Intelligent human motion intention recognition method for lower limb prosthesis based on fried meat optimization algorithm

Abstract

The application is suitable for the technical field of motion intention recognition, and provides an intelligent lower limb prosthesis human motion intention recognition method based on a stir-fried meat optimization algorithm. The application proposes a feature weight dynamic optimization framework based on the stir-fried meat optimization algorithm, realizes real-time adaptive distribution of motion feature weights through two links of offline weight vector training and real-time feature optimization execution. The algorithm achieves optimal results on 7 standard unimodal test functions and has good migration; through the above-knee amputation patient data set verification, the optimized weight matrix has the lowest number of recognition errors and the fastest optimization speed, significantly improves the classifier prediction accuracy and the environmental anti-interference ability of intention recognition, and meets the real-time requirement. The application solves the adaptability problem of intelligent lower limb prosthesis in complex life scenes, lays a foundation for high-reliability product development, has important practical application value in the motion scene recognition of lower limb prosthesis and walking aid robots, and has significant industrialization transformation potential.

Description

Technical Field

[0001] This invention belongs to the field of motion intention recognition technology, and particularly relates to a method for recognizing human motion intention in intelligent lower limb prostheses based on a stir-fry optimization algorithm. Background Technology

[0002] In today's complex and ever-changing living environment, intelligent lower limb prostheses need to accurately analyze the user's movement intentions in real time and rapidly adjust the knee joint's biomechanical properties (such as stiffness / damping). However, existing technologies have insufficient adaptability in this regard, directly leading to a higher risk of falls for amputees using prostheses during daily activities, severely limiting their mobility and autonomy. Therefore, achieving highly robust perception and response of prostheses to the environment and user intentions is of significant practical importance and clinical application value for ensuring amputee safety, improving their confidence in mobility, and enhancing their quality of life.

[0003] Currently, intelligent lower limb prostheses employ a hierarchical control strategy, with the human motion intention recognition system acting as the high-level controller. This system is responsible for perceiving the amputee's motion intentions based on signals from the user, the environment, and the prosthesis. Ideally, the output of the high-level controller should allow the robot to autonomously switch between different motion modes without requiring any conscious input from the user. However, current intelligent lower limb prostheses face core challenges in real-life scenarios: insufficient accuracy in intention recognition during motion mode transitions, coupled with time-consuming prediction algorithms leading to poor real-time system response and discontinuous motion output. These technical bottlenecks collectively restrict the control accuracy and naturalness of prosthetic movements, significantly impacting the patient's daily user experience and functional effectiveness, representing critical challenges that urgently need to be overcome.

[0004] In real-life scenarios, real-time intent recognition technology for intelligent lower limb prostheses faces systemic challenges. The core problem stems from the lack of feature optimization: different sensor features (such as the vertical component of ground reaction force and the rate of change of joint angle) contribute significantly to motion classification, but existing algorithms lack an effective dynamic feature weighting mechanism. This leads to key features being overwhelmed by noise or secondary features being over-reliant on, directly weakening the foundation of recognition accuracy. This deficiency, combined with environmental signal noise (such as road vibration affecting inertial measurements), forces the system to increase computational complexity to compensate for accuracy. However, high-complexity processing inevitably introduces real-time latency (50-200 milliseconds) in multi-source heterogeneous data streams, causing delays in the response of critical commands such as emergency stops and steering. The coupling effect of noise and latency further exposes the bottleneck of insufficient motion pattern generalization—when users face scenarios not covered in the laboratory (such as continuously crossing two steps or overcoming obstacles on a slippery surface), the statically trained model cannot adaptively adjust the decision boundary, leading to a surge in the probability of misjudgment. The aforementioned triple deficiency creates a cascading risk: missing feature optimization reduces noise tolerance, noise amplifies delay, delay limits generalization ability, and generalization failure directly increases the false recognition rate. The resulting safety consequences are particularly serious; for example, switching from "slow down" to "stair climbing mode" in a crowded subway station can cause collisions. Clearly, the lack of a dynamic feature weighting mechanism is the root cause of the current system's performance limitations. It amplifies environmental noise interference, induces a chain reaction of delay and generalization failure, ultimately threatening user safety and becoming a key barrier to technology implementation.

