A human gait recognition method based on particle swarm optimization BP neural network

Abstract

This invention provides a human gait recognition method based on a particle swarm optimization (PSO) backpropagation (BP) neural network. The method uses gait data, after filtering, feature extraction, and segmentation based on gait phase division criteria, as input. The weights and thresholds of the BP neural network are optimized by iteratively updating the velocity and position of particles. The optimized BP neural network is then used to train the input data, deriving the trained model parameters, which are then embedded into an exoskeleton prototype for real-time gait phase recognition testing. This method exhibits good real-time performance and high accuracy, enabling accurate gait phase recognition.

Description

Technical Field

[0001] This invention relates to the field of human gait recognition technology, specifically to a human gait recognition method based on a particle swarm optimization BP neural network. Background Technology

[0002] Gait recognition, as a key technology in wearable exoskeleton system research, is of great significance for improving human quality of life and work efficiency. With the continuous advancement of technology, exoskeleton technology has moved from concept to reality, gradually becoming an important tool to assist humans in daily activities and special tasks.

[0003] In the field of medical rehabilitation, gait recognition technology helps doctors and therapists more accurately assess patients' walking patterns and motor abilities, thereby developing more personalized rehabilitation plans. Through precise gait analysis, exoskeleton systems can provide necessary support for patients with mobility impairments, helping them regain their walking ability and improve their quality of life.

[0004] In the military and emergency rescue fields, gait recognition technology is equally crucial. Soldiers and rescue personnel often need to perform tasks in complex and ever-changing environments, and exoskeleton systems can provide additional strength and endurance to help them carry heavy equipment or walk in harsh conditions. The accuracy of gait recognition is directly related to the performance of the exoskeleton system and the safety of the user.

[0005] In industrial production, exoskeleton systems can assist workers in lifting heavy objects and performing precision operations, reducing physical strain and the risk of injury. Gait recognition technology ensures that the exoskeleton system synchronizes with the user's movements, improving work efficiency and safety.

[0006] Furthermore, with the development of artificial intelligence and machine learning technologies, the accuracy and efficiency of gait recognition will be further improved. Through big data analysis and pattern recognition, exoskeleton systems can more intelligently adapt to changes in the user's gait, achieving more natural and fluid human-computer interaction.

[0007] Applying a robust gait recognition algorithm in wearable smart devices can better couple human posture and movement, and predict motion information. The reference "Lim DH, Kim WS, Kim H J. Development of real-time gait phase detection system for a lower extremity exoskeleton robot[J]. International Journal of Precision Engineering and..." Manufacturing, 2017, 18(5): 681-687 (Lin Dehui, Jin Weisen, Jin Hongjie. Development of real-time gait phase detection system for lower limbs and research on exoskeleton robots [J]. Journal of Precision Engineering, 2017, 18(5): 681-687) Tsinghua University used a third-order Butterworth filter to filter the gait acceleration data of human movement, and then used wavelet transform to extract gait features from the filtered data. Finally, the DTW algorithm was used to identify the human gait phase. The literature "Hu Huilian. Design of gait recognition system based on plantar pressure and posture signal [D]. Harbin Institute of Technology, 2019" proposed an improved method based on force-sensitive resistor to measure plantar pressure to determine gait phase in order to address the problem of poor adaptability of traditional threshold methods under different individuals and different walking speeds. This improved method uses the proportional method and self-correcting threshold algorithm for gait phase detection. In addition, the literature "Martinez HU, Dehghani S A. Adaptive Bayesian inference system for recognition of walking activities and prediction of gait events" Martinez HU, Dehghani SA, "Gait Recognition for Mobile Robots Based on Bayesian Inference [J]. Neural Networks, 2018, 102(1): 107-119" proposed a gait recognition method based on Adaboost using multi-class support vector machine (SVM). After testing and analysis, it was shown that the method improved the gait classification accuracy to a certain extent. Summary of the Invention

[0008] The purpose of this invention is to provide a human gait recognition method based on particle swarm optimization BP neural network, which aims to solve the problems of efficiency and accuracy in classifying active pulmonary tuberculosis images under conditions of small sample size and sample imbalance.

[0009] To achieve the above objectives, this invention provides a human gait recognition method based on a particle swarm optimization (BP) neural network, comprising the following steps:

[0010] Step S1: Data acquisition and preprocessing. Collect plantar pressure and IMU joint angle data for the set gait cycle as the training and testing sets for the neural network, and preprocess the collected data.

