Intelligent prosthesis control method and system with feedback mechanism

Abstract

The application discloses a kind of intelligent prosthesis control method and system with feedback mechanism, and the application relates to intelligent prosthesis control technical field, comprising the following steps: S1: acquisition step: through the sensor cluster of prosthesis foot bottom, ankle joint and knee joint deployment, real-time acquisition terrain feature, joint motion parameter and residual limb electromyogram signal.The intelligent prosthesis control method and system with feedback mechanism, through the double driving architecture of biomechanics cognitive model and reinforcement learning strategy model, in combination with the double verification mechanism of mutual inspection arbitration unit, realizes the adaptive accurate matching of feedback signal type, intensity and timing under complex dynamic environment.Biomechanics model ensures that feedback parameters meet clinical safety boundary, reinforcement learning model optimizes environmental adaptability in real time, mutual inspection arbitration resolves model conflict through physiological tolerance screening and control stability prediction, solves the problems of control misalignment, too high user cognitive load and walking confidence decline caused by traditional static rule feedback.

Description

Technical Field

[0001] This invention relates to the field of intelligent prosthetic control technology, specifically to an intelligent prosthetic control method and system with a feedback mechanism. Background Technology

[0002] Significant progress has been made in modern intelligent prosthetic control technology, particularly in the integration of feedback mechanisms. Existing solutions generally acquire prosthetic status or environmental information through integrated sensors and transmit basic feedback signals (such as pressure, position, or simple vibrations) to the user, aiming to bridge the gap in user perception and assist control decisions. However, these technologies have significant limitations when dealing with complex and dynamic real-world environments. Users' daily activities involve varied and unstructured terrain, such as uneven surfaces, slopes, or stairs, while gait requirements rapidly switch between walking, running, and turning. This results in high-speed, non-linear, and complex changes in the interaction between the prosthesis and the environment, including force patterns, kinematic and dynamic parameters. The core deficiency of current feedback mechanisms lies in their lack of intelligence and adaptability. Feedback strategies typically rely on preset fixed rules or simple linear mapping relationships, failing to perform intelligent analysis and decision-making based on real-time perceived multidimensional and dynamic interactive states (such as simultaneous impacts, lateral slips, and joint angle changes). The modal selection, intensity adjustment, application location, and output timing of feedback signals often remain static or have slow responses, making it difficult to accurately match the rapidly changing real-world situation. The result is that the feedback information received by the user is distorted, delayed, or overloaded, failing to accurately reflect the true state and deviations during the execution of the prosthesis.

[0003] This mismatch between feedback and the current state directly leads to inaccurate decision-making by the user or control system, increases the user's cognitive burden, and severely weakens walking confidence and control effectiveness in complex environments. Therefore, it is urgent to break through the static feedback model of existing technologies and improve the real-time intelligence and adaptability of the feedback mechanism. The key technical problem to be solved is: how to achieve adaptive real-time adjustment of the type, intensity, and timing of intelligent prosthetic feedback signals in complex dynamic environments, so as to accurately match the instantaneous state of the prosthetic-environment interaction with the user's control needs. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides a method for controlling an intelligent prosthesis with a feedback mechanism, comprising the following steps:

[0005] S1: Data Acquisition Steps: Real-time acquisition of terrain features, joint motion parameters, and residual limb electromyography signals through a cluster of sensors deployed on the soles of the prosthetic foot, ankle, and knee joints;

[0006] S2: Decision-making steps: Input multi-source data into the dynamic feedback decision-making module, and simultaneously execute the biomechanical cognitive model and the reinforcement learning strategy model to generate the theoretically optimal feedback parameter set and the real-time adaptive feedback parameter set, respectively.

[0007] S3: Mutual inspection arbitration step: physiological tolerance screening and control stability simulation verification are performed on the two parameter sets. When the parameter difference is less than the safety threshold, a fusion command is output; otherwise, a weighted arbitration based on user operation habits is triggered.

[0008] S4: Execution steps: Drive the multimodal actuator array of the residual limb interface according to the final instruction, and adjust the feedback type, intensity, position and output timing.

[0009] Preferably, the dynamic feedback decision module includes:

[0010] Biomechanical cognitive unit, storing a database of mapping relationships between joint kinematic chains and neural responses;

[0011] The reinforcement learning policy unit employs a deep deterministic policy gradient algorithm with a hierarchical reward mechanism.

[0012] The mutual inspection arbitration unit connects the two aforementioned units and includes a physiological tolerance screening device and a control stability assessment device.

[0013] Preferably, the operation of the biomechanical cognitive unit includes:

[0014] Match predefined feedback mode combination rules based on terrain features;

[0015] The output contains a set of parameters including feedback type priority, intensity safety interval, and time series reference window.

[0016] Preferably, the operation of the reinforcement learning policy unit includes:

[0017] Real-time reception of a six-dimensional tensor representing the environment, prosthesis, and user interaction state;

[0018] Output feedback type switching threshold, intensity gradient coefficient, and timing offset.

[0019] Preferably, the mutual inspection arbitration unit step includes the following steps when used:

[0020] Step a: Physiological tolerance screening: Compare feedback intensity parameters with a user pain threshold database;

[0021] Step b: Control stability simulation: Predict the plantar pressure trajectory after executing parameters;

[0022] Step c: When the outputs of both models pass the validation and the Euclidean distance between the parameters is less than the tolerance, generate a weighted fusion instruction;

[0023] Step d: If verification fails or the distance exceeds the limit, then assign weighted arbitration based on the user's high-frequency operation scenario preferences.