[0005] To address the core deficiency of existing human motion intention recognition systems in real-life scenarios—the lack of dynamic feature weight optimization mechanisms—this invention proposes an intelligent lower limb prosthesis human motion intention recognition method based on a stir-fry optimization algorithm. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent lower limb prosthesis human movement intention recognition method based on the stir-frying optimization algorithm, which aims to solve the problems mentioned in the background art.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for recognizing human movement intent in intelligent lower limb prostheses based on a stir-fry optimization algorithm includes the following steps:

[0009] Offline weight vector training: First, construct and preprocess a multi-motion mode dataset to form an equal sample size dataset and five unbalanced datasets constructed according to prior probabilities. Then, design and train a stir-fry optimization algorithm to simulate the Chinese stir-fry process and adjust the feature weights in the machine learning model to obtain weight vectors corresponding to different motion scenarios.

[0010] Real-time feature optimization execution: Collect sensor signals from the intelligent lower limb prosthesis, extract feature values ​​to form a feature vector; load the weight vector obtained from the offline weight vector training, and optimize the weight of the feature vector; input the optimized feature vector into the intention recognition classifier to predict the movement pattern; depending on whether the movement pattern is switched, adaptively select semi-regular weight vector optimization or use the stir-fry optimization algorithm to retrain and optimize the weight vector.

[0011] Furthermore, the stir-frying optimization algorithm includes the following process:

[0012] Construct a temperature distribution model for a wok;

[0013] The basic heat regulation process of meat slices based on the hot field and cooling effect of the wok;

[0014] Based on the heat transfer process between meat slices guided by elite meat slices;

[0015] The process of flipping meat slices, which involves repositioning the meat slices in the search space through a flipping operation;

[0016] Elite update process used to update Alpha, Beta, and Delta elites and monitor the algorithm's convergence status.

[0017] Furthermore, the temperature distribution of the wok temperature distribution model is calculated using the following formula:

[0018] ;

[0019] in, It refers to the temperature at the location where the meat slice is located; It is the base temperature; It is the Euclidean distance from the meat slice to the origin; It is a dimension.

[0020] Furthermore, during the basic calorie adjustment process of the meat slices, the initial heat renewal of the meat slices... The calculation formula is as follows:

[0021] ;

[0022] in, It's sliced ​​meat. Calories; It is a pot heating item; It's a cooling item.

[0023] Furthermore, during the heat transfer process between the meat slices, the meat slices The final heat obtained The calculation formula is as follows:

[0024] ;

[0025] in, It's sliced ​​meat. Initial heat replenishment; It is the total heat transfer term.

[0026] Furthermore, the process of turning the meat slices includes:

[0027] The flipping action is triggered based on the probability of the flipping selection;

[0028] Based on the relationship between the fitness of meat slices and the median fitness of all meat slices, meat slices are divided into two categories: poor meat slices and good meat slices. The position of the poor meat slices is updated according to the direction of the historical elite combination or the current elite combination. The position of the good meat slices is updated through elite guidance or Levi's flight strategy.

[0029] Furthermore, the elite update process includes:

[0030] The top 3 elite candidates are selected by fitness ranking, corresponding to Alpha, Beta and Delta elites respectively;

[0031] A stall counter is used to monitor the convergence status of the algorithm. The counter increases when no better solution is found for a continuous period of time, and is reset when a new Alpha elite is found.

[0032] Furthermore, during the real-time feature optimization process, when the motion mode has not switched and the motion has not terminated due to recognition errors, a semi-regular weight vector optimization is performed, with the following convergence formula:

[0033] ;

[0034] in: These are the newly calculated weight vector values; As a convergence factor, ; This represents the current weight vector value; The weight vector to be approached; It is a random factor; In order to be in A random number uniformly distributed within a range;

[0035] When the motion mode changes and the motion does not terminate due to recognition error, the retrained dataset is used. It is expressed as follows:

[0036] ;

[0037] in: The number of motion states; The current movement number; For the first A dataset of pre-stored motion states; A dataset consisting of the previous few steps of the current gait;

[0038] Total sample size for:

[0039] ;

[0040] in: The amount of data pre-stored for each motion state; The amount of historical motion data pre-stored before the current gait.

[0041] Furthermore, the sensor signals include IMU sensor signals, the vertical component of the ground reaction force, and the knee joint angle signal.