[0011] Step S2, Feature Extraction and Data Classification: Using a temporal feature extraction method, the peak value, standard deviation, minimum value, variance, and mean value of each gait phase in the preprocessed data from Step S1 are extracted from the plantar pressure and IMU joint angle data to form feature vectors. According to the defined gait phase classification criteria, the acquired feature data are classified into four phases: heel strike, flat stance, toe lift, and swing phase. The segmented phases are labeled as Class 1, Class 2, Class 3, and Class 4, respectively. Each phase data is regarded as an input sample, and each input sample contains 8 feature vectors. The sample class ratio is 1:1:1:1.

[0012] Step S3, Neural Network Model Optimization: Optimize the BP neural network model using the PSO algorithm on the samples.

[0013] In step S1, the preprocessing uses a combination of Kalman filtering and sliding window processing.

[0014] The formula for extracting the feature vector using the time-domain feature extraction method in step S2 is as follows:

[0015] (1) Peak value: The maximum value of the data:

[0016] X max =max|x i | (3)

[0017] In the formula, x i This indicates the number of data points, and `max` indicates extracting the maximum value among all numbers.

[0018] (2) Variance: The average deviation of the data.

[0019]

[0020] In the formula, This represents the average value of the data. This means finding each x i The sum of squares of the differences between the mean and the average;

[0021] (3) Standard deviation: The degree of dispersion of the numerical values.

[0022]

[0023] In the formula, S n σ represents the average deviation of the data, while σ describes the degree of dispersion of the data.

[0024] (4) Average: The average value:

[0025]

[0026] In the formula, When x is displayed i The sum of all values ​​divided by the total number n is the average.

[0027] (5) Minimum value: The minimum value of the numerical value:

[0028] X min =min|x i | (7)

[0029] In the formula, x i This indicates the number of data points, and min indicates extracting the minimum value among all numbers.

[0030] The method for classifying feature data is as follows: First, the Findpeaks function is used to find the location and number of peak points in each cycle; then, the number of sampling points in each cycle, i.e., the cycle length, is estimated based on the ratio of the total number of sampling points to the number of peak points; second, the sampling segment in which each cycle is located is determined based on the location of the peak points and the cycle length; finally, based on the forefoot pressure threshold and the heel pressure threshold, the gait data of each cycle is divided into four parts: heel strike (HS), flat stance (FF), toe lift (TO), and swing phase (SW), and sample data of each category are randomly selected to form the training set and the test set.

[0031] Neural network model optimization steps:

[0032] Step 1: Determine the BP neural network structure and the input layer x. i Hidden layer b i and output layer y i The number of neurons a, b, and c; training set samples and test set samples P and T;

[0033] Step 2: Initialize particle population parameters. First, initialize the population size N and the maximum number of iterations E. max The inertia coefficient ω, learning factors c1 and c2, and particle dimension d = a × b + b × c are used to initialize the particle's velocity and position using a Gaussian distribution.

[0034]

[0035] In the formula: x i,j (t) represents the initial position of the particle, v i,j (t) represents the initial velocity of the particle, N(μ,σ)2 ) indicates that the sample follows a pattern with mean μ and variance σ. 2 Gaussian distributed random numbers;

[0036] Step 3: Calculate the particle's fitness value (fit) and individual extreme value (p). best and the group extreme value g best To reduce the error between the predicted and expected values, the mean squared error formula is chosen as the fitness function for the PSO algorithm:

[0037]

[0038] In the formula: For the predicted value, y i (k) represents the expected value, N represents the number of particles, and m represents the number of features for each sample;

[0039] Step 4: Update the particle's velocity and position using the velocity formula (10) and position formula (11):

[0040] v i,j (t+1)=ωv i,j (t)+c1r1(p i,j (t)-x i,j (t))+c2r2(g i (t)-x i,j (t)) (10)

[0041] x i,j (t+1)=x i,j (t)+v i,j (t+1) (11)

[0042] In the formula: i = 1, 2, ..., m; j = 1, 2, ..., n; v i,j (t) and v i,j (t+1) represents the current velocity and the velocity at the next moment, respectively; ω represents the inertia coefficient; c1 and c2 are learning factors; r1 and r2 are random numbers between [0,1]; x i,j (t) and x i,j (t+1) represent the particle's current position and next position, respectively. i,j (t) represents the historical best position of the i-th particle in the j-th dimension, g i (t) The global best position of all particles in the j-th dimension so far;

[0043] Step 5: Determine whether the minimum error requirement has been met. If the requirement has been met, end the algorithm and initialize the initial weights and thresholds of the BP neural network using the globally optimal particle position. Otherwise, return to step 3 to continue updating the particle's velocity and position.