[0024] Preferably, the execution steps include:

[0025] Modulate the tactile vibration frequency to a preset vibration frequency range within the gait cycle, and / or control the electrical stimulation current within a preset current intensity range;

[0026] The phase deviation of the feedback pulse sequence is less than the preset phase threshold ratio of the gait period.

[0027] Preferably, it also includes a closed-loop verification step:

[0028] Real-time monitoring and feedback of plantar pressure trajectory after intervention; if the trajectory deviates from the set threshold three times in a row, it will roll back to the previous valid parameters.

[0029] Regularly optimize the parameters of the decision-making model based on users' walking confidence scores.

[0030] An intelligent prosthetic control system for implementing an intelligent prosthetic control method with a feedback mechanism includes:

[0031] The sensor cluster module includes a three-dimensional force sensor for the sole of the foot, a joint inertial measurement unit, and electromyography (EMG) acquisition electrodes.

[0032] The dynamic feedback decision-making module, which is connected to the sensor cluster module, includes a biomechanical cognitive unit, a reinforcement learning strategy unit, and a mutual inspection and arbitration unit.

[0033] A multimodal actuator array module, with a piezoelectric vibrating pad and nerve electrical stimulation electrodes embedded in the residual limb receiving cavity.

[0034] Preferably, the mutual inspection arbitration unit includes:

[0035] Physiological tolerance screener, accesses a database of users' historical pain thresholds;

[0036] Control stability evaluator with built-in gait trajectory prediction algorithm;

[0037] The weight allocator stores a table of user scenario operation preference weights.

[0038] Preferably, it also includes a safety switching module:

[0039] When the mutual inspection arbitration unit outputs conflict commands three times in a row, it switches to the conservative mode that only enables the biomechanical cognitive unit and activates the remote expert communication interface.

[0040] This invention provides an intelligent prosthetic limb control method and system with a feedback mechanism. It has the following beneficial effects:

[0041] This intelligent prosthetic limb control method and system with a feedback mechanism achieves adaptive and precise matching of feedback signal type, intensity, and timing in complex dynamic environments through a dual-drive architecture of a biomechanical cognitive model and a reinforcement learning strategy model, combined with a dual verification mechanism of a mutual inspection and arbitration unit. The biomechanical model ensures that feedback parameters meet clinical safety boundaries, the reinforcement learning model optimizes environmental adaptability in real time, and the mutual inspection and arbitration resolves model conflicts through physiological tolerance screening and control stability prediction. This solves the problems of control inaccuracy, excessive user cognitive load, and decreased walking confidence caused by traditional static rule-based feedback. Users can receive tactile / electrical stimulation feedback synchronized in real time with the prosthetic limb-environment interaction state, effectively improving control accuracy and safety in complex terrain. Attached Figure Description

[0042] Figure 1 This is an interaction diagram of the intelligent prosthetic control module with feedback mechanism of the present invention;

[0043] Figure 2 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1 and Figure 2 This invention provides a technical solution: an intelligent prosthetic limb control method with a feedback mechanism, comprising the following steps:

[0046] S1: Data Acquisition Steps: Real-time acquisition of terrain features, joint motion parameters, and residual limb electromyography signals through a cluster of sensors deployed on the soles of the prosthetic foot, ankle, and knee joints;

[0047] S2: Decision-making steps: Input multi-source data into the dynamic feedback decision-making module, and simultaneously execute the biomechanical cognitive model and the reinforcement learning strategy model to generate the theoretically optimal feedback parameter set and the real-time adaptive feedback parameter set, respectively.

[0048] S3: Mutual inspection arbitration step: physiological tolerance screening and control stability simulation verification are performed on the two parameter sets. When the parameter difference is less than the safety threshold, a fusion command is output; otherwise, a weighted arbitration based on user operation habits is triggered.

[0049] S4: Execution steps: Drive the multimodal actuator array of the residual limb interface according to the final instruction, and adjust the feedback type, intensity, position and output timing.

[0050] It should be further explained that, in the specific implementation process, the three-dimensional force sensor on the sole of the prosthetic foot, the inertial measurement unit at the joint, and the electromyographic electrodes on the surface of the residual limb socket are used to collect data on terrain ruggedness, joint angle and angular velocity changes, and user muscle activation signals in real time. The above-mentioned multi-source heterogeneous data is input into the dynamic feedback decision module, and the biomechanical cognitive model and the reinforcement learning strategy model are launched simultaneously. The biomechanical cognitive model matches the current terrain features and motion state according to the pre-constructed joint kinematic chain-neural response mapping library, and outputs the theoretically optimal set of feedback parameters, including the suggested combination of feedback types, the intensity safety range, and the timing reference window. The feedback type combination includes vibration + electrical stimulation, the intensity safety range includes electrical stimulation of 0.1-2mA, and the timing reference window includes triggering 50ms after heel strike. The reinforcement learning strategy model is based on the real-time generated six-dimensional spatiotemporal state tensor and dynamically calculates the feedback type switching threshold, intensity gradient coefficient, and timing offset through a hierarchical reward mechanism to generate a real-time adaptive parameter set.

[0051] The two parameter sets undergo mutual verification and arbitration: First, physiological tolerance is screened by comparing the feedback intensity value with the user's historical pain threshold database. If the tolerance limit is exceeded, the parameter instruction is automatically blocked. Second, control stability simulation is performed by predicting the trajectory of the plantar pressure center after the parameters are executed based on the current prosthesis status. If the trajectory deviation exceeds the gait stability tolerance, the scheme is rejected. When both model outputs pass verification and the parameter difference is less than the preset safety tolerance, the weighted average of the two parameter sets is used to generate a fusion instruction. If the verification fails or the difference exceeds the limit, weights are assigned according to the user's high-frequency operation scenario preferences. For example, in the scenario of going up and down stairs, the reinforcement learning strategy has a significantly higher weight than the biomechanical model, and a weighted arbitration instruction is generated.

