Artificial limb fixing pressure optimization method based on neural network

Abstract

The invention discloses an artificial limb fixing pressure optimization method based on a neural network. The method comprises the following steps: S1, outputting a standardized artificial limb interface pressure data set; s2, generating a prediction target pressure distribution diagram and a comfort level evaluation score; s3, constructing a grey wolf optimization algorithm pressure regulation model, and generating an initial grey wolf population solution vector set by taking a multi-target optimization framework jointly formed by the predicted target pressure distribution diagram, the comfort evaluation score and a preset medical safety threshold as a search space; s4, driving a pressure adjusting device of each section of the artificial limb multi-section pressure control execution module to perform synchronous pressure adjustment by taking the optimal pressure distribution parameter set as a set value; and S5, judging whether the artificial limb interface pressure data set subjected to synchronous pressure adjustment falls into a medical safety threshold value interval or not and whether the comfort level evaluation score reaches a preset threshold value or not. The regulation stability and the physiological matching degree are remarkably improved, and the problems that traditional centralized regulation response is lagged and regulation is uneven are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of prostheses, and in particular to a method for optimizing prosthesis fixing pressure based on a neural network. Background Art

[0002] With the development of biomedical engineering and intelligent control technology, prosthetic assistive systems are gradually evolving towards personalization and self-adaptation. Fine adjustment of prosthetic fixation pressure has become a key link in improving the wearing comfort, safety and functional stability of prosthetics. The prosthetic fixation methods currently commonly used in clinical and commercial applications mainly rely on static structural design or manual adjustment mode. The pressure at the prosthetic interface is adjusted based on the subjective experience of medical technicians or users. It has certain effects when dealing with static wearing states, but it is often difficult to effectively adapt to the dynamic changes of the user's limbs during movement, resulting in frequent problems of local pressure being too high or too low.

[0003] The main defects of existing prosthetic fixation technology are manifested in the following aspects: First, traditional prosthetic pressure adjustment methods are mostly manual adjustment or fixed structures, which cannot perceive changes in the user's limb status in real time, have low adjustment frequency and delayed response, and are prone to produce continuous high-pressure or low-pressure areas in specific areas when the user is walking, going up and down hills, or running, causing skin redness, ulcers, and even causing safety risks such as loosening and slipping of the prosthesis. Secondly, some systems that attempt to introduce electronic feedback control rely on simple threshold settings or finite logic control to understand the pressure data at a superficial level, making it difficult to capture the deep-level spatiotemporal evolution characteristics and biomechanical laws, and unable to achieve adaptive response to individual differences of the wearer. Third, there is currently a lack of an efficient and robust global optimization mechanism to determine the optimal pressure distribution strategy based on the actual wearing state, which makes the system adjustment strategy prone to falling into local optimality, affecting the adjustment accuracy and consistency of the effect.

[0004] In summary, there is an urgent need for a fusion method that combines dynamic perception, deep feature modeling and global optimization capabilities to improve the adaptability and reliability of prosthetic systems in complex application scenarios, so as to better meet the comprehensive needs of modern prosthetic users for comfort, safety and intelligent experience. Summary of the Invention

[0005] One purpose of the present invention is to propose a prosthetic fixation pressure optimization method based on a neural network. The present invention significantly improves the stability and physiological matching of the adjustment, and effectively overcomes the problems of response lag and uneven adjustment of traditional centralized control.

[0006] A method for optimizing prosthetic fixation pressure based on a neural network according to an embodiment of the present invention includes the following steps:

[0007] S1. Continuously collect prosthetic interface pressure data sets and combine and package them into a time-windowed prosthetic interface pressure data set, perform noise suppression, missing complementation, and normalization on the time-windowed prosthetic interface pressure data set, and output a standardized prosthetic interface pressure data set;

[0008] S2. Input the standardized prosthetic interface pressure data set into the MLP-Mixer network pressure feature model, extract the prosthetic interface pressure depth feature representation through the spatial mixing layer and channel mixing layer, and generate the predicted target pressure distribution map and comfort assessment score;

[0009] S3. Construct a gray wolf optimization algorithm pressure regulation model, using the multi-objective optimization framework formed by the predicted target pressure distribution map, comfort assessment score and preset medical safety threshold as the search space, generate an initial gray wolf population solution vector set, and use the individualized initialization strategy obtained by training the historical prosthetic interface pressure data set to set the position of the initial gray wolf population solution vector set so that the initial gray wolf population solution vector set is consistent with the predicted target pressure distribution map. In each round of iteration of the gray wolf optimization algorithm pressure regulation model, the positions of the remaining gray wolf solution vectors are updated according to the current optimal gray wolf solution vector and the predicted target pressure distribution map to obtain the gray wolf population solution vector set after iterative update, and determine whether the gray wolf population solution vector set after iterative update meets the convergence conditions of the multi-objective optimization framework. If so, the current optimal gray wolf solution vector is written into the optimal pressure distribution parameter set, otherwise the iteration continues;

[0010] S4. Each section of the prosthesis driving multi-segment pressure control execution module pressure regulating device to the optimal pressure distribution parameter set as the set value for synchronous pressure adjustment;

[0011] S5. Determine whether the prosthetic interface pressure data set after synchronous pressure adjustment falls within the medical safety threshold range and the comfort assessment score reaches the preset threshold. If so, execute the maintenance control mode; otherwise, return to step S2 and re-execute real-time closed-loop adjustment.

[0012] Optionally, the S1 includes the following steps:

[0013] S11. Collect the original prosthetic interface pressure data stream collected by the multi-channel flexible pressure sensor array deployed between the prosthetic limb and the human body, and construct an original prosthetic interface pressure data set. Each data in the original prosthetic interface pressure data set includes the acquisition timestamp t of the pressure data. i , instantaneous pressure value vector p i , and the corresponding flexible pressure sensor number s i ,The instantaneous pressure value vector represents the interface pressure value collected by each flexible pressure sensor at the same moment, and the flexible pressure sensor number indicates which flexible pressure sensor channel each data comes from;

[0014] S12. Segment the raw prosthetic interface pressure data set in chronological order to construct a time-windowed prosthetic interface pressure data set. Each time window contains all raw prosthetic interface pressure data collected within a set time interval. All raw prosthetic interface pressure data with a timestamp falling within the time interval are included in the time window.

