Implementation method of intelligent robot based on integrated care framework

Abstract

The application discloses an implementation method of an intelligent robot based on integrated care framework, and relates to the technical field of artificial intelligence, comprising: scene perception and intelligent triggering; through a finite state machine model, the user state is continuously monitored, and only when a preset evaluation sensitive state is entered, intelligent triggering of deep analysis is performed; multi-layer evidence fusion and an old-age integrated care framework evaluation engine; after intelligent triggering of deep analysis, the collected data is processed, through a dynamic Bayesian network model, the daily behavior of the user is converted into a quantitative and probabilistic evaluation of the inherent ability of the user; closed-loop adaptive intervention and feedback; according to the evaluation result, a personalized intervention task is planned and executed, and the task difficulty is adjusted in real time through a closed-loop adaptive mechanism; the core of the application lies in its unique and logically closed computing architecture, which integrates scene perception, multi-layer time-series evidence fusion and closed-loop adaptive intervention into one, and constitutes a complete and continuously self-optimizing workflow.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly to an implementation method of an intelligent robot based on an integrated care framework. BACKGROUND

[0002] Traditional care framework intelligent robots mainly include the following categories:

[0003] The first category: general-purpose companion robots, representative product: Temi personal robot (Temi Global Ltd.); structure and principle: this type of robot usually has a mobile chassis, a display screen, a voice assistant and a camera, and the main functions are video calls, information queries, schedule reminders and simple following navigation. Its core is to serve as a mobile intelligent sound box and communication terminal. Disadvantages: ① Lack of active health assessment capability: its perception system is mainly used for obstacle avoidance and face recognition, and it cannot perform professional and quantitative analysis on health indicators such as user gait speed, sitting and standing ability, and cognitive state. ② Interaction stays in instruction response: the robot behavior is driven by user instructions, and it is a passive service provider, and it does not have the ability to actively initiate intervention based on health insights. ③ No integration of professional care framework: its functions are scattered and do not follow any systematic medical or care theoretical framework, and it cannot provide integrated care.

[0004] The second category: single-point health monitoring devices, representative products: millimeter wave radar vital sign monitors, fall alarm devices, intelligent mattresses, etc.; structure and principle: through a single sensor (such as radar, accelerometer, pressure sensor) to monitor specific indicators (such as heart rate, respiration, whether to fall, in / out of bed state). Disadvantages: ① Single data dimension, unable to integrate evaluation: only isolated physiological indicators or events can be monitored, for example, a fall alarm device can only alarm after a fall, but it cannot evaluate the trend of decreased activity ability that leads to increased fall risk. ② Lack of interaction and intervention capability: these devices are pure data collection or alarm tools, and cannot interact with users, let alone perform any health promotion or rehabilitation intervention measures.

[0005] The third category: telemedicine software platforms, representative products: various health management Apps and remote consultation platforms; structure and principle: through software applications, allowing users to manually enter health data (such as blood pressure, blood sugar), and conduct video consultations with doctors. Disadvantages: ① Dependence on user active participation: the continuity and objectivity of data collection are poor, and it is heavily dependent on user compliance. ② Lack of contextual data: unable to obtain user behavior data in real-life environments, such as home activity volume, social interactions, and nutritional intake. ③ No physical execution capability: unable to provide companionship, guided training, delivery of goods, and other services in the physical world.

[0006] In summary, existing technologies are either single-function monitoring tools or general-purpose robots lacking in-depth health insights. Specifically, this manifests in the following ways: ① Passivity and lag in health monitoring: Traditional health monitoring is mostly event-driven (e.g., fall alarms) or periodic physical examinations, lacking continuous and proactive quantitative monitoring of the elderly's intrinsic abilities (e.g., activity and cognitive abilities), making effective intervention in the early stages of functional decline impossible. ② Intrusiveness and fragmentation of the assessment process: Existing functional assessments (e.g., gait analysis, cognitive scales) typically require professionals to conduct assessments in specific scenarios, disrupting the elderly's daily lives. Furthermore, the assessment data is isolated and discontinuous, making it difficult to form a complete health trend view. ③ Impersonalized and static care plans: Care plans are often standardized, lacking mechanisms for dynamic adjustment based on continuously changing individual data, resulting in poor care efficiency and effectiveness. ④ Fragmentation of data and services: Data collected by various smart devices is scattered and fails to be integrated with a systematic, internationally recognized health management framework (e.g., ICOPE), making it difficult to translate into actionable, clinically guiding care actions. It is evident that no technology can combine a mobile, multimodal sensing robot with an authoritative, multidimensional integrated care framework for the elderly (ICOPE) to form a closed-loop system from "unobtrusive assessment" to "intelligent intervention" and then to "collaborative management". Summary of the Invention