[0052] The final decision command drives the multimodal actuator array module embedded in the residual limb receiving cavity: if the command requires tactile feedback, the frequency of the piezoelectric vibrating plate is modulated to the range of 100-300 Hz within 20 ms; if the command includes electrical stimulation, the current intensity is precisely controlled within the safe range of 0.1-5 mA; the deviation between the trigger time of the feedback pulse sequence and the reference time of the gait cycle is strictly less than one-twentieth of the gait cycle.

[0053] By employing a progressive logic of dual-model generation, dual-verification, and dynamic arbitration, the system ensures real-time and accurate matching of feedback parameters under complex environments, thus resolving the problem of control inaccuracy caused by static feedback lag.

[0054] The dynamic feedback decision-making module includes:

[0055] Biomechanical cognitive unit, storing a database of mapping relationships between joint kinematic chains and neural responses;

[0056] The reinforcement learning policy unit employs a deep deterministic policy gradient algorithm with a hierarchical reward mechanism.

[0057] The mutual inspection arbitration unit connects the two aforementioned units and includes a physiological tolerance screening device and a control stability assessment device.

[0058] It should be further explained that, in the specific implementation process, the intelligent prosthetic control system includes a multi-physics sensor cluster module. Three-dimensional force sensors deployed on the sole of the prosthetic foot capture the distribution of ground reaction forces and terrain tilt in real time. Inertial measurement units at the ankle and knee joints collect joint angular acceleration and motion trajectory. Electromyographic electrodes on the surface of the residual limb socket monitor the user's muscle activation patterns. All sensor data is transmitted to the dynamic feedback decision module via a high-speed bus. The biomechanical cognition unit of this module has a built-in joint kinematic chain-neural response mapping library, which calls a predefined set of feedback rules based on terrain classification results. The reinforcement learning strategy unit runs a deep deterministic policy gradient algorithm with a hierarchical reward mechanism, inputting the fused spatiotemporal state tensor into the policy network and dynamically outputting feedback adjustment parameters. The spatiotemporal state tensor includes terrain risk level, joint load coefficient, and user intent label.

[0059] The mutual inspection and arbitration unit, acting as the core hub, receives output parameters from both units in real time. Its physiological tolerance screener accesses the user's historical pain threshold database, such as the individualized electrical stimulation upper limit for amputees. The control stability evaluator predicts the deviation of the plantar pressure trajectory based on the gait dynamics model. When the difference between the parameters of the two models exceeds the safety tolerance, a weighted arbitration command is generated according to the user's high-frequency operation scenario preferences, such as prioritizing reinforcement learning strategies when going up and down stairs. Through the system architecture of multi-source perception, dual-model parallel decision-making, and mutual inspection and arbitration closed loop, the feedback control is ensured to both comply with biomechanical safety constraints and adapt to complex environmental changes in real time.

[0060] The operation of the biomechanical cognitive unit includes:

[0061] Match predefined feedback mode combination rules based on terrain features;

[0062] The output contains a set of parameters including feedback type priority, intensity safety interval, and time series reference window.

[0063] It should be further explained that, in the specific implementation process, the operation of the biomechanical cognitive unit is as follows: When the sensor cluster module detects that the terrain feature is an uneven road surface, it automatically activates the preset "rough terrain mode", retrieves the corresponding feedback rule set from the joint kinematic chain-neural response mapping library, and if the plantar pressure distribution shows a sudden increase in load in the forefoot area and the ankle joint inversion angle exceeds the limit, it determines that there is a risk of lateral slippage, and the priority of the output feedback type is "high frequency vibration + medium intensity electrical stimulation"; For the stair climbing scenario, when it is recognized that the knee joint flexion angle is continuously greater than 50 degrees and the angular velocity is positive, the "stair climbing mode" rule is triggered, and the timing reference window is set to start vibration feedback 30 milliseconds after the toes touch the ground, and an electrical stimulation confirmation signal is superimposed when the knee joint extends to the maximum angle; All output parameters include an intensity safety range. For example, the upper limit of electrical stimulation intensity is dynamically set according to the user's historical pain data. If the user's maximum tolerated current last week was 2.1 mA, then the output range this time is automatically limited to 0.1 to 2.0 mA.

[0064] Through a four-layer progressive logic of terrain modeling, risk diagnosis, temporal anchoring, and individualized constraints, clinical medical knowledge is transformed into executable control rules, binding the temporal reference window with the action phase and solving the problem of disconnect between feedback and action.

[0065] The operations of the reinforcement learning policy unit include:

[0066] Real-time reception of a six-dimensional tensor representing the environment, prosthesis, and user interaction state;

[0067] Output feedback type switching threshold, intensity gradient coefficient, and timing offset.

[0068] It should be further explained that, in the specific implementation process, the reinforcement learning strategy unit receives the fused six-dimensional spatiotemporal state tensor in real time, which includes interactive features such as terrain risk level, joint load coefficient, and user electromyographic intention label; when the user's electromyographic signal shows an acceleration intention and the terrain radar identifies a wet subway staircase, the deep deterministic strategy gradient algorithm of the hierarchical reward mechanism initiates multi-objective optimization: the first-layer reward function calculates the deviation between the trajectory of the plantar pressure center and the ideal gait, and the smaller the deviation, the higher the reward value; the second-layer reward function evaluates the user's real-time comfort score and dynamically adjusts the weights based on pain feedback records in similar scenarios in historical data; the third-layer reward function correlates the energy consumption efficiency of the prosthetic motor and suppresses high-power parameter combinations.