[0015] S13. Perform noise suppression on the prosthetic interface pressure data in each time window. Smooth the pressure values within each flexible pressure sensor channel using a sliding average filter to obtain a noise-suppressed time window data set. During the smoothing process, the local average value of the pressure value at each sampling point is calculated in the time domain to replace the original pressure sampling value. The noise-suppressed pressure value is used to suppress data noise caused by high-frequency fluctuations.

[0016] S14. Perform missing completion operations on the noise-suppressed time window data set, so that each flexible pressure sensor channel has valid data in each time window. In the case of missing data, use the interpolation algorithm to estimate and complete the pressure value of the missing point, and perform normalization processing to obtain the standardized prosthetic interface pressure data set D norm During the normalization process, the completed pressure value of each flexible pressure sensor in each time window is converted into a standardized value. The standardized value is calculated by subtracting the minimum pressure value historically collected by the flexible pressure sensor from the current pressure value and dividing it by the difference between the maximum and minimum values historically collected by the flexible pressure sensor. The result is limited to between 0 and 1.

[0017] Optionally, the S2 includes the following steps:

[0018] S21. Standardize the prosthetic interface pressure data set D norm The improved MLP-Mixer network pressure feature model is input, and the normalized prosthetic interface pressure data in each time window is represented as the pressure state tensor X k ;

[0019] S22. According to the improved MLP-Mixer network pressure feature model, the pressure state tensor X k Perform local-global dual-scale spatial mixing processing. The local-scale spatial mixing operation uses an adaptive attention mechanism to dynamically calculate the pressure correlation weight of each flexible pressure sensor and its adjacent flexible pressure sensors to form a local spatial weighted matrix. And the pressure state tensor X k Perform local spatial blending to obtain the local scale-space blending tensor:

[0020]

[0021] Local spatial weighting matrix Elements in represents the relationship between flexible pressure sensor channel j and adjacent flexible pressure sensor channel j ′ The dynamic pressure correlation between the two in the current window pressure state enables the improved MLP-Mixer network pressure feature model to capture the local pressure abnormality area under the prosthesis wearing state;

[0022] S23. At the same time, perform global spatial mixing operations at the global scale, introduce a multi-head adaptive channel attention module to calculate the contribution weight of each flexible pressure sensor channel to the overall pressure balance in the current wearing scenario, and form a global spatial weighted matrix And obtain the global scale-space mixing tensor:

[0023]

[0024] Global spatial weighting matrix Chinese elements represents the relationship between flexible pressure sensor channel j and adjacent flexible pressure sensor channel j ′ The contribution weights of the two to the global pressure balance of the prosthetic interface enable the improved MLP-Mixer network pressure feature model to identify and optimize the overall pressure equilibrium state of the prosthetic interface;

[0025] S24. Mix the local scale space tensor Mixing tensors with global scale space Perform fusion operations to form a local-global fusion tensor:

[0026]

[0027] Among them, γ is the fusion coefficient, which is dynamically calculated according to the complexity of the current prosthetic interface pressure distribution. When the pressure distribution is relatively abnormal and complex, more attention is paid to the local pressure abnormal area, and when the pressure distribution is relatively stable, more attention is paid to the global pressure equilibrium state.

[0028] S25. Local-global fusion tensor Perform dynamic gated channel mixing processing, and introduce the gated recurrent unit GRU to dynamically mine and learn the pressure change trend of the flexible pressure sensor channel between different time steps, and obtain the pressure feature representation tensor F after channel mixing. k :

[0029]

[0030] The gated recurrent unit GRU combines the hidden state vector h of the previous time windowk-1 , dynamically update the hidden state vector h of the current time window k ,capturing the temporal evolution trend and subtle changes of pressure during prosthesis wearing, enabling the improved MLP-Mixer network pressure feature model to predict upcoming pressure anomalies or imbalance trends;

[0031] S26. Representing tensor F based on pressure characteristics k Calculate the predicted target pressure distribution map for the current time window and overall comfort evaluation score S k , the predicted target pressure value of each flexible pressure sensor channel in the predicted target pressure distribution diagram Based on the prediction of the stability and comfort of prosthesis wearing in the short term in the future, the overall comfort assessment score is generated by pressure feature representation tensor mapping, which is used to quantitatively characterize the real-time impact of the current pressure distribution on the overall comfort of the prosthesis wearer.

[0032] Optionally, the predicted target pressure distribution map is calculated as a pressure feature representation tensor F k Perform weighted averaging in the time dimension to obtain the time-sensitive pressure response mean vector μ of each flexible pressure sensor channel in the current window k =(μ k,1 ,μ k,2 ,...,μ k,M ),in:

[0033]

[0034] Among them, f t,j represents the depth characteristic value of the flexible pressure sensor channel j at time step t, ω t is a dynamic weight that decreases with time, which is used to highlight the dominant influence of pressure mutation on regulation in the short term, and T is the total time step;

[0035] The pressure response mean vector μ k The predicted target pressure distribution map is generated through the learnable nonlinear mapping function φ(·)

[0036]

[0037] Among them, W p is the trainable weight matrix, b p is the bias vector, σ(·) is the Sigmoid activation function, which is used to compress the output to the [0,1] interval so that each predicted target pressure value Indicates the pressure intensity that should be applied by the j-th flexible pressure sensor target in the current window.

[0038] Optionally, the overall comfort evaluation score is calculated as the pressure feature representation tensor F k Calculate its time-channel joint fluctuation tensor V k , where each element v t,j =|f t,j -μ k,j |, represents the instantaneous pressure deviation degree of each channel at each time step;

[0039] The time-channel joint fluctuation tensor V k Summarized into a single comfort impact index η k , and then mapped to the overall comfort evaluation score S through the nonlinear mapping function ψ(·) k :

[0040]

[0041] S k =ψ(η k )=exp(-α·η k );

[0042] Where M represents the total number of flexible pressure sensor channels, α>0 is the empirical adjustment coefficient, which represents the suppression strength of pressure fluctuation on comfort, and the overall comfort evaluation score S k Values closer to 1 indicate more stable pressure and a more comfortable fit, while values closer to 0 indicate severe imbalance or discomfort.