[0007] To address the problems existing in the prior art, the purpose of this invention is to provide a method for implementing an intelligent robot based on an integrated care framework. This invention deeply integrates the robot's multimodal perception capabilities with the Integrated Care for the Elderly (ICOPE) framework for proactive, non-invasive assessment, intervention, and management of the elderly's intrinsic abilities.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is: a method for implementing an intelligent robot based on an integrated care framework, comprising the following steps:

[0009] Step 1, Context Awareness and Intelligent Triggering: Continuously monitor the user's state through a finite state machine model, and intelligently trigger in-depth analysis only when the user enters a preset evaluation-sensitive state;

[0010] Step 2, Multi-layered evidence fusion and integrated elderly care framework assessment engine: After intelligent triggering and deep analysis, the collected data is processed, and the user's daily behavior is transformed into a quantitative and probabilistic assessment of their intrinsic abilities through a dynamic Bayesian network model.

[0011] Step 3, Closed-loop adaptive intervention and feedback: Based on the assessment results, plan and execute personalized intervention tasks, and adjust the task difficulty in real time through a closed-loop adaptive mechanism.

[0012] As a further improvement of the application, the step 1 is specifically as follows:

[0013] The multi-modal raw data stream of the robot terminal is input into a finite state machine model, and the finite state machine model is defined as a five-tuple finite state machine M = (S, Σ, δ, , F), wherein S represents a home scene set, Σ represents a sensor input symbol, δ is a state transition function, is an initial state, and F is a termination state set; the finite state machine model continuously judges whether the user enters a pre-defined evaluation sensitive state; if not, only low-frequency recording or ignoring is performed; if yes, deep analysis is triggered, and a current scene context is taken as a data label and transmitted to a next process stage.

[0014] As a further improvement of the application, in the sensor input symbol, each input symbol is a vector σ =<t, l, v, a, b>, wherein t represents a time stamp, l represents a robot position, v represents a visual event, a represents an auditory event, and b represents a biological signal.

[0015] As a further improvement of the application, the state transition function δ is specifically as follows:

[0016] δ: S × Σ → S, which is used to define how the robot is transferred to a next state according to a current state and a sensor input.

[0017] As a further improvement of the application, the step 2 specifically comprises the following steps:

[0018] Step 2.1, when a trigger signal is received, first, a continuous activity stream of the user is decomposed into quantifiable parameters by a behavior primitive extraction module, each primitive B_i is associated with a parameter vector V_i, and the parameters are taken as evidence and sent to an evidence update module of a dynamic Bayesian network model;

[0019] Step 2.2, the evidence update module iteratively updates a risk probability stored in an internal ability probability model of the user through a Bayesian formula, and when the risk probability and confidence exceed a threshold value, a structured evaluation finding is generated;

[0020] Step 2.3, a parallel cross-domain causal correlation analysis module deeply mines long-term data to generate a deep insight report.

[0021] As a further improvement of the application, the step 2.2 is specifically as follows:

[0022] The internal ability domain H_k of the user is regarded as a hidden variable, each observed behavior primitive parameter V_i(t) at time t is regarded as an evidence E_t, and a belief on the hidden variable state is iteratively updated using a Bayesian theorem.

[0023] P(H_k | E_1,..., E_t) = η * P(E_t | H_k) * P(H_k | E_1,..., E_{t-1});

[0024] where P(H_k | E_1,..., E_t) is the posterior probability that the capability domain H_k is in the risk state after observing the evidence at time t; P(E_t | H_k) is the likelihood function, representing the probability of observing the evidence E_t when H_k is in the risk state; P(H_k | E_1,..., E_{t-1}) is the posterior probability at time t-1, which is taken as the prior probability at the current time; η is the normalization constant.