[0042] The intelligent lower limb prosthesis human motion intention recognition system based on the stir-frying optimization algorithm includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the intelligent lower limb prosthesis human motion intention recognition method based on the stir-frying optimization algorithm described above is implemented.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention addresses the problem of dynamic optimization of motion feature weights in human motion intent recognition for intelligent lower limb prostheses. It proposes for the first time a dynamic feature weight optimization framework based on a "stir-fried meat" optimization algorithm. This framework achieves real-time adaptive allocation of motion feature weights through two main stages: offline weight vector training and real-time feature optimization execution. The "stir-fried meat" optimization algorithm achieves optimal results on seven standard single-peak test functions, demonstrating good transferability. Furthermore, validation using a dataset of above-knee amputees shows that optimizing the weight matrix results in the lowest number of recognition errors and the fastest optimization speed, significantly improving classifier prediction accuracy and environmental interference resistance for intent recognition, while also meeting real-time requirements. This invention not only advances the understanding of human movement mechanisms (having scientific value) but also solves the adaptability challenge of intelligent lower limb prostheses in complex life scenarios, laying the foundation for the development of highly reliable products. It has significant practical application value in motion scene recognition for lower limb prostheses and assistive robots, with significant potential for industrialization. Attached Figure Description

[0045] Figure 1 A flowchart for optimizing execution for real-time features.

[0046] Figure 2The diagram illustrates the principle of the stir-frying optimization algorithm; (a) is a schematic diagram of the non-uniform thermal field of the wok; (b) is a schematic diagram of different heat states of the meat slices; (c) is a schematic diagram of heat transfer between meat slices; (d) is a schematic diagram of meat slice cooling; (e) is a schematic diagram of overlapping meat slices; and (f) is a schematic diagram of the flipping operation to redistribute the position of the meat slices.

[0047] Figure 3 The flowchart shows the algorithm for optimizing stir-fried meat.

[0048] Figure 4 The graph shows the iterative trend changes of five optimization algorithms; (a) is the test result on the sphere function; (b) is the test result on the Schweffer 2.22 function; (c) is the test result on the Schweffer 1.2 function; (d) is the test result on the Schweffer 2.21 function; (e) is the test result on the Rosenblock function; (f) is the test result on the step function; and (g) is the test result on the noisy quartic function.

[0049] Figure 5 The graph shows the changes in the optimization results of five optimization algorithms based on motion data of patients with above-knee amputations. Detailed Implementation

[0050] In order to provide a clearer understanding of the technical features, objectives and beneficial effects of the present invention, the technical solution of the present invention will now be described in detail below, but it should not be construed as limiting the scope of implementation of the present invention.

[0051] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0052] One embodiment of the present invention provides an intelligent lower limb prosthesis human movement intention recognition method based on a stir-fry optimization algorithm. The core of this method is achieved through the combined efforts of offline weight vector training and real-time feature optimization. The specific process of the real-time feature optimization system is as follows (e.g., ...). Figure 1 As shown, the weight vector obtained from offline training is required as the core support. Therefore, before executing the real-time steps, it is necessary to complete the construction and preprocessing of the multi-motion mode dataset to lay the foundation for subsequent weight training.

[0053] I. Offline weight vector training;

[0054] Offline training is the foundation for real-time optimization. It requires sequentially constructing the dataset, training the optimization algorithm, and finally obtaining multi-scene weight vectors. The specific process is as follows:

[0055] 1. Construction and preprocessing of multi-motion mode datasets;

[0056] Users wearing intelligent lower limb prostheses collected data on various movement patterns (e.g., walking on flat ground, uphill, downhill, climbing stairs, and downstairs) to construct a multi-movement pattern dataset. Research showed that movement pattern recognition was more effective after the foot left the ground; therefore, the data was processed, extracting only the data from the first 200ms after foot liftoff in each gait cycle for analysis. The data were used to construct an equal-size dataset and five imbalanced datasets constructed according to prior probabilities (i.e., high probability datasets for horizontal walking, climbing stairs, descending stairs, uphill, and downhill). In these imbalanced datasets, the amount of data for each movement pattern was configured according to the probability of its occurrence in the real environment.

[0057] 2. Design and training of an optimization algorithm for stir-frying meat;

