A Personalized Prediction Method for Human Lower Limb Gait Phase Based on Informer

Abstract

This invention relates to a personalized prediction method for human lower limb gait phase based on Informer, belonging to the field of exoskeleton robot technology. The method includes the following steps: collecting human lower limb gait motion information through a data acquisition device to form a dataset; training a pre-model; saving the trained pre-model; collecting individual user lower limb data; training a personalized model based on the pre-model; saving the personalized model; collecting the swing angle of the right thigh of the human lower limb through the data acquisition device; and using the model based on the wearer's real-time data. Unlike traditional gait phase prediction methods, this method increases model complexity while avoiding the use of numerous sensors, reducing hardware usage and increasing model accuracy. It then uses prior information such as specific human-machine coupling characteristics to more accurately set the corresponding controller structure and control parameters, effectively improving the wearing comfort of exoskeleton users during human-machine coupled collaborative walking movements.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton robot technology, and in particular to a personalized prediction method for human lower limb gait phase based on Informer. Background Technology

[0002] Lower limb exoskeleton rehabilitation robots are a novel type of wearable bionic device primarily used to enable patients with lower limb motor dysfunction to stand and walk, and are currently a hot research topic in rehabilitation medicine. Gait phase refers to the relationship between the movement states of different limbs during human walking. Lower limb exoskeleton rehabilitation robots need to control their movement based on gait phase to achieve an effect similar to human gait.

[0003] Currently, there are several main methods for gait phase control. First, sensor control involves installing sensors on the robot, such as pressure sensors, gyroscopes, and accelerometers, to acquire information about the robot and its environment, thereby controlling the robot's movement. Second, EMG signal control uses electromyography (EMG) signals to control the robot's movement. This method requires attaching electrodes to muscles and controlling the robot by measuring the electrical signals generated during muscle contraction. Finally, model predictive control uses mathematical models to predict the robot's movement and controls it based on the predictions. This method requires modeling the robot and its environment and using optimization algorithms to solve for the optimal control strategy. Despite these methods enabling gait phase estimation for lower limb exoskeleton rehabilitation robots, several drawbacks remain. For example, human gait information perception relies heavily on numerous sensors, and gait phase prediction models have poor generalization ability across different wearers. Summary of the Invention

[0004] The present invention aims to solve the technical problems in the prior art by providing a personalized prediction method for human lower limb gait phase based on Informer.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A personalized prediction method for human lower limb gait phase based on Informer includes the following steps:

[0007] Step S1: Collect human lower limb gait movement information through a data acquisition device and form a dataset;

[0008] Step S2: Train the pre-model;

[0009] Data standardization of human lower limb gait data;

[0010] Foot force information is used to mark heel strike events, and gait phase is mapped proportionally to the angle range;

[0011] Window sliding cropping processing: A sliding window is used to crop the standard data matrix to obtain several training data matrices;

[0012] Design the loss function and learning rate for the Informer model;

[0013] Training and validation sets are used for training neural network models;

[0014] The test set evaluates the model's generalization ability;

[0015] Step S3: Save the trained Informer pre-model;

[0016] Step S4: Collect individual user lower limb data;

[0017] Collect information on the swing angle of the right thigh and plantar force of the wearer's lower limb under different speeds and movements;

[0018] Step S5: Train a personalized model based on the Informer pre-model;

[0019] Step S6: Save the Informer personalized model;

[0020] Step S7: Collect the swing angle of the right thigh of the human lower limb using a data acquisition device;

[0021] Collect real-time user data to obtain the real-time swing angle of the right thigh of the user's lower limb under different speeds and movements;

[0022] Step S8: Use the model based on the wearer's real-time data;

[0023] The user's real-time data is standardized to obtain a standardized real-time user data vector. The standardized real-time user data vector is then truncated using a sliding window to obtain several real-time right thigh swing angle vectors of the user's lower limb. Using the Informer personalization model, the user's real-time right thigh swing angle vectors are input into the Informer personalization model to obtain the sine and cosine values ​​of the user's real-time gait phase, and inverse trigonometric function calculations are performed to obtain the user's real-time gait phase prediction value.