[0025] As a further improvement of the present application, step 2.3 is specifically as follows:

[0026] Given the time series data of two capability domains, X(t) and Y(t), two autoregressive models are established:

[0027] Restricted model: Y(t) = Σ_{i=1}^{p} α_i * Y(t-i) + ε_t;

[0028] Unrestricted model: Y(t) = Σ_{i=1}^{p} α_i * Y(t-i) + Σ_{j=1}^{q} β_j * X(t-j) + η_t;

[0029] where p is the lag order of Y(t), which is an integer representing how many past time points need to be backtracked when predicting the current value Y(t); Y(t-i) is the lag term of Y(t), representing the value of Y at the i-th past time point; α_i is the autoregressive coefficient, representing the influence weight or importance of Y at the i-th past time point Y(t-i) on the current value Y(t); ε_t is the residual term, representing the difference between the predicted value and the true value Y(t) at time point t; q is the lag order of X(t), which is an integer representing how many past time points of X data need to be introduced when predicting Y(t); X(t-j) is the lag term of X(t), representing the value of X at the j-th past time point; β_j is the Granger causality coefficient, representing the influence weight of X at the j-th past time point X(t-j) on the current value Y(t); η_t is the residual term, representing the difference between the new predicted value and the true value Y(t) at time point t after adding the historical information of X;

[0030] By comparing the residual sum of squares of the two models, an F statistic is constructed to test whether X is the Granger cause of Y.

[0031] As a further improvement of the present application, the step 3 is specifically as follows:

[0032] The error between the real-time performance of the user and the preset optimal challenge point is taken as an input to dynamically adjust the difficulty parameter of the next intervention task; the difficulty level D(t+1) of the next intervention is determined by the following formula:

[0033] D(t+1) = D(t) + K_p * e(t) + K_i * Σ_{i=0}^{t} e(i) + K_d * (e(t) -e(t-1));

[0034] Wherein: D(t) is the current difficulty level, e(t) = Perf_target - Perf(t) is the error between the current performance and the target performance, Perf(t) is the quantified real-time performance score, Perf_target is the preset optimal performance target, K_p, K_i, K_d are three control parameters of proportion, integral and differential, respectively, for responding to the current error, cumulative error and error trend.

[0035] As a further improvement of the present application, the calculation method of the real-time performance score Perf(t) is specifically as follows:

[0036] Perf(t) = w1·C + w2·B - w3·F;

[0037] Wherein: C is the completion degree, B is the standard degree, F is the fatigue degree, and w1, w2, w3 are dimensionless weight coefficients.

[0038] As a further improvement of the present application, the step 3 further comprises: feeding back the performance data in the intervention process to the evaluation engine as new evidence to form a closed loop of continuous learning and optimization.

[0039] The present application has the following beneficial effects:

[0040] 1. The change from passive response to active prevention is realized: through continuous and unconscious evaluation based on the ICOPE framework, the risk can be identified at an early stage when the internal ability of the elderly shows signs of decline, and proactive intervention can be carried out to effectively delay the occurrence of disability.

[0041] 2. The objectivity, continuity and convenience of the evaluation are greatly improved: the professional evaluation is integrated into daily life, avoiding the "white coat effect", obtaining more real and continuous health data, and completely not interfering with the normal life of the user.

[0042] 3. Achieving highly personalized and dynamic care: The care plan is generated based on the individual's continuously updated data, truly realizing "different faces for different people", and dynamically adjusting according to the intervention effect, greatly improving the accuracy and effectiveness of care.

[0043] 4. Building a systematic integrated care model: For the first time, the authoritative ICOPE integrated care framework is landed into specific technical products and service processes, making fragmented data and functional services serve a unified and clear health management goal, providing strong technical support for the realization of medical care combination.

[0044] 5. Reducing care costs and improving service efficiency: Through robot automation to perform most of the daily monitoring and intervention tasks, freeing professional manpower from repetitive labor, so that they can focus more on complex decision-making and humanistic care, optimizing the overall allocation of care resources. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flowchart of the embodiment of the present application. DETAILED DESCRIPTION

[0046] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0047] Embodiment

[0048] As shown in the figure, an implementation method of an intelligent robot based on an integrated care framework, comprising the following steps: Figure 1 Step 1, scenario perception and intelligent triggering: The system continuously monitors the user's state through a finite state machine (FSM) model, and only when entering the pre-set "evaluation sensitive state", the intelligent triggering of deep analysis is triggered, solving the problem of "when to evaluate and what to evaluate".

