Rehabilitation nursing robot interaction regulation and control method based on man-machine trust feedback mechanism

Abstract

The invention discloses a rehabilitation nursing robot interaction regulation and control method based on a man-machine trust feedback mechanism. The method comprises the steps that expression, voice, physiology and behavior data of a user are collected through a multi-mode sensor; constructing a trust evaluation model, and predicting a user trust value in real time; according to trust level matching interaction strategy parameters, dynamically adjusting information transparency, control right distribution and interaction frequency; a rehabilitation task is executed, feedback signals are collected, the model is optimized, and self-adaptive learning is achieved. The system forms a closed loop structure of perception-evaluation-regulation-execution-feedback-optimization, and has humanization, individuation and dynamic adaptive capacity. According to the method, the trust degree and participation sense of the user on the rehabilitation nursing robot are improved, the rehabilitation compliance is improved, and the method has good clinical application prospects and popularization value.

Description

Technical Field

[0001] The present invention relates to the technical field of rehabilitation nursing, and in particular to a method for interactive control of a rehabilitation nursing robot based on a human-machine trust feedback mechanism. Background Art

[0002] With the aging population and the growing number of patients with chronic diseases, rehabilitation nursing plays an increasingly prominent role in the modern healthcare system. In particular, the demand for assistive services provided by smart devices is growing in scenarios such as postoperative recovery, neurological rehabilitation, and intervention for functional impairment in the elderly. Rehabilitation nursing robots, due to their programmability, repeatability, and high efficiency, are widely used in various rehabilitation training, daily care, and emotional support tasks.

[0003] Currently, mainstream rehabilitation and nursing robots on the market mostly perform tasks based on fixed scripted processes or simple sensor feedback control. Their interaction strategies are primarily static and preset. For example, when a user requires physical activity training, the robot issues standard movement instructions via voice prompts and determines task completion status based on timing or position sensors. While this type of interaction can achieve basic guidance functions, it has significant technical limitations.

[0004] First, existing technologies lack the ability to perceive users' subjective states in real time, particularly cognitive aspects like emotions, trust, and willingness to participate. Users may develop resistance to robots due to fear, suspicion, or misunderstanding, leading to interruptions in training, reduced participation, and even refusal to cooperate, directly impacting rehabilitation outcomes. These subjective states are dynamic and influenced by multiple factors, including verbal expression, facial expressions, and physiological reactions. Relying solely on behavioral feedback alone cannot accurately capture a user's true feelings.

[0005] Secondly, current robot interaction strategies cannot adaptively adjust to the user's trust level. When user trust is low, high autonomy and frequent command output are often used, often leading to user discomfort and excessive cognitive burden. However, when trust is high, redundant prompts and repeated confirmations are still used, disrupting the recovery rhythm and reducing the user's sense of efficiency and autonomy. This lack of dynamic adjustment makes it difficult for robots to balance the efficiency of human-robot collaboration with a user-friendly experience, resulting in generally low user satisfaction.

[0006] Third, current learning mechanisms mostly rely on static model updates, lacking the ability to continuously optimize based on interactive feedback. As users' cognitive states change, the system should be able to dynamically learn and adjust models to achieve continuous service over long periods of time and across multiple scenarios. However, existing systems often lack a closed loop of data, feedback, and regulation, making it difficult to continuously improve personalized service levels over multiple rounds of interaction.

[0007] In summary, the existing rehabilitation nursing robot systems generally have the following problems:

[0008] Lack of real-time modeling and identification of user-machine trust status;

[0009] Unable to adjust interaction behaviors and task execution methods based on trust levels;

[0010] Lack of continuous self-learning and behavior optimization mechanisms based on user feedback;

[0011] Human-computer interaction behavior is rigid and lacks flexibility and adaptability, which reduces user stickiness and rehabilitation compliance.

[0012] Therefore, there is an urgent need for an interactive control method for rehabilitation nursing robots based on a human-machine trust feedback mechanism. Through multimodal perception, dynamic trust evaluation, strategy matching and behavior feedback closed-loop control, the robot can accurately identify the user's trust status and dynamically adjust the interactive behavior in actual rehabilitation scenarios, thereby improving user experience, enhancing rehabilitation efficiency and promoting the long-term and stable construction of the human-machine relationship. Summary of the Invention

[0013] The purpose of the present invention is to provide a rehabilitation nursing robot interactive control method based on a human-machine trust feedback mechanism to solve the problems raised by the above background technology.