[0069] Based on the weighted sum of triple rewards, the policy network outputs a feedback type switching threshold. For example, in the scenario of slippery stairs in the rain, the vibration warning threshold is adjusted from the usual 15 Newtons to 8 Newtons to achieve earlier risk warning. At the same time, an intensity gradient coefficient is generated. If the current ankle joint load reaches 80% of the safety limit, the electrical stimulation intensity is automatically increased to 90% of the upper limit of the tolerance range. For the timing offset, when the user's cadence suddenly increases by 20%, the algorithm compresses the feedback window and triggers the pulse sequence 10 milliseconds in advance to ensure synchronization.

[0070] Through a reinforcement learning decision chain of environmental perception, hierarchical optimization, and dynamic output, real-time and accurate adaptation of feedback parameters is achieved within the biomechanical safety boundary. The environmental response logic of the three-layer reward function is highlighted, particularly the temporal compression of threshold adjustment and step frequency-driven mechanisms in rainy / slippery scenarios, which solves the problem that fixed rules cannot cope with sudden environmental changes. The load correlation design of the intensity gradient coefficient reflects safe collaboration with the biomechanical model, laying the foundation for subsequent mutual inspection and arbitration.

[0071] The steps for using the mutual inspection arbitration unit include the following:

[0072] Step a: Physiological tolerance screening: Compare feedback intensity parameters with a user pain threshold database;

[0073] Step b: Control stability simulation: Predict the plantar pressure trajectory after executing parameters;

[0074] Step c: When the outputs of both models pass the validation and the Euclidean distance between the parameters is less than the tolerance, generate a weighted fusion instruction;

[0075] Step d: If verification fails or the distance exceeds the limit, then assign weighted arbitration based on the user's high-frequency operation scenario preferences.

[0076] It should be further explained that, in the specific implementation process, when there is a conflict between the theoretically optimal parameter set output by the biomechanical cognitive unit and the adaptive parameter set generated by the reinforcement learning strategy unit, the mutual inspection arbitration unit initiates a dual verification process. The theoretically optimal parameter set includes a suggested heel vibration intensity of level 3 + electrical stimulation of 1.5mA in staircase mode, while the adaptive parameter set generated by the reinforcement learning strategy unit includes a suggested vibration intensity of level 4 + electrical stimulation of 2.0mA in slippery rain conditions.

[0077] The mutual inspection and arbitration unit initiates a dual verification process, including: first, performing a physiological tolerance screening, retrieving the user's electrical stimulation pain threshold record from the previous week; if historical data shows that 2.0mA stimulation has caused a pain score exceeding 7 points (out of 10), then the parameter instruction is blocked; after passing the screening, the system enters a stability control simulation, predicting the plantar pressure trajectory after vibration intensity level 4 based on the current ankle joint angle and ground friction coefficient; if the simulation shows that the pressure center offset exceeds the gait stability tolerance, such as a sudden 30% drop in load in the forefoot area, then the reinforcement learning scheme is rejected.

[0078] When both schemes pass verification and the parameter difference is less than the safety tolerance, such as a vibration intensity difference of 1 level or a current difference of 0.3mA, the weighted instruction is generated according to the user's set staircase scenario preference, i.e., reinforcement learning weights dominate, such as a final vibration intensity of 3.8 level + electrical stimulation of 1.7mA; if the simulation shows that the biomechanical scheme will cause trajectory deviation, such as insufficient vibration intensity failing to warn of slippage, then the reinforcement learning scheme is fully adopted and online fine-tuning is triggered to reduce the vibration intensity switching threshold for subsequent similar scenarios.

[0079] By employing a four-order decision chain of "physiological screening - stability verification - scenario arbitration - model evolution," risk response is achieved in sudden cases. The key features are "stability prediction based on physical simulation" and "dynamic arbitration driven by scenario preferences," addressing the pain point that fixed rules cannot resolve model conflicts.

[0080] The execution steps include:

[0081] The tactile vibration frequency is modulated to a preset vibration frequency range of 100 to 300 Hz during the gait cycle, and / or the electrical stimulation current is controlled within a preset current intensity range of 0.1 to 5 mA.

[0082] The phase deviation of the feedback pulse sequence is less than one-twentieth of the preset phase threshold ratio of the gait cycle.

[0083] It should be further explained that, in the specific implementation process, the multimodal actuator array module is driven according to the final arbitration instruction: when the instruction requires tactile feedback, the piezoelectric vibrating pad is activated at the heel strike phase of the gait cycle, and the vibration frequency is precisely modulated to the range of 100 to 300 Hz within 20 milliseconds through the phase-locking circuit. If the arbitration instruction specifies a "rain and slippery warning mode", then a high-frequency vibration of 280 Hz is used to enhance risk perception. For the electrical stimulation channel, the current intensity is strictly limited to the safe range of 0.1 to 5 mA. When the upper limit of the output intensity of the biomechanical model is 2.0 mA and the arbitration instruction requires 1.7 mA, the electrical stimulation controller automatically loads the gradient ascent algorithm to avoid sudden stabs.

[0084] The core of timing control lies in the synchronous anchoring of the pulse sequence and gait phase: In the scenario of climbing stairs, vibration feedback is strictly constrained to trigger within a window of 30±5 milliseconds after the toes touch the ground, and the timing reference is calibrated in real time through a plantar switch sensor; when the user's gait frequency increases, causing the gait cycle to shorten to 80% of the original cycle, the feedback window is automatically compressed to one-twentieth of the gait cycle, ensuring that the vibration pulse is output before the knee flexion peak. Through the timing control chain of "event anchoring - environmental adaptation - safety boundary", the abstract time precision is transformed into detectable biomechanical events and physical tolerances. This solves the problem that fixed timing cannot adapt to sudden changes in gait.