[0024] In the above technical solution, step S2 involves performing a sliding window truncation on the standard data matrix to obtain several training data matrices, specifically as follows:

[0025] Obtain the truncation width of the desired training data matrix, which is smaller than the width of the standard data matrix; select multiple adjacent columns of data from the standard data matrix with the truncation width to obtain the training data matrix; construct the training set, validation set, and test set; design the Informer model network parameters, loss function, optimizer, and learning rate; input the training data matrix into the Informer model to obtain the Informer pre-model, and calculate the loss value by comparing the true gait phase and estimated gait phase corresponding to the training data matrix; then optimize the Informer model based on the loss value.

[0026] In the above technical solution, step S5 specifically includes:

[0027] Based on the user's plantar force information, the heel strike event is marked and the gait phase at this moment is marked as 0%. The phase is gradually increased linearly to 100% until the next heel strike event occurs and is remarked as 0%. The gait phase is then mapped proportionally to the angle range, and the corresponding sine and cosine values ​​are calculated.

[0028] The swing angle of the right thigh of the user's lower limb is standardized to obtain a standardized user data matrix;

[0029] A sliding window is used to truncate the standard user data matrix to obtain several training matrices. This includes: obtaining the truncation width of the desired training user data matrix, which is smaller than the width of the standard user data matrix; selecting multiple adjacent columns of data from the standard user data matrix with the truncation width to obtain the training user data matrix; constructing user training sets, user validation sets, and user test sets; loading the Informer pre-model and freezing all layer parameters except for the fully connected layers; inputting the training user data matrix into the Informer pre-model to obtain the Informer personalized model; calculating the loss value by comparing the true gait phase and the estimated gait phase corresponding to the training data matrix; and then optimizing the Informer model based on the loss value.

[0030] In the above technical solution, in step S5, when constructing the user training set, user verification set and user test set, the dataset is divided into the user training set, user verification set and user test set in a ratio of 7:2:1.

[0031] In the above technical solution, in step S2 and / or step S5, the gait phase is mapped to the angle range proportionally, specifically: the 0%-100% gait phase is mapped to the 0-2π angle range proportionally.

[0032] In the above technical solution, in step S4 and / or step S7, an IMU and a foot force measuring shoe are used to measure the swing angle of the thigh on one side of the human lower limb and / or the plantar force information, respectively.

[0033] In the above technical solution, in step S4 and / or step S7, the IMU sensor is bound to the right thigh above the knee joint in the coronal plane of the human body using a nylon strap with elasticity; the sampling frequency is 200Hz and the communication rate is 2MB / s.

[0034] In the above technical solution, the data acquisition device in step S1 and / or step S7 is: an inertial measurement unit and a thin-film pressure sensor.

[0035] In the above technical solution, step S3 specifically involves saving the parameters and structure of the Informer pre-model, including the loss function used, the optimizer, the dimensions of the input data, and the dimensions of the output data.

[0036] In the above technical solution, step S6 specifically involves saving the network structure, network parameters, and optimizer of the Informer personalized model.

[0037] The present invention has the following beneficial effects:

[0038] This invention presents an Informer-based personalized gait phase prediction method for the human lower limbs. It employs an inertial measurement unit (IMU) and a thin-film pressure sensor to collect information on the unilateral thigh swing angle and plantar force of different individuals of varying ages and body types under different movements and walking speeds. Gait is then segmented based on the plantar force information to obtain gait phase variables. An Informer model is trained using the collected data to obtain an Informer pre-model. Next, user data is collected and the Informer pre-model is trained to obtain a personalized Informer model. Finally, when using the personalized Informer model, real-time user data is collected and input into the personalized Informer model to obtain the user's real-time gait phase prediction value. This method differs from traditional gait phase prediction methods by increasing model complexity while avoiding the use of numerous sensors. This reduces hardware usage and increases model accuracy, allowing for more accurate setting of the corresponding controller structure and control parameters based on prior information such as specific human-machine coupling characteristics. This effectively improves the wearing comfort of exoskeleton users during human-machine coupled collaborative walking movements. Attached Figure Description

[0039] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0040] Figure 1This is a schematic diagram of the Informer neural network model structure of the Informer-based personalized prediction method for human lower limb gait phase according to the present invention.

[0041] Figure 2 This diagram illustrates the steps of Informer pre-model training in the Informer-based personalized prediction method for human lower limb gait phase of the present invention.

