A method and system for predicting lower limb joint angles

Abstract

This invention discloses a method and system for predicting lower limb joint angles, comprising the following steps: collecting first original joint angle time series from multiple healthy test subjects continuously performing various typical movements; collecting second original joint angle time series from multiple healthy test subjects performing any type of typical movement; standardizing the first original joint angle time series to obtain standard samples, wherein a portion of the standard samples constitutes the training set, and the remaining standard samples serve as the first samples; standardizing the second original joint angle time series to obtain second samples; using both the first and second samples as test samples, and all test samples constitute the test set. This invention can effectively improve the prediction accuracy and real-time computation efficiency of the prediction model, and enhances the cross-scene adaptability of the prediction model while ensuring prediction accuracy, facilitating practical use.

Description

Technical Field

[0001] This invention relates to a prediction method and system, and more particularly to a method and system for predicting lower limb joint angles. Background Technology

[0002] Accurate prediction of lower limb joint angles is a core technological prerequisite for achieving efficient and compliant assisted control of exoskeleton robots, directly determining the coordination and comfort of human-robot interaction. As wearable robot technology develops towards active control, inertial measurement unit (IMU)-based sensing solutions have become the mainstream technology for lower limb joint angle prediction due to their advantages such as ease of wear and strong robustness. Its core principle is to collect inertial data such as acceleration and angular velocity of lower limb movements, combine this data with an algorithm model to establish a mapping relationship between the data and joint angles, and achieve real-time prediction to compensate for sensing delays.

[0003] However, most existing lower limb joint angle prediction methods use deep learning models to capture gait temporal features. In pursuit of accuracy, deep learning models typically design complex network structures, resulting in a large number of parameters and high computational overhead. This leads to delayed calculation results, making it difficult to guarantee real-time performance and inconvenient for real-time operation on resource-constrained exoskeleton hardware. Furthermore, most existing lower limb joint angle prediction methods are used for single movement scenarios, such as walking on flat ground. However, users' actual life scenarios are complex, and single movement scenarios are insufficient to meet users' actual life needs, resulting in low applicability. Summary of the Invention

[0004] Purpose of the invention: The first purpose of this invention is to provide a method for predicting lower limb joint angles that can achieve real-time calculation, high prediction accuracy, and strong applicability.

[0005] The second objective of this invention is to provide a lower limb joint angle prediction system.

[0006] Technical solution: This invention discloses a method for predicting lower limb joint angles, comprising the following steps:

[0007] The time series of the first original joint angles were collected from multiple healthy test subjects during the continuous completion of various typical movements, and the time series of the second original joint angles were collected from multiple healthy test subjects during the completion of any type of typical movement.

[0008] The first original joint angle time series is standardized to obtain standard samples, some of which constitute the training set, and the remaining standard samples are used as the first samples; the second original joint angle time series is standardized to obtain the second samples; the first and second samples are used as test samples, and all test samples constitute the test set.

[0009] A basic model is constructed based on an adaptive radial basis function neural network model and an enhanced proportional differential feedback module. The basic model is trained using standard joint time angle sequences from the training set to obtain a prediction model for predicting lower limb joint angles.

[0010] The test set is input into the prediction model, and the evaluation index used to evaluate the prediction model is calculated based on the prediction results of the prediction model and its corresponding test samples.

[0011] Furthermore, the test subject remained seated at the start of the test, and then the test subject continuously completed static standing, walking on flat ground, running on flat ground, climbing stairs, descending stairs, climbing a slope, and descending a slope.

[0012] Completion of any typical movement means that the test subject randomly and non-continuously completes one or more of the following: maintaining a seated posture, standing statically, walking on flat ground, running on flat ground, climbing stairs, descending stairs, climbing up a slope, and descending a slope.

[0013] The method for acquiring the first and second original joint angles is as follows: inertial sensors for real-time acquisition of joint angles are arranged on both lower limbs of the test subject, with two inertial sensors symmetrically deployed on the outer thighs of both lower limbs, two inertial sensors symmetrically deployed on the middle of the lower legs of both lower limbs, and two symmetrically deployed on the ankle joint surfaces of both lower limbs.

[0014] Furthermore, before standardizing the time series of the first and second original joint angles, the time series of the first and second original joint angles are first preprocessed by filtering, denoising and baseline correction, and the standardization method for the time series of the first and second original joint angles is Z-score standardization.