[0043] Optionally, S3 includes the following steps:

[0044] S31. According to the predicted target pressure distribution map Overall comfort evaluation score S k And the preset medical safety threshold pair (L, U), where L represents the lowest allowable pressure value of each flexible pressure sensor channel and U represents the highest allowable pressure value of each channel, construct a multi-objective optimization function F(Z) that integrates the predicted target pressure value, threshold crossing degree and comfort information:

[0045]

[0046] Among them, w1, w2, w3 are non-negative weight coefficients, l j Indicates the medical safety lower limit pressure value of channel j, u j represents the medical safety upper pressure limit value of channel j, z j represents the prosthesis fixation pressure value that should be applied by each flexible pressure sensor channel in channel j;

[0047] S32. Based on the historical prosthetic interface pressure data set D hist, respectively count the historical average pressure value and historical fluctuation variance of each flexible pressure sensor channel, construct an individualized initialization strategy, add Gaussian perturbations generated based on historical fluctuation variance to each channel with the predicted target pressure distribution map as the center, generate several initial gray wolf solution vectors, and form the initial gray wolf solution vectors into an initial gray wolf population solution vector set;

[0048] S33. Calculate the fitness function value of each initial gray wolf solution vector under the multi-objective optimization function F(Z), and sort them from small to large according to the fitness function value. Select the three gray wolves with the smallest fitness value as guide individuals, named α, β, and δ respectively, and their position vectors are

[0049] S34. During each iteration, the gray wolf optimization algorithm uses the positions of the three guide individuals α, β, and δ to guide the search direction of the remaining individuals. The position difference between each gray wolf and the three guide individuals is calculated, and the update amplitude is controlled by a random coefficient. The new position vector of each gray wolf after the update is the weighted average of its position with the positions of α, β, and δ.

[0050] In this process, a linearly decreasing convergence factor a is used. t , which is used to balance the exploration ability in the early stage of the search and the convergence ability in the later stage; at the same time, a random vector r generated by uniform distribution is used i,t To enhance the diversity of search, so that the gray wolf population can better escape the local optimum during the update;

[0051] S35. After completing the update of the gray wolf population solution vector in each round, recalculate the fitness function values of all gray wolf individuals under the multi-objective optimization function F(Z). Based on the updated fitness function values, reselect the three leading gray wolf individuals for the new round, i.e., the first three individuals with the smallest fitness values. At the same time, record the fitness function value of the best gray wolf in the current round.

[0052] S36. At the end of each iteration, determine whether any of the following conditions are met: first, whether the fitness function value of the current optimal gray wolf is less than or equal to a preset threshold; second, whether the current iteration round has reached the maximum number of iterations. If any of these conditions are met, the iteration is terminated; otherwise, the next iteration is entered.

[0053] S37. When the iteration of the gray wolf optimization algorithm ends, the position vector of the gray wolf individual with the minimum current fitness function value is The optimal pressure distribution parameter set as the final output The optimal pressure distribution parameter set represents the most suitable prosthetic interface pressure regulation target within the current time window, combined with comfort, prediction accuracy and medical threshold constraints.

[0054] Optionally, the S4 includes the following steps:

[0055] S41. Obtaining the optimal pressure distribution parameter set in, Indicates the optimal pressure value corresponding to the flexible pressure sensor channel j in the current time window, which serves as the pressure control setting target;

[0056] S42. Set the optimal pressure distribution parameters The target control pressure value set Q mapped to each independent control segment in the prosthesis multi-segment pressure control execution module k =(q k,1 ,q k,2 ,...,q k,N ), where N represents the actual number of segments of the prosthetic control execution module, q k,n Indicates the pressure setting value that should be applied to the nth pressure regulation section, according to the mapping relationship matrix M between the flexible pressure sensor channel and the physical control section map Determine as follows:

[0057]

[0058] Among them, m n,j Indicates whether sensor channel j belongs to control section n, which is used to achieve sensor-to-control unit aggregation;

[0059] S43. Control pressure value set Q according to target k The pressure of all the segment pressure regulating devices under the prosthesis multi-segment pressure control execution module is adjusted synchronously. The controller sets each pressure setting value q k,n As the setting target of the nth segment, the contact interface is dynamically loaded or unloaded by adjusting its corresponding driving unit to form a physical pressure output state consistent with the optimal pressure distribution parameter set;

[0060] S44. During the pressure adjustment process of each section, the execution module provides real-time feedback on the actual pressure value of the current section. and the pressure setting value q k,n Perform difference comparison, if there is a value that exceeds the allowable error threshold ∈ p deviation, that is The fine-tuning closed-loop control mechanism is immediately triggered to perform pressure adjustment and compensation operations on the section until the pressure in all sections is stabilized within the set range;

[0061] S45. After completing the synchronous pressure adjustment, the current actual pressure state is recorded as the adjusted prosthetic interface pressure data set in, is the jth flexible pressure sensor in time window tk The actual pressure value detected within.

[0062] The beneficial effects of the present invention are:

[0063] (1) The present invention introduces a local-global dual-scale attention mechanism into the MLP-Mixer network structure. The local spatial mixing module can accurately capture the local interaction relationship between flexible pressure sensors and is used to discover local high-pressure areas or mutation boundaries. The global spatial mixing module extracts the pressure coordination characteristics between all channels and establishes the perception capability of overall pressure balance. At the same time, a gated recurrent unit is introduced for time series modeling to dynamically capture the evolution law of pressure distribution with the change of wearing state and integrate it into a unified deep feature representation. It effectively solves the defect that traditional neural networks cannot take into account both local mutation response and global trend modeling at the same time, and has higher accuracy and response sensitivity in predicting the target pressure distribution of the prosthetic interface in the short term in the future.

[0064] (2) The present invention designs a multi-objective optimization structure that is highly adaptable to the prosthetic pressure regulation task, which is composed of a joint optimization objective function including a predicted pressure deviation term, a comfort penalty term, and a medical safety threshold penalty term. At the same time, a personalized Gaussian perturbation initialization strategy based on historical individual data is proposed, which enables the gray wolf population to be efficiently initialized around the predicted target distribution and shortens the convergence time. The iterative update mechanism in the gray wolf algorithm is combined with a dynamic convergence factor and a collaborative search strategy of three guiding individuals. While meeting the global optimal search capability, it improves the convergence stability to the physiological reasonable boundary, which is significantly better than the regulation performance of traditional genetic algorithms or particle swarm algorithms in high-dimensional, multi-constrained search spaces.

