Wearable robot adaptive assistance control method based on just noticeable difference

Abstract

The application discloses a wearable robot adaptive assistance control method based on a minimum perceptible difference, and steps include: JND calibration and physiological correction, establishing user individualized perception benchmark, including force sense minimum perceptible difference and gesture recognition minimum perceptible difference; using a non-symmetrical dead zone control mechanism based on control increment error e(t) to correct the torque instruction of the wearable robot; if gesture recognition confidence >= and <, the current output is kept; if gesture recognition confidence >= and >=, the non-symmetrical dead zone control mechanism in the second step is used for correction; if gesture recognition confidence <, a triggered progressive safety response is executed; after executing the output, user subjective evaluation is collected by using a user feedback and adaptive module, and closed-loop adaptive calibration is executed according to the subjective evaluation. Through the non-symmetrical dead zone and the adaptive mechanism based on subjective feedback, the system is more and more in line with the habits of the user.

Description

Technical Field

[0001] This invention relates to the field of wearable robot human-computer interaction control technology, specifically to an adaptive assist control method for wearable robots based on minimum perceptible difference. Background Technology

[0002] In existing wearable robots, such as myoelectric prostheses, exoskeletons, and rehabilitation training equipment, the research and development focus of their human-computer interaction control systems is generally on improving the accuracy of gesture recognition and the precision of force control. Evaluation metrics often use objective values ​​such as classification accuracy and root mean square error. However, the end users of these systems are humans, and the user's experience is the key to evaluating the system's quality. For example, a prosthesis with a 99% recognition rate but whose errors cause embarrassment or danger to the user is far less valuable than a system with only a 95% recognition rate but whose error patterns are stable, predictable, and tolerable to the user.

[0003] The existing technology has the following main shortcomings:

[0004] 1. Control strategy is disconnected from human perception: Existing systems pursue "error minimization" without distinguishing between "perceptible errors" and "imperceptible errors." This leads to the system frequently making unnecessary corrections without the user's awareness, causing equipment vibration, increased energy consumption, and ultimately reducing user trust in the system.

[0005] 2. Lack of personalized perception benchmarks in parameter settings: The parameter adjustments of adaptive controllers are mostly based on objective performance indicators, lacking an effective method to quantify user subjective feelings into controller parameters. For example, some solutions, although using adaptive controllers, still adjust their parameters based on the goal of minimizing errors, and use a symmetrical dead zone structure, failing to take advantage of the asymmetrical nature of human perception of errors in positive and negative directions, and also lacking a mechanism for allowing user feedback to adjust parameters in a closed loop.

[0006] 3. Limited error handling: Existing systems use the same correction strategy for different types of errors, such as minor deviations, incorrect orientations, and recognition failures. This makes it impossible to distinguish which errors users can tolerate and which are serious problems that must be dealt with immediately, resulting in the system's response often not meeting user expectations.

[0007] 4. Ignoring changes in user physiological state: Existing technology does not consider the impact of changes in physiological state such as muscle fatigue and decreased attention on the user's perception and control ability, and cannot dynamically adjust control parameters when the user is in a poor state, resulting in an unstable control experience. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive assist control method for wearable robots based on minimum perceptible difference, comprising the following steps:

[0009] Step 1) Perform JND calibration and dynamic compensation based on physiological state to establish a user-personalized perception benchmark, including the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition ;

[0010] Step 2) Based on the minimum perceptible difference of force perception An asymmetric dead-zone control mechanism based on the control increment error e(t) is used to correct the torque command of the wearable robot.

[0011] Step 3) Compare the confidence level of gesture recognition with the minimum perceptible difference of gesture recognition. Compare;

[0012] If the confidence level of gesture recognition is ≥ and Then the torque command from the previous control cycle will be maintained. constant.

[0013] If the confidence level of gesture recognition is ≥ and Then, the asymmetric dead zone control mechanism in step 2) is used for correction;

[0014] If the confidence level of gesture recognition is < If so, a progressive security response will be triggered.

[0015] Step 4) After the output is executed, the user's subjective evaluation is collected using the user feedback and adaptive module, and closed-loop adaptive calibration is performed based on the subjective evaluation.

[0016] Furthermore, JND calibration can be performed using any of the following methods;

[0017] The first method: use the pre-stored default thresholds as the initial minimum perceptible difference in force perception and the initial minimum perceptible difference in gesture recognition;

[0018] The initial minimum perceptible difference in force perception and the initial minimum perceptible difference in gesture recognition are calibrated using a recursive least squares algorithm. The steps include:

[0019] Initialize the estimated values ​​and covariance;

[0020] Monitor users’ implicit or explicit feedback behavior to obtain observations y(k);

[0021] Calculate Kalman gain and update the parameter estimates. Covariance Where λ is the forgetting factor; R is the covariance of the previous time step; R is the variance.

[0022] When the covariance is less than a preset threshold and the relative change of the estimated value is less than a preset threshold multiple times in a row, the calibration is determined to be converged.

[0023] The method for obtaining the observed value y(k) is as follows:

[0024] When calibrating the initial minimum perceptible force error, when the system detects that the user is actively applying force in the opposite direction within the dead zone, the system will take the current absolute value of the force error |e(t)| as the observed value.

[0025] When multiple consecutive over-dead-zone corrections in opposite directions are detected, let the observed values... ; For coefficients;

[0026] Minimum perceptible difference for gesture recognition During calibration, when the gesture confidence falls within a preset neighborhood of the threshold θ(k−1) and the user does not intervene within the time Tmax, the current gesture confidence value is taken as the observation value; if the user reports an error, the observation value is the sum of the current gesture confidence value and the preset compensation amount.

[0027] The second approach is to use the adaptive step method to test the minimum perceptible difference of force perception at multiple force levels, and to use a combination of offline playback and online verification to test the minimum perceptible difference of gesture recognition.

[0028] The third method involves testing at a force level within the range of 30% to 50% of the user's maximum voluntary contraction force, and then combining this with historical user data to complete the JND calibration.

[0029] Furthermore, the physiological correction steps are as follows: real-time monitoring of the user's fatigue level through operation duration and median frequency of electromyography signals, and detection of attention through response delay, thereby generating a fatigue correction coefficient. Attention correction coefficient Dynamic compensation is performed on the initial perceptual reference calibrated by JND to obtain the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition ;

[0030] Minimum perceptible difference in force perception Minimum perceptible difference in gesture recognition As shown below:

[0031]

[0032]

[0033] In the formula, , The initial perceptual baseline calibrated for JND.

[0034] Furthermore, the torque commands of wearable robots include auxiliary force, resistive force, human-machine interaction force, gravity compensation force, and guiding force.

[0035] Furthermore, the steps for correcting the torque commands of the wearable robot using an asymmetric dead-zone control mechanism include:

[0036] Step 2.1) Set the positive dead zone boundary Negative dead zone boundary ;

[0037] Step 2.2) Determine whether the control increment error e(t) satisfies If satisfied, then issue the torque command. Otherwise, proceed to step 2.3).