[0085] It also includes a closed-loop verification step:

[0086] Real-time monitoring and feedback of plantar pressure trajectory after intervention; if the trajectory deviates from the set threshold three times in a row, it will roll back to the previous valid parameters.

[0087] Regularly optimize the parameters of the decision-making model based on users' walking confidence scores.

[0088] It should be further explained that, in the specific implementation process, after the feedback command is executed, the change in the trajectory of the plantar pressure center is monitored in real time. If the trajectory deviation of three consecutive steps exceeds the dynamically set threshold, such as a lateral drift of the pressure center greater than 15 mm in a stair climbing scenario, the system automatically rolls back to the previous valid decision parameters and marks the current situation as abnormal. Simultaneously, decision tracing analysis is initiated. When the rollback event is caused by biomechanical model parameters, the rule base for that scenario is frozen and the weight of the reinforcement learning strategy is increased to 90%. If the deviation is caused by reinforcement learning parameters, a high penalty term in its reward function is triggered and the exploration rate is reduced. The user's walking confidence score is aggregated weekly and compared with the baseline value of prosthetic energy consumption under the same terrain. When the average score of a specific scenario is below 6 points for two consecutive weeks and the energy consumption increases by more than 20%, the rule base of the biomechanical cognitive unit is optimized offline. The side slip judgment criterion for gravel roads is expanded to include the ground humidity factor, and the hierarchical reward weight of the reinforcement learning strategy is recalibrated so that the proportion of the comfort factor is increased to the extreme value of the safety limit. The user's walking confidence score is recorded based on a 10-point scale.

[0089] By employing a three-tiered closed loop of "real-time monitoring - failure attribution - long-term iteration," the dynamic optimization of the core decision-making mechanism is visualized. Utilizing terrain-adaptive dynamic tolerance and model failure differentiation compensation, the problem of static systems being unable to continuously adapt to user changes is addressed.

[0090] An intelligent prosthetic control system for implementing an intelligent prosthetic control method with a feedback mechanism includes:

[0091] The sensor cluster module includes a three-dimensional force sensor for the sole of the foot, a joint inertial measurement unit, and electromyography (EMG) acquisition electrodes.

[0092] The dynamic feedback decision-making module, which is connected to the sensor cluster module, includes a biomechanical cognitive unit, a reinforcement learning strategy unit, and a mutual inspection and arbitration unit.

[0093] A multimodal actuator array module, with a piezoelectric vibrating pad and nerve electrical stimulation electrodes embedded in the residual limb receiving cavity.

[0094] It should be further explained that, in the specific implementation process, the system uses an embedded three-dimensional force sensor matrix on the sole of the foot to capture the distribution of ground reaction forces in the forefoot, heel, and lateral regions in real time. Combined with nine-axis inertial measurement units at the ankle and knee joints, it simultaneously acquires three-axis acceleration, three-axis angular velocity, and three-axis magnetic force data. An electromyography (EMG) electrode array covers the proximal nerve-rich area of ​​the residual limb's receiving cavity, extracting surface EMG signal features at a sampling rate of 2000 Hz. The aforementioned multi-source data is transmitted via anti-interference shielded cables to the core processor of the dynamic feedback decision module. This processor has a built-in dual-core architecture: the biomechanical cognition unit runs on a real-time operating system kernel, calling a pre-compiled joint kinematic chain mapping library; the reinforcement learning strategy unit is deployed on a Linux kernel, executing a deep deterministic policy gradient algorithm; the two cores exchange spatiotemporal state tensors and parameter proposals via shared memory.

[0095] The mutual testing arbitration unit, acting as an independent coprocessor, receives the dual-core output and activates the physiological tolerance screener and the control stability evaluator. The arbitration result is sent via the CAN bus to the piezoelectric vibrating pad array and programmable electrical stimulation electrode embedded in the receiving cavity. The electrical stimulation channel is equipped with an overcurrent protection circuit to ensure that the output is strictly limited to the safe range of 0.1 to 5 mA. The physiological tolerance screener is used to access the encrypted user pain database, and the stability evaluator is used to call the gait dynamics simulation engine.

[0096] The mutual inspection and arbitration unit includes:

[0097] Physiological tolerance screener, accesses a database of users' historical pain thresholds;

[0098] Control stability evaluator with built-in gait trajectory prediction algorithm;

[0099] The weight allocator stores a table of user scenario operation preference weights.

[0100] It should be further explained that, in the specific implementation process, the mutual inspection arbitration unit accesses the encrypted user pain threshold curve database in real time through the physiological tolerance screener. When the reinforcement learning strategy unit outputs electrical stimulation intensity parameters, it automatically matches the historical pain records of the same anatomical points of the current user. If the parameter value exceeds the highest value of the three most recent painless stimuli, such as 1.8mA being the safe upper limit for the user's gastrocnemius muscle point, the instruction is immediately intercepted and marked as a high-risk operation. The parameter set screened is input into the control stability evaluator, which calls the gait dynamics lightweight engine to simulate the plantar pressure trajectory within the next 200 milliseconds based on the current ankle joint angle, ground friction coefficient, and feedback intensity parameters. If the prediction shows that performing vibration intensity level 4 in an ice environment will cause the lateral shift of the pressure center to exceed the ice surface tolerance limit, the scheme is automatically rejected and a high penalty term of the reinforcement learning reward function is triggered. The weight allocator dynamically adjusts the arbitration weights based on the pre-stored risk-scenario matrix: when a user is detected to be in a peak scenario of a subway transfer staircase, the "efficiency-safety" preference label is activated, and the weight of the reinforcement learning strategy is increased to 3 times that of the biomechanical model; if the environmental radar identifies a quiet home environment, it switches to the "comfort priority" mode, and the weight of the biomechanical model increases to 2.5 times that of the reinforcement learning.