[0014] To achieve the above objectives, the present invention provides the following technical solution: a method for interactive control of a rehabilitation nursing robot based on a human-machine trust feedback mechanism, comprising the following steps:

[0015] (1) Multimodal data acquisition: The multimodal sensors deployed on the rehabilitation nursing robot are used to collect the user’s facial images, voice intonation, physiological signals and operation behavior data in real time during the interaction process, forming a multi-source data set D = {x1, x2, ..., x n};

[0016] (2) Feature fusion and extraction: The multi-source data are processed through feature engineering methods to form a unified feature vector F t =[f1,f2,...,f m ], used to characterize the user's current emotional state, physiological load, and interactive behavior pattern;

[0017] (3) Trust evaluation model construction and prediction: The fusion feature F t Input into the trust prediction model to calculate the user's instant trust value T for the robot t ∈[0,1], the model is a nonlinear regression model based on gradient boosting tree XGBoost or lightweight neural network, and its output satisfies:

[0018] T t =σ(w T φ(Ft )+b)

[0019] Among them, T t : The trust value of the user on the robot at time t, ranging from [0,1]; F t : The multimodal fusion feature vector at time t, including image, voice, physiological signal and behavioral features, with a dimension of m, where the multimodal fusion feature vector Its dimension m represents the total number of fusion features, which is specifically composed of expression recognition features, voice emotion parameters, physiological index parameters and behavioral statistical features. The actual value can be configured according to the system design and is generally between 10 and 100 dimensions. φ(·): feature transformation function, which performs nonlinear mapping on the original features and can be represented by a tree model or a neural network. w: model weight coefficient vector, the dimension is the same as φ(F t ), b: model bias term, used for linear superposition correction, σ(x): Sigmoid function, used to normalize the model output to [0,1], defined as

[0020] (4) Trust level classification and interaction strategy matching: According to the trust value T t The user status is divided into three trust levels: low (0-0.4), medium (0.4-)0.7) and high (0.7-1.0), and the user status is divided into three trust levels according to the preset policy library. Dynamically match the corresponding interactive control parameter set Θ = {θ1, θ2, θ3, θ4};

[0021] (5) Interaction strategy control and task execution: Applying strategy parameters to the rehabilitation task execution module to adjust the level of prompt information detail, control rights allocation ratio, user confirmation process and robot autonomous behavior, and finally generate interactive output;

[0022] (6) Feedback collection and adaptive optimization: Construct a loss function based on multi-dimensional feedback such as user expression changes, command interruption frequency, and satisfaction ratings:

[0023]

[0024] Where L: model training loss value, representing the deviation between the predicted trust value and the actual feedback; T i : The confidence value of the model's prediction for the i-th sample; R i : The real trust score obtained from user feedback, which comes from questionnaires or sentiment analysis; n: the number of samples; λ: the regularization coefficient, which controls the weight amplitude to prevent model overfitting; ||w|| 2 : The second norm of the weight vector, used for regularization calculation.

[0025] As a preferred technical solution of the present invention, the multimodal data acquisition module includes: an image acquisition unit: using a high frame rate camera to acquire a facial image sequence I t , and extract the emotion feature vector x through convolutional neural network (img) ;

[0026] Audio acquisition unit: collects voice in real time through a microphone and extracts acoustic features such as speaking speed, pitch, and pause duration. (aud) ;

[0027] Physiological signal acquisition unit: acquires the user's HRV heart rate variability and EDA skin conductance signals in real time through the ECG belt and skin electrical sensor, forming x (bio) ;

[0028] Behavioral data collection: monitor the user's operation behavior on the robot, such as the frequency of instructions, cancellation rate, interruption points, etc., to form x (act) , the final eigenvector is expressed as: F t =[x (img) ,x (aud) ,x (bio) ,x (act) ].

[0029] As a preferred technical solution of the present invention, the trust assessment model is constructed using a supervised learning method, and its training process includes:

[0030] Data annotation: Quantify users' subjective trust in different interaction scenarios based on the Human-Computer Trust Scale, which serves as a model supervision signal.