[0038] Step 2.3) If The torque command correction for the wearable robot is as follows:

[0039]

[0040] like The torque command correction for the wearable robot is as follows:

[0041]

[0042] in, and This is a correction factor.

[0043] Furthermore, the control increment error e(t) is as follows:

[0044]

[0045] in, The desired torque command given by the upper-level decision-maker. e(t) represents the torque command actually output by the controller in the previous control cycle; e(t) represents the torque increment expected in the current control cycle.

[0046] Furthermore, in a single control flow, if the number of iterations of the correction loop used in steps 2 and 3 to handle directional deviations is greater than the preset maximum number of iterations and the gesture recognition confidence level is ≥ , If the convergence fails, it is determined to be a failure, and a progressive safety response is executed.

[0047] Furthermore, the progressive security response includes four response phases, each with a set response duration;

[0048] In the first stage, maintain output and wait for user correction; user correction methods include actively applying reverse force, re-execution by gesture, physical button reset, or voice command.

[0049] In the second stage, the output decays linearly and provides tactile cues; output ;in, =0.5× ; The torque command value that triggers the safety response instantaneously. Output the target value; This is the total preset duration for the second phase. The timestamp for the start of the second phase is t, and the current timestamp is t.

[0050] In the third stage, the output decays to zero and a voice alarm is provided.

[0051] The fourth stage involves a complete shutdown, awaiting system reset.

[0052] If a user-initiated corrective action is detected at any stage, the security response is terminated and normal control is restored.

[0053] Furthermore, subjective evaluations include being too high, too low, or incorrect;

[0054] The triggering conditions for closed-loop adaptive calibration are: the number of times the same type of subjective evaluation appears greater than or equal to m in the most recent n user action execution cycles, the error accumulation ratio is greater than the preset threshold, or the number of times the same type of subjective evaluation appears greater than or equal to k in the first h operations of the system's first use.

[0055] Furthermore, the steps of closed-loop adaptive calibration include:

[0056] S1) Determine the adjustment direction based on the feedback type:

[0057] If the subjective evaluation is too high, then increase it. And reduce the negative dead zone boundary according to the dead zone conservation rule. This reduces positive sensitivity and increases negative sensitivity, creating an asymmetric structure that conforms to human perception; the dead zone conservation rule is, after adjustment... and The sum equals the original amount. and The sum of ;

[0058] If the subjective evaluation is too small, then reduce it. And in accordance with the dead zone conservation rule, the negative dead zone boundary is increased accordingly. ;

[0059] The adjustment step size is as follows:

[0060] (6)

[0061] in This is the dead zone boundary before adjustment. The adjustment amount is calculated based on user feedback; It is the positive excess recorded during the N most recent instances of excessive feedback that triggered this adaptive calibration. The average value; when the subjective evaluation is "too small", It is the negative excess in the recent N small feedback operations | The average value of |; This refers to the sensing sensitivity coefficient;

[0062] If the subjective evaluation is incorrect, then adjust the minimum perceptible difference of the gesture recognition. Or it may trigger recalibration, separating gesture recognition errors from force control errors;

[0063] The minimum perceptible difference adjustment method for gesture recognition is as follows: = + · ,in Direction coefficient; , This is the preset maximum adjustment limit per transaction; Confidence level;

[0064] Recalibration is triggered when error feedback occurs more than the preset number of times or when the complaint rate does not decrease after adjustment.

[0065] S2) Monitor subsequent feedback; if the complaint rate decreases, retain the adjustment and update the perception sensitivity coefficient. = ×

[0066] If the complaint rate increases or complaints in the opposite direction occur, the adjustment will be rolled back, and the perceived sensitivity coefficient will be updated. = × ; This is the step size amplification factor. This is the step size reduction factor.

[0067] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0068] Improved smoothness and naturalness of operation: By incorporating minute errors imperceptible to the user into the dead zone and ignoring them, unnecessary system jitter and "fragmented movements" are effectively suppressed, making the device movement smoother and more natural.

[0069] Reduced system energy consumption and mechanical wear: The reduction of ineffective corrections directly reduces the operating frequency and amplitude of the actuator, thereby saving energy and extending the service life of key components such as motors and reducers.

[0070] Enhanced user trust and sense of control: Through asymmetric dead zones and adaptive mechanisms based on subjective feedback, the system's behavior increasingly aligns with user habits. Simultaneously, intelligent error grading ensures safe responses at critical moments, further increasing user trust in the system.

[0071] It provides a seamless experience of personalization and adaptation: from the initial multi-mode calibration to physiological state compensation during operation, and then to parameter self-learning based on feedback, the entire system process realizes refined and personalized adaptation from "a thousand faces for a thousand people" to "one policy for a time".

[0072] This invention achieves a balance between theoretical completeness and engineering practicality: It provides a Lyapunov stability proof for an asymmetric dead-zone controller and offers engineering solutions such as rapid calibration, calibration-free mode, and hardware safety redundancy, combining theoretical advancement with engineering practicality. Attached Figure Description

[0073] Figure 1 To implement the process of establishing a user-personalized perception benchmark using JND calibration and physiological correction;

[0074] Figure 2 This is an asymmetric dead-zone control logic diagram;

[0075] Figure 3 The input-output characteristic curves of an asymmetric dead-time controller are shown.

[0076] Figure 4 A flowchart for error classification and progressive safety response;

[0077] Figure 5 This is a nonlinear adjustment curve based on the perceptual saturation law, using a logarithmic law as an example;

[0078] Figure 6 Workflow diagram for user feedback and adaptive modules;

[0079] Figure 7 Hardware structure block diagram of the secure execution module;

[0080] Figure 8 This is a time-series diagram showing the output and time relationship at each stage of a progressive safety response.

[0081] Figure 9 This is the overall process. Detailed Implementation

[0082] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0083] Example 1:

[0084] See Figures 1 to 9 An adaptive assist control method for wearable robots based on minimum perceptible difference includes the following steps:

[0085] Step 1) Perform JND calibration and physiological correction to establish a user-personalized perception baseline, including the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition It should be noted that "the minimum perceptible difference in gesture recognition" This paper extends the application of the classical psychophysical concept of the "minimum perceptible difference" (JND). The classical JND reflects the human sensory threshold for changes in the intensity of physical stimuli, while the JND in this paper... This reflects the user's subjective judgment threshold regarding the reliability of gesture recognition results. Essentially, it refers to the level of confidence the system needs to achieve to make the user feel that "this recognition is trustworthy." Both describe the boundary of "just-perceptible differences" in human-computer interaction, hence this paper refers to them collectively as "gesture recognition JND."

[0086] Step 2) Based on the minimum perceptible difference of force perception An asymmetric dead-zone control mechanism based on the control increment error e(t) is used to correct the torque command of the wearable robot.