[0042] Figure 3 This diagram illustrates the steps of training the Informer personalized model in the Informer-based personalized prediction method for human lower limb gait phase of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the steps of using the Informer model in the Informer-based personalized prediction method for human lower limb gait phase of the present invention.

[0044] Figure 5 This is a schematic diagram of the data acquisition device involved in the Informer-based personalized prediction method for human lower limb gait phase of the present invention. Detailed Implementation

[0045] The inventive concept of this invention is as follows:

[0046] The Informer-based personalized prediction method for human lower limb gait phase of this invention trains a human gait phase estimation model by using real-time measurements of the swing angle of the right thigh and plantar force information of the human lower limb from two sensors. Then, it uses only one sensor to measure the swing angle of the right thigh of the human lower limb in real time to predict the future gait phase of the exoskeleton robot, and trains a personalized network model according to a specific user individual, thereby achieving personalized gait phase estimation and prediction.

[0047] The present invention will now be described in detail with reference to the accompanying drawings.

[0048] like Figure 1-5 As shown, the Informer-based personalized prediction method for human lower limb gait phase of the present invention includes the following steps:

[0049] Step S1: Collect human lower limb gait movement information through a data acquisition device and form a dataset;

[0050] The swing angle of the right thigh and the plantar force information of different body types of people under different speeds and movements were collected by an inertial measurement unit and a thin-film pressure sensor.

[0051] To ensure the estimation performance of the trained Informer neural network model, the data collection experiments involved healthy subjects of different body types (height, weight, gender, and age, etc.). The tested activity patterns included basic daily activities such as walking, uphill, downhill, climbing stairs, and descending stairs. Multiple experiments were conducted at different movement speeds for each activity pattern. For pre-model data collection, volunteers were recruited and wore data collection devices to complete specific walking patterns, resulting in the collection of a large amount of lower limb movement data to train the Informer pre-model.

[0052] This invention does not impose any particular limitations on the acquisition device used to measure the swing angle of one thigh and plantar force information of the human lower limb; any device capable of performing the aforementioned measurement functions within the art is acceptable. In the experimental process, this invention employs an IMU (Internal Measurement Unit) and a plantar force measuring shoe to measure the swing angle of one thigh and plantar force information of the human lower limb, respectively.

[0053] To accurately measure the swing angle of the unilateral thigh and plantar force information of the human lower limb during walking, an IMU sensor is strapped to the right thigh, approximately 4 cm above the knee joint in the coronal plane, using an elastic nylon strap. The sampling frequency is 200Hz, and the communication rate is 2MB / s. To ensure simultaneous data acquisition by the IMU and the plantar force-measuring shoe, a trigger switch is implemented on the hardware controller (STM32F103) and connected to an external interrupt. When the switch is on, all sensors begin operating; when the switch is off, all sensors cease operation. A schematic diagram of the data acquisition device is shown below. Figure 5 As shown.

[0054] Step S2, train the Informer pre-model, the specific steps of which include:

[0055] First, the data of human lower limb gait is standardized; the swing angle of the right thigh is standardized to obtain a standardized data matrix.

[0056] Specifically, Z-score standardization is performed on human lower limb gait data to obtain standardized data. In this embodiment, the mean and standard deviation of each variable at n time sampling points are first calculated. Then, historical data are Z-score standardized, and the standardized data is used as standardized data X, where X is an n×m matrix. The Z-score standardization process is calculated according to the following formula:

[0057]

[0058] Among them, X m D represents the value of the m-th variable in the standardized data X. mLet mean(D) represent the m-th variable in historical data D. m std(D) represents the average value of the m-th variable in historical data D. m ) represents the standard deviation of the m-th variable in the historical data D.

[0059] Then, the plantar force information is used to mark heel strike events, and the gait phase is mapped proportionally to the angle range;

[0060] Foot plantar force information is used to mark heel strike events, and the phase variable P (0-100%) is mapped to the angle range (0-2π). This is because daily human movements such as walking, running, climbing stairs, and descending stairs are periodic movements. Therefore, within a single cycle, the phase variable P represents the gait progress at the current moment, i.e., P ranges from 0-100%. However, if the phase variable P is directly used as the output of the Informer neural network model, it will cause the phase variable to directly enter the next gait phase cycle before reaching 100% in the current cycle, resulting in discontinuity in the gait phase. Therefore, the phase variable P is converted from rectangular coordinates to polar coordinates and represented by the angle α.