[0101] Through a triple mutual inspection mechanism of "physiological real-time interception - physical simulation prediction - scenario-based weight allocation", abstract decisions are transformed into quantifiable safe operations. Emphasizing "painless dynamic defense based on historical highest values" and "multi-parameter gait simulation veto", it solves the problem that a single safety mechanism cannot cope with complex risks.

[0102] It also includes a secure switching module:

[0103] When the mutual inspection and arbitration unit outputs conflicting commands three times consecutively, it switches to a conservative mode that only activates the biomechanical cognitive unit and activates the remote expert communication interface. It should be further noted that in specific implementation, when the mutual inspection and arbitration unit outputs conflicting commands three times consecutively, such as the biomechanical model continuously recommending 1.6mA electrical stimulation while the reinforcement learning strategy insists on a 2.0mA scheme and both verifications fail, the safety switching module immediately initiates a level-three emergency response: first, it forcibly disconnects the decision-making authority of the reinforcement learning strategy unit, downgrading the system to a conservative mode that only runs the biomechanical cognitive unit. In this mode, the feedback type only retains the basic vibration channel, and the upper limit of electrical stimulation intensity is locked at 80% of the user's historical pain-free average. For example, if the user's average tolerance is 1.8mA, it is limited to 1.44mA. The timing window is expanded to one-tenth of the gait cycle to reduce synchronization accuracy requirements; at the same time, the remote expert communication interface is activated to transmit the current environment multimodal data stream through an encrypted channel, including terrain radar point cloud, joint motion trajectory and failure decision log, and to receive remotely issued temporary rule patches, such as a special vibration frequency scheme for the metal grid terrain of subway gates; if no manual instruction is received within 10 minutes, a minimalist feedback strategy based on terrain classification is automatically activated, and a 200 Hz standard vibration at 50% phase point of each step cycle is triggered when a smooth hard ground is detected, until the system self-check passes and then the dual-model decision is gradually restored.

[0104] By adopting a four-stage emergency response system of "conflict diagnosis, security degradation, human-machine collaboration, and autonomous recovery", the risk of core decision failure is controlled within a clinically acceptable range, thus solving the problem of user loss of control caused by system stagnation.

[0105] It should be further explained that, in the specific implementation process, the intelligent prosthetic control system captures environmental terrain features, prosthetic kinematic parameters, and user electromyographic intention signals in real time through sensor cluster modules deployed on the soles of the feet, joints, and residual limb sockets. Multi-source heterogeneous data are fused to generate a spatiotemporal tensor input dynamic feedback decision module representing the interaction state between the environment, prosthesis, and user. This module runs a biomechanical cognitive model and a reinforcement learning strategy model in parallel: the biomechanical model matches the theoretically optimal feedback parameters for the current scene based on a pre-constructed joint kinematic chain and neural response mapping library, including type combination intensity safety intervals and time-series benchmark windows; the reinforcement learning model dynamically calculates the feedback type switching threshold, intensity gradient coefficient, and time-series offset based on a deep deterministic policy gradient algorithm combined with a hierarchical reward function.

[0106] The parameter sets output by the two models are input into a mutual verification arbitration unit for dual validation: a physiological tolerance screener retrieves the user's historical pain threshold database and compares the intensity parameters; if the intensity exceeds the individual's pain-free upper limit, the command is immediately intercepted; a control stability evaluator predicts the plantar pressure trajectory after the parameters are executed based on a gait dynamics model, and rejects the scheme if the offset exceeds the terrain adaptive tolerance. If the difference in the validated parameter set is less than the safety tolerance, it is fused and output according to the user's scenario preference weights; if the difference exceeds the limit or validation fails, a weight allocator is activated to dynamically tilt the arbitration weights according to the high-frequency operation scenario type. For example, in the subway transfer staircase scenario, the reinforcement learning weights are significantly higher than those of the biomechanical model.

[0107] The final arbitration command drives the multimodal actuator array module: the piezoelectric vibrator is modulated to the target frequency range within twenty milliseconds, and the trigger electrode outputs a safe current intensity under the constraint of the gradient loading algorithm. The feedback timing is strictly anchored to key gait phase events, such as triggering within a fixed delay window after the toes touch the ground, and the window length is dynamically compressed with the user's step frequency to ensure synchronization with the action.

[0108] The system monitors and provides real-time feedback on the plantar pressure trajectory after intervention. If multiple consecutive steps deviate beyond the limit, it rolls back to the previous valid parameters and traces the source to mark the failed model. It aggregates user walking confidence scores and prosthetic energy consumption data over a long period. When the score consistently falls below a set threshold and energy consumption rises abnormally in a specific scenario, it offline optimizes the biomechanical rule base and recalibrates the reinforcement learning reward weights.

[0109] When the mutual inspection arbitration unit outputs conflicting commands multiple times consecutively, the safety switching module forcibly downgrades to a conservative mode that only enables the biomechanical cognitive unit, locking the basic vibration channel and limiting the intensity upper limit to the historical painless average ratio, relaxing the timing accuracy requirements, and simultaneously activating the remote expert interface to transmit environmental data streams and receive manual rule patches. If there is no response within the timeout period, the terrain classification simplified strategy is activated to maintain basic tactile feedback until the system self-check passes, after which the dual-model decision weights are gradually restored.