[0087] Step 3) Calculate the confidence level and minimum perceptible difference in gesture recognition. Compare;

[0088] If the confidence level of gesture recognition is ≥ and If so, then maintain the current output;

[0089] If the confidence level of gesture recognition is ≥ and Then, the asymmetric dead zone control mechanism in the second step will be used for correction;

[0090] If the confidence level of gesture recognition is < If so, a progressive security response will be triggered.

[0091] Step 4) After the output is executed, the user's subjective evaluation is collected using the user feedback and adaptive module, and closed-loop adaptive calibration is performed based on the subjective evaluation.

[0092] Example 2:

[0093] The adaptive assist control method for wearable robots based on minimum perceptible difference has the same technical content as in Example 1. Furthermore, JND calibration is performed using any of the following methods.

[0094] The first method: use the pre-stored default thresholds as the initial minimum perceptible difference in force perception and the initial minimum perceptible difference in gesture recognition;

[0095] The initial minimum perceptible difference of force perception and the initial minimum perceptible difference of gesture recognition are calibrated using a recursive least squares algorithm;

[0096] The second approach is to use the adaptive step method to test the minimum perceptible difference of force perception at multiple force levels, and to use a combination of offline playback and online verification to test the minimum perceptible difference of gesture recognition.

[0097] The third method: Perform a test at a medium force level and complete the JND calibration by combining the user's historical data.

[0098] Example 3:

[0099] The wearable robot adaptive assist control method based on minimum perceptible difference has the same technical content as any one of embodiments 1-2. Further, the physiological correction steps are as follows: the user's fatigue level is monitored in real time by measuring the operation duration and the median frequency of electromyography signals; attention is detected by detecting the response delay, thereby generating a correction coefficient. and Dynamic compensation is performed on the initial perceptual reference calibrated by JND to obtain the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition ;

[0100] Among them, fatigue correction coefficient The calculation method is as follows:

[0101] Fatigue coefficient based on operation time Let: Measure the duration of continuous operation Given preset thresholds T1=30 minutes and T2=60 minutes, then when When ≤T1 =1, when T1 < When ≤T2 =1.05, when >T2 =1.10;

[0102] Fatigue coefficient based on median electromyography frequency Let the current median frequency be: Baseline frequency ,but =1 / max( / ,0.85);

[0103] Comprehensive fatigue correction factor = × ;

[0104] Among them, attention correction coefficient The calculation method is as follows:

[0105] Assume the current response delay Baseline response delay Then when ≥ hour, = / And the lower limit is 0.85, when < hour, =1;

[0106] Minimum perceptible difference in force perception Minimum perceptible difference in gesture recognition As shown below:

[0107]

[0108]

[0109] In the formula, , The initial perceptual baseline calibrated for JND.

[0110] Example 4:

[0111] The adaptive assist control method for wearable robots based on minimum perceptible difference has the same technical content as any one of embodiments 1-3. Furthermore, the torque commands of the wearable robot include assist force, impedance force, human-machine interaction force, gravity compensation force, and guiding force.

[0112] Example 5:

[0113] The adaptive assist control method for wearable robots based on minimum perceptible difference has the same technical content as any one of embodiments 1-4. Further, the step of correcting the torque command of the wearable robot using an asymmetric dead-zone control mechanism includes:

[0114] Step 2.1) Set the positive dead zone boundary Negative dead zone boundary ;

[0115] Step 2.2) Determine whether the control increment error e(t) satisfies If satisfied, then issue the torque command. Otherwise, proceed to step 2.3).

[0116] Step 2.3) If The torque command correction for the wearable robot is as follows:

[0117]

[0118] like The torque command correction for the wearable robot is as follows:

[0119]

[0120] in, and This is a correction factor.

[0121] Example 6:

[0122] The adaptive assist control method for wearable robots based on minimum perceptible difference has the same technical content as any one of embodiments 1-5. Furthermore, the control increment error e(t) is as follows:

[0123]

[0124] in, For the expected change, This is the actual output from the previous moment.

[0125] Example 7:

[0126] The wearable robot adaptive assist control method based on minimum perceptible difference has the same technical content as any one of embodiments 1-6. Further, in one control flow, if the number of iterations in steps 2 and 3 is greater than the preset maximum number of iterations and the gesture recognition confidence level is ≥ , If so, a progressive safety response will be implemented.

[0127] Example 8:

[0128] The wearable robot adaptive assist control method based on minimum perceptible difference has the same technical content as any one of embodiments 1-7. Furthermore, the progressive safety response includes four response stages.

[0129] The first phase involves maintaining output and waiting for user correction.

[0130] The second stage features linearly decaying output accompanied by tactile cues;

[0131] In the third stage, the output decays to zero and is accompanied by a voice alarm;

[0132] The fourth stage involves a complete shutdown, awaiting system reset.

[0133] Example 9:

[0134] The wearable robot adaptive assist control method based on minimum perceptible difference has the same technical content as any one of embodiments 1-8. Furthermore, the subjective evaluation includes being too large, too small, or incorrect.

[0135] The triggering conditions for closed-loop adaptive calibration are: the number of times the same type of subjective evaluation appears in the most recent n operations is greater than or equal to m, the cumulative error ratio is greater than a preset threshold, or the number of times the same type of subjective evaluation appears in the first h operations before the system is first used is greater than or equal to k.

[0136] Example 10:

[0137] The wearable robot adaptive assist control method based on minimum perceptible difference has the same technical content as any one of embodiments 1-9. Further, the closed-loop adaptive calibration step includes:

[0138] S1) Determine the adjustment direction based on the feedback type:

[0139] If the subjective evaluation is too high, then increase it. And reduce the negative dead zone boundary according to the dead zone conservation rule. This reduces positive sensitivity and increases negative sensitivity, creating an asymmetric structure that conforms to human perception; the dead zone conservation rule is, after adjustment... and The sum equals the original amount. and The sum of ;

[0140] If the subjective evaluation is too small, then reduce it. And in accordance with the dead zone conservation rule, the negative dead zone boundary is increased accordingly. ;

[0141] The adjustment step size is as follows:

[0142] (6)

[0143] in This is the dead zone boundary before adjustment. The adjustment amount is calculated based on user feedback; It is the positive excess recorded during the N most recent instances of excessive feedback that triggered this adaptive calibration. The average value; when the subjective evaluation is "too small", It is the negative excess in the recent N small feedback operations | The average value of |; This refers to the sensing sensitivity coefficient;

[0144] If the subjective evaluation is incorrect, then adjust the minimum perceptible difference of the gesture recognition. Or it may trigger recalibration, separating gesture recognition errors from force control errors;

[0145] The minimum perceptible difference adjustment method for gesture recognition is as follows: = + · ,in Direction coefficient; , This is the preset maximum adjustment limit per transaction; Confidence level;

[0146] Recalibration is triggered when error feedback occurs more than the preset number of times or when the complaint rate does not decrease after adjustment.