[0065] (3) The present invention constructs a mapping matrix between the flexible sensor channel and the physical pressure regulation section, calculates the pressure target of each control unit based on the target pressure distribution parameters, and monitors the actual pressure value in real time through a parallel pressure control feedback network. If there is an error in the pressure of a local section that deviates from the optimal set value, the system can immediately enter the fine-tuning compensation state to ensure that the pressure distribution of all pressure sections always converges to the optimal solution during dynamic movement, significantly improving the stability and physiological matching of the regulation, and effectively overcoming the problems of response lag and uneven regulation of traditional centralized regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0067] Figure 1 This is a flow chart of a neural network-based prosthetic fixation pressure optimization method proposed by the present invention. DETAILED DESCRIPTION

[0068] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0069] refer to Figure 1 , a prosthesis fixation pressure optimization method based on neural network, comprising the following steps:

[0070] S1. Continuously collect prosthetic interface pressure data sets and combine and package them into a time-windowed prosthetic interface pressure data set, perform noise suppression, missing complementation, and normalization on the time-windowed prosthetic interface pressure data set, and output a standardized prosthetic interface pressure data set;

[0071] S2. Input the standardized prosthetic interface pressure data set into the MLP-Mixer network pressure feature model, extract the prosthetic interface pressure depth feature representation through the spatial mixing layer and channel mixing layer, and generate the predicted target pressure distribution map and comfort assessment score;

[0072] S3. Construct a gray wolf optimization algorithm pressure regulation model, using the multi-objective optimization framework formed by the predicted target pressure distribution map, comfort assessment score and preset medical safety threshold as the search space, generate an initial gray wolf population solution vector set, and use the individualized initialization strategy obtained by training the historical prosthetic interface pressure data set to set the position of the initial gray wolf population solution vector set so that the initial gray wolf population solution vector set is consistent with the predicted target pressure distribution map. In each round of iteration of the gray wolf optimization algorithm pressure regulation model, the positions of the remaining gray wolf solution vectors are updated according to the current optimal gray wolf solution vector and the predicted target pressure distribution map to obtain the gray wolf population solution vector set after iterative update, and determine whether the gray wolf population solution vector set after iterative update meets the convergence conditions of the multi-objective optimization framework. If so, the current optimal gray wolf solution vector is written into the optimal pressure distribution parameter set, otherwise the iteration continues;

[0073] S4. Each section of the prosthesis driving multi-segment pressure control execution module pressure regulating device to the optimal pressure distribution parameter set as the set value for synchronous pressure adjustment;

[0074] S5. Determine whether the prosthetic interface pressure data set after synchronous pressure adjustment falls within the medical safety threshold range and the comfort assessment score reaches the preset threshold. If so, execute the maintenance control mode; otherwise, return to step S2 and re-execute real-time closed-loop adjustment.

[0075] In this embodiment, S1 includes the following steps:

[0076] S11. Collect the original prosthetic interface pressure data stream collected by the multi-channel flexible pressure sensor array deployed between the prosthetic limb and the human body, and construct an original prosthetic interface pressure data set. Each data in the original prosthetic interface pressure data set includes the acquisition timestamp t of the pressure data. i , instantaneous pressure value vector p i , and the corresponding flexible pressure sensor number s i ,The instantaneous pressure value vector represents the interface pressure value collected by each flexible pressure sensor at the same moment, and the flexible pressure sensor number indicates which flexible pressure sensor channel each data comes from;

[0077] S12. Segment the raw prosthetic interface pressure data set in chronological order to construct a time-windowed prosthetic interface pressure data set. Each time window contains all raw prosthetic interface pressure data collected within a set time interval. All raw prosthetic interface pressure data with a timestamp falling within the time interval are included in the time window.

[0078] S13. Perform noise suppression on the prosthetic interface pressure data in each time window. Smooth the pressure values within each flexible pressure sensor channel using a sliding average filter to obtain a noise-suppressed time window data set. During the smoothing process, the local average value of the pressure value at each sampling point is calculated in the time domain to replace the original pressure sampling value. The noise-suppressed pressure value is used to suppress data noise caused by high-frequency fluctuations.

[0079] S14. Perform missing completion operations on the noise-suppressed time window data set, so that each flexible pressure sensor channel has valid data in each time window. In the case of missing data, use the interpolation algorithm to estimate and complete the pressure value of the missing point, and perform normalization processing to obtain the standardized prosthetic interface pressure data set D norm During the normalization process, the completed pressure value of each flexible pressure sensor in each time window is converted into a standardized value. The standardized value is calculated by subtracting the minimum pressure value historically collected by the flexible pressure sensor from the current pressure value and dividing it by the difference between the maximum and minimum values historically collected by the flexible pressure sensor. The result is limited to between 0 and 1.

[0080] In this embodiment, S2 includes the following steps:

[0081] S21. Standardize the prosthetic interface pressure data set D norm The improved MLP-Mixer network pressure feature model is input, and the normalized prosthetic interface pressure data in each time window is represented as the pressure state tensor X k ;

[0082] S22. According to the improved MLP-Mixer network pressure feature model, the pressure state tensor X k Perform local-global dual-scale spatial mixing processing. The local-scale spatial mixing operation uses an adaptive attention mechanism to dynamically calculate the pressure correlation weight of each flexible pressure sensor and its adjacent flexible pressure sensors to form a local spatial weighted matrix. And the pressure state tensor X k Perform local spatial blending to obtain the local scale-space blending tensor:

[0083]

[0084] Local spatial weighting matrix Elements in represents the relationship between flexible pressure sensor channel j and adjacent flexible pressure sensor channel j ′ The dynamic pressure correlation between the two in the current window pressure state enables the improved MLP-Mixer network pressure feature model to capture the local pressure abnormality area under the prosthesis wearing state;

[0085] S23. At the same time, perform global spatial mixing operations at the global scale, introduce a multi-head adaptive channel attention module to calculate the contribution weight of each flexible pressure sensor channel to the overall pressure balance in the current wearing scenario, and form a global spatial weighted matrix And obtain the global scale-space mixing tensor:

[0086]

[0087] Global spatial weighting matrix Chinese elements represents the relationship between flexible pressure sensor channel j and adjacent flexible pressure sensor channel j ′ The contribution weights of the two to the global pressure balance of the prosthetic interface enable the improved MLP-Mixer network pressure feature model to identify and optimize the overall pressure equilibrium state of the prosthetic interface;

[0088] S24. Mix the local scale space tensor Mixing tensors with global scale space Perform fusion operations to form a local-global fusion tensor:

[0089]

[0090] Among them, γ is the fusion coefficient, which is dynamically calculated according to the complexity of the current prosthetic interface pressure distribution. When the pressure distribution is relatively abnormal and complex, more attention is paid to the local pressure abnormal area, and when the pressure distribution is relatively stable, more attention is paid to the global pressure equilibrium state.