[0049] The starting point of this embodiment is an efficient scenario perception and event triggering mechanism. The multi-modal raw data stream of the robot terminal is sent to a DL-FSM state monitoring module. The module is formally defined as a five-tuple finite state machine M = (S, Σ, δ,

[0050] , F), where S represents the set of home scenarios, Σ represents the sensor input symbol, and δ is the state transition function. The FSM continuously judges whether the user has entered the pre-defined "evaluation sensitive state". If not, the system only records at low frequency or ignores it; if it enters (for example, recognizes the user's action of getting up from a chair), the FSM will trigger deep analysis and pass the current scenario context as a data label to the next process stage.

[0051] ​Specifically, the embodiment proposes a Context-Aware & Event-Triggered data collection and analysis mechanism to achieve the optimization of computing resources and the precision of data analysis. Formal definition: The working state of the robot is modeled as a five-tuple finite state machine (FSM) M = (S, Σ, δ, , F), where: S: a finite state set representing typical home scenarios of the elderly. For example: S = {s_sleep, s_meal, s_walk, s_sit, s_exercise,...}, where s_sleep = sleep state; s_meal = eating state; s_walk = walking state; s_sit = sitting state; s_exercise = exercise state. Σ: a finite input symbol set composed of multi-modal sensor data. Each input symbol is a vector σ = <t, l, v, a, b>, where t = timestamp, l = robot position, v = visual event, a = auditory event, and b = biological signal. δ: state transition function, δ: S × Σ → S. It defines how the robot transitions to the next state according to the current state and sensor input. For example, a transition rule can be expressed as: δ(s_sit, σ) → s_walk Condition: σ.v ='stand-up action' ∧ σ.t ∈ [06:00, 22:00] This rule is an "IF-THEN" logic. It describes a scenario in formal language: if (IF) the algorithm believes that the user is in the'sitting static' state, and (AND) receives a new set of sensor signals that meet two conditions: 1) the visual system recognizes a'stand-up action', 2) the action occurs within the normal activity time (e.g. between 6am and 10pm); then (THEN) the understanding of the user's state should be immediately updated from'sitting static' to 'walking state'; in short, this rule is an intelligent trigger that switches the system from a passive, low-power observation state to an "evaluation-sensitive state" ready for evaluation; : initial state, for example = s_idle (standby). F: termination state set (negligible in this application).

[0052] Only when the FSM enters the pre-defined "evaluation-sensitive state" (such as s_walk, s_meal), the system will trigger the corresponding deep analysis module. This mechanism ensures that high-energy data processing is only performed in the most valuable scenarios, avoiding excessive analysis of routine information.

[0053] Step 2, multi-layer evidence fusion and ICOPE assessment: After being triggered, the system processes the collected data and converts the user's daily behavior into a quantitative and probabilistic assessment of their intrinsic ability through a dynamic Bayesian network (DBN) model, solving the problem of "how to accurately and deeply assess."

[0054] This step is the core analysis engine of the system, responsible for converting raw behavior data into clinically meaningful insights. After receiving the trigger signal from the previous step, the system first decomposes activities into quantifiable parameters through the behavior primitive extraction module. These parameters are sent as evidence to the dynamic Bayesian network (DBN) evidence update module, which iteratively updates the risk probability stored in the user's intrinsic ability probability model through the Bayesian formula P(H_k | E_t) =.... When the risk probability and confidence exceed the threshold, the system generates structured "assessment findings." At the same time, a parallel cross-domain causal association analysis module performs deep mining on long-term data to generate "deep insight reports." The "assessment findings" and "deep insight reports" produced in this phase are the basis for decision-making to start the next phase of intervention.

[0055] The assessment engine of this embodiment is a multi-level temporal evidence fusion model, with a dynamic Bayesian network (Dynamic Bayesian Network, DBN) as its core for ICOPE assessment. Specifically, it includes:

[0056] First layer: Behavioral Primitives extraction module: decomposes the user's continuous activity stream into standardized and quantifiable "behavioral primitives". For example, a "timed up and go test (TUG)" can be decomposed into [sit-to-stand transfer], [forward walking 3 meters], [180-degree turn], [walk back to chair], and [stand-to-sit transfer]. Each primitive B_i is associated with a parameter vector V_i.