[0031] Feature processing: standardize, denoise and reduce the dimension of the collected multimodal data;

[0032] Model training: Use the cross-validation method to train the XGBoost model, and the loss function is the mean square error L MSE ;

[0033] Model iteration: Dynamically introduce new data during the deployment process and update model parameters in real time through online learning.

[0034] As a preferred technical solution of the present invention, the interaction strategy parameter θ includes the following variables: θ1: information explanatory power, i.e., the level of detail of the prompt content; θ2: control power ratio, i.e., the ratio of robot autonomous decision-making to user control; θ3: confirmation frequency, i.e., the frequency of user confirmation required during interaction; θ4: active interaction frequency, i.e., whether the robot actively initiates prompts. Each parameter is a function of the trust value T:

[0035]

[0036] Among them, θ1: Information interpretation, represents the level of detail of the robot's prompt content. Higher values indicate more detail. θ2: Control right ratio, represents the degree of autonomous execution of the robot. Higher values indicate stronger autonomy. θ3: Command confirmation frequency. Higher values indicate more user confirmation is required. θ4: Interaction initiative frequency. Higher values indicate that the robot is more proactive in initiating interactions. k: Adjustment factor, controls the slope of the initiative growth curve. An empirical value is k = 2.

[0037] As a preferred technical solution of the present invention, the rehabilitation task module dynamically adjusts the task intensity and guidance method according to the trust level. The task intensity control formula is as follows:

[0038] A t =α·T t +β

[0039] Among them A t It represents the output intensity of the robot's rehabilitation action at that time. α and β are constants set by experience. The higher the trust, the greater the training intensity.

[0040] As a preferred technical solution of the present invention, the system further constructs a trust control map, maps the trust value to a four-dimensional parameter vector in the policy space, and constructs a mapping function:

[0041]

[0042] This function is embedded into the robot behavior scheduling system to coordinate the strategy and participate in the behavior execution in real time.

[0043] As a preferred technical solution of the present invention, the method further includes a trust trend modeling module, which models the dynamic trend of user trust in the continuous interaction process and uses a long short-term memory network (LSTM) to predict the trust value of several steps in the future:

[0044] T t+1 =LSTM([F t-n ,...,F t ])

[0045] Used to adjust interaction behaviors in advance to prevent a sudden drop in trust.

[0046] As a preferred technical solution of the present invention, when the uncertainty of the trust model prediction is high, the system triggers the fuzzy reasoning module and uses fuzzy rules to make auxiliary judgments: IF expression = "due to confusion" AND HRV = "abnormal" THEN trust = "low",

[0047] IF operation success rate = "high" AND expression = "smile" THEN trust = "high",

[0048] Fuzzy rules take precedence over model output for compensation adjustment, thus enhancing the robustness of the system.

[0049] As a preferred technical solution of the present invention, the user behavior feedback information is used to train or optimize the online learning process of the trust assessment model, and the feedback information includes at least one or more of the following:

[0050] User satisfaction scores obtained based on the questionnaire scoring module;

[0051] The positive-negative ratio of the user's facial expression changes before and after a single round of tasks;

[0052] the number of manual interruptions that occurred during a single round of the interactive task;

[0053] The average number of times users modify system instructions in each round of interaction.

[0054] As the preferred technical solution of the present invention, the method is deployed on a rehabilitation nursing robot platform, which includes: a data perception and fusion module; a real-time trust assessment subsystem; a policy template library and mapping module; a task scheduling and action control executor; an interactive UI presentation and voice control module; and an online feedback collection and model optimization module.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] This invention, by introducing a dynamic human-machine trust assessment mechanism and feedback control strategy, achieves the first real-time perception and precise response to a user's trust status in a rehabilitation and nursing robot. Compared to traditional interaction systems based on fixed processes or static rules, this invention dynamically adjusts interaction parameters, such as information transparency, control allocation, and interaction rhythm, based on the user's current psychological state, emotional feedback, and operational behavior. This improves the robot's ability to adapt to the user's subjective feelings, thereby enhancing the user's trust, sense of security, and willingness to cooperate.