[0044] Step 6: Input samples, train the BP neural network, calculate the output and error of each layer, calculate the output of the BP neural network based on the current weights and thresholds, and calculate the error between the output and the actual label;

[0045] Step 7: Determine whether the error of all samples has reached the expected target. If the preset termination condition is met, end the algorithm; otherwise, return to step 6 to continue training the BP neural network.

[0046] Compared with traditional methods, this invention uses a PSO-BP neural network model for human gait recognition, which can solve the problem that traditional BP neural networks are prone to getting trapped in local optima. Through experimental verification, real-time walking data of the human body is collected and the current gait phase is identified. Online test results show that the model has a high recognition rate, good real-time performance, and can accurately identify the human gait phase. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the phase division method;

[0048] Figure 2 It is an IMU joint angle curve;

[0049] Figure 3 This is a schematic diagram after data processing;

[0050] Figure 4 This is a schematic diagram of gait phase division;

[0051] Figure 5 This is a diagram of the PSO-BP algorithm;

[0052] Figure 6 This is a diagram of the network training process;

[0053] Figure 7 This is the ROC curve of the PSO-BP neural network;

[0054] Figure 8 It is a confusion matrix diagram;

[0055] Figure 9 This is the result of gait recognition. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0057] Example 1: A human gait recognition method based on particle swarm optimization BP neural network, comprising the following steps:

[0058] Step S1: Data acquisition and preprocessing. Collect plantar pressure and IMU joint angle data for the set gait cycle as the training and testing sets for the neural network, and preprocess the collected data.

[0059] The wearable multi-sensor gait data acquisition method is not only simple to acquire information and highly reliable, but also low in cost and has strong real-time performance. The hip exoskeleton robot prototype platform used in this invention includes the following components: a processor (STM32F765IIK6), an encoder, an IMU (JY931), a hip joint motor (BLDC), and a plantar pressure shoe. After the system is powered on, data from the IMU sensor and the plantar pressure sensor are sent to the PC for real-time storage and processing.

[0060] During data acquisition, numerous uncertainties can affect the data and introduce noise. To filter out noise, preprocessing employs a combination of Kalman filtering and sliding window processing. This not only improves the accuracy and stability of filtering but also reduces data latency and lag. Figure 3 As shown, the filtered data curve is smoother and more even than the original data curve.

[0061] Step S2, Feature Extraction and Data Classification: Using the temporal feature extraction method, the peak value, standard deviation, minimum value, variance and mean value of each gait phase of the preprocessed data in step S1 are extracted from the plantar pressure and IMU joint angle data to form a feature vector. According to the defined gait phase classification criteria, the acquired feature data are classified into four phases: heel strike, flat stance, toe lift-off and swing phase.

[0062] For the application of lower limb hip joint exoskeleton robots, this paper divides the gait cycle into four phases based on the characteristics of heel-to-toe pressure changes: heel strike, level stance, toe lift-off, and swing phase. Figure 1 As shown. Taking a single cycle of the left leg as an example, the phase division criteria are as follows:

[0063] Heel strike (HS): Starts when the left heel touches the ground and ends when the toes touch the ground;

[0064] Horizontal FF (Mid-Support): Start from the left toes touching the ground and end when the left heel leaves the ground;

[0065] Toe liftoff (TO): From the moment the left heel lifts off the ground until the moment the toes lift off the ground;

[0066] Swing phase (SW): begins when the left toes leave the ground and ends when the left heel makes contact with the ground again.

[0067] The segmented phases are labeled as class 1, class 2, class 3 and class 4 respectively. Each phase data is regarded as an input sample. Each input sample contains 8 feature vectors. The sample class ratio is 1:1:1:1.