[0147] S2) Monitor subsequent feedback; if the complaint rate decreases, retain the adjustment and update the perception sensitivity coefficient. = ×

[0148] If the complaint rate increases or complaints in the opposite direction occur, the adjustment will be rolled back, and the perceived sensitivity coefficient will be updated. = × ; This is the step size amplification factor. This is the step size reduction factor.

[0149] Example 11:

[0150] The adaptive assist control method for wearable robots based on minimum perceptible difference has the following steps:

[0151] Step 1: Establish personalized user perception benchmarks using JND calibration and physiological correction.

[0152] The system first obtains the user's personalized baseline threshold through the JND testing and calibration module. This module provides three modes: one is the uncalibrated mode, which is used when a new user uses the system for the first time. It can load a safe default threshold and calibrate it online during subsequent use using a recursive least squares algorithm to achieve seamless personalization.

[0153] This RLS calibration is performed separately for force perception JND and gesture recognition JND, using the same algorithm framework for both, differing only in the source of observations. Specific steps:

[0154] Step 1, Initialization: Load the default threshold as the initial estimate. = Initialize covariance Set a forgetting factor =0.95 and observation noise variance R. For example, in force calibration, R=0.1, and in gesture JND calibration, R=0.01. The smaller the R value, the greater the Kalman gain, and the more sensitive the estimated value is to the current observation value.

[0155] The second step is to acquire observations: For force perception JND, when a user is detected actively applying force in the opposite direction within the dead zone, the current absolute error value ‖e‖ is used as the observation (indicating the threshold is too high); when multiple consecutive out-of-dead-zone corrections in opposite directions are detected, half of the current estimate is used as the observation (indicating the threshold is too low). For gesture recognition JND, when the confidence level falls near the current threshold and the user does not intervene, the current confidence level is used as the observation (indicating the threshold is too high); when the user presses the "error" button, the current confidence level plus a compensation amount (e.g., 5%) is used as the observation. If there is no triggering event, the update is skipped.

[0156] The third step is to calculate the gain: The value of R for explicit feedback is smaller than that for implicit feedback.

[0157] Step 4: Update the estimated values: .

[0158] Step 5, update the covariance: .

[0159] Step 6, Convergence Judgment: Calibration is considered successful when all three of the following conditions are met: ① P(k) < 10; ② Relative change of 5 consecutive adjacent estimates < 2%; ③ Observation residuals of 5 consecutive times < 5% of the current estimate. After successful calibration, the update frequency is reduced to save computational resources.

[0160] It should be noted that this RLS calibration updates the JND perception baseline itself, addressing the question of "how much the user can perceive," which complements the fourth step of adjusting the dead zone boundary to address "how the system should respond."

[0161] Secondly, a complete calibration mode was implemented, employing an adaptive step method to test force perception JND at multiple force levels, such as 20%, 35%, and 50% MVC. Furthermore, a combination of offline playback and online verification was used to test gesture recognition JND, ultimately yielding the desired results. and .

[0162] use Rules: Initial amplitude The step size is 10% of the reference force, with an adjustment coefficient α = 1.2. In a single test, the reference force and a variable force are randomly presented; the user determines which force is greater. Two consecutive correct presses... = / Reduce the range, one incorrect press = × Increase the magnitude; record the magnitude of the turning point. After completing the preset number of repetitions, the average value of the last N turning points is taken as the force level JND, i.e. = The maximum value is taken in the end. =max( , , ).

[0163] Gesture recognition JND uses a combination of offline playback and online verification: in the offline stage, the user tests a preset candidate threshold set. ={50%,55%,…,90%} Each component is scored individually, and the highest score is selected. One enters the candidate set During the online phase, candidate thresholds were blindly tested in actual operation, and the recognition rejection rate at each threshold was recorded. User error correction rate and task completion time ;

[0164] Finally, from the candidate set In the middle, choose to simultaneously meet the user's error correction rate. ≤ and task completion time ≤ Among all thresholds, the rejection rate is... The highest threshold is used as the minimum perceptible difference for gesture recognition. ,Right now ,Right now =

[0165] Thirdly, there is a fast calibration mode, which allows calibration to be completed within 3 minutes when the user has historical data and only tests at a medium force level.

[0166] The historical data includes at least: the force JND values ​​and Weber fractional constants measured at 20%, 35%, and 50% MVC during the initial complete calibration; and updated data after each adaptive calibration. and Values; historical average response delay, fatigue change slope, and other operational habit data. During rapid calibration, the system loads historical archives. The user only needs to test once at a medium force level. The system compares the historical benchmark to calculate the offset coefficient, calculates the full-range JND value, and inherits the historical personalized parameters, achieving "test only once, update the entire range".

[0167] The confidence score for gesture recognition is output by a classifier. This patent provides two options: Option 1 uses traditional Softmax, taking the highest probability as the confidence score; Option 2 uses evidence-based deep learning, first filtering through uncertainty, then taking the highest probability of the Dirichlet mean as the confidence score. Regardless of the option used, a confidence score value ranging from 0% to 100% is obtained and compared with JND_gesture. Option 1 offers high real-time performance, while Option 2 offers higher reliability. Those skilled in the art can choose according to application requirements. While selecting the adaptation mode and performing calibration, the physiological state correction module monitors the user's fatigue level in real time through operation duration and median frequency of electromyography signals, and detects attention through response delay, generating correction coefficients. and Dynamic compensation is applied to the initial JND threshold.

[0168] In the physiological state correction module of this invention, the method for measuring response delay is as follows: After the robot performs a specific action, tactile or auditory cues are issued at random time intervals, requiring the user to make a predefined gesture as soon as possible. The system records the time interval between issuing the cues and the classifier recognizing the electromyographic signal pattern of the gesture as the response delay for this instance. Baseline response delay. This is the average of multiple measurements taken by the user while they are awake and calm.

[0169] The implementation process is as follows Figure 1 As shown.

[0170] This invention defines the following parameter relationships:

[0171] (1)

[0172] (2)

[0173] When a decrease in user attention is detected, i.e. When the value is less than 1, it indicates that the user's ability to detect and correct misidentification is weakened, and the system should be more conservative. By multiplying by this coefficient which is less than 1, the threshold of the minimum perceptible difference in gesture recognition is reduced, causing more gesture recognition results to be judged as "unreliable" and triggering a safety response, so as to make up for the lack of user attention with the system's conservative strategy.

[0174] At this point, JND calibration and physiological correction are complete, and the output is complete. and .

[0175] This step, for the first time, quantifies the user's subjective perception threshold and real-time physiological state as the core parameter benchmark of the controller, solving the fundamental problem of "disconnect between control and perception".

[0176] The second step: asymmetric dead-zone control based on the perception benchmark, which is the core control logic of this invention.