[0058] This invention proposes a stir-frying meat optimization algorithm. By simulating the Chinese stir-frying meat process, it adaptively adjusts the feature weights in a machine learning model to achieve optimal solution search, obtaining weight vectors applicable to different scenarios. The principle of the stir-frying meat optimization algorithm is as follows: Figure 2 As shown in (a), the wok provides a non-uniform thermal field, with a high temperature at the center and a low temperature at the edges, exhibiting non-linear fluctuations to simulate the heat distribution of a real wok. The meat slices will experience different "heat levels" at different locations within the wok, affecting their exploration and processing behavior. For example... Figure 2 As shown in (b), the meat slices have different heat states, categorized as raw, medium, well-done, and charred, with each state also having different weight values. Figure 2 As shown in (c), heat transfer also occurs between the meat slices. The "heat source" for this heat transfer comes from the three best-performing meat slices in the current iteration. These three slices have the highest heat content and thus the "capital" to influence the other meat slices. Figure 2 As shown in (d), the meat slices themselves cool down with the airflow. This algorithm performs nonlinear forced cooling based on the fourth power of the meat slices' own heat, and reduces the cooling intensity as the algorithm progresses. It also includes an "anti-burning" mechanism, which uses an attenuation factor inversely proportional to the cube of the heat to prevent the meat slices from receiving excessive external heat, thus preventing them from overheating and "burning." Figure 2 As shown in (e), if overlapping meat slices are all in the same area of ​​the pot, they will be affected by the common temperature of that area. Closely overlapping slices will also experience more frequent heat exchange. If meat slices overlap in an overheated area for a long time, they will burn. The most direct way to solve the problem of overlapping meat slices is to flip them. In the algorithm, flipping redistributes the positions of the meat slices, forcing them to explore new areas. This is a key operation to prevent the population from prematurely clustering and getting trapped in local optima, such as... Figure 2 As shown in (f).

[0059] (1) The mathematical principle of the stir-frying optimization algorithm is as follows:

[0060] ① Temperature distribution model of a wok:

[0061] (1);

[0062] in, It refers to the temperature at the location where the meat slice is located; It is the base temperature; It is the Euclidean distance from the meat slice to the origin; The first term is the dimension; the second and third terms are nonlinear perturbations composed of sine and cosine functions, simulating the non-uniformity of the wok temperature.

[0063] ② Adjusting the basic calories of the meat slices:

[0064] The basic heat regulation of meat slices simulates the heating and cooling of the "pot" by a "stove fire." It adaptively adjusts the basic temperature field based on the meat slice performance and algorithm status, providing a foundation for subsequent heat transfer. It's sliced ​​meat. The initial heat of regeneration is calculated using the following formula:

[0065] (2);

[0066] Pot heating function:

[0067] (3);

[0068] Cooling:

[0069] (4);

[0070] in, It's sliced ​​meat. Calories; It is an adaptive heating factor used to control the ranking-based heating intensity during individual heat update; It's sliced ​​meat. Rank percentage; It refers to the temperature at the location where the meat slice is located; This represents the current iteration number; This represents the maximum number of iterations. It is the cooling rate; It is the charring threshold of the meat slices; The stagnation calculation parameter is calculated using the following formula:

[0071] (5);

[0072] in, This is a stall counter for the algorithm, recording the number of consecutive times no better solution is found; This is the restart threshold; the default value is 20.

[0073] ③ Heat transfer between meat slices:

[0074] Simulating heat exchange between "meat slices," elite meat slices transfer heat to each other, incorporating anti-burning and heat balance mechanisms. The elite meat slices receive additional heating to maintain their activity. The final heat obtained The calculation formula is as follows:

[0075] (6);

[0076] in, It's sliced ​​meat. Initial heat replenishment; It is the total heat transfer term.

[0077] Total heat transfer term The calculation formula is as follows:

[0078] (7);

[0079] in, It's sliced ​​meat. From Elite Meat Slices Heat transfer; This indicates the sequence number of the elite meat slice, ranging from 0 to 2; For meat slices The function that performs the judgment, if the meat slices If it is an elite meat slice, it will be heated further. The judgment formula is shown in (11).

[0080] From Elite Meat Slices The heat transfer is calculated using the following formula:

[0081] (8);

[0082] in, It's elite meat slices The weight, Corresponding to the top three elites; Thermal conductivity; It's sliced ​​meat. With elite meat slices The distance; and They are sliced ​​meat And elite meat slices Initial heat replenishment; and Meat slices And elite meat slices fitness value; It is a very small constant to prevent division by zero; and These are the anti-burning parameter and the heat balance parameter, respectively, and the calculation formulas are shown in (9) and (10) respectively.

[0083] (9);

[0084] (10);

[0085] in, This represents the average initial calories of all meat slices.

[0086] Elite Meat Slices Additional heating The calculation formula is as follows:

[0087] (11);

[0088] Constraints:

[0089] (12);

[0090] in, A gathering of elites.

[0091] The basic heat regulation of meat slices and the heat transfer between meat slices together realize a two-level heat regulation system, which can control the overall exploration-development balance through basic heat regulation, and realize a more refined search strategy through local interactions between meat slices.