[0057] Second layer: temporal evidence accumulation and confidence update: the user's intrinsic ability H_k (for example, k represents "activity ability") is considered as a hidden variable. Each observed behavior primitive parameter V_i(t) (at time t) is considered as an evidence E_t. The belief in the state of the hidden variable is iteratively updated using Bayes' theorem.

[0058] Formula: The posterior probability update of the risk state of any ICOPE ability domain H_k follows the following formula:

[0059] P(H_k | E_1,..., E_t) = η * P(E_t | H_k) * P(H_k | E_1,..., E_{t-1})

[0060] where P(H_k | E_1,..., E_t) is the posterior probability of the capability domain H_k being in the risk state after observing the evidence at time t. P(E_t | H_k) is the likelihood function, representing the probability of observing the evidence E_t (e.g. step speed < 0.8 m / s) given that H_k is in the risk state. This model is pre-trained with clinical data. P(H_k | E_1,..., E_{t-1}) is the posterior probability at time t-1, serving as the prior probability at the current time. η is the normalization constant.

[0061] This iterative process smoothly integrates discrete, possibly noisy observation data into a stable, reliable long-term assessment of the user's intrinsic capability state.

[0062] Third layer: Cross-domain causal association analysis module: Vector Autoregression (VAR) and Granger causality test are used to explore the potential lead-lag relationships between different ICOPE capability domain indicators.

[0063] Formula: Suppose there are two time series data of capability domains, X(t) (such as nutritional intake) and Y(t) (such as activity capability score). Two autoregressive models are established:

[0064] 1. Restricted model (Restricted): Y(t) = Σ_{i=1}^{p} α_i * Y(t-i) + ε_t;

[0065] 2. Unrestricted model (Unrestricted): Y(t) = Σ_{i=1}^{p} α_i * Y(t-i) + Σ_{j=1}^{q} β_j * X(t-j) + η_t;

[0066] p: Lag order of Y(t). This is an integer that represents how many past time points we need to go back to when predicting the current value Y(t). For example, if p=3, it means we use the data from the past 3 days (Y(t-1), Y(t-2), Y(t-3)) to predict today's data. Y(t-i): Lagged term of Y(t). It represents the value of Y at the i-th past time point. For example, Y(t-1) is the value of Y at the previous time point (yesterday). a_i: Autoregressive coefficient. This is a parameter that the model needs to learn, which represents the weight or importance of the value Y(t-i) at the i-th past time point on the current value Y(t). e_t: Residual term. It represents the difference between the model's predicted value and the true value Y(t) at time point t. e_t contains all the information that is not explained by the history of Y itself.

[0067] q: Lag order of X(t). Similar to p, this is an integer that represents how many past time points of X we need to introduce when predicting Y(t). X(t-j): Lagged term of X(t). It represents the value of X at the j-th past time point. For example, X(t-1) is the value of X at the previous time point (yesterday). b_j: Granger causality coefficient. This is a key parameter that the model needs to learn. It represents the weight of the value X(t-j) at the j-th past time point on the current value Y(t). If these b coefficients are statistically significant and not zero, it indicates that the history of X is helpful for predicting Y. e_t: Residual term. It represents the difference between the new model's predicted value and the true value Y(t) at time point t after incorporating the historical information of X.

[0068] By comparing the residual sum of squares (RSS) of the two models, we can construct an F-statistic to test "whether X is a Granger cause of Y". If it is statistically significant, it indicates that a change in one ability domain may predict a future change in another, providing data support for fundamental intervention.

[0069] Step 3, closed-loop adaptive intervention and feedback: the system plans and executes personalized intervention tasks according to the evaluation results, and adjusts the task difficulty in real time through a closed-loop adaptive mechanism (such as a PID controller). More importantly, the performance data during the intervention process will be fed back as new evidence to the evaluation engine in step 2, forming a closed loop of continuous learning and optimization.

[0070] This step is responsible for converting the assessment insights into specific care actions and forming a closed loop. The intervention decision and task planning module plans the intervention task according to the input of the previous step. The human-computer interaction module is responsible for executing the task. During the execution process, the system will collect the user's performance data Perf(t) in real time and input it into the dynamic difficulty adjustment (DDA) engine. The engine adopts the PID controller idea, adjusts the task difficulty in real time through the formula D(t+1) = D(t) +..., and forms a tight internal adaptive loop through the adjustment loop.