[0177] The core of the system's execution control is the asymmetric dead-zone controller. Its fundamental difference from the traditional symmetric dead-zone controller lies in utilizing the asymmetric nature of human perception of errors in positive and negative directions. For example, a user may be more sensitive to "insufficient force output" than "excessive force output." Positive dead-zone boundaries are independently set for these characteristics. and negative dead zone boundary Both are initialized to the corrected result output in the first step. The asymmetric dead-time control logic of this system is as follows: Figure 2 As shown.

[0178] Specifically, this controller does not directly handle the absolute error, but rather the control increment error e(t), which satisfies:

[0179] (3)

[0180] in, The desired torque command given by the upper-level decision-maker. e(t) represents the torque command actually output by the controller at the previous moment; e(t) represents the expected torque increment for the current control cycle. The expected change... The control mode is obtained through one of the following methods: error-driven based on trajectory tracking. = ;in For trajectory tracking gain, For the desired motion trajectory, The actual motion trajectory can be obtained by measuring sensors such as joint angles and end-effector positions; or it can be directly set based on external task commands. = ,in Torque commands directly given by an external task planner or operator; or methods based on decoding user physiological signals to enable... =f(s(t)), where s(t) is the collected physiological signals such as electromyography and electroencephalography, and f(·) is a pre-trained decoding / regression model from physiological signals to torque commands.

[0181] For linear motion actuators, the The unit is N, consistent with the unit of the torque command u(t); for rotary joints, The torque threshold has been converted to an equivalent torque value using a known lever arm during calibration, or it can be calibrated directly in the torque domain.

[0182] Its control logic is as follows: if e(t) is within the dead zone, i.e. ≤e(t)≤ This indicates that the expected change is less than the threshold that the user can perceive. At this point, the controller does not perform any action. Specifically, when step 3 (error classification module) determines the error to be "tolerable" and requires maintaining the current output, the controller in step 2 will directly execute... The value of e(t) will no longer be subject to a secondary judgment on the dead zone boundary.

[0183] This avoids a large number of "ineffective corrections" that are imperceptible to users, thus significantly reducing system jitter and power consumption. Conversely, if e(t) exceeds the dead zone, there are two possibilities:

[0184] First, e(t)> This indicates that the increase in expected value is clearly perceptible to the user, and the controller performs a positive correction.

[0185] (4)

[0186] Second The controller performs a negative correction.

[0187] (5)

[0188] in, and These determine the strength or sensitivity of the system's response to correcting "increase output" and "decrease output" when the control intention exceeds the user's perception range. and Must satisfy 0 < <2 / ( The electromechanical time constant of the actuator is used to ensure system stability. Configured to be greater than or equal to This is designed to match the asymmetrical perception of the human body regarding the increase in positive force and the decrease in negative force. There are three setting methods: preset fixed value, synchronous measurement with JND calibration, and online adaptive adjustment. The characteristic curve of the relationship between the input e(t) and output u(t) of the asymmetric dead-zone controller is shown in the figure below. Figure 3 As shown. In this invention, "positive correction" refers to increasing the amplitude of the output torque, and "negative correction" refers to decreasing the amplitude of the output torque, in order to adapt to the adjustment of the flexion, extension, or gripping force of the wearable robot joint.

[0189] This step will personalize the perception benchmark. It directly serves as the dead zone boundary and adopts an asymmetric structure, achieving anthropomorphic control at the control law level of "moving when it should move and not moving when it shouldn't move," rather than blindly pursuing the elimination of tiny errors.

[0190] Step 3: Intelligent Error Classification and Response Based on Perception Benchmarks

[0191] The system's error classification module bases the confidence level of gesture recognition on the results obtained in the first step. The comparison results demonstrate the implementation of differentiated security strategies, addressing the problems of traditional single-strategy approaches. The error grading and progressive security response process is as follows: Figure 4 As shown. The gesture recognition confidence level here refers to the classifier's degree of certainty about its output, expressed as a percentage. A higher value indicates greater confidence in the system's judgment of the current gesture. It can be divided into the following three error levels:

[0192] The first is the tolerable error, when the identification confidence level is ≥ And the control increment error |e(t)|≤min( , When the system determines that a minor fluctuation is tolerable, the decision module in step 3 directly determines to maintain the torque command calculated and output in step 2 of the previous control cycle. If unchanged, skip the dead zone determination and torque update calculation in step 2 of this cycle. This design aims to avoid responding to minute fluctuations in user intent that are imperceptible, fundamentally suppressing system jitter.

[0193] Second, there is directional bias, when the identification confidence level is ≥ However, when e(t) exceeds the dead zone, the system determines that the deviation needs to be corrected and assigns it to the asymmetric dead zone controller in the second step for correction. This controller has asymptotic convergence characteristics, and the error is reduced after correction, and it re-enters the error classification. This correction directly calls formula (4) or formula (5) in step 2 when the dead zone is exceeded, and the dead zone boundary is substituted. / Both the confidence threshold JND_gesture_final and the current value are the latest values, which may have been updated in real time by the physiological correction module in step one or the adaptive calibration module in step four. The above process will be repeated until the error converges to the dead zone. The convergence principle is: each correction in step 2 only moves out of the portion exceeding the dead zone by a certain proportion (because...). <1), causing e(t+1) to be recalculated in step 3 of the next cycle to gradually decrease, iterating until the error falls into the dead zone. If the number of iterations exceeds the preset upper limit and convergence is still not achieved, it is downgraded to "unreliable identification" and a gradual safety response is triggered. To prevent infinite loops in extreme cases, the system presets a maximum number of iterations, such as 10. If convergence is still not achieved after exceeding the limit, it is downgraded to "unreliable identification" and a gradual safety response is triggered.

[0194] Third is unreliable identification, when the identification confidence level < When the system determines a high-risk uncertain state, it immediately overwrites the calculation results of the second step and triggers a progressive safety response: the first stage maintains the torque command from the previous moment. The first stage remains unchanged, awaiting user correction; the second stage involves linearly decreasing output with tactile feedback; the third stage decreases to zero with a voice alarm; and the fourth stage stops completely, awaiting system reset.

[0195] The typical timing sequence is as follows: Phase 1: 0-0.5 seconds; Phase 2: 0.5-2 seconds; Phase 3: 2-5 seconds; Phase 4: Lock after 5 seconds. User correction methods include actively applying reverse force, re-executing gestures, and resetting via physical buttons.

[0196] The linear decay formula for the second stage is: , The third stage continues to decay to zero; in terms of tactile cues, the second stage is an intermittent vibration of 1Hz, the third stage is a continuous vibration of 2Hz, and the fourth stage is locked with a long vibration followed by silence.

[0197] in, The torque command value that triggers the safety response instantaneously. Set the target output value to 0.5× , This is the total preset duration for the second phase. t is the timestamp at the start of the second phase, and t is the current timestamp.

[0198] The hardware architecture block diagram of the progressive security response and the timing relationship diagram of the output at each stage are shown below. Figure 7 and Figure 8 As shown.