[0092] ④ Meat slice turning mechanism:

[0093] The "flipping" action during cooking is simulated, causing meat slices to redistribute within the search space. Based on the relationship between the fitness of the meat slices and the median, the slices are divided into "poor" and "good" categories, each employing a different update strategy. Poor meat slices move towards elite combinations, incorporating information from historical and current elites. Better meat slices either conduct local searches near elites or explore using Levi's flight path. The flipping selection probability formula is:

[0094] (13);

[0095] in: Choose a probability function for the flipping motion; For the first A slice of meat that needs to be turned over; This represents the flip rate, with a default value of 0.4. These are uniformly distributed random numbers.

[0096] Iteration factor:

[0097] (14);

[0098] in: For the first The iteration factor of the generation; This represents the current iteration number. This represents the maximum number of iterations.

[0099] The stagnation condition is:

[0100] (15);

[0101] in: For stagnation counting, This is the restart threshold; the default value is 20.

[0102] Poor meat slice placement update strategy:

[0103] (16);

[0104] in: Indicates the first A new location for each slice of meat; For the first The location of each slice of meat; Step size; It is the direction vector; This is a disturbance term.

[0105] Step length The calculation formula is:

[0106] (17);

[0107] Direction vector The calculation formula is:

[0108] (18);

[0109] in: It is a very small constant to prevent division by zero; There are two methods for selecting the target elites: historical elite combinations (probability 0.8) and current elite combinations (probability 0.2).

[0110] A team of historical elites:

[0111] (19);

[0112] in: To randomly select up to 3 different historical elite points from the historical elite pool, For historical elite points, if there are fewer than 3 historical elites, select all available historical elites. The weight vector is randomly generated, and the weights are sampled from a symmetric Dirichlet distribution; For the selected historical elite points, the selected historical elite points are linearly combined according to their corresponding weights to obtain a new, virtual guiding point.

[0113] Current elite combination:

[0114] (20);

[0115] in: , and These are the current optimal solution, the second-best solution, and the third-best solution, respectively. , and These are the corresponding weights, sampled from an asymmetric Dirichlet distribution. .

[0116] Better meat slice placement update strategy:

[0117] The better meat slice position update strategies are Elite Guidance (probability 0.6) and Levi Flight (probability 0.4), with the following position update formulas:

[0118] Elite Guidance:

[0119] (twenty one);

[0120] in, Indicates the first A new location for each slice of meat; ; ; It is a dimension.

[0121] Levi Flight:

[0122] (twenty two);

[0123] (twenty three);

[0124] (twenty four);

[0125] in, Indicates the first A new location for each slice of meat; Step size; Iteration factor; , , Indicates the interval A random number that is uniformly distributed on the upper surface; For stability parameters. Levi's flight is conducive to a balance between exploration and development, performing local searches around the current position with small steps and making long-distance jumps with large steps to explore new areas.

[0126] ⑤ Elite Refresh Mechanism:

[0127] Candidate selection:

[0128] (25);

[0129] (26);

[0130] (27);

[0131] in, The top 3 indexes after fitness sorting; It is the fitness vector of all meat slices; It is a matrix of candidate elite positions; It is a matrix representing the positions of all the meat slices; It is the candidate elite fitness vector.

[0132] Stasis counter update:

[0133] (28);

[0134] in, It acts as a stall counter to monitor whether the algorithm gets stuck in a local optimum (i.e., convergence state monitoring) and records the number of iterations that fail to find a better solution. The adaptability of the current best candidate elites This refers to the current Alpha elites' fitness level.

[0135] Elite Pool Update:

[0136] Alpha Elite Update: When When assigning values ​​to Alpha elites, (in This is the current optimal solution. (The top candidate among current elites) It also manages historical elites and updates their weights.

[0137] When the number of historical elites If it is less than 15, then:

[0138] (29);

[0139] (30);

[0140] In other cases:

[0141] (31);

[0142] (32);

[0143] (33);

[0144] in, A gathering of historical elites; For historical elite weight vector; This is the index with the lowest weight.

[0145] Beta Elite Update, if ,but:

[0146] (34);

[0147] (35);

[0148] in, This is the current suboptimal solution; He is the second among the current elite candidates; The current Beta elite's adaptability; The fitness of the current second-ranked elite candidate.

[0149] Delta Elite update, if ,but:

[0150] (36);

[0151] (37);

[0152] in, This is the current third best solution; He is currently ranked third among the elite candidates. The current Delta elite's adaptability; The fitness of the current third-ranked elite candidate.