[0068] The formula for extracting feature vectors using the temporal feature extraction method is as follows:

[0069] (1) Peak value: The maximum value of the data:

[0070] X max =max|x i | (3)

[0071] In the formula, x i This indicates the number of data points, and `max` indicates extracting the maximum value among all numbers.

[0072] (2) Variance: The average deviation of the data.

[0073]

[0074] In the formula, This represents the average value of the data. This means finding each x i The sum of squares of the differences between the mean and the average;

[0075] (3) Standard deviation: The degree of dispersion of the numerical values.

[0076]

[0077] In the formula, S n σ represents the average deviation of the data, while σ describes the degree of dispersion of the data.

[0078] (4) Average: The average value:

[0079]

[0080] In the formula, When x is displayed i The sum of all values ​​divided by the total number n is the average.

[0081] (5) Minimum value: The minimum value of the numerical value:

[0082] X min =min|x i | (7)

[0083] In the formula, x i This indicates the number of data points, and min indicates extracting the minimum value among all numbers.

[0084] The method for classifying feature data is as follows: First, the Findpeaks function is used to find the location and number of peak points in each cycle; then, the number of sampling points in each cycle, i.e., the cycle length, is estimated based on the ratio of the total number of sampling points to the number of peak points; second, the sampling segment in which each cycle is located is determined based on the location of the peak points and the cycle length; finally, based on the forefoot pressure threshold and the heel pressure threshold, the gait data of each cycle is divided into four parts: heel strike (HS), flat stance (FF), toe lift (TO), and swing phase (SW), and sample data of each category are randomly selected to form the training set and the test set.

[0085] Step S3, Neural Network Model Optimization: Optimize the BP neural network model using the PSO algorithm on the samples.

[0086] Neural network model optimization steps:

[0087] Step 1: Determine the BP neural network structure and the input layer x. i Hidden layer b i and output layer y i The number of neurons a, b, and c; training set samples and test set samples P and T;

[0088] Step 2: Initialize particle population parameters. First, initialize the population size N and the maximum number of iterations E. max The inertia coefficient ω, learning factors c1 and c2, and particle dimension d = a × b + b × c are used to initialize the particle's velocity and position using a Gaussian distribution.

[0089]

[0090] In the formula: x i,j (t) represents the initial position of the particle, v i,j (t) represents the initial velocity of the particle, N(μ,σ) 2 ) indicates that the sample follows a pattern with mean μ and variance σ. 2 Gaussian distributed random numbers;

[0091] Step 3: Calculate the particle's fitness value (fit) and individual extreme value (p). best and the group extreme value g best To reduce the error between the predicted and expected values, the mean squared error formula is chosen as the fitness function for the PSO algorithm:

[0092]

[0093] In the formula: For the predicted value, y i (k) represents the expected value, N represents the number of particles, and m represents the number of features for each sample;

[0094] Step 4: Update the particle's velocity and position using the velocity formula (10) and position formula (11):

[0095] v i,j (t+1)=ωv i,j (t)+c1r1(p i,j (t)-x i,j (t))+c2r2(g i (t)-x i,j (t)) (10)

[0096] x i,j (t+1)=x i,j (t)+v i,j (t+1) (11)

[0097] In the formula: i = 1, 2, ..., m; j = 1, 2, ..., n; v i,j (t) and v i,j (t+1) represents the current velocity and the velocity at the next moment, respectively; ω represents the inertia coefficient; c1 and c2 are learning factors; r1 and r2 are random numbers between [0,1]; x i,j (t) and x i,j (t+1) represent the particle's current position and next position, respectively. i,j (t) represents the historical best position of the i-th particle in the j-th dimension, g i (t) The global best position of all particles in the j-th dimension so far;

[0098] Step 5: Determine whether the minimum error requirement has been met. If the requirement has been met, end the algorithm and initialize the initial weights and thresholds of the BP neural network using the globally optimal particle position. Otherwise, return to step 3 to continue updating the particle's velocity and position.

[0099] Step 6: Input samples, train the BP neural network, calculate the output and error of each layer, calculate the output of the BP neural network based on the current weights and thresholds, and calculate the error between the output and the actual label;

[0100] Step 7: Determine whether the error of all samples has reached the expected target. If the preset termination condition is met, end the algorithm; otherwise, return to step 6 to continue training the BP neural network.