[0091] S25. Local-global fusion tensor Perform dynamic gated channel mixing processing, and introduce the gated recurrent unit GRU to dynamically mine and learn the pressure change trend of the flexible pressure sensor channel between different time steps, and obtain the pressure feature representation tensor F after channel mixing. k :

[0092]

[0093] The gated recurrent unit GRU combines the hidden state vector h of the previous time window k-1 , dynamically update the hidden state vector h of the current time window k ,capturing the temporal evolution trend and subtle changes of pressure during prosthesis wearing, enabling the improved MLP-Mixer network pressure feature model to predict upcoming pressure anomalies or imbalance trends;

[0094] S26. Representing tensor F based on pressure characteristics k Calculate the predicted target pressure distribution map for the current time window and overall comfort evaluation score S k , the predicted target pressure value of each flexible pressure sensor channel in the predicted target pressure distribution diagram Based on the prediction of the stability and comfort of prosthesis wearing in the short term in the future, the overall comfort assessment score is generated by pressure feature representation tensor mapping, which is used to quantitatively characterize the real-time impact of the current pressure distribution on the overall comfort of the prosthesis wearer.

[0095] In this embodiment, the predicted target pressure distribution map is calculated as the pressure feature tensor F k Perform weighted averaging in the time dimension to obtain the time-sensitive pressure response mean vector μ of each flexible pressure sensor channel in the current window k =(μ k,1 ,μ k,2 ,...,μ k,M ),in:

[0096]

[0097] Among them, f t,j represents the depth characteristic value of the flexible pressure sensor channel j at time step t, ω t is a dynamic weight that decreases with time, which is used to highlight the dominant influence of pressure mutation on regulation in the short term, and T is the total time step;

[0098] The pressure response mean vector μ k The predicted target pressure distribution map is generated through the learnable nonlinear mapping function φ(·)

[0099]

[0100] Among them, W p is the trainable weight matrix, b p is the bias vector, σ(·) is the Sigmoid activation function, which is used to compress the output to the [0,1] interval so that each predicted target pressure value Indicates the pressure intensity that should be applied by the j-th flexible pressure sensor target in the current window.

[0101] In this embodiment, the overall comfort evaluation score is calculated as the pressure feature tensor F k Calculate its time-channel joint fluctuation tensor V k , where each element v t,j =|f t,j -μ k,j |, represents the instantaneous pressure deviation degree of each channel at each time step;

[0102] The time-channel joint fluctuation tensor V k Summarized into a single comfort impact index η k , and then mapped to the overall comfort evaluation score S through the nonlinear mapping function ψ(·) k :

[0103]

[0104] S k =ψ(η k )=exp(-α·η k );

[0105] Where M represents the total number of flexible pressure sensor channels, α>0 is the empirical adjustment coefficient, which represents the suppression strength of pressure fluctuation on comfort, and the overall comfort evaluation score S k Values closer to 1 indicate more stable pressure and a more comfortable fit, while values closer to 0 indicate severe imbalance or discomfort.

[0106] In this embodiment, S3 includes the following steps:

[0107] S31. According to the predicted target pressure distribution map Overall comfort evaluation score S k And the preset medical safety threshold pair (L, U), where L represents the lowest allowable pressure value of each flexible pressure sensor channel and U represents the highest allowable pressure value of each channel, construct a multi-objective optimization function F(Z) that integrates the predicted target pressure value, threshold crossing degree and comfort information:

[0108]

[0109] Among them, w1, w2, w3 are non-negative weight coefficients, l j Indicates the medical safety lower limit pressure value of channel j, u j represents the medical safety upper pressure limit value of channel j, z j represents the prosthesis fixation pressure value that should be applied by each flexible pressure sensor channel in channel j;

[0110] S32. Based on the historical prosthetic interface pressure data set D hist , respectively count the historical average pressure value and historical fluctuation variance of each flexible pressure sensor channel, construct an individualized initialization strategy, add Gaussian perturbations generated based on historical fluctuation variance to each channel with the predicted target pressure distribution map as the center, generate several initial gray wolf solution vectors, and form the initial gray wolf solution vectors into an initial gray wolf population solution vector set;

[0111] S33. Calculate the fitness function value of each initial gray wolf solution vector under the multi-objective optimization function F(Z), and sort them from small to large according to the fitness function value. Select the three gray wolves with the smallest fitness value as guide individuals, named α, β, and δ respectively, and their position vectors are

[0112] S34. During each iteration, the gray wolf optimization algorithm uses the positions of the three guide individuals α, β, and δ to guide the search direction of the remaining individuals. The position difference between each gray wolf and the three guide individuals is calculated, and the update amplitude is controlled by a random coefficient. The new position vector of each gray wolf after the update is the weighted average of its position with the positions of α, β, and δ.

[0113] In this process, a linearly decreasing convergence factor a is used. t , which is used to balance the exploration ability in the early stage of the search and the convergence ability in the later stage; at the same time, a random vector r generated by uniform distribution is used i,t To enhance the diversity of search, so that the gray wolf population can better escape the local optimum during the update;

[0114] S35. After completing the update of the gray wolf population solution vector in each round, recalculate the fitness function values of all gray wolf individuals under the multi-objective optimization function F(Z). Based on the updated fitness function values, reselect the three leading gray wolf individuals for the new round, i.e., the first three individuals with the smallest fitness values. At the same time, record the fitness function value of the best gray wolf in the current round.

[0115] S36. At the end of each iteration, determine whether any of the following conditions are met: first, whether the fitness function value of the current optimal gray wolf is less than or equal to a preset threshold; second, whether the current iteration round has reached the maximum number of iterations. If any of these conditions are met, the iteration is terminated; otherwise, the next iteration is entered.

[0116] S37. When the iteration of the gray wolf optimization algorithm ends, the position vector of the gray wolf individual with the minimum current fitness function value is The optimal pressure distribution parameter set as the final output The optimal pressure distribution parameter set represents the most suitable prosthetic interface pressure regulation target within the current time window, combined with comfort, prediction accuracy and medical threshold constraints.