[0061] In this embodiment, the phase variable and angle are first mapped, and then the angle is substituted into the sine and cosine functions and labeled as y1 and y2, which are used as the output of the Informer neural network model.

[0062]

[0063] y1=sin(α)

[0064] y2=cos(α)

[0065] Next, a sliding window cropping process is performed; a sliding window is used to crop the standard data matrix to obtain several training data matrices.

[0066] In the above embodiments, the historical data matrix is ​​truncated to obtain several training data matrices. First, the truncated width of the desired training matrix is ​​determined, and this width is smaller than the width of the historical data matrix. Then, adjacent columns of data within the truncated historical data matrix are selected to obtain the training data matrices. It should be noted that the truncated width is generally much smaller than the width of the historical data matrix to ensure that a sufficient number of training data matrices are generated. For example, standardized data X of size m×n can be reorganized into m×d×(n-d+1) training data T. T contains (n-d+1) m×d training data matrices as input data. The dataset is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0067] Next, the loss function and learning rate of the Informer model are designed; the loss function is the RMSE loss used for training, and the Adam algorithm is used to optimize the model parameters.

[0068] Depend on Figure 1 As shown, this is the architecture diagram of the Informer model. The Informer model includes an Encoder module and a Decoder module. The training data matrix is ​​input into the Encoder module to obtain the encoded features.

[0069] In this embodiment, the Encoder can generate a feature matrix based on the input training data matrix. For example... Figure 1 As shown, the Encoder consists of K modules stacked together for feature extraction. Each sub-module includes Multi-head Prob Sparse Self-attention and Self-attention distillation. The Self-attention structure includes W... Q W K W V Given a matrix and input data X, then:

[0070] Q = W Q X

[0071] K = W K X

[0072] V = W V X

[0073] Q, K, and V represent query, key, and value, respectively. Then, features are extracted using the following formula:

[0074]

[0075] In Multi-head ProbSarse and Self-attention, feature extraction is performed using a sparse matrix Q. For each vector q in Q... i Calculate the sparsity metric M(q) i K):

[0076]

[0077] Select the n q values ​​with the largest sparsity metric. i Reassemble Then feature extraction calculations are performed. Only the selected n q values ​​are calculated. i The other q i We directly take the mean(V), and then merge the results of S1 and S0 to obtain the final output S.

[0078]

[0079] In the Self-attention distilling operation, higher weights are assigned to dominant features that have primary characteristics, and a focused Self-attention feature matrix is ​​generated in the next layer. Self-attention distilling includes one-dimensional convolution (Conv1d), ELU activation function, and max pooling operation.

[0080] I i+1 =MaxPool(ELU(Conv1d[S i ]))

[0081] It should be noted that in this embodiment, in order to predict the gait phase at the future prediction time, the training data matrix also needs to be padded. For example, the historical time length is d, and the prediction time length is t. Furthermore, the Decoder's input uses "generative inference," where the predicted future phase sequence is padded with 0 values ​​of length t, and the input vector is (d+t). Figure 1 As shown, the input vector of the Decoder passes through a Masked Multi-head Prob Sparse Self-attention layer, which prevents each position from focusing on the future position, thus avoiding autoregression. The output of the Masked Multi-head Prob Sparse Self-attention layer serves as the query, and the output feature map of the Encoder serves as the key and value, which are then input to a Multi-head Attention layer. The output of the Decoder passes through a fully connected layer, outputting the gait phase prediction result at a future time length t.

[0082] In this embodiment, firstly, the input historical human lower limb data corresponding to all matrix elements of the training data matrix are obtained. Then, the training data matrix is ​​input into the Encoder module to obtain the feature matrix. Next, the feature matrix and the input historical gait phases are input into the Decoder module to generate predicted gait phases. Finally, the Informer model is trained based on the predicted gait phases. Thus, by inputting historical gait phases and the training data matrix into the Informer model, and through continuous iteration and optimization, an Informer model capable of accurately predicting lower limb gait phases can be obtained.

[0083] Next, the training and validation sets are used to train the Informer neural network model, and the test set is used to test the model's generalization ability. If the generalization ability does not meet the requirements, the model is retrained with parameter tuning until its generalization ability meets the requirements. The trained model is then saved as an Informer pre-model.