[0101] This invention proposes a BP neural network recognition method based on particle swarm optimization. This method uses data that has undergone filtering, feature extraction, and segmentation based on gait phase division criteria as input. It iteratively updates the velocity and position of particles to optimize the weights and thresholds of the BP neural network. The optimized BP neural network is then trained on the input data, and the trained model parameters are derived and embedded into an exoskeleton prototype for real-time gait phase recognition testing. Results show that the model exhibits good real-time performance and high accuracy, accurately identifying gait phases. This solves the problem of traditional BP neural networks easily getting trapped in local optima during gait phase recognition, and enhances the accuracy and efficiency of BP neural networks in gait phase recognition.

[0102] Simulation Experiment: Data Acquisition. Ten healthy men aged 20 to 30, weighing between 62 kg and 78 kg, and tall between 162 cm and 180 cm, were invited to wear the exoskeleton device for data acquisition and testing. Plantar pressure and IMU joint angle data from 120 gait cycles (480 gait phases) for each participant were selected as the training and testing sets for the neural network. The IMU joint angle data is shown below. Figure 2 As shown.

[0103] Data Classification: The acquired feature data is classified according to the defined gait phase division criteria. First, the Findpeaks function is used to identify the locations and number of peak points within each cycle. Next, the number of sampling points per cycle (i.e., cycle length) is estimated based on the ratio of the total number of sampling points to the number of peak points. Then, the sampling segment for each cycle is determined based on the peak point locations and cycle length. Finally, based on the forefoot pressure threshold (1 kg) and the rearfoot pressure threshold (2 kg), the gait data for each cycle is divided into four parts: HS, FF, TO, and SW. The division results are as follows: Figure 4 As shown, the segmented phases were labeled as class 1, class 2, class 3, and class 4. Each phase data was considered as an input sample, and each input sample contained 8 feature vectors. The sample class ratio was 1:1:1:1. Ten experimenters provided a total of 4800 samples. 800 samples from each class were randomly selected to form a training set of 3200 samples, and the remaining 1600 samples were used as the test set.

[0104] Construction of the PSO-BP model: The experiment was conducted on the MATLAB R2022b platform. Labeled sample data was used to construct sample vectors to train the PSO-BP network. The input and output layers of the BP neural network had 8 and 4 neurons, respectively. Through multiple experiments, the best results were achieved when the number of particles N was 20, learning factors c1 and c2 were 1.5 and 2, the inertia coefficient ω was 0.7, the number of hidden layer neurons b was 20, the target error was set to 0.1, and the number of iterations was 100. The training process of the network at this time is as follows: Figure 6 As shown in the figure, the green circle indicates that the network required 10 iterations to achieve the optimal mean squared error value on the validation set. The ROC curve of the PSO-BP neural network model is shown below. Figure 7 As shown. The confusion matrix of the sample data recognition results is as follows. Figure 8 As shown, the overall recognition rate reached 96.1%.

[0105] Online testing: First, the parameters of the PSO-BP neural network model were exported and embedded into the exoskeleton prototype control program using C language. Real-time acquisition of plantar pressure and IMU joint angle data was used to construct sample vectors, which were then input into the control program for recognition. The four phases HS, FF, TO, and SW in the recognition results were labeled with 5, 10, 15, and 20 respectively. A line graph was constructed based on the labels, and the line graph and acquired data were output in real time. Figure 9 As shown in the figure, the model has a high recognition rate and good real-time performance, and can accurately identify the phase of human gait.

[0106] In summary, this invention focuses on gait phase recognition during normal human walking. First, the collected data is filtered and classified. Then, addressing the problem of traditional BP neural networks easily getting trapped in local optima, a PSO-BP neural network model is constructed, trained, and tested on the classified data. Finally, the neural network model parameters are exported and embedded into the control program to collect real-time human walking data and identify the current gait phase. Online testing results show that the model has a high recognition rate, good real-time performance, and can accurately identify human gait phase.

[0107] The present invention provides a detailed description of a human gait recognition method based on a particle swarm optimization BP neural network. Specific examples are used to illustrate the principle and implementation of the invention. The above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​the invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the invention. Therefore, the content of this specification should not be construed as a limitation of the invention.