[0117] In this embodiment, S4 includes the following steps:

[0118] S41. Obtaining the optimal pressure distribution parameter set in, Indicates the optimal pressure value corresponding to the flexible pressure sensor channel j in the current time window, which serves as the pressure control setting target;

[0119] S42. Set the optimal pressure distribution parameters The target control pressure value set Q mapped to each independent control segment in the prosthesis multi-segment pressure control execution module k =(q k,1 ,q k,2 ,...,q k,N ), where N represents the actual number of segments of the prosthetic control execution module, q k,n Indicates the pressure setting value that should be applied to the nth pressure regulation section, according to the mapping relationship matrix M between the flexible pressure sensor channel and the physical control section map Determine as follows:

[0120]

[0121] Among them, m n,j Indicates whether sensor channel j belongs to control section n, which is used to achieve sensor-to-control unit aggregation;

[0122] S43. Control pressure value set Q according to target k The pressure of all the segment pressure regulating devices under the prosthesis multi-segment pressure control execution module is adjusted synchronously. The controller sets each pressure setting value q k,n As the setting target of the nth segment, the contact interface is dynamically loaded or unloaded by adjusting its corresponding driving unit to form a physical pressure output state consistent with the optimal pressure distribution parameter set;

[0123] S44. During the pressure adjustment process of each section, the execution module provides real-time feedback on the actual pressure value of the current section. and the pressure setting value q k,n Perform difference comparison, if there is a value that exceeds the allowable error threshold ∈ p deviation, that is The fine-tuning closed-loop control mechanism is immediately triggered to perform pressure adjustment and compensation operations on the section until the pressure in all sections is stabilized within the set range;

[0124] S45. After completing the synchronous pressure adjustment, the current actual pressure state is recorded as the adjusted prosthetic interface pressure data set in, is the jth flexible pressure sensor in time window t k The actual pressure value detected within.

[0125] Example 1:

[0126] During routine outpatient training, the Prosthetic Assistance Experimental Center at Rehabilitation Hospital A saw a 37-year-old male prosthesis wearer, surnamed Chen, who had suffered a three-year right mid-calf amputation and was fitted with a carbon fiber modular below-knee prosthesis. The patient reported experiencing significantly increased pressure on the medial tibial bone of the prosthetic interface during his morning commute, with repeated bouts of skin erythema and even mild ulceration. Traditional manual adjustment of the silicone liner proved inconsistent. To address the issues of prosthetic comfort and stability, the center decided to conduct a full-process adjustment experiment using the fusion method described in this invention.

[0127] The patient began pressure regulation training. A total of 64 flexible pressure sensors are embedded in the prosthesis, which collect data in real time at a frequency of 100Hz. The system detected that the pressure value of channel 12 (located at the front and lower part of the calf) has exceeded 72.4kPa for 7 consecutive seconds and triggered a local high-pressure alarm. The MLP-Mixer network receives standardized time window data, and the extracted local depth features show that the fluctuation frequency of channel 12 in the past 3 seconds is 2.7 times per second, and the average pressure rise rate is +6.1kPa per second. The GRU channel trend analysis module determines that the current pressure fluctuation is not short-term and sporadic, and predicts that the pressure peak will rise to 78.6kPa in the next 2 seconds, and the comfort index will drop to 0.33 (lower than the preset threshold of 0.6).

[0128] The Gray Wolf optimization algorithm initiated an iterative adjustment process. The predicted target pressure distribution map, as input, indicated that the pressure in the control section corresponding to channel 12 needed to be reduced by 12.7%. Channels 9 and 13, due to their force-related interactions, needed to be adjusted by +4.3% and +5.1%, respectively. In the first iteration, the Gray Wolf population generated 30 candidate pressure distribution solution vectors, using historical training mean data to guide the algorithm's rapid approach to the target distribution. By the seventh iteration, the system converged to the optimal solution, achieving a mean squared error of only 0.019 between the target solution and the predicted target pressure distribution map. The algorithm completed the process in 1.84 seconds.

[0129] The prosthetic's multi-segment pneumatic pressure actuator module began synchronized adjustments. The pressure in actuator zone 3 (corresponding to channel 12) was dynamically adjusted from 62.1 kPa to 54.2 kPa. The remaining segments were then adjusted in tandem according to the mapping matrix. The system's internal feedback loop indicated that all segments had achieved dynamic steady-state lock, with an error of no more than ±1.2 kPa.

[0130] The system updates the actual pressure data of the current time window. The pressure monitoring graph shows that the pressure curve of channel 12 tends to be stable after self-adjustment, and the maximum value drops to 56.4kPa. No abnormal mutation occurs in a short period of time. The patient subjectively reports that the pressure feeling at the anterior edge of the tibia is significantly reduced, and the gait test is repeated.

[0131] In subsequent scenario tests (including 10 sets of stairs, 12 standing switches, and 4 minutes of brisk walking), the system triggered 5 local pressure abnormality warnings, 3 of which were active prediction adjustments and 2 were adjustment feedback compensation. All processing times did not exceed 2.5 seconds. Compared with traditional adjustment methods, the following performance data was obtained:

[0132] Table 1 Comparative test of the present invention and the traditional adjustment method

[0133]

[0134]

[0135] In addition, the system of the present invention automatically generates an adjustment cycle report:

[0136] Adjustment time period: 10:10:21-10:10:35, trigger type: local high pressure warning (channel 12), predicted pressure peak: 78.6kPa, optimal adjustment target: 54.2kPa (reduced by 12.7%), actual steady-state value: 56.4kPa, total adjustment time: 4.2 seconds, predicted MSE: 0.019, comfort improvement value: increased from 0.33 to 0.91.

[0137] After completing the training that day, the patient gave feedback: "This is the first time I've felt that the prosthetic system 'understands my pressure'. Even before I felt any obvious discomfort, the system had already adjusted itself."

[0138] In summary, the process of Example 1 truly demonstrates the method of the present invention's ability to quickly predict and respond to dynamic pressure fluctuations, optimize and adjust capabilities, and execute a closed-loop control effect in actual wearing scenarios. It demonstrates real-time performance, comfort, and safety that are superior to traditional solutions in multiple motion state switching and multi-segment pressure linkage adjustment, and has significant value in clinical promotion and engineering implementation.