[0084] It is understandable that model training is an iterative process, involving continuous adjustment of the model's network parameters until the overall loss function value is less than a preset value, or the overall loss function value stops changing or changes slowly, indicating model convergence and a well-trained model is obtained. In this embodiment, the preset value can be changed according to actual design needs, and no limitations are imposed here.

[0085] Step S3: Save the trained Informer pre-model;

[0086] After training is complete, save the parameters and structure of the Informer pre-model, including the loss function used, the optimizer, the dimensions of the input data, and the dimensions of the output data.

[0087] Step S4: Based on the designed acquisition device, acquire the swing angle of the wearer's right thigh and the plantar force signal.

[0088] To ensure the predictive performance of the Informer personalized model, the data collection experiments used personalized data from wearer users. The tested activity patterns included basic daily activities such as walking, uphill, downhill, climbing stairs, and descending stairs. Each activity pattern was tested at a different speed. For personalized model data collection, users wore the data collection device and completed specific walking patterns to collect lower limb movement data for training the Informer personalized model.

[0089] During the experiment, an IMU (Inertial Measurement Unit) and a plantar force-measuring shoe were used to measure the swing angle of the thigh and plantar force information of one side of the human lower limb, respectively. To accurately measure the swing angle of the thigh and plantar force information of one side of the human lower limb during walking, an elastic nylon strap was used to attach the IMU sensor approximately 4 cm above the knee joint on the right thigh in the coronal plane. The sampling frequency was 200 Hz, and the communication rate was 2 MB / s.

[0090] Step S5: Train the Informer personalized model based on user data.

[0091] In this embodiment, firstly, the mean and standard deviation of n_user time sampling points for each user data variable are calculated. Then, the historical data are Z-score standardized, and the standardized data is used as the standardized data X. user X user Let n be an n_use × m matrix. The Z-score standardization process is calculated according to the following formula:

[0092]

[0093] Among them, X m U represents the value of the m-th variable in the standardized data X. m Let m be the m-th variable in the user's historical data U, and mean(U) m std(U) represents the average of the m-th variable in the historical data U. m ) represents the standard deviation of the m-th variable in the historical data U.

[0094] Then, the user's plantar force information is used to mark the heel strike event, and the phase variable P is set... user (0-100%) mapped to the angle range α user (0-2π) is used because daily human movements such as walking, running, climbing stairs, and going down stairs are periodic movements. Therefore, within a single cycle, the phase variable P is used. user This represents the gait progress at the current moment, i.e., P. user The range is 0-100%. However, if the phase variable P is directly... user As the output of the Informer neural network model, this can cause the phase variable to directly enter the next phase cycle before reaching 100% in the current cycle, resulting in discontinuities in the gait phase. Therefore, the phase variable P... user Using angle α in the conversion from rectangular coordinates to polar coordinates user express,

[0095] In this embodiment, the phase variable and angle are first mapped, and then the angle is substituted into the sine and cosine functions and denoted as y. u1 and y u2 This serves as the output of the Informer personalization model.

[0096]

[0097] y u1 =sin(α) user )

[0098] y u2 =cos(α) user )

[0099] In the above embodiment, the window sliding cropping process involves cropping the user's historical data matrix to obtain several training data matrices. First, the cropping width of the desired training data matrix is ​​determined, and this width is smaller than the width of the user's historical data matrix. Then, multiple adjacent columns of data within the cropped user's historical data matrix are selected to obtain the user training matrix. The dataset is then divided into a user training set, a user validation set, and a user test set in a 7:2:1 ratio.

[0100] In this embodiment, when modifying network parameters, only the network parameters of the fully connected layer are modified. That is, the transfer learning method is used to freeze the parameters of the Informer except for the fully connected layer, so that the parameters remain unchanged during training, ensuring that the learning ability of the model is not destroyed. Only the output of the model is changed, so that the model can more accurately predict the user's personalized gait phase, thereby ensuring the real-time performance and compliance in the human-machine coupling process.

[0101] Step S6: Save the training parameters, network structure, and optimizer of the Informer personalized model;

[0102] The Informer personalized model training is complete; its network structure, network parameters, and optimizer are saved.