Claims

1. A human gait recognition method based on particle swarm optimization BP neural network, characterized in that, Includes the following steps: Step S2: Data acquisition and preprocessing. Collect plantar pressure and IMU joint angle data for the set gait cycle as the training and testing sets for the neural network, and preprocess the acquired data. Step S2, Feature Extraction and Data Classification: Using a temporal feature extraction method, the peak value, standard deviation, minimum value, variance, and mean value of each gait phase in the preprocessed data from Step S1 are extracted from the plantar pressure and IMU joint angle data to form feature vectors. According to the defined gait phase classification criteria, the acquired feature data are classified into four phases: heel strike, flat stance, toe lift, and swing phase. The segmented phases are labeled as Class 1, Class 2, Class 3, and Class 4, respectively. Each phase data is regarded as an input sample, and each input sample contains 8 feature vectors. The sample class ratio is 1:1:1:

1. Step S3, Neural Network Model Optimization: Optimize the BP neural network model using the PSO algorithm on the samples; Neural network model optimization steps: Step 1: Determine the BP neural network structure and the input layer. Hidden layer and output layer Number of neurons Training set samples and test set samples and ; Step 2: Initialize particle population parameters, starting with the population size. Maximum number of iterations coefficient of inertia Learning factor and and particle dimensions Secondly, the velocity and position of the particles are initialized using a Gaussian distribution: （8）， In the formula: Indicates the initial position of the particle. Represents the initial velocity of the particle. Indicates that it follows the mean. The variance is Gaussian distributed random numbers; Step 3: Calculate the fitness value of the particles. Individual extreme values and group extreme values To reduce the error between the predicted and expected values, the mean squared error formula is chosen as the fitness function of the PSO algorithm: （9）， In the formula: For predicted values, For the expected value, This represents the number of particles, where m represents the number of features per sample. Step 4: Update the particle's velocity and position using the velocity formula (10) and position formula (11): （10）， （11）， In the formula: ; ; and These represent the current speed and the speed at the next moment, respectively. Represents the coefficient of inertia. and As a learning factor, and It is a random number between [0,1]. and These represent the particle's current position and its next position, respectively. Indicates the first The particle in the first The best historical location on the vehicular, So far, all particles are in the first... The globally optimal position in the dimension; Step 5: Determine whether the minimum error requirement has been met. If the requirement has been met, end the algorithm and initialize the initial weights and thresholds of the BP neural network using the globally optimal particle position. Otherwise, return to step 3 to continue updating the particle's velocity and position. Step 6: Input samples, train the BP neural network, calculate the output and error of each layer, calculate the output of the BP neural network based on the current weights and thresholds, and calculate the error between the output and the actual label; Step 7: Determine whether the error of all samples has reached the expected target. If the preset termination condition is met, end the algorithm; otherwise, return to step 6 to continue training the BP neural network.

2. The human gait recognition method based on a particle swarm optimization BP neural network according to claim 1, characterized in that, In step S1, the preprocessing uses a combination of Kalman filtering and sliding window processing.

3. The human gait recognition method based on particle swarm optimization BP neural network as described in claim 1, characterized in that, The formula for extracting the feature vector using the time-domain feature extraction method in step S2 is as follows: (1) Peak value: The maximum value of the data: （3）， In the formula, Indicates the number of data points. This indicates extracting the maximum value among all numbers; (2) Variance: The average deviation of the data: (4)， In the formula, This represents the average value of the data. This means finding each The sum of squares of the differences between the mean and the average; (3) Standard deviation: The degree of dispersion of the numerical values: （5）， In the formula, Indicates the average deviation of the data. Describes the degree of dispersion of the data; (4) Average: The average value: (6)， In the formula, When the table will be The sum of all values ​​divided by the total number n is the average. (5) Minimum value: The minimum value of the numerical value: (7)， In the formula, Indicates the number of data points. This means extracting the minimum value among all numbers.

4. The human gait recognition method based on particle swarm optimization BP neural network according to claim 1, characterized in that, The method for classifying feature data is as follows: First, the Findpeaks function is used to find the location and number of peak points in each cycle; then, the number of sampling points in each cycle, i.e., the cycle length, is estimated based on the ratio of the total number of sampling points to the number of peak points; second, the sampling segment in which each cycle is located is determined based on the location of the peak points and the cycle length; finally, based on the forefoot pressure threshold and the heel pressure threshold, the gait data of each cycle is divided into four parts: heel strike (HS), flat stance (FF), toe lift (TO), and swing phase (SW), and sample data of each category are randomly selected to form the training set and the test set.

Images