[0110] It should be further explained that, in the specific implementation process, the biomechanical cognitive model calls a predefined rule set based on the terrain classification results. For example, when an uneven road surface and abnormal plantar pressure distribution are detected, the rugged terrain mode is activated to output a combination of high-frequency vibration and electrical stimulation; in the stair climbing scenario, the vibration trigger window is set by binding the toe-touching event. The reinforcement learning strategy model dynamically adjusts parameters based on the real-time spatiotemporal tensor. For example, in rainy and slippery environments, the vibration warning threshold is automatically lowered, and the gradient coefficient of electrical stimulation intensity is increased when walking on ice. When the outputs of the two models conflict, the mutual inspection arbitration unit prioritizes physiological tolerance screening: if the electrical stimulation intensity proposed by the reinforcement learning exceeds the highest value of the user's painless record at the same anatomical point last week, it is intercepted and triggers a reduction in the output upper limit at that point.

[0111] The stability evaluator constructs a virtual gait environment by inputting the current ankle angle, ground friction coefficient, and feedback intensity parameters. A lightweight dynamics engine simulates the plantar pressure trajectory over the next 200 milliseconds. If the prediction indicates that executing the parameters will cause the pressure center to shift beyond the ground deviation limit (e.g., a shift greater than 15 mm on an ice surface), the scheme is automatically rejected, and the shift is converted into a penalty term in the reinforcement learning reward function. Simulation results are simultaneously used for temporal optimization. When the user's cadence increases, the feedback window duration is proportionally compressed to ensure that the vibration pulse is output before the knee flexion peak.

[0112] Three consecutive failed mutual inspections trigger a Level 3 response: Immediately strip reinforcement learning decision-making authority and switch to a biomechanical conservative mode. In this mode, only a single vibration channel is retained, electrical stimulation is disabled, and the intensity limit is locked at a fixed percentage of the individual's historical pain-free mean. Simultaneously, a remote expert collaboration interface is activated, encrypting and transmitting terrain point cloud data, joint motion trajectories, and decision failure logs. After remote manual analysis of special scenarios such as metal grid terrain, a vibration frequency adjustment patch is issued. If communication times out without response, the terrain type is automatically matched, a minimalist feedback strategy is enabled, and on smooth, hard surfaces, standard frequency vibrations are triggered at a fixed phase point in each cycle until the system self-checks and confirms the environmental risk has been eliminated. Then, the dual-model weights are restored according to a three-stage gradient.

[0113] The real-time rollback mechanism is linked to model responsibility determination: if the trajectory deviation is caused by biomechanical parameters, the rule base for that scene is frozen, and the reinforcement learning weight is elevated to a dominant position; if the reinforcement learning parameters cause failure, its exploration penalty coefficient is increased. Long-term optimization is based on a multi-dimensional evaluation matrix. When the user's walking confidence score is consistently below a set threshold and the energy consumption on the same terrain is significantly higher than the baseline value, offline rule base iteration is triggered. For example, the criteria for gravel roads are expanded to include surface humidity factors, and the comfort level of reinforcement learning is improved, with the reward weight increased to a safe and permissible extreme value.

[0114] The architecture, employing a "dual-model parallel decision-making, dual-safety verification, scenario-based dynamic arbitration, and fault-state human-machine collaboration" approach, covers the entire technology chain from environmental perception to safety control. It emphasizes individualized defense mechanisms based on physiological tolerance screening and stability prediction technology based on physical simulation, transforming clinical medical knowledge into computable safety rules and achieving real-time optimization of feedback parameters in complex environments through reinforcement learning.

[0115] A method for controlling an intelligent prosthesis with a feedback mechanism includes the following steps:

[0116] Step S1: Real-time acquisition of terrain features, joint motion parameters and user electromyographic intention signals through a three-dimensional force sensor on the sole of the foot, a joint inertial measurement unit and electromyographic electrodes of the residual limb, and fusion to generate a spatiotemporal state tensor;

[0117] Step S2: Synchronously input the spatiotemporal state tensor into the biomechanical cognitive model and the reinforcement learning strategy model:

[0118] The biomechanical cognitive model calls the joint kinematic chain-neural response mapping library to match the terrain scene and output the theoretically optimal feedback parameter set, including type combination, intensity safety interval and time series benchmark window; the reinforcement learning strategy model dynamically calculates the feedback type switching threshold, intensity gradient coefficient and time series offset based on the hierarchical reward mechanism to generate a real-time adaptive parameter set.

[0119] Step S3: Perform mutual arbitration on the output parameter sets of the two models:

[0120] Physiological tolerance screening: Compare feedback intensity parameters with the user's historical pain threshold database to block commands that exceed the limit;

[0121] Control stability simulation: Input joint angle and ground friction coefficient to predict the deviation of plantar pressure trajectory, and reject over-tolerance schemes;

[0122] When the parameter difference is less than the safety tolerance, the output is fused according to the user scenario preference weight; if the difference exceeds the limit or the verification fails, the arbitration weight is dynamically allocated according to the high-frequency operation type.

[0123] Step S4: Drive the piezoelectric vibrating plate and electrical stimulation electrodes according to the final instruction:

[0124] The vibration frequency is modulated to the target range within twenty milliseconds, and the timing is anchored to key phase events of the gait, such as the delay window after the toe touches the ground.

[0125] The electrical stimulation intensity is safely output using a gradient loading algorithm, and the window duration is compressed proportionally with the step frequency acceleration.

[0126] Step S5: Real-time monitoring and feedback of plantar pressure trajectory after intervention:

[0127] If the offset exceeds the dynamic threshold for multiple consecutive steps, roll back to the previous valid parameter.