[0103] Step S7: The design-based acquisition device collects the swing angle of the wearer's right thigh in real time.

[0104] After the Informer personalized model is trained, the saved Informer personalized model can be directly used for gait phase prediction of the human lower limb. Before real-time prediction, it is necessary to collect the swing angle of the user's right thigh in real time. The data collection experiment uses real-time user data and collects data during basic daily activities such as walking, uphill, downhill, upstairs, and downstairs. The requirement is to collect real-time user data and wear the data collection device to complete specific walking patterns, thereby completing the real-time acquisition of lower limb motion data for real-time prediction of the user's gait phase. Let the collected data be represented by R, where R is n. real_time A vector of size ×1.

[0105] During data acquisition, an IMU (Inertial Measurement Unit) was used to collect the real-time swing angle of the user's right thigh. To accurately collect the swing angle of the right thigh during walking, an elastic nylon strap was used to attach the IMU sensor approximately 4 cm above the knee joint on the right thigh in the coronal plane of the user. The sampling frequency was 200 Hz, and the communication rate was 2 MB / s.

[0106] After the Informer personalized model is trained, the saved Informer personalized model can be directly used for gait phase prediction of the human lower limb. Before real-time prediction, it is necessary to collect the swing angle of the right thigh of the human lower limb in real time.

[0107] Step S8: Develop a usage model based on the wearer's real-time data.

[0108] In this embodiment, the user's real-time lower limb thigh swing angle is standardized using the standard deviation and mean of the user's thigh swing angle through Z-score normalization. The normalized data is then used as the user's real-time standardized data R. Z R Z For n real_time A vector of ×1. The Z-score normalization process for real-time user data is calculated according to the following formula:

[0109]

[0110] Among them, R Z U represents user-standardized real-time data. m This represents user data; in this embodiment, m is the first variable, i.e., U. m This indicates the angle of the user's thigh swing. mean(U m ) represents the average angle of the user's thigh swing, std(U m ) represents the standard deviation of the user's thigh swing angle.

[0111] In the above embodiment, the window sliding cropping process involves cropping the user's real-time data matrix to obtain several user input data vectors. The cropping width is consistent with the cropping width of the user's historical data. Then, the user's real-time data within the cropping width is retrieved.

[0112] After data processing is completed, the data is input into the Informer personalized model, and the predicted gait phase representation variables y1 and y2 are obtained. Then, the inverse trigonometric function is used to calculate the gait phase variable P.

[0113] In this embodiment, during actual model deployment, the corresponding equivalent gait phase variable is calculated according to the following formula:

[0114]

[0115] The training termination condition for a deep neural network model is set when the current model achieves the required performance on the validation set or when the number of training epochs reaches the set training limit.

[0116] This invention presents an Informer-based personalized gait phase prediction method for the human lower limbs. It employs an inertial measurement unit (IMU) and a thin-film pressure sensor to collect information on the unilateral thigh swing angle and plantar force of different individuals of varying ages and body types under different movements and walking speeds. Gait is then segmented based on the plantar force information to obtain gait phase variables. An Informer model is trained using the collected data to obtain an Informer pre-model. Next, user data is collected and the Informer pre-model is trained to obtain a personalized Informer model. Finally, when using the personalized Informer model, real-time user data is collected and input into the personalized Informer model to obtain the user's real-time gait phase prediction value. This method differs from traditional gait phase prediction methods by increasing model complexity while avoiding the use of numerous sensors. This reduces hardware usage and increases model accuracy, allowing for more accurate setting of the corresponding controller structure and control parameters based on prior information such as specific human-machine coupling characteristics. This effectively improves the wearing comfort of exoskeleton users during human-machine coupled collaborative walking movements.