[0139] The present invention introduces a local-global dual-scale attention mechanism into the MLP-Mixer network structure. The local space mixing module can accurately capture the local interaction relationship between flexible pressure sensors, and is used to discover local high-pressure areas or mutation boundaries. The global space mixing module extracts the pressure coordination characteristics between all channels and establishes the perception capability of overall pressure balance. At the same time, a gated recurrent unit is introduced for time series modeling to dynamically capture the evolution law of pressure distribution with changes in wearing status, and integrate it into a unified deep feature representation, which effectively solves the defect that traditional neural networks cannot take into account both local mutation response and global trend modeling at the same time, and has higher accuracy and response sensitivity in predicting the target pressure distribution of the prosthetic interface in the short term in the future.

[0140] The present invention designs a multi-objective optimization structure that is highly adaptable to the prosthetic pressure regulation task, consisting of a joint optimization objective function including a predicted pressure deviation term, a comfort penalty term, and a medical safety threshold penalty term. At the same time, a personalized Gaussian perturbation initialization strategy generated based on historical individual data is proposed, so that the gray wolf population can be efficiently initialized around the predicted target distribution, shortening the convergence time. The iterative update mechanism in the gray wolf algorithm is combined with a dynamic convergence factor and a collaborative search strategy of three guiding individuals, which improves the convergence stability to the physiological reasonable boundary while meeting the global optimal search capability, and is significantly better than the regulation performance of traditional genetic algorithms or particle swarm algorithms in high-dimensional, multi-constrained search spaces.

[0141] The present invention constructs a mapping matrix between the flexible sensor channel and the physical pressure regulation section, calculates the pressure target of each control unit based on the target pressure distribution parameters, and monitors the actual pressure value in real time through a parallel executed pressure control feedback network. If there is an error in the pressure of a local section that deviates from the optimal set value, the system can immediately enter the fine-tuning compensation state to ensure that the pressure distribution of all pressure sections always converges to the optimal solution during dynamic movement, significantly improving the stability and physiological matching of the regulation, and effectively overcoming the problems of response lag and uneven regulation of traditional centralized regulation.

[0142] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for optimizing prosthetic fixation pressure based on a neural network, characterized in that: The steps include: S1. Continuously collect prosthetic interface pressure data sets, combine and package them into time-windowed prosthetic interface pressure data sets, perform preprocessing on them, and output standardized prosthetic interface pressure data sets; S2. Input the standardized prosthetic interface pressure data set into the improved MLP-Mixer network pressure feature model to generate a predicted target pressure distribution map and comfort assessment score; S3. Construct a pressure regulation model for the gray wolf optimization algorithm. Use the multi-objective optimization framework formed by the predicted target pressure distribution map, comfort assessment score, and preset medical safety threshold as the search space to generate an initial gray wolf population solution vector set. In each iteration of the gray wolf optimization algorithm pressure regulation model, update the positions of the remaining gray wolf solution vectors based on the current optimal gray wolf solution vector and the predicted target pressure distribution map to obtain the iteratively updated gray wolf population solution vector set. Determine whether the iteratively updated gray wolf population solution vector set meets the convergence conditions of the multi-objective optimization framework. If so, write the current optimal gray wolf solution vector into the optimal pressure distribution parameter set. S4. Each section of the prosthesis driving multi-segment pressure control execution module pressure regulating device to the optimal pressure distribution parameter set as the set value for synchronous pressure adjustment; S5. Determine whether the prosthetic interface pressure data set after synchronous pressure adjustment falls within the medical safety threshold range and the comfort assessment score reaches the preset threshold. If so, execute the maintenance control mode; otherwise, return to step S2 and re-execute real-time closed-loop adjustment.

2. The neural network-based prosthesis fixation pressure optimization method according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Collect the original prosthetic interface pressure data stream collected by the multi-channel flexible pressure sensor array deployed between the prosthetic limb and the human body, and construct an original prosthetic interface pressure data set. Each data in the original prosthetic interface pressure data set includes the acquisition timestamp t of the pressure data. i , instantaneous pressure value vector p i , and the corresponding flexible pressure sensor number s i ,The instantaneous pressure value vector represents the interface pressure value collected by each flexible pressure sensor at the same moment, and the flexible pressure sensor number indicates which flexible pressure sensor channel each data comes from; S12. Segment the raw prosthetic interface pressure data set in chronological order to construct a time-windowed prosthetic interface pressure data set. Each time window contains all raw prosthetic interface pressure data collected within a set time interval. All raw prosthetic interface pressure data with a timestamp falling within the time interval are included in the time window. S13. Perform noise suppression on the prosthetic interface pressure data in each time window, smoothing the pressure values within each flexible pressure sensor channel using a sliding average filter to obtain a noise-suppressed time window data set. During the smoothing process, the local average value of the pressure value at each sampling point in the time domain is calculated to replace the original pressure sampling value. S14. Perform missing completion operations on the noise-suppressed time window data set, so that each flexible pressure sensor channel has valid data in each time window. In the case of missing data, use the interpolation algorithm to estimate and complete the pressure value of the missing point, and perform normalization processing to obtain the standardized prosthetic interface pressure data set D norm .

3. The neural network-based prosthesis fixation pressure optimization method according to claim 2, characterized in that: The S2 comprises the following steps: S21. Standardize the prosthetic interface pressure data set D norm The improved MLP-Mixer network pressure feature model is input, and the normalized prosthetic interface pressure data in each time window is represented as the pressure state tensor X k S22. According to the improved MLP-Mixer network pressure feature model, the pressure state tensor X k Perform local-global dual-scale spatial mixing processing. The local-scale spatial mixing operation uses an adaptive attention mechanism to dynamically calculate the pressure correlation weight of each flexible pressure sensor and its adjacent flexible pressure sensors to form a local spatial weighted matrix. And the pressure state tensor X k Perform local spatial mixing to obtain a local scale-space mixing tensor S23. At the same time, perform global spatial mixing operations at the global scale, introduce a multi-head adaptive channel attention module to calculate the contribution weight of each flexible pressure sensor channel to the overall pressure balance in the current wearing scenario, and form a global spatial weighted matrix And obtain the global scale-space mixing tensor S24. Mix the local scale space tensor Mixing tensors with global scale space Perform fusion operations to form a local-global fusion tensor: Among them, γ is the fusion coefficient, which is dynamically calculated according to the complexity of the current prosthetic interface pressure distribution; S25. Local-global fusion tensor Perform dynamic gated channel mixing processing, and introduce the gated recurrent unit GRU to dynamically mine and learn the pressure change trend of the flexible pressure sensor channel between different time steps, and obtain the pressure feature representation tensor F after channel mixing. k , the gated recurrent unit GRU combines the hidden state vector h of the previous time window k-1 , dynamically update the hidden state vector h of the current time window k ; S26. Representing tensor F based on pressure characteristics k Calculate the predicted target pressure distribution map for the current time window and overall comfort evaluation score S k .