[0071] Core closed-loop feedback: The user performance data collected during the intervention is itself a valuable new evidence about the user's ability. This data will be transmitted back to the DBN evidence update module in step 2 to more accurately update the user ability probability model. It is this feedback mechanism that integrates the three independent innovations into a continuously learning and dynamically evolving organic whole, which constitutes the core technical barrier of this embodiment. Finally, all data and reports will be synchronized to the cloud platform for use by family members and doctors.

[0072] This embodiment introduces the PID (Proportional-Integral-Derivative) controller idea in classical control theory to adjust the difficulty of the intervention task in a closed loop. The error between the user's real-time performance and the preset "optimal challenge point" is taken as the input to dynamically adjust the difficulty parameter of the next intervention task.

[0073] Formula: The difficulty level D(t+1) of the next intervention is determined by the following formula:

[0074] D(t+1) = D(t) + K_p * e(t) + K_i * Σ_{i=0}^{t} e(i) + K_d * (e(t) -e(t-1));

[0075] wherein: D(t) is the current difficulty level (e.g. training repetition number, cognitive game reaction time requirement, etc.). e(t) = Perf_target - Perf(t) is the error between the current performance and the target performance. Perf(t) is the quantified real-time performance score, calculated by w1•C + w2•B - w3•F (C = completion, B = standard, F = fatigue). w1, w2, w3 are configurable, dimensionless weight coefficients. Their core role is to adjust the relative importance of the three dimensions of “completion (C)”, “standard (B)” and “fatigue (F)” in the final performance score, so that the DDA (Dynamic Difficulty Adjustment) engine can make the most appropriate difficulty adjustment for tasks of different natures and users of different abilities. These weight coefficients are all positive numbers (w1, w2, w3 > 0), and their sum is not necessarily 1. Perf_target is a preset optimal performance target (e.g. 0.85), representing the state of “slightly challenging but achievable”. K_p, K_i, K_d are three control parameters of proportion, integration and differentiation, respectively, responsible for responding to the current error, cumulative error and error trend, ensuring the rapidity, stability and predictability of difficulty adjustment. Technical advantages: This mechanism ensures that the intervention task is always in the user's “zone of proximal development”, avoiding both invalidity due to being too simple and user frustration or injury due to being too difficult, and achieving a scientific, safe and personalized rehabilitation closed loop.

[0076] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A method for implementing an intelligent robot based on an integrated care framework, characterized in that, Includes the following steps: Step 1, Context Awareness and Intelligent Triggering: Continuously monitor the user's state through a finite state machine model, and intelligently trigger in-depth analysis only when the user enters a preset evaluation-sensitive state; Step 1 is described in detail as follows: The multimodal raw data stream from the robot terminal is input into the finite state machine model, which is then formally defined as a quintuple finite state machine M = (S, Σ, δ, ... , F), where S represents the set of home scenarios, Σ represents the sensor input symbol, and δ is the state transition function, F is the initial state, and F is the set of final states. The finite state machine model continuously determines whether the user enters a predefined evaluation sensitive state. If not, only low-frequency recording or ignoring is performed. If the user enters, deep analysis is triggered, and the current context is used as a data label and passed to the next process stage. In the sensor input symbols, each input symbol is a vector σ =<t,l,v,a,b> Where t represents a timestamp, l represents the robot's position, v represents a visual event, a represents an auditory event, and b represents a biosignal; Step 2, Multi-layered evidence fusion and integrated elderly care framework assessment engine: After intelligent triggering and deep analysis, the collected data is processed, and the user's daily behavior is transformed into a quantitative and probabilistic assessment of their intrinsic abilities through a dynamic Bayesian network model. Step 2 specifically includes the following steps: Step 2.1: Upon receiving the trigger signal, the user's continuous activity flow is first decomposed into quantifiable parameters through the behavior primitive extraction module, and each primitive B_i is associated with a parameter vector V_i. The parameters are then sent as evidence to the evidence update module of the dynamic Bayesian network model. Step 2.2: The evidence update module iteratively updates the risk probability stored in the user's intrinsic ability probability model using Bayesian formula. When the risk probability and confidence level exceed the threshold, a structured assessment finding is generated. Step 2.3: The parallel cross-domain causal relationship analysis module performs in-depth mining of long-term data and generates in-depth insight reports; Step 3, Closed-loop adaptive intervention and feedback: Plan and execute personalized intervention tasks based on the assessment results, and adjust the task difficulty in real time through a closed-loop adaptive mechanism; Step 3 is as follows: The error between the user's real-time performance and the preset optimal challenge point is used as input to dynamically adjust the difficulty parameters of the next intervention task; Step 3 further includes: feeding performance data from the intervention process back to the evaluation engine as new evidence, forming a closed loop of continuous learning and optimization.