[0153] The algorithm's elite update mechanism ensures that elite quality remains monotonically constant, limits the number of historical elites (maximum 15), prevents unlimited memory growth, and increments the stagnation count when the current best candidate is no better than an Alpha elite. When a new Alpha elite is found, the stagnation count is reset, ensuring the algorithm can adaptively balance exploration and development, maintaining a diverse elite population to guide the search direction. The flowchart of the optimized algorithm is shown below. Figure 3 As shown.

[0154] (2) Algorithm training and execution;

[0155] Using the aforementioned six datasets (one dataset with equal sample size and five datasets with imbalance) as inputs to the stir-frying optimization algorithm, the algorithm iteratively searches for the optimal solution, ultimately obtaining weight vectors corresponding to different scenarios. Specifically, this includes weight vectors obtained by training on the dataset with equal sample size. Used for initial algorithm startup; trained on five imbalanced datasets respectively. , , , and Weight convergence optimization is used for specific motion scenarios.

[0156] (3) Probability characteristics of motion state switching;

[0157] The transition between motion states is probabilistic. By analyzing real-world motion scenarios, the probability of transitioning between motion states can be determined. For example, the probability of switching from "going upstairs" to "going downstairs" is extremely low, indicating that not all motion modes can be directly converted to each other. Taking walking on flat ground, going uphill, going downhill, going upstairs, and going downstairs as examples, the probabilities of the next motion state are shown in Table 1.

[0158] Table 1. Probability Table of Next Motion State

[0159]

[0160] As shown in Table 1, the model tends to maintain its current state rather than switch to other states when in the "going upstairs" state. If the weight vector trained on an equally sized dataset is still used in this situation, the recognition accuracy will decrease. Therefore, the model's weights should be closer to those trained on a majority of the "going upstairs" dataset. The approximation formula is shown below to improve the accuracy of intent recognition in such real-world scenarios.

[0161] (38);

[0162] in: These are the newly calculated weight vector values; For the approaching factor ( ); This represents the current weight vector value; The weight vector to be approached; This is a random factor (to control the amplitude of the disturbance). In order to be in Random numbers that are uniformly distributed within a range.

[0163] The weighted convergence method employs a linear weighted superposition of random perturbations to ensure smooth and gradual changes, while adjusting... and These two hyperparameters can precisely control the behavior of the system and introduce controllable stochastic exploration to avoid local optima.

[0164] II. Real-time feature optimization execution;

[0165] After completing offline weight vector training, the real-time feature optimization process begins. The core of this process is a dual-mode dynamic feature weight optimization mechanism (flowchart shown in Figure 1). The optimization strategy is adaptively selected based on whether the motion mode switches. The specific steps are as follows:

[0166] a. Acquire IMU sensor signals, vertical component of ground reaction force, and knee joint angle during continuous motion;

[0167] b. Extract feature values ​​such as the maximum value of motion data to form a feature vector, thereby reducing the number of data channels;

[0168] c. Load the weight vector obtained from the offline training of the stir-fry optimization algorithm, optimize the feature vector by assigning higher weights to features with high contribution, and input them into the intent recognition classifier for motion state prediction. Each time the algorithm restarts, the first step uses the weight vector obtained from the offline training of the algorithm with an equal sample size dataset. ;

[0169] d. Use an improved ELM classifier for pattern recognition to predict the movement pattern that is about to enter the gait;

[0170] e. If the movement mode is not switched and the movement is not terminated due to recognition error (the movement anomaly recognition module recognizes the movement mode recognition error), then semi-regular weight vector optimization is performed. Based on the current gait state, the weight values ​​trained on the current gait majority dataset are used as the approximation formula (38).

[0171] f. If the motion mode changes and the motion does not terminate due to recognition error, the optimized weight vector is retrained using the "stir-fried meat" optimization algorithm to ensure recognition accuracy in the new scene. The retraining dataset includes: pre-stored data for each motion state. Data points, and the current gait before Sequential motion data, dataset This is represented as shown in formula (39). The weight vector of the previous state cycle is still used for optimization before the training process ends.

[0172] (39);

[0173] in: The number of motion states; The current movement number; For the first A dataset of pre-stored motion states; A dataset consisting of the previous few steps of the current gait. Total number of samples. for:

[0174] (40);

[0175] Before the training process is complete, the weight vector from the previous state cycle is still used for optimization.