[0015] Furthermore, the base model consists of an adaptive radial basis function neural network model and an enhanced proportional-differential feedback module. The standard samples of the training set are used as input to the radial basis function neural network model, which outputs the feedforward prediction result. The error between the standard samples of the training set and the output prediction result of the base model is used as input to the enhanced proportional-differential feedback module, which outputs the error correction amount. The sum of the feedforward prediction result and the error correction amount is the final prediction result of the base model.

[0016] Furthermore, the adaptive radial basis function neural network model consists of an input layer, a hidden layer, and an output layer;

[0017] The method of training the basic model using standard samples of the training set is as follows: K-means clustering algorithm is used to cluster the angle time series data points of the standard samples contained in the training set to obtain the initial center vector and Gaussian function width of each radial basis function neuron in the hidden layer; the weight matrix and bias vector of the adaptive radial basis function neural network model are iteratively optimized by combining mini-batch gradient descent and Adam optimization algorithms.

[0018] Furthermore, the steps to obtain the initial center vector and Gaussian function width of each radial basis function neuron in the hidden layer are as follows:

[0019] S1: Set the total number of radial basis function neurons to the number of clusters;

[0020] S2: Calculate the Euclidean distance from each standard sample in the training set to the center of each cluster, and assign the corresponding standard sample to the cluster with the closest Euclidean distance;

[0021] S3: Calculate the mean of the standard samples contained in each cluster, and use the mean as the cluster center of the corresponding cluster;

[0022] S4: Return to step S2 until the difference between the cluster centers calculated in two consecutive iterations of all clusters is less than the preset value, or the number of iterations reaches the predicted number, then proceed to the next step;

[0023] S5: Use the obtained cluster centers as the initial center vectors of the corresponding radial basis function neurons;

[0024] S6: For each cluster, calculate the mean Euclidean distance of all standard samples contained within it to its cluster center, and then... The mean is used as the initial Gaussian function width for the corresponding radial basis function neuron.

[0025] Furthermore, the hidden layer The activation function of the radial basis function neurons is: ,in Refers to the input vector, and The vector expression for the standard sample; , The number of radial basis function neurons in an adaptive radial basis function neural network model; Refers to the first The center vector of each radial basis function neuron; Refers to the first The Gaussian function width of each radial basis function neuron Input vector With the center vector The Euclidean distance;

[0026] The method for iteratively optimizing the weight matrix and bias vector of the adaptive radial basis function neural network model by combining mini-batch gradient descent and Adam optimization algorithms is as follows:

[0027] A subset of time-series data points from the training set are used to construct an optimized dataset.

[0028] Define the loss function for an adaptive radial basis function neural network model. ,and ,in This refers to optimizing the number of standard samples included in the dataset. Optimize the first in the dataset A standard sample, The output of the adaptive radial basis function neural network model is related to... The corresponding feedforward prediction results;

[0029] The weight matrix is ​​calculated based on the mini-batch gradient descent algorithm. and bias vector gradient, , ;in Weight matrix No. Line number The elements of the column, and ; For prediction error The 1 element, and , ; For bias vector The One element;

[0030] The Adam optimization method is used to set the weight matrix. and bias vector The update rules are used to obtain the weight matrices respectively. Adaptive law and bias vector Adaptive law ;in refer to Time weight matrix No. Line number Column elements, refer to Time weight matrix No. Line number Column elements, Learning rate Refers to small, stable positive constants. refer to Moment The first moment estimate, refer to Moment The second moment estimate; and , ,in The exponential decay rate of the first moment. The exponential decay rate of the second-order moment coefficients, and ;

[0031] Set a preset value for the loss function and adopt an adaptive approach. and For the weight matrix respectively and bias vector Perform parameter optimization until the loss function is less than the preset value, then stop iterative optimization to obtain the prediction model.

[0032] Furthermore, enhance the input of the proportional-derivative feedback module. Before inputting into the enhanced proportional-derivative feedback module, for The wave error is obtained by performing moving average filtering. Wave error Differentiation yields ;

[0033] Set the correction rules for the enhanced proportional-derivative feedback module in the base model and prediction model:

[0034] The output of the enhanced proportional-derivative feedback module is ,and ,in This refers to proportional control gain. Refers to the differential control gain. Indicates gain;

[0035] Robust control terms are used to control gain. Adaptive update, and gain The adaptive update rule is ,in Refers to the linear growth rate. Refers to the error threshold, and .