2. The implementation method of the intelligent robot based on the integrated care framework according to claim 1, characterized in that, The state transition function δ is specifically: δ: S × Σ → S, is used to define how the robot transitions to the next state based on the current state and sensor inputs.

3. The implementation method of the intelligent robot based on the integrated care framework according to claim 1, characterized in that, Step 2.2 is as follows: Treating the user's intrinsic capability domain H_k as a latent variable, and the behavioral primitive parameter V_i(t) observed at time t as a piece of evidence E_t, the belief in the state of the latent variable is iteratively updated using Bayes' theorem: P(H_k | E_1, ..., E_t) = η * P(E_t | H_k) * P(H_k | E_1, ..., E_{t-1}); Where: P(H_k | E_1, ..., E_t) is the posterior probability that the capability domain H_k is in a risky state after observing evidence at time t; P(E_t | H_k) is the likelihood function, representing the probability of observing evidence E_t when H_k is in a risky state; P(H_k | E_1, ..., E_{t-1}) is the posterior probability at time t-1, serving as the prior probability at the current time; and η is a normalization constant.

4. The implementation method of the intelligent robot based on the integrated care framework according to claim 1, characterized in that, Step 2.3 is as follows: Given time series data with two capability domains, X(t) and Y(t), establish two autoregressive models: Constrained model: Y(t) = Σ_{i=1}^{p} α_i * Y(ti) + ε_t; Unrestricted model: Y(t) = Σ_{i=1}^{p} α_i * Y(ti) + Σ_{j=1}^{q} β_j * X(tj) + η_t; Where: p is the lag order of Y(t), an integer representing how many past time points need to be looked back when predicting the current value Y(t); Y(ti) is the lag term of Y(t), representing the value of Y at the i-th past time point; α_i is the autoregressive coefficient, representing the weight or importance of the value Y(ti) at the i-th past time point on the current value Y(t); ε_t is the residual term, representing the difference between the predicted value and the true value Y(t) at time point t; q is the lag order of X(t), an integer representing how many past time points of X need to be included when predicting Y(t); X(tj) is the lag term of X(t), representing the value of X at the j-th past time point; β_j is the Granger causality coefficient, representing the weight of the value X(tj) at the j-th past time point on the current value Y(t); η_t is the residual term, representing the value of X added at time point t. Based on historical information, the difference between the new predicted value and the true value Y(t); An F-statistic is constructed by comparing the sum of squared residuals of the two models to test whether X is a Granger cause of Y.

5. The implementation method of the intelligent robot based on the integrated care framework according to claim 1, characterized in that, In step 3, the difficulty level D(t+1) of the next intervention is determined by the following formula: D(t+1) = D(t) + K_p * e(t) + K_i * Σ_{i=0}^{t} e(i) + K_d * (e(t) - e(t-1)); Where: D(t) is the current difficulty level, e(t) = Perf_target - Perf(t) is the error between the current performance and the target performance, Perf(t) is the quantified real-time performance score, Perf_target is the preset optimal performance target, and K_p, K_i, K_d are three control parameters: proportional, integral, and derivative, which are used to respond to the current error, cumulative error, and error change trend, respectively.

6. The implementation method of the intelligent robot based on the integrated care framework according to claim 5, characterized in that, The calculation method for the real-time performance score Perf(t) is as follows: Perf(t)=w1·C + w2·B - w3·F; Where: C is the degree of completion, B is the degree of standardization, F is the degree of fatigue, and w1, w2, and w3 are dimensionless weighting coefficients.

Images