[0176] Example 1: Performance Verification;

[0177] Based on seven standard single-peak test functions (as shown in Table 2), the proposed stir-frying optimization algorithm is compared with the hippocampus optimization algorithm, the red-bellied tragopan optimization algorithm, the cloud drift optimization algorithm, and the slime mold optimization algorithm. The training results are shown in the figure below. Figure 4 As shown in Table 3, the performance of the algorithm was comprehensively evaluated by the mean fitness, standard deviation, and convergence speed. The mean fitness, standard deviation, and convergence speed of the optimal algorithm for each test function are bolded. The mean fitness mainly reflects the solution quality of the algorithm; a smaller value is better. The standard deviation mainly reflects the stability of the algorithm; a smaller value is better. The convergence speed is the ratio of the number of iterations required to reach the target fitness to the total number of iterations; a smaller value is better, ideally close to 0, indicating that the algorithm reaches a near-optimal solution in a relatively small number of iterations.

[0178] Table 2. Seven Standard Single-Peak Test Functions

[0179]

[0180] Table 3. Comparison of Experimental Results

[0181]

[0182] From Table 3 and Figure 4 It can be seen that the stir-frying algorithm performs best among all test functions. For example... Figure 4 As shown in (a), the stir-fry optimization algorithm performed best in the spherical function test. This algorithm converged quickly in the early stages of iteration, rapidly approaching the optimal solution with a convergence rate of only 1.0%, indicating it was very close to the final solution in the early stages. Furthermore, its mean fitness and standard deviation were the lowest among all compared algorithms, demonstrating excellent stability. Figure 4As shown in (b), in the Schweffer 2.22 function test, the stir-fry optimization algorithm also performed best, with a mean fitness of only 3.762 × 10⁻², and possessed fast convergence ability, approaching the optimal solution in the early stages of iteration. Furthermore, the algorithm continued to significantly improve the quality of the solution in the later stages of iteration. While the slime mold optimization algorithm performed second best, it showed a significant difference in convergence speed compared to the stir-fry optimization algorithm. Figure 4 As shown in (c), in the Schweffer 1.2 function test, the convergence speed of the stir-fry optimization algorithm is only 1.0%, which can quickly approach the optimal solution in the early iterations. Although the results fluctuate to some extent, the mean fitness performance is good, and the effect is far superior to other comparison algorithms. Figure 4 As shown in (d), in the Schweffer 2.21 function test, the stir-fry optimization algorithm performed best, quickly approaching the optimal solution in the early iterations and exhibiting good stability, with a mean fitness of only 7.644 × 10⁻⁶. -3 The standard deviation is 2.633 × 10⁻⁶. -3 Even in later stages, it can significantly improve the solution quality. For example... Figure 4 As shown in (e), in the Rosenblock function test, the average fitness of the stir-fry optimization algorithm is only 8.958, far lower than other algorithms, indicating that the algorithm's performance is very stable with minimal fluctuations. The convergence speed is only 1.0%, indicating that the algorithm approaches the final solution early in the search process. Figure 4 As shown in (f), in the step function test, the convergence rate is only 1.0%, approaching the final solution early on. The results are very stable with minimal fluctuations. Compared to other algorithms, it exhibits the best overall performance. Figure 4 As shown in Figure (g), the stir-frying optimization algorithm performs best in the noisy quartic function test. This is mainly reflected in the fact that the algorithm quickly approaches the optimal solution in the early iterations, the results are very stable with very small fluctuations, and the average fitness is only 1.190 × 10⁻⁶. -3 It can still significantly improve the solution quality in the later stages.

[0183] The results above demonstrate that, in terms of solution accuracy and stability, the "Stir-fried Meat" optimization algorithm consistently converges to an optimal solution with near-machine accuracy in multiple independent runs. The mean and standard deviation of the results are significantly smaller than those of other comparative algorithms, exhibiting superior robustness. Validation using a single-peaked test function shows that the "Stir-fried Meat" optimization algorithm possesses very strong convergence ability and optimization accuracy, quickly and accurately finding the unique optimal point of the single-peaked test function, proving the effectiveness of its core optimization mechanisms (such as iterative strategies and population diversity preservation). Furthermore, the "Stir-fried Meat" optimization algorithm can quickly find a "decent" solution, helping to quickly assess the difficulty of the problem and the approximate behavior of the objective function, and its convergence speed is significantly better than other optimization algorithms. In summary, due to its fast convergence, high accuracy, and strong stability, the "Stir-fried Meat" optimization algorithm is a highly effective and reliable solution for handling single-peaked optimization problems.