[0057] Secondly, the present invention constructs a multimodal trust assessment model and achieves high-precision prediction and personalized evolution of trust values through supervised learning and continuous optimization mechanisms. By unifying the modeling of heterogeneous information such as images, voice, physiological signals, and behavioral data, and introducing time series trend analysis and fuzzy inference compensation mechanisms, the system not only adapts to the individual differences of different users but also makes forward-looking predictions about the changing trends of users' trust status, effectively avoiding interaction failures caused by sudden drops in trust and improving the stability and continuity of the rehabilitation training process.

[0058] Finally, the present invention establishes a closed-loop architecture for the entire process: "perception-assessment-control-execution-feedback-optimization," enabling the rehabilitation and nursing robot to adapt, self-learn, and evolve its personality. This closed-loop mechanism enables the robot to continuously improve human-machine collaboration efficiency and user experience over the long term, driving the transformation of rehabilitation services from "program-controlled" to "trust-driven." This approach is particularly suitable for elderly patients and neurological rehabilitation populations with high sensitivity, low compliance, or long trust-building cycles, and holds broad application value and industrial potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces and describes the drawings required for use in the embodiments of the present invention or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0060] Figure 1 : Overall flow chart of the method of the present invention;

[0061] Figure 2 : Figure 2 : Schematic diagram of the human-machine trust evaluation model structure;

[0062] Figure 3 : Interaction strategy regulation logic diagram. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In order to make a clearer explanation and description of the technical solutions and implementation methods of the present invention, the following introduces a preferred specific embodiment for implementing the technical solutions of the present invention.

[0064] See Figure 1-Figure 3In order to facilitate the understanding of the technical solution of the present invention, the "interactive control method of rehabilitation nursing robots based on human-machine trust feedback mechanism" described in the present invention is described in detail in combination with specific embodiments. It should be understood that the following implementation content is only for technical description and does not constitute a limitation on the scope of protection of the present invention. The method described in the present invention relies on the rehabilitation nursing robot platform, and aims to achieve dynamic perception, real-time evaluation and interactive behavior control of human-machine trust, so as to enhance user compliance and rehabilitation participation, and improve the humanization and intelligence level of robot-assisted rehabilitation services. The system consists of five major parts: multimodal perception, trust evaluation, strategy matching, task execution and feedback learning, forming a complete closed loop during operation.

[0065] First, multimodal sensors installed on the rehabilitation nursing robot are used to collect user interaction status information in real time. The data sources include user facial image sequences (used to extract facial emotional features), voice signals (used to analyze speech rate, pitch, semantic keywords, etc.), physiological signals (such as heart rate variability HRV and skin conduction EDA), and user operation behaviors on the robot (such as task confirmation, cancellation, and interruption frequency).

[0066] After feature extraction, the collected data forms a unified multi-dimensional feature vector, denoted as F t =[x (img) ,x (aud) ,x (bio) ,x (act) ].

[0067] The system then inputs the feature vector into the trained trust evaluation model. The model is implemented using a nonlinear regression structure based on XGBoost or a lightweight neural network structure, and outputs the user's trust value T for the robot at the current moment. t , the calculation formula is:

[0068] T t =σ(w T φ(F t )+b)

[0069] Among them, φ(·) represents the mapping function after feature transformation, is a Sigmoid function used to normalize the trust value to the range of [0, 1]. w and b are model parameters. Once the trust value is output, it will be immediately used to guide the robot's current interaction strategy decision.

[0070] According to the set rules, trust values are divided into three levels: low trust (0≤T<0.4), medium trust (0.4≤T<0.7), and high trust (0.7≤T≤1.0). Based on the level, the system extracts the corresponding interaction strategy template from the strategy template library and adjusts the control parameter set Θ = {θ1, θ2, θ3, θ4}. The four parameters represent the information interpretation degree, the control right allocation ratio, the user confirmation frequency, and the degree of interaction initiative. The mapping relationship between them and the trust value is as follows:

[0071]

[0072] This design ensures that when trust is low, the robot prioritizes enhancing instruction interpretability and interaction redundancy; while when trust is high, the robot gradually enhances its autonomous execution capabilities, reduces interference prompts, and improves rehabilitation efficiency.

[0073] Control parameters will act on the execution process of rehabilitation tasks. For example, when the robot performs lower limb training tasks, it will adjust the training intensity and prompt content based on the trust level. The adjustment of training intensity is determined by the following formula:

[0074] A t =α·T t +β

[0075] Among them, A t It represents the physical or prompt behavior indicators such as the output force and rhythm control parameters of the robot per step. α and β are system preset constants used to set the basic task intensity and trust influence weight.