[0199] This step uses the JND threshold of gesture recognition as a benchmark to judge whether the recognition result is reliable, and combines it with the severity of control error to achieve a graded and progressive safety response, which greatly improves the security of the system and the user's sense of control.

[0200] Step 4: Closed-loop adaptive learning based on user subjective feedback, serving as the core learning mechanism of the system.

[0201] This is crucial for the system to form a closed loop and achieve continuous optimization. The system collects subjective user feedback through user feedback and an adaptive module. Users provide feedback at any time via a physical three-button feedback module integrated into the prosthesis or exoskeleton. The module includes three buttons: "Too Small," "Too Large," and "Incorrect." It should be noted that the user's judgment of whether the force output is "too large" or "too small" is based on a pre-set force intention target in their mind (e.g., wanting to apply exactly 10N of force) or an external task target (e.g., being able to just pick up a paper cup without deforming it) as a reference, and is compared with the actual output of the system. Users do not need to evaluate after each operation; they only actively press the button when dissatisfied. A single button press does not trigger adjustment; the system accumulates feedback until a trigger condition is met before initiating adaptive calibration. The following conditions are included; meeting any one of them will trigger the adjustment. The values ​​in the following trigger conditions are exemplary parameters, and those skilled in the art can adjust them according to the application scenario. For example, the same type of subjective feedback appears ≥2 times in the last 5 operations; or ≥4 times in the last 20 operations; or the overall complaint rate (including “error” feedback) exceeds 15%; or any single feedback appears ≥2 times in the first 10 operations when the system is first used.

[0202] After triggering automatic calibration, the system first determines the adjustment direction based on the feedback type. If "too large" feedback is received, the positive dead zone boundary is increased. And correspondingly reduce the negative dead zone boundary. This is done to increase negative sensitivity while decreasing positive sensitivity, creating an asymmetrical structure that conforms to human perception; if "too small" feedback is received, the positive dead zone boundary is reduced. And correspondingly increase the negative dead zone boundary. If an "error" feedback is received, adjustments will be made first. Or trigger recalibration to avoid incorrect gesture recognition affecting force control parameters. According to the formula = + · Adjustments, among which This is the direction coefficient; it is incremented by 1 when an increase is needed and decremented by 1 when a decrease is needed. , , This is the upper limit for a single adjustment. Recalibration is triggered when "error" feedback occurs more than the preset number of times consecutively or when the complaint rate does not decrease after adjustment. Here, "positive sensitivity" refers to the system's responsiveness to the intention to increase output; the larger the dead zone boundary, the lower the sensitivity; "negative sensitivity" is the opposite.

[0203] The adjustment range follows a dynamic adjustment based on a perceptual saturation-type nonlinear law. and / or Boundary. The system automatically selects and adjusts the law based on preset user scenarios: logarithmic law is used for prosthetic control and rehabilitation training scenarios to ensure smoothness and safety; power law is used for exoskeleton assistance and virtual reality interaction scenarios to make the response more sensitive; piecewise linear law is used as a general alternative and is automatically switched when the above laws fail to converge.

[0204] For example, when using logarithmic patterns:

[0205] (6)

[0206] in This is the dead zone boundary before adjustment. The adjustment amount is calculated based on user feedback; when the subjective evaluation is "too large", In the most recent N "excessively large" feedback operations, the positive excess quantity e(t)− The average value; when the subjective evaluation is "too small", In the most recent N "slightly small" feedback operations, the negative excess quantity |e(t)+ | is the average value; k is the perceptual sensitivity coefficient. The core feature of this formula is that it is monotonically increasing with a decreasing rate of increase, meaning that the more the user feels "badly", the more the dead zone boundary needs to be adjusted, but the adjustment range will gradually saturate, which conforms to the human perception law described by the Weber-Fechner law. The nonlinear adjustment curve based on the perceptual saturation law is shown in the figure below. Figure 5 As shown. The adjustment step size adopts the same as The same logarithmic formula is used for calculation, and the adjustment direction is determined based on the above linkage rules.

[0207] To prevent blind adjustments and oscillations, the system continues to monitor subsequent feedback after each adjustment: if the complaint rate decreases, the adjustment is retained and the step size is appropriately increased; if the complaint rate increases or complaints in the opposite direction appear, the adjustment is rolled back and the step size is decreased. An adjustment dead zone is also introduced, allowing reverse adjustments only when the current adjustment direction differs from the previous one, avoiding alternating oscillations. For example, if the complaint rate is below 5% and there is no continuous similar feedback after 30 consecutive operations, the system determines that it has converged and pauses adaptive adjustment. The workflow of this step is as follows: Figure 6 As shown.

[0208] This step establishes a complete closed loop: "User subjective experience -> Quantitative feedback -> Direction perception -> Scene adaptive law selection -> Nonlinear adjustment -> Effect evaluation -> Convergence judgment." This enables the controller to continuously adapt to user preferences and habits, like a learning human assistant, rather than simply optimizing an objective error metric.

[0209] Thus, the above four steps form a complete intelligent cycle of "perception-decision-execution-feedback-learning," as follows: Figure 9 As shown.

[0210] The core of this invention lies in using "whether the user can perceive the change" as the trigger threshold for the control action, rather than the "error minimization" pursued by traditional control systems. This is achieved through the calibration in the first step. and The system obtained two key indicators: the smallest change in force that the user can feel, and the lowest confidence level that the gesture recognition result is acceptable.

[0211] In actual control, these two indicators operate at different levels:

[0212] Force control level: Asymmetric dead zone controller used and Let its initial value be equal to The system assesses the error increment for each control step. When the change in user intent is less than the perceived threshold, the system maintains its output, avoiding invalid corrections and system jitter. Only when the change exceeds the threshold does the system perform a smooth correction. This makes the movement of prostheses or exoskeletons more natural, more energy-efficient, and less irritating.

[0213] Security aspects: Error classification module usage The system determines the reliability of gesture recognition results. When the recognition confidence level falls below a certain threshold, the system considers the result unreliable and triggers a progressive safety response, rather than blindly executing potentially erroneous instructions.

[0214] Application Value: This invention can be embedded in various myoelectric prostheses, exoskeleton robots, rehabilitation training equipment, and virtual reality interactive systems. Compared with existing control systems, the advantages of this invention are: it does not rely on high-precision sensors or complex algorithms; simply by introducing the new indicator of "user perception threshold," it can significantly improve the smoothness, naturalness, and user trust in operation. The system can also automatically adjust the dead zone boundary based on user subjective feedback, achieving a personalized experience that becomes more intuitive with use.

[0215] Example 12:

[0216] An adaptive assistive control method for wearable robots based on minimum perceptible difference, applied to myoelectric prosthetic control, comprises the following steps:

[0217] 1. System initialization and JND calibration

[0218] When a user uses a myoelectric prosthesis equipped with this invention for the first time, the system automatically starts the JND test and calibration module.