[0184] After optimizing the weight matrix using the stir-frying optimization algorithm, the optimization results were validated on a dataset of above-knee amputees. The variation of the optimization results of the five optimization algorithms based on the motion data of above-knee amputees is shown in the figure below. Figure 5 As shown, the fitness index reflects the number of recognition errors (the smaller the value, the fewer the recognition errors). From Figure 5 As can be seen, compared with the other four optimization algorithms, the stir-frying algorithm has the lowest number of recognition errors and finds the optimal solution the fastest. This indicates that the present invention can meet real-time requirements while ensuring recognition accuracy, demonstrating good overall performance.

[0185] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for recognizing human movement intent in intelligent lower limb prostheses based on a stir-fry optimization algorithm, characterized in that, Includes the following steps: Offline weight vector training: First, construct and preprocess a multi-motion mode dataset to form an equal sample size dataset and five unbalanced datasets constructed according to prior probabilities. Then, design and train a stir-fry optimization algorithm to simulate the Chinese stir-fry process and adjust the feature weights in the machine learning model to obtain weight vectors corresponding to different motion scenarios. Real-time feature optimization execution: Collect sensor signals from the intelligent lower limb prosthesis and extract feature values ​​to form a feature vector; Load the weight vector obtained from the offline weight vector training and optimize the weight of the feature vector; input the optimized feature vector into the intention recognition classifier to predict the motion pattern; depending on whether the motion pattern is switched, adaptively select semi-regular weight vector optimization or use the stir-fry optimization algorithm to retrain and optimize the weight vector. The optimized algorithm for stir-frying meat includes the following process: Construct a temperature distribution model for a wok; The basic heat regulation process of meat slices based on the hot field and cooling effect of the wok; Based on the heat transfer process between meat slices guided by elite meat slices; The process of flipping meat slices, which involves repositioning the meat slices in the search space through a flipping operation; Elite update process used to update Alpha, Beta, Delta elites and monitor the convergence status of the algorithm; The temperature distribution of the wok temperature distribution model is calculated using the following formula: ； in, It refers to the temperature at the location where the meat slice is located; It is the base temperature; It is the Euclidean distance from the meat slice to the origin; It is a dimension; During the basic calorie adjustment process of the meat slices, the initial heat renewal of the meat slices... The calculation formula is as follows: ； in, It's sliced ​​meat. Calories; It is a pot heating item; It's a cooling item; During the heat transfer process between the meat slices The final heat obtained The calculation formula is as follows: ； in, It's sliced ​​meat. Initial heat replenishment; It is the total heat transfer term; The process of turning the meat slices includes: The flipping action is triggered based on the probability of the flipping selection; Based on the relationship between the fitness of meat slices and the median fitness of all meat slices, meat slices are divided into two categories: poor meat slices and good meat slices. The position of the poor meat slices is updated according to the direction of the historical elite combination or the current elite combination. The position of the good meat slices is updated through elite guidance or Levi's flight strategy. The elite update process includes: The top 3 elite candidates are selected by fitness ranking, corresponding to Alpha, Beta and Delta elites respectively; A stall counter is used to monitor the convergence status of the algorithm. The counter increases when no better solution is found for a continuous period of time, and resets when a new Alpha elite is found. During the real-time feature optimization process, when the motion mode has not switched and the motion has not terminated due to recognition errors, a semi-regular weight vector optimization is performed, with the following convergence formula: ； in: These are the newly calculated weight vector values; As a convergence factor, ; This represents the current weight vector value; The weight vector to be approached; It is a random factor; In order to be in A random number that is uniformly distributed within a range; When the motion mode changes and the motion does not terminate due to recognition error, the retrained dataset is used. It is expressed as follows: ； in: The number of motion states; The current movement number; For the first A dataset of pre-stored motion states; A dataset consisting of the previous few steps of the current gait; Total sample size for: ； in: The amount of data pre-stored for each motion state; The amount of historical motion data pre-stored before the current gait.

2. The intelligent lower limb prosthesis human movement intention recognition method based on the stir-frying optimization algorithm according to claim 1, characterized in that, The sensor signals include IMU sensor signals, vertical components of ground reaction force, and knee joint angle signals.

3. An intelligent lower limb prosthesis human motion intention recognition system based on a stir-fry optimization algorithm, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent lower limb prosthesis human movement intention recognition method based on the stir-fry optimization algorithm as described in claim 1 or 2 is implemented.

Images