[0117] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A personalized prediction method for human lower limb gait phase based on Informer, characterized in that, Includes the following steps: Step S1: Collect human lower limb gait movement information using a data acquisition device and generate a dataset; Step S2: Train the pre-model; Data standardization of human lower limb gait data; Foot force information is used to mark heel strike events, and gait phase is mapped proportionally to the angle range; Window sliding capture processing; A sliding window is used to truncate the standard data matrix to obtain several training data matrices; Design the loss function and learning rate for the Informer model; Training and validation sets are used for training neural network models; The test set evaluates the model's generalization ability; Step S3: Save the trained Informer pre-model; Step S4: Collect individual user lower limb data; Collect information on the swing angle of the right thigh and plantar force of the wearer's lower limb under different speeds and movements; Step S5: Train a personalized model based on the Informer pre-model; Step S6: Save the Informer personalized model; Step S7: Collect the swing angle of the right thigh of the human lower limb using a data acquisition device; Collect real-time user data to obtain the real-time swing angle of the right thigh of the user's lower limb under different speeds and movements; Step S8: Use the model based on the wearer's real-time data; The user's real-time data is standardized to obtain a standardized real-time user data vector. The standardized real-time user data vector is then truncated using a sliding window to obtain several real-time right thigh swing angle vectors of the user's lower limb. Using the Informer personalization model, the user's real-time right thigh swing angle vectors are input into the Informer personalization model to obtain the sine and cosine values ​​of the user's real-time gait phase, and inverse trigonometric function calculations are performed to obtain the user's real-time gait phase prediction value.

2. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1, characterized in that, In step S2, a sliding window is used to truncate the standard data matrix to obtain several training data matrices, specifically as follows: Obtain the truncation width of the desired training data matrix, which is smaller than the width of the standard data matrix; select multiple adjacent columns of data from the standard data matrix with the truncation width to obtain the training data matrix; construct the training set, validation set, and test set; design the Informer model network parameters, loss function, optimizer, and learning rate; input the training data matrix into the Informer model to obtain the Informer pre-model, and calculate the loss value by comparing the true gait phase and estimated gait phase corresponding to the training data matrix; then optimize the Informer model based on the loss value.

3. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1, characterized in that, Step S5 is as follows: Based on the user's plantar force information, the heel strike event is marked and the gait phase at this moment is marked as 0%. The phase is gradually increased linearly to 100% until the next heel strike event occurs and is remarked as 0%. The gait phase is then mapped proportionally to the angle range, and the corresponding sine and cosine values ​​are calculated. The swing angle of the right thigh of the user's lower limb is standardized to obtain a standardized user data matrix; A sliding window is used to truncate the standard user data matrix to obtain several training matrices, including: obtaining the truncation width of the training user data matrix to be generated, where the truncation width is smaller than the width of the standard user data matrix; The training user data matrix is ​​obtained by selecting multiple adjacent columns of data from the standard user data matrix with a truncated width; user training set, user validation set and user test set are constructed; the Informer pre-model is loaded and all layer parameters except for the fully connected layer are frozen; the training user data matrix is ​​input into the Informer pre-model to obtain the Informer personalized model; the true gait phase and estimated gait phase corresponding to the training data matrix are calculated to determine the loss value; and then the Informer model is optimized based on the loss value.

4. The personalized prediction method for human lower limb gait phase based on Informer according to claim 3, characterized in that, In step S5, when constructing the user training set, user validation set, and user test set, the dataset is divided into the user training set, user validation set, and user test set in a ratio of 7:2:

1.

5. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1 or 3, characterized in that, In step S2 and / or step S5, the gait phase is mapped proportionally to the angle range, specifically: the 0%-100% gait phase is mapped proportionally to the 0-2π angle range.

6. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1, characterized in that, In steps S4 and / or S7, an IMU and a plantar force measuring shoe are used to measure the swing angle of the thigh on one side of the lower limb and / or plantar force information, respectively.

7. The personalized prediction method for human lower limb gait phase based on Informer according to claim 6, characterized in that, In steps S4 and / or S7, the IMU sensor is strapped to the right thigh above the knee joint in the coronal plane of the human body using a nylon strap with elasticity; the sampling frequency is 200Hz and the communication rate is 2MB / s.

8. The personalized prediction method for human lower limb gait phase based on Informer according to claim 6, characterized in that, The data acquisition devices in step S1 and / or step S7 are: an inertial measurement unit and a thin-film pressure sensor.

9. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1, characterized in that, Step S3 specifically involves saving the parameters and structure of the Informer pre-model, including the loss function used, the optimizer, the dimensions of the input data, and the dimensions of the output data.

10. The personalized prediction method for human lower limb gait phase based on Informer according to claim 1, characterized in that, Step S6 specifically involves saving the network structure, network parameters, and optimizer of the Informer personalized model.

Images