[0219] First, force-sensing JND calibration was performed. The system guided the user to perform three maximum voluntary contractions, and the measured MVC was 120N. Subsequently, at three force levels—24N (20% MVC), 42N (35% MVC), and 60N (50% MVC)—the following force was applied... The adaptive ladder method was used for testing. Each force level was tested 20 times, and the average of the last 10 turning points was taken as the JND for that force level. Finally, the maximum value among the three force levels was taken as the JND. It can be set to be measured. =8%MVC, which is 9.6N.

[0220] Following gesture recognition and JND calibration, the system collects electromyographic signals from user gestures such as clenching a fist, extending a palm, and making an OK sign, training a classifier capable of outputting confidence scores. In the offline phase, users experience the system's performance at confidence thresholds set sequentially from 50% to 90%, and rate it. In the online phase, the system with the highest rating is selected. The actual operation is performed using several thresholds, and the minimum threshold that the user ultimately accepts is... It can be set to be measured. =75%.

[0221] Meanwhile, the initial dead zone boundary of the asymmetric dead zone controller is set as Based on the preset proportional gain for the prosthetic limb control scenario: =0.65, =0.55. The safety limit is set to 144N, which is 120% of MVC.

[0222] The following discussion covers real-time control and adaptive processes in four scenarios.

[0223] Scenario 1: User Fatigue. When a user uses the prosthesis continuously for more than 60 minutes, the physiological state correction module detects that the operation time has exceeded the limit and generates a fatigue coefficient. =1.10. The system automatically corrects the force sensory JND to: =9.6N × 1.10 = 10.56N. The controller dead zone boundary expands accordingly, reducing the precision requirements for fatigued users, decreasing false triggers, and improving user comfort.

[0224] Scenario 2: Tolerable Error - User Intends to Generate a 10N Grip Force Increment (e(t) = 10N). Currently... =9.6N, since e(t)> The controller determines that the increased demand is perceptible to the user and executes a correction: The output increases smoothly. However, if the user intends to increase the output by only 5N (e(t) = 5N), since 5N < The controller detects the slight fluctuation as imperceptible and does not perform any action, keeping the prosthesis stable at the current grip strength and avoiding unnecessary shaking.

[0225] Scenario 3: User Feedback and Adaptation. A user felt the prosthesis was applying too much force during certain fine motor operations and pressed the "too much" feedback button four times consecutively. The user feedback and adaptation module detected an excessive complaint rate and triggered automatic calibration. Based on the modified adaptive logic, the system not only increased the positive dead zone boundary... To reduce positive sensitivity, and at the same time, according to the dead zone conservation rule Negative dead zone boundary The torque was simultaneously reduced by 1.6N, from 9.6N to 8.0N, to increase sensitivity to force reduction requirements, creating an asymmetric structure that is positively insensitive and negatively sensitive. This was based on the average error from recent instances exceeding the dead zone. Using logarithmic patterns (k=0.10) Calculate the adjustment amount and move it to the positive dead zone boundary. The force was adjusted from 9.6N to 11.2N. After that, the prosthesis became more "insensitive" in similar scenarios and more "sensitive" to force reduction, which avoided excessive force output while ensuring the user's flexibility in reducing force, thus meeting user preferences.

[0226] Scenario 4: Unreliable recognition of the user's "OK" gesture. Due to poor electrode contact, the classifier's output confidence level is only 60%, lower than... =75%. The error classification module immediately determines it as "unreliable identification," triggering a progressive safety response: Hold the current action for 2 seconds; Within seconds, the grip force is linearly reduced to 50%, while the haptic motor vibrates at a frequency of 1Hz; Within seconds, the grip force is reduced to 0, the vibration frequency is increased to 2Hz, and a voice prompt "Gesture recognition unclear" is played; after 5 seconds, the system locks, waiting for the user to manually reset it. This effectively prevents dangerous actions caused by misrecognition.

[0227] Example 13:

[0228] An adaptive assistive control method for wearable robots based on minimum perceptible difference, applied to rehabilitation training, takes stroke patients as an example. For patients in the rehabilitation phase, the system adopts a conservative strategy. Initially... The default setting is 12% MVC. The preset value is 85%, and the sensing sensitivity coefficient k is preset to 0.05. Controller gain. =0.25, =0.35, resulting in a wider dead zone and a smoother response, ensuring patient safety. One asymmetry of this embodiment lies in the gain setting: Greater than This means the system corrects for "reducing output" more quickly than "increasing output," which aligns with the safety principle in rehabilitation training: ensuring the patient is not overburdened. When the system detects that the patient has been training continuously for more than 30 minutes, the physiological state correction module automatically expands the dead zone boundary to 1.05 times, i.e. =1.05, further reducing the demands on fatigued patients and preventing them from experiencing frustration due to feeling overwhelmed. During use, if the therapist or patient indicates via the feedback button that the assist force is "too high," the system will increase... At the same time, the dead zone conservation rule will be applied to reduce the linkage. This makes it easier for the system to withdraw its assistive force, thereby gradually transitioning the rehabilitation mode from "passive assistance" to "active induction." This adaptive process constitutes the unique personalized rehabilitation strategy of this embodiment.

[0229] Example 14:

[0230] The adaptive assist control method for wearable robots based on minimum perceptible difference has the same technical content as any one of Examples 1-13. An elderly user is using the system for the first time, with no prior user history. The system automatically loads a safety default threshold. =8% MVC, =75%. The system continuously runs a lightweight online calibration in the background, using the Recursive Least Squares (RLS) algorithm. The forgetting factor is included here. Designated as =0.95, updated every time user feedback is received, such as clicking the "too large" button. The estimated value, i.e., the user's force perception baseline, is gradually converged towards the user's true perceptual threshold. It should be noted that the RLS calibration here differs from the adaptive adjustment in step four: RLS calibration updates the... The first step, the perceived baseline, addresses the question of "how much the user can actually perceive"; while the fourth step adjusts... / The dead zone boundary addresses the question of "how the system should respond." The two complement each other, working together to achieve personalized adaptation.

[0231] Typical parameter configuration example

[0232]

[0233] The asymmetric dead-time controller in this invention can be proven to be stable. (Using a Lyapunov function...) When the control increment error e exceeds the dead zone, for example... At this time, the controller performs a positive correction: Under the quasi-static assumption (udes(t) changes sufficiently slowly), we can obtain... Below, the closed-loop error dynamics can be approximated as: The derivative of its corresponding Lyapunov function .because Therefore The error will converge. Similarly, it can be proven that... The situation.

[0234] When the error is within the dead zone, the system is open-loop, but due to the dead zone boundary... and Since the proportional gain is bounded, the entire error system is proven to be globally uniform and ultimately bounded. According to the circle criterion in nonlinear control theory, as long as the proportional gain... satisfy ,in The electromechanical time constant of the actuator guarantees this boundedness, and the error will eventually converge to the dead zone. , Within the range.