[0076] During the entire interaction process, the system continuously collects user feedback information as the evaluation criteria for the control results. The feedback includes the user's task execution success rate, the ratio of positive and negative emotion changes, the number of task interruptions, and satisfaction questionnaire scores, etc., which constitute the feedback dataset {Rt}.

[0077] The trust model training loss function is defined as:

[0078]

[0079] The model continuously fine-tunes the parameter w through optimization algorithms such as gradient descent, forming an adaptive learning closed loop to enhance the accuracy of long-term trust prediction.

[0080] To further enhance the system's foresight, the system also introduces a long short-term memory network (LSTM) model to predict future trust trends. The trust sequence input is the feature vector sequence in the historical window [F t-n ,...,F t ], output the predicted value Used to provide early warning of trust decline and prevent the risk of interaction interruption.

[0081] In addition, when the system determines that the current prediction model has a high degree of uncertainty (such as a large span in the confidence interval of the input sample), the fuzzy reasoning module will be activated. This module uses the "IF-THEN" rule to judge the user's emotional behavior combination and infer the trust level to compensate for the scenario where the model confidence is insufficient.

[0082] For example, if a "confused expression + abnormal HRV + command interruption" is detected, the credibility rule determines "low trust" and automatically invokes the low-trust template strategy. The system is deployed within a comprehensive rehabilitation robot platform, which includes six modules: a perception fusion processing unit, a trust estimation module, a policy mapping module, an execution controller, a voice interaction interface, and a user feedback manager. This completes the technical loop. During operation, the system follows a closed-loop logic of "perception → evaluation → policy matching → behavior regulation → user feedback → model optimization," enabling personalized, dynamic, and intelligent interaction behavior regulation. This effectively enhances user experience and rehabilitation outcomes, making it particularly suitable for elderly individuals with unstable cognition or those who require extended trust development.

[0083] Example: Interactive control process for lower limb rehabilitation training of elderly patients after surgery;

[0084] This embodiment takes an elderly patient who has undergone knee replacement surgery (referred to as User A) as the subject. The rehabilitation nursing robot provides him with personalized lower limb functional rehabilitation training services. During the training process, the system dynamically adjusts the interaction strategy based on User A's human-computer trust level to achieve precise rehabilitation and humanized interaction.

[0085] The implementation steps are as follows:

[0086] Step 1: Initialize the system and user registration;

[0087] Before user A uses the rehabilitation nursing robot for the first time, medical staff assists him in entering basic information, including age, number of days after surgery, and medical history. The system initializes the user configuration and starts with an initial trust value of T0 = 0.5.

[0088] Step 2: Start the multimodal perception module and collect initial state features:

[0089] Before user A starts training, the robot activates the multimodal perception module and collects the following data:

[0090] The facial image is captured by a camera and fed into the expression recognition model to obtain an emotion score, such as a perplexity score of 0.7;

[0091] The voice input captured a low tone and slow speaking speed, containing the semantic keyword "Can you speak slower?"

[0092] HRV is lower than the baseline, and the skin electrode signal EDA fluctuates greatly, indicating emotional tension;

[0093] The user frequently canceled the training preparation instructions, resulting in two interruptions. All information was fused to form a feature vector F1, which was then input into the trust prediction model.

[0094] Step 3: Calculate the current trust value and determine the level:

[0095] Model output trust value:

[0096] T1=σ(w T φ(F1)+b)=0.31

[0097] If it is judged as "low trust level", the system will start the low trust policy template P L , and calculate the interaction strategy parameters:

[0098] Step 4: Interaction task adjustment and execution according to low-trust strategy;

[0099] The system prompt changes to:

[0100] "We are now going to do the knee bend exercise. I will guide you step by step and will wait for your confirmation at each step."

[0101] The training pace is slowed down, and the execution of commands requires manual confirmation by the user;

[0102] Range of motion setting intensity: The system starts the first stage of training at a lower physical intensity.

[0103] Step 5: Collect feedback and update mid-way;

[0104] T2=0.57

[0105] A1=α·T1+β=0.5·0.31+0.3=0.455

[0106] Users gradually adapted to the system rhythm and began to actively cooperate. Their facial expressions changed from confusion to neutrality, their command response time shortened, and the training was not interrupted.