4. The neural network-based prosthesis fixation pressure optimization method according to claim 3, characterized in that: The predicted target pressure distribution map is calculated as the pressure feature representation tensor F k Perform weighted averaging in the time dimension to obtain the time-sensitive pressure response mean vector μ of each flexible pressure sensor channel in the current window k =(μ k,1 ,μ k,2 ,...,μ k,M ),in: Among them, f t,j represents the depth characteristic value of the flexible pressure sensor channel j at time step t, ω t is a dynamic weight that decreases with time, which is used to highlight the dominant influence of pressure mutation on regulation in the short term, and T is the total time step; The pressure response mean vector μ k The predicted target pressure distribution map is generated through the learnable nonlinear mapping function φ(·) Among them, W p is the trainable weight matrix, b p is the bias vector, σ(·) is the Sigmoid activation function, which is used to compress the output to the [0,1] interval so that each predicted target pressure value Indicates the pressure intensity that should be applied by the j-th flexible pressure sensor target in the current window.

5. The neural network-based prosthesis fixation pressure optimization method according to claim 3, characterized in that: The overall comfort evaluation score is calculated as the pressure feature tensor F k Calculate its time-channel joint fluctuation tensor V k , where each element v t,j =|f t,j -μ k,j |, represents the instantaneous pressure deviation degree of each channel at each time step; The time-channel joint fluctuation tensor V k Summarized into a single comfort impact index η k , and then mapped to the overall comfort evaluation score S through the nonlinear mapping function ψ(·) k : S k =ψ(η k )=exp(-a·h k ); Where M represents the total number of flexible pressure sensor channels, α>0 is the empirical adjustment coefficient, which represents the suppression strength of pressure fluctuation on comfort, and the overall comfort evaluation score S k Values closer to 1 indicate more stable pressure and a more comfortable fit, while values closer to 0 indicate severe imbalance or discomfort.

6. The neural network-based prosthetic fixation pressure optimization method according to claim 3, characterized in that: The S3 includes the following steps: S31. According to the predicted target pressure distribution map P k (target) , Overall comfort evaluation score S k And the preset medical safety threshold pair (L, U), where L represents the lowest allowable pressure value of each flexible pressure sensor channel and U represents the highest allowable pressure value of each channel, construct a multi-objective optimization function F(Z): Among them, w1, w2, w3 are non-negative weight coefficients, l j Indicates the medical safety lower limit pressure value of channel j, u j represents the medical safety upper pressure limit value of channel j, z j represents the prosthesis fixation pressure value that should be applied by each flexible pressure sensor channel in channel j; S32. Based on the historical prosthetic interface pressure data set D hist , respectively count the historical average pressure value and historical fluctuation variance of each flexible pressure sensor channel, construct an individualized initialization strategy, add Gaussian perturbations generated based on historical fluctuation variance to each channel with the predicted target pressure distribution map as the center, generate several initial gray wolf solution vectors, and form the initial gray wolf solution vectors into an initial gray wolf population solution vector set; S33. Calculate the fitness function value of each initial gray wolf solution vector under the multi-objective optimization function F(Z), and sort them from small to large according to the fitness function value. Select the three gray wolves with the smallest fitness value as guide individuals, named α, β, and δ respectively, and their position vectors are S34. During each iteration, the gray wolf optimization algorithm uses the positions of the three guide individuals α, β, and δ to guide the search direction of the remaining individuals. The position difference between each gray wolf and the three guide individuals is calculated, and the update amplitude is controlled by a random coefficient. The new position vector of each gray wolf after the update is the weighted average of its position with the positions of α, β, and δ. S35. After completing the update of the gray wolf population solution vector in each round, recalculate the fitness function values of all gray wolf individuals under the multi-objective optimization function F(Z). Based on the updated fitness function values, reselect the three leading gray wolf individuals for the new round, i.e., the first three individuals with the smallest fitness values. At the same time, record the fitness function value of the best gray wolf in the current round. S36. At the end of each iteration, determine whether any of the following conditions are met: first, whether the fitness function value of the current optimal gray wolf is less than or equal to a preset threshold; second, whether the current iteration round has reached the maximum number of iterations. If any of these conditions are met, the iteration is terminated; otherwise, the next iteration is entered. S37. When the iteration of the gray wolf optimization algorithm ends, the position vector of the gray wolf individual with the minimum current fitness function value is The optimal pressure distribution parameter set as the final output 7. The neural network-based prosthetic fixation pressure optimization method according to claim 6, characterized in that: The S4 comprises the following steps: S41. Obtaining the optimal pressure distribution parameter set in, Indicates the optimal pressure value corresponding to the flexible pressure sensor channel j in the current time window, which serves as the pressure control setting target; S42. Set the optimal pressure distribution parameters The target control pressure value set Q mapped to each independent control segment in the prosthesis multi-segment pressure control execution module k =(q k,1 ,q k,2 ,...,q k,N ), where N represents the actual number of segments of the prosthetic control execution module, q k,n Indicates the pressure setting value that should be applied to the nth pressure regulation section, according to the mapping relationship matrix M between the flexible pressure sensor channel and the physical control section map Determine as follows: Among them, m n,j Indicates whether sensor channel j belongs to control section n, which is used to achieve sensor-to-control unit aggregation; S43. Control pressure value set Q according to target k The pressure of all the segment pressure regulating devices under the prosthesis multi-segment pressure control execution module is adjusted synchronously. The controller sets each pressure setting value q k,n As the setting target of the nth segment, the contact interface is dynamically loaded or unloaded by adjusting its corresponding driving unit to form a physical pressure output state consistent with the optimal pressure distribution parameter set; S44. During the pressure adjustment process of each section, the execution module provides real-time feedback on the actual pressure value of the current section. and the pressure setting value q k,n Perform difference comparison, if there is a value that exceeds the allowable error threshold ∈ p deviation, that is The fine-tuning closed-loop control mechanism is immediately triggered to perform pressure adjustment and compensation operations on the section until the pressure in all sections is stabilized within the set range; S45. After completing the synchronous pressure adjustment, the current actual pressure state is recorded as the adjusted prosthetic interface pressure data set in, is the jth flexible pressure sensor in time window t k The actual pressure value detected within.

Images