction robot (concentric circles) and the multimodal data flow it emits (wavy lines) intuitively correspond to the real-time acquisition of the child's state and environmental data by the first data processing unit.

[0229] The multimodal data stream enters the process frame Multimodal Data Flow Causal Attribution (first data processing unit), and after processing, it enters the Emotional State Trend Prediction (second data processing unit), finally reaching the core adaptive adjustment and strategy switching (regulator unit). The path differentiation here is the key to this invention: on the one hand, the perception sensitivity feedback (the dashed arrow returning from box 3 to box 1) vividly illustrates how the feedback control unit dynamically adjusts the operating parameters of the first data processing unit according to the regulator instructions, realizing the adaptive sharpening of perception; on the other hand, the forward path enters the linkage behavior instruction generation, and through the behavior instructions (the solid arrow returning from box 4 to the robot), it illustrates how the decision output unit controls the robot's interaction strategy based on the updated data; the structure clearly visualizes the closed-loop feedback and synchronous coupling architecture driven by the unified decision core for perception adjustment (inward control) and behavior decision (outward control).

[0230] To quantitatively verify the advancements of the preschool emotional interaction robot system based on multimodal context perception and adaptive decision-making compared to existing technologies, this embodiment constructs a high-fidelity digital twin simulation environment. A Markov decision process is used to simulate the transitions in a child's emotional state, with the transition probability influenced by a sequence of external environmental events. The system of this invention and the existing technology system (serving as a control group) are configured to simultaneously receive the event sequence output from the simulation environment and make their respective judgments. The control group employs a common single-modal high-intensity threshold scheme, where an alarm signal is only output when the intensity of a single modality exceeds a fixed high threshold. Simulation data is shown in Table 1 below.

[0231] Table 1: Simulation data comparing the performance of the system of the present invention and the control system under three typical scenarios.

[0232] Parameter Name / Scenario Comparison Scenario 1: Sudden strong stimulus (loud thud when toy falls to the ground) Scenario 2: Gradual accumulation of negative emotions (persistent refusal to eat) Scenario 3: Complex situation (minor discomfort in an unfamiliar environment) Causal relaxation time 1st time 4.8 0.9 2.5 Causal relaxation time 2nd time 4.5 1.1 2.3 Modal mutual information gain, first time 0.05 1.20 0.45 Modal mutual information gain, second time 0.08 1.15 0.50 This invention: Emotional breakdown risk index, first time. 0.12 0.85 0.41 This invention: The second edition of the Emotional Breakdown Risk Index. 0.15 0.81 0.45 Control group: First alarm signal 1 (Trigger) 0 (Not triggered) 0 (Not triggered) Control group: Second alarm signal 1 (Trigger) 0 (Not triggered) 0 (Not triggered) Effective early warning gain for the first time -0.88 0.85 0.41 Effective early warning gain for the second time -0.85 0.81 0.45

[0233] To quantitatively evaluate the superiority of the present invention, this embodiment introduces a derivative effect evaluation index:

[0234] The parameter sign of the effective warning gain is set to The following logic is used to calculate the output emotional breakdown risk index. The binary alarm signal output by the control group Perform the difference operation, that is ;

[0235] This is used to quantify the overall gain of the invention in early warning capability compared to the prior art; a value greater than 0 indicates that the invention successfully identified the risk when the control group was aphonic, and the larger the value, the higher the early warning value; a value less than 0 indicates that the invention correctly assessed the risk as low when the control group had false alarms, and the larger the absolute value, the better the effect of suppressing false alarms.

[0236] Scenario 1 (Sudden Strong Stimulus) Analysis: In this scenario, the control group triggered an alarm (output 1) due to sound intensity exceeding the threshold. The average causal relaxation time calculated by the system of this invention is as high as 4.65 seconds, indicating that the system has extremely strong inherent stability. At the same time, the average modal mutual information gain is only 0.065 bits, indicating that this is an isolated auditory event and has not triggered a cross-modal chain reaction. The average emotional breakdown risk index output by this invention is only 0.135, correctly judging that this is a one-time, non-continuous fright, and no excessive intervention is required. The effective warning gain of this invention is negative (average value is -0.865), quantitatively proving that this invention can effectively suppress false alarms caused by isolated strong stimuli, and has overwhelming robustness compared with the control group.