[0107] The system records this process and recalculates: the trust level rises to "medium level", and the policy is updated to medium trust template P M , the parameters are changed as follows:

[0108] θ1=0.43, θ2=0.57, θ3=0.43, θ4≈0.33

[0109] The system moderately reduces the frequency of confirmation, increases the task pace, guides users to perform more independent operations, and appropriately simplifies voice prompts.

[0110] Step 6: Post-training feedback collection and model fine-tuning;

[0111] After completing the training, the user evaluated the satisfaction through the UI interface, with a score of 0.8. The system collected 0 training interruptions and the operation fluency was high.

[0112] Construct a feedback signal R = 0.8, which the system uses as a supervisory signal and compares with the estimated value during training to minimize the following loss function

[0113]

[0114] Step 7: Continue learning and prepare for the next training optimization;

[0115] The system uses the micro-gradient method to adjust the model weights and complete an online model update to achieve personalized calibration of trust prediction for user A. The training records and parameter adjustment logs are retained to support task intensity prediction, rhythm adjustment and behavior guidance for subsequent training, so that the robot can gradually adapt to user preferences and build a long-term trust relationship.

[0116] Through the above embodiments, it can be seen that the present invention has good adaptability and intelligent response capabilities in real rehabilitation scenarios, can dynamically adjust human-computer interaction strategies, thereby effectively improving user cooperation, rehabilitation efficiency and system reliability, and has clear clinical value and engineering implementation basis.

[0117] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for interactive control of rehabilitation nursing robots based on human-machine trust feedback mechanism, characterized in that: The steps include: (1) Multimodal data acquisition: The multimodal sensors deployed on the rehabilitation nursing robot are used to collect the user’s facial images, voice intonation, physiological signals and operation behavior data in real time during the interaction process, forming a multi-source data set D = {x1, x2, ..., x n }; (2) Feature fusion and extraction: The multi-source data are processed through feature engineering methods to form a unified feature vector F t =[f1,f2,...,f m ], used to characterize the user's current emotional state, physiological load, and interactive behavior pattern; (3) Trust evaluation model construction and prediction: The fusion feature F t Input into the trust prediction model to calculate the user's instant trust value T for the robot t ∈[0,1], the model is a nonlinear regression model based on gradient boosting tree XGBoost or lightweight neural network, and its output satisfies: Among them, T t : The trust value of the user on the robot at time t, ranging from [0,1]; F t : The multimodal fusion feature vector at time t, including image, voice, physiological signal and behavioral features, with a dimension of m, where the multimodal fusion feature vector Its dimension m represents the total number of fusion features, which is specifically composed of expression recognition features, voice emotion parameters, physiological index parameters and behavioral statistical features. The actual value can be configured according to the system design and is between 10 and 100 dimensions. φ(·): feature transformation function, which performs nonlinear mapping on the original features and can be represented by a tree model or a neural network. w: model weight coefficient vector. The dimension is the same as φ(F t ), b: model bias term, used for linear superposition correction, σ(x): Sigmoid function, used to normalize the model output to [0,1], defined as (4) Trust level classification and interaction strategy matching: According to the trust value T t The user status is divided into three trust levels: low (0-0.4), medium (0.4-0.7) and high (0.7-1.0), and the user status is divided into three trust levels according to the preset policy library. Dynamically match the corresponding interactive control parameter set Θ = {θ1, θ2, θ3, θ4}; (5) Interaction strategy control and task execution: Applying strategy parameters to the rehabilitation task execution module to adjust the level of prompt information detail, control rights allocation ratio, user confirmation process and robot autonomous behavior, and finally generate interactive output; (6) Feedback collection and adaptive optimization: Construct a loss function based on multi-dimensional feedback such as user expression changes, command interruption frequency, and satisfaction ratings: Where L: model training loss value, representing the deviation between the predicted trust value and the actual feedback; T i : The confidence value of the model's prediction for the i-th sample; R i : The real trust score obtained from user feedback, which comes from questionnaires or sentiment analysis; n: the number of samples; λ: the regularization coefficient, which controls the weight amplitude to prevent model overfitting; ||w|| 2 : The second norm of the weight vector, used for regularization calculation.

2. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1 is characterized in that: The multimodal data acquisition module includes: an image acquisition unit that uses a high frame rate camera to acquire a facial image sequence I t , and extract the emotion feature vector x through convolutional neural network (img) ; Audio acquisition unit: collects voice in real time through a microphone and extracts acoustic features such as speaking speed, pitch, and pause duration. (aud) ; Physiological signal acquisition unit: acquires the user's HRV heart rate variability and EDA skin conductance signals in real time through the ECG belt and skin electrical sensor, forming x (bio) ; Behavioral data collection: monitor the user's operation behavior on the robot, such as the frequency of instructions, cancellation rate, interruption points, etc., to form x (act) , the final eigenvector is expressed as: F t =[x (img) ,x (aud) ,x (bio) ,x (act) ].

3. The interactive control method of the rehabilitation nursing robot based on the human-machine trust feedback mechanism according to claim 1 is characterized in that: The trust assessment model is constructed using a supervised learning method, and its training process includes: Data annotation: Quantify users' subjective trust in different interaction scenarios based on the Human-Computer Trust Scale, which serves as a model supervision signal. Feature processing: standardize, denoise and reduce the dimension of the collected multimodal data; Model training: Use the cross-validation method to train the XGBoost model, and the loss function is the mean square error L MSE ; Model iteration: Dynamically introduce new data during the deployment process and update model parameters in real time through online learning.

4. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The interaction strategy parameters θ include the following variables: θ1: information explanatory power, i.e., the level of detail of the prompt content; θ2: control power ratio, i.e., the ratio of robot autonomous decision-making to user control; θ3: confirmation frequency, i.e., the frequency of user confirmation required during interaction; θ4: active interaction frequency, i.e., whether the robot actively initiates prompts. Each parameter is a function of the trust value T: Among them, θ1: Information interpretation, represents the level of detail of the robot's prompt content. Higher values indicate more detail. θ2: Control right ratio, represents the degree of autonomous execution of the robot. Higher values indicate stronger autonomy. θ3: Command confirmation frequency. Higher values indicate more user confirmation is required. θ4: Interaction initiative frequency. Higher values indicate that the robot is more proactive in initiating interactions. k: Adjustment factor, controls the slope of the initiative growth curve. An empirical value is k = 2.

5. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The rehabilitation task module dynamically adjusts the task intensity and guidance method according to the trust level. The task intensity control formula is as follows: A t =α·T t +b Among them A t It represents the output intensity of the robot's rehabilitation action at that time. α and β are constants set by experience. The higher the trust, the greater the training intensity.

6. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The system further constructs a trust regulation graph, maps the trust value to the four-dimensional parameter vector of the policy space, and constructs a mapping function: This function is embedded into the robot behavior scheduling system to coordinate the strategy and participate in the behavior execution in real time.

7. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The method also includes a trust trend modeling module that models the dynamic trend of user trust during continuous interaction and uses a long short-term memory network to predict the trust value for several steps in the future: T t+1 =LSTM([F t-n ,...,F t ]) Used to adjust interaction behaviors in advance to prevent a sudden drop in trust.

8. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: When the uncertainty of the trust model's prediction of the fused feature vector exceeds a preset threshold, the system triggers the fuzzy reasoning module to use preset fuzzy rules to assist in determining the user status, where the fuzzy rules include: IF expression="confused" ANDHRV="abnormal" THEN trust="low", IF operation success rate = "high" AND expression = "smile" THEN trust = "high", Fuzzy rules take precedence over model output for compensation adjustment, thus enhancing the robustness of the system.

9. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The user behavior feedback information is used to train or optimize the online learning process of the trust assessment model. The feedback information includes at least one or more of the following: User satisfaction scores obtained based on the questionnaire scoring module; The positive-negative ratio of the user's facial expression changes before and after a single round of tasks; the number of manual interruptions that occurred during a single round of the interactive task; The average number of times users modify system instructions in each round of interaction.

10. The interactive control method of a rehabilitation nursing robot based on a human-machine trust feedback mechanism according to claim 1, characterized in that: The method is deployed on a rehabilitation nursing robot platform, which includes: a data perception and fusion module; a real-time trust assessment subsystem; a policy template library and mapping module; a task scheduling and action control executor; an interactive UI presentation and voice control module; and an online feedback collection and model optimization module.

Images