Claims

1. An adaptive assist control method for wearable robots based on minimum perceptible difference, characterized in that, Includes the following steps: Step 1) Perform JND calibration and dynamic compensation based on physiological state to establish a user-personalized perception benchmark, including the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition ; Step 2) Based on the minimum perceptible difference of force perception An asymmetric dead-zone control mechanism based on the control increment error e(t) is used to correct the torque command of the wearable robot. Step 3) Compare the confidence level of gesture recognition with the minimum perceptible difference of gesture recognition. Compare; If the confidence level of gesture recognition is ≥ and Then the torque command from the previous control cycle will be maintained. constant; If the confidence level of gesture recognition is ≥ and Then, the asymmetric dead zone control mechanism in step 2) is used for correction; If the confidence level of gesture recognition is < If so, a progressive security response will be triggered. Step 4) After the output is executed, the user's subjective evaluation is collected using the user feedback and adaptive module, and closed-loop adaptive calibration is performed based on the subjective evaluation.

2. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, JND calibration can be performed using any of the following methods; The first method: use the pre-stored default thresholds as the initial minimum perceptible difference in force perception and the initial minimum perceptible difference in gesture recognition; The initial minimum perceptible difference in force perception and the initial minimum perceptible difference in gesture recognition are calibrated using a recursive least squares algorithm. The steps include: Initialize the estimated values ​​and covariance; Monitor users’ implicit or explicit feedback behavior to obtain observations y(k); Calculate Kalman gain and update the parameter estimates. Covariance Where λ is the forgetting factor; R is the covariance of the previous time step; R is the variance. When the covariance is less than a preset threshold and the relative change of the estimated value is less than a preset threshold multiple times in a row, the calibration is determined to be converged. The method for obtaining the observed value y(k) is as follows: When calibrating the initial minimum perceptible force error, when the system detects that the user is actively applying force in the opposite direction within the dead zone, the system will take the current absolute value of the force error |e(t)| as the observed value. When multiple consecutive over-dead-zone corrections in opposite directions are detected, let the observed values... ; For coefficients; Minimum perceptible difference for gesture recognition During calibration, when the gesture confidence falls below the threshold Within a preset neighborhood and the user is If no intervention occurs within a given time period, the current gesture confidence value is used as the observation value; if the user reports an error, the observation value is the sum of the current gesture confidence value and the preset compensation amount. The second approach is to use the adaptive step method to test the minimum perceptible difference of force perception at multiple force levels, and to use a combination of offline playback and online verification to test the minimum perceptible difference of gesture recognition. The third method involves testing at a force level within the range of 30% to 50% of the user's maximum voluntary contraction force, and then combining this with historical user data to complete the JND calibration.

3. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, The physiological correction steps are as follows: The user's fatigue level is monitored in real time by measuring the operation duration and the median frequency of electromyography signals; attention is detected by detecting response delay; and a fatigue correction coefficient is generated accordingly. Attention correction coefficient Dynamic compensation is performed on the initial perceptual reference calibrated by JND to obtain the minimum perceptible difference in force perception. Minimum perceptible difference in gesture recognition ; Minimum perceptible difference in force perception Minimum perceptible difference in gesture recognition As shown below: In the formula, , The initial perceptual baseline calibrated for JND.

4. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, The torque commands of wearable robots include auxiliary force, resistive force, human-computer interaction force, gravity compensation force, and guiding force.

5. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, The steps for correcting the torque commands of a wearable robot using an asymmetric dead-zone control mechanism include: Step 2.1) Set the positive dead zone boundary Negative dead zone boundary ; Step 2.2) Determine whether the control increment error e(t) satisfies If satisfied, then issue the torque command. Otherwise, proceed to step 2.3). Step 2.3) If The torque command correction for the wearable robot is as follows: like The torque command correction for the wearable robot is as follows: in, and This is a correction factor.

6. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, The control increment error e(t) is shown below: in, The desired torque command given by the upper-level decision-maker. e(t) represents the torque command actually output by the controller in the previous control cycle; e(t) represents the torque increment expected in the current control cycle.

7. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, In a single control flow, if the number of iterations of the correction loop used in steps 2 and 3 to handle directional deviations is greater than the preset maximum number of iterations and the gesture recognition confidence level is ≥ , If the convergence fails, it is determined to be a failure, and a progressive safety response is executed.

8. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, The progressive security response consists of four response phases, each with a set response duration. The first stage involves maintaining output and waiting for user correction. In the second stage, the output is linearly decayed and haptic cues are provided. Output ;in, =0.5× ; The torque command value that triggers the safety response instantaneously. Output the target value; This is the total preset duration for the second phase. The timestamp for the start of the second phase is t, and the current timestamp is t. In the third stage, the output decays to zero and a voice alarm is provided. The fourth stage involves a complete shutdown, awaiting system reset. If a user-initiated corrective action is detected at any stage, the security response is terminated and normal control is restored.

9. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 1, characterized in that, Subjective evaluations include being too high, too low, or incorrect; The triggering conditions for closed-loop adaptive calibration are: the number of times the same type of subjective evaluation appears greater than or equal to m in the most recent n user action execution cycles, the error accumulation ratio is greater than the preset threshold, or the number of times the same type of subjective evaluation appears greater than or equal to k in the first h operations of the system's first use.

10. The adaptive assist control method for wearable robots based on minimum perceptible difference according to claim 9, characterized in that, The steps of closed-loop adaptive calibration include: S1) Determine the adjustment direction based on the feedback type: If the subjective evaluation is too high, then increase it. And reduce the negative dead zone boundary according to the dead zone conservation rule. This reduces positive sensitivity and increases negative sensitivity, creating an asymmetric structure that conforms to human perception; the dead zone conservation rule is, after adjustment... and The sum equals the original amount. and The sum of ; If the subjective evaluation is too small, then reduce it. And in accordance with the dead zone conservation rule, the negative dead zone boundary is increased accordingly. ; The adjustment step size is as follows: (6) in This is the dead zone boundary before adjustment. The adjustment amount is calculated based on user feedback; It is the positive excess recorded during the N most recent instances of excessive feedback that triggered this adaptive calibration. The average value; when the subjective evaluation is "too small", It is the negative excess in the recent N small feedback operations | The average value of |; This refers to the sensing sensitivity coefficient; If the subjective evaluation is incorrect, then adjust the minimum perceptible difference of the gesture recognition. Or it may trigger recalibration, separating gesture recognition errors from force control errors; The minimum perceptible difference adjustment method for gesture recognition is as follows: = + · ,in Direction coefficient; ), This is the preset maximum adjustment limit per transaction; Confidence level; Recalibration is triggered when error feedback occurs more than the preset number of times or when the complaint rate does not decrease after adjustment. S2) Monitor subsequent feedback; if the complaint rate decreases, retain the adjustment and update the perception sensitivity coefficient. = × If the complaint rate increases or complaints in the opposite direction occur, the adjustment will be rolled back, and the perceived sensitivity coefficient will be updated. = × ; This is the step size amplification factor. This is the step size reduction factor.

Images