[0237] Scenario 2 (Progressive Negative Accumulation) Analysis: This scenario simulates a slow but continuously deteriorating process. No single stimulus intensity reaches the alarm threshold of the control group, therefore the control group never triggers an alarm (output is 0). The system of this invention accurately captures the changes in the inherent laws: the mean causal relaxation time drops sharply to 1.0 second, indicating that the system stability has been severely compromised; at the same time, the mean modal mutual information gain soars to 1.175 bits, revealing that a highly correlated negative feedback has formed between visual (seeing food), auditory (hearing urging), and physical (physical resistance); the mean emotional breakdown risk index output by this invention is as high as 0.83, issuing a strong high-risk warning; in this critical scenario, the effective warning gain of this invention is positive (mean value of 0.83), quantitatively proving that this invention has a unique ability to identify complex risk patterns with low intensity and high synergy, solving the fundamental technical problem that existing technologies cannot warn of emotional breakdowns;

[0238] Scenario 3 (Complex Situation) Analysis: In this scenario, the control group also did not trigger an alarm. The average emotional breakdown risk index output by the system of this invention is 0.43, which is at a moderate level. Thanks to its comprehensive assessment of causal relaxation time (average 2.4 seconds) and modal mutual information gain (average 0.475 bits), it correctly identifies this as a potential developing risk rather than an emergency.

[0239] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, Min-Max Normalization and Z-Score standardization.

[0240] The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0241] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0242] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A preschool emotional interaction robot system based on multimodal situational perception and adaptive decision-making, characterized in that, It includes a first data processing unit, a second data processing unit, a regulator unit, a feedback control unit, and a decision output unit; The first data processing unit is configured to acquire multimodal data streams in real time and generate first data for immediate emotion attribution based on cross-modal causal inference processing of the multimodal data streams. The second data processing unit is configured to receive the first data and, based on the dynamic analysis of the first data, generate second data for predicting trend sentiment states. The regulator unit is configured to receive the second data and map the second data to the configuration instructions of the diagnostic model according to a preset diagnostic strategy mapping table. The feedback control unit is configured to receive configuration instructions from the diagnostic model and dynamically adjust the operating parameters of the first data processing unit based on the configuration instructions; The decision output unit is configured to control the interaction strategy of the emotional interaction robot based on the second data used to trigger this decision and the first data newly generated after dynamic adjustment.

2. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 1, characterized in that: The first data processing unit identifies events in the multimodal data stream as nodes in a time sequence graph, and generates first data by calculating the causal transmission relationship between nodes, wherein the first data is a cognitive focus graph sequence.

3. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 2, characterized in that: Multimodal data streams include visual data streams, speech data streams, tactile data streams, and environmental state data streams.

4. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 3, characterized in that: The second data processing unit performs statistical analysis on the topological structure of the cognitive focus map sequence within a sliding time window to calculate macro-dynamic indicators, and generates second data based on these macro-dynamic indicators, whereby the second data is an emotional breakdown risk index.

5. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 4, characterized in that: The calculation process for the emotional breakdown risk index is as follows: The input is a multimodal data stream provided by a complete set of multi-channel sensors integrated into the emotional interaction robot system, synchronized with time stamps. The system performs event recognition on each data stream in parallel, and continuously generates event nodes with modal labels, timestamps and feature vectors based on preset visual event trigger thresholds, voice event trigger thresholds, contact thresholds and environmental fluctuation thresholds. All modal event nodes generated in the recent period are placed into a unified graph structure; the graph neural network model is used to calculate the causal transit relationship between any two temporally adjacent nodes, and the output is used as the weight of the directed edge connecting the two nodes. All nodes and weighted edges together constitute the cognitive focus graph sequence at the current time point; Calculate and output the causal relaxation time for the input baseline cognitive focus map sequence; The system maintains a buffer that stores the cognitive focus map sequence of the past minute; Based on the cognitive focus map sequence of the past minute, calculate and output the full-spectrum modal mutual information gain; Causal relaxation time and full-spectrum modal mutual information gain are used as observation variables and input into the pre-trained dynamic Bayesian network inference engine; The pre-trained dynamic Bayesian network inference engine infers the posterior probability distribution of state nodes representing the true emotional stability of young children based on the observed causal relaxation time and the value of full-spectrum modal mutual information gain. The posterior probability value of the unstable state is normalized and then output as the final emotional breakdown risk index.

6. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 5, characterized in that: The regulator unit uses an emotional homeostasis adaptive regulator, and the built-in diagnostic strategy mapping table defines a non-linear mapping relationship between the emotional breakdown risk index and the event detection threshold in the first data processing unit.

7. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 6, characterized in that: The calculation process for the emotion steady-state adaptive regulator to generate the final configuration instruction after receiving the emotion collapse risk index is as follows: Enter the real-time updated risk index of emotional breakdown; Call the non-linear mapping function to convert the input emotional breakdown risk index into the diagnostic resource budget for the current period; initialize the set of configuration instructions to be executed, containing the current diagnostic parameter values; The process enters a cyclical decision-making process, which terminates when the remaining diagnostic resource budget is insufficient to cover any effective parameter adjustments. In each iteration, all adjustable diagnostic parameters are iterated over, and candidate adjustment actions are generated for each parameter; then, the benefit-cost ratio is calculated for each candidate adjustment action. From all candidate adjustment actions, select the adjustment action with the highest benefit-cost ratio as the optimal action for this cycle; Determine whether the parameter adjustment cost of the optimal action is less than or equal to the current remaining diagnostic resource budget; If the determination is yes, execute the optimal action, that is, update the value of the corresponding parameter in the set of configuration instructions to be executed; subtract the cost of the optimal action from the current diagnostic resource budget; the process returns to the loop stage and begins the next optimization loop; If the result is negative, it indicates that the remaining budget is insufficient to cover any worthwhile adjustments, and the optimization loop terminates. The final output is the set of configuration instructions to be executed when the optimization loop terminates; the set of configuration instructions to be executed contains a set of new diagnostic parameter target values ​​that have been co-optimized and will be sent to the feedback control unit to update the actual operating status of the first data processing unit.

8. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 7, characterized in that: The feedback control unit is configured to apply the configuration instructions of the diagnostic model to the first data processing unit; and the decision output unit is configured to select the corresponding interaction strategy mode based on the emotional breakdown risk index, and generate specific robot behavior instructions based on the updated first data.

9. The preschool emotional interaction robot system based on multimodal situational awareness and adaptive decision-making according to claim 8, characterized in that: The complete calculation and execution steps for robot behavior instructions are as follows: The input consists of two parallel, real-time updated data streams: a set of configuration instructions from the diagnostic model of the emotion homeostasis adaptive regulator, and an emotion breakdown risk index from the second data processing unit. After receiving the configuration instruction set of the diagnostic model, the feedback control unit immediately applies the configuration instruction set to the first data processing unit; this is an atomic operation, that is, the new parameter values ​​in the configuration instruction set are overwritten to the corresponding configuration register inside the first data processing unit; this ensures that the diagnostic engine can immediately run with higher sensitivity or a deeper level of analysis. At the same time, the decision output unit executes the following two parallel sub-steps: Based on the input emotional breakdown risk index, the threshold state machine is invoked to determine the current strategic interaction mode; Obtain the latest cognitive focus map sequence output by the updated first data processing unit, and calculate the attribution focus vector; Entering the core decision-making logic is achieved through a preset strategic and tactical decision matrix. Based on the current strategic interaction mode and attribution focus vector, the corresponding cell is found in the strategic and tactical decision matrix. The cell stores the final action instruction primitive to be executed. Before outputting the final instruction, perform a sanity check; check whether the timestamp of the emotional breakdown risk index is within the preset data freshness window. If the determination is yes, the process continues normally; If the judgment is negative, it indicates that the prediction has failed; the degradation mechanism is triggered; the strategic interaction mode is forced to be set to the predefined standard response mode, but the decision output unit still executes two parallel sub-steps; during decision-making, the most conservative action instruction primitive is selected only in the row corresponding to the standard response mode of the strategic and tactical decision matrix according to the attribution focus vector. The final output is to send the selected behavioral instruction primitives to the underlying behavior executor of the emotional interaction robot through a standardized interface.

Images