[0036] Furthermore, the evaluation metrics include the root mean square error, which measures the deviation between the predicted result and its corresponding test sample input. Mean absolute error, used to reflect the prediction error of the prediction model. and the coefficient of determination used to measure the accuracy of the prediction model fit. ;

[0037] ,

[0038] ,

[0039] ,

[0040] in This refers to the number of test samples contained in the test set. Refers to the h-th test sample. Pointing to The corresponding prediction results The mean of all test samples.

[0041] Based on the same inventive concept, this invention also discloses a lower limb joint angle prediction system, comprising,

[0042] The raw data acquisition unit collects the first raw joint angle time series during multiple healthy test subjects continuously completing various typical movements, and collects the second raw joint angle time series during multiple healthy test subjects completing any kind of typical movement.

[0043] The raw data processing unit standardizes the first raw joint angle time series to obtain standard samples, some of which constitute the training set and the remaining standard samples serve as the first samples; it also standardizes the second raw joint angle time series to obtain the second samples. Both the first and second samples are test samples, and all the test samples constitute the test set.

[0044] Prediction model training unit: A basic model is constructed based on an adaptive radial basis function neural network model and an enhanced proportional differential feedback module. The basic model is trained using the standard joint time angle sequence of the training set to obtain a prediction model for predicting lower limb joint angles.

[0045] The prediction model evaluation unit inputs the test set into the prediction model and calculates the evaluation index for evaluating the prediction model based on the prediction output of the prediction model and its corresponding test samples.

[0046] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The prediction model of the present invention captures the nonlinear characteristics of joint angles through an adaptive radial basis function neural network model, and combines it with an enhanced proportional differential feedback module to correct errors in real time, which can effectively improve the prediction accuracy of the prediction model; moreover, the prediction model of the present invention has a simple structure and relatively few parameters, which reduces the computational cost and can effectively improve the real-time computational efficiency of the prediction model.

[0047] This invention trains a prediction model using joint angle data from multiple typical movements, which not only enables the prediction model to adapt to gait changes and individual differences, but also improves the cross-scene adaptability of the prediction model while ensuring prediction accuracy, making it convenient for practical use. Attached Figure Description

[0048] Figure 1 This is a schematic diagram showing a test subject continuously performing various typical movements according to the present invention;

[0049] Figure 2 This is a schematic diagram showing the arrangement of inertial sensors when acquiring joint angle time series data according to the present invention.

[0050] Figure 3 This is a schematic diagram illustrating the definition of joint angles in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the predictive model of the present invention;

[0052] Figure 5 This is an experimental diagram of a test subject wearing the exoskeleton device in an embodiment of the present invention;

[0053] Figure 6 This is a joint angle curve and a typical motion classification diagram of the first original joint angle time series in this embodiment of the invention;

[0054] Figure 7 This is a comparison chart of the angle curves between the predicted right lower limb and the actual input test sample in an embodiment of the present invention.

[0055] Figure 8 This is a graph showing the results of an online comparative experiment in an embodiment of the present invention;

[0056] Figure 9 This is a diagram showing the results of an offline comparative experiment in an embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0058] Example 1

[0059] This invention discloses a method for predicting lower limb joint angles, comprising the following steps:

[0060] S1: Collect the time series of the first original joint angles of multiple healthy test subjects during a series of typical movements, and collect the time series of the second original joint angles of multiple healthy test subjects during any type of typical movement.

[0061] The number of first raw joint angle time series collected is consistent with the number of test subjects who continuously completed multiple typical movements; the number of second raw joint angle time series collected is consistent with the number of test subjects who completed any type of typical movement. Preferably, the number of first raw joint angle time series obtained is greater than or equal to 10, and the number of second raw joint angle time series obtained is greater than or equal to 5. Preferably, the healthy test subjects are aged 23 ± 2 years, height 172.8 cm ± 10.8 cm, and weight 64.2 kg ± 16 kg.

[0062] like Figure 1 As shown, the continuous completion of multiple typical movements refers to: at the start of the test, the test subject remains seated (ST), and then the test subject continuously performs static standing (SS), walking on flat ground (LW), running on flat ground (LR), climbing stairs (US), descending stairs (DS), ascending an incline (UR), and descending an incline (DR). Maintaining the seated posture (ST) is as follows... Figure 1 As shown in (a) in the figure, static standing (SS) is as follows Figure 1 As shown in (b) in the diagram, going up the stairs (US) is as follows Figure 1 As shown in (e), the downstairs staircase