[0128] The failure responsibility model is traced to determine the source, triggering corresponding weight compensation or rule freezing;

[0129] Step S6: Regularly aggregate user walking confidence scores and prosthetic energy consumption data:

[0130] When the score for a specific scenario continues to fall below the set threshold and energy consumption rises abnormally, the biomechanical rule base is optimized offline and the reinforcement learning reward weights are calibrated.

[0131] Step S7: If mutual arbitration fails multiple times consecutively, then activate the safety fallback:

[0132] Forced switch to only biomechanical conservative mode to lock the basic vibration channel and individualized intensity limit;

[0133] Activate the remote expert interface to transmit environmental data streams and load manually patched rules.

[0134] Step S8: If remote collaboration times out and fails to respond, activate the simplified terrain classification strategy: trigger fixed-phase standard vibration on smooth, hard surfaces; maintain basic feedback until the system self-test passes.

[0135] Step S9: Gradually restore the decision weights of the two models and increase the reinforcement learning authority in three stages.

[0136] Through a dual-driven architecture combining a biomechanical cognitive model and a reinforcement learning strategy model, along with a dual verification mechanism using a mutual arbitration unit, adaptive and precise matching of feedback signal type, intensity, and timing is achieved in complex dynamic environments. The biomechanical model ensures that feedback parameters meet clinical safety boundaries, the reinforcement learning model optimizes environmental adaptability in real time, and the mutual arbitration resolves model conflicts through physiological tolerance screening and control stability prediction. This addresses the issues of control inaccuracy, excessive user cognitive load, and decreased walking confidence caused by traditional static rule-based feedback. Users receive tactile / electrical stimulation feedback synchronized in real time with the prosthesis-environment interaction state, effectively improving control accuracy and safety in complex terrains.

[0137] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of intelligent prosthesis control with feedback mechanism, characterized in that, Comprising the following steps: S1: Acquisition step: Real-time acquisition of terrain features, joint motion parameters and residual limb electromyographic signals through a sensor cluster deployed on the prosthetic foot sole, ankle and knee joint; S2: Decision step: Input multi-source data into the dynamic feedback decision module, simultaneously execute the biomechanics cognitive model and the reinforcement learning strategy model, respectively generate the theoretical optimal feedback parameter set and the real-time adaptive feedback parameter set; S3: Mutual inspection arbitration step: Physiological tolerance screening and control stability simulation verification are performed on the two parameter sets, when the parameter difference is less than the safety threshold, the fusion instruction is output, otherwise the weighted arbitration based on user operation habit is triggered; S4: Execution step: According to the final instruction, drive the multi-modal actuator array of the residual limb interface, adjust the feedback type, intensity, position and output timing; The dynamic feedback decision module comprises: A biomechanics cognitive unit stores the mapping relationship database of joint motion chain and neural response; A reinforcement learning strategy unit adopts a deep deterministic policy gradient algorithm with a hierarchical reward mechanism; A mutual inspection arbitration unit is connected to the above two units and contains a physiological tolerance screener and a control stability evaluator; The operation of the biomechanics cognitive unit comprises: According to the terrain features, match the pre-defined feedback mode combination rule; Output the parameter set containing feedback type priority, intensity safety interval and timing reference window; The operation of the reinforcement learning strategy unit comprises: Real-time receive six-dimensional tensor representing the environment-prosthesis-user interaction state; Output feedback type switching threshold, intensity gradient coefficient and timing offset; The mutual inspection arbitration step comprises the following steps: Step a: Physiological tolerance screening: Compare the feedback intensity parameter with the user's pain threshold database; Step b: Control stability simulation: Predict the sole pressure trajectory after executing the parameter; Step c: When both model outputs pass the verification and the parameter Euclidean distance is less than the tolerance, generate a weighted fusion instruction; Step d: If the verification fails or the distance is out of limit, then allocate weights according to the user's high-frequency operation scene preference; The execution step comprises: Modulate the tactile vibration frequency to the preset vibration frequency range within the gait cycle, and / or control the electric stimulation current within the preset current intensity range; The phase deviation of the feedback pulse sequence is less than the preset phase threshold proportion of the gait cycle.

2. The intelligent prosthesis control method with feedback mechanism according to claim 1, characterized in that, Further comprising a closed-loop verification step: Real-time monitor the sole pressure trajectory after feedback intervention, if the trajectory deviation exceeds the set threshold for three consecutive times, roll back to the previous valid parameter; Periodically optimize the decision model parameters based on the user's walking confidence score.

3. An intelligent prosthesis control system implementing the method of any of claims 1-2, characterized by Comprise: Sensor cluster module, including three-dimensional force sensor on the sole, joint inertial measurement unit and electromyographic acquisition electrode; Dynamic feedback decision module, communication connection with the sensor cluster module and containing biomechanics cognitive unit, reinforcement learning strategy unit and mutual inspection arbitration unit; Multi-modal actuator array module, embedded in the residual limb receiving cavity, piezoelectric vibration piece and nerve electric stimulation electrode.

4. The system of claim 3, wherein: The mutual inspection arbitration unit comprises: Physiological tolerance screener, access to user historical pain threshold database; Control stability evaluator, built-in gait trajectory prediction algorithm; Weight distributor, store user scene operation preference weight table.

5. The intelligent prosthesis control system of claim 4, wherein, Further comprising a safety switching module: When the mutual inspection arbitration unit outputs the conflict instruction for three times in succession, switch to the conservative mode of enabling only the biomechanics cognition unit, and activate the remote expert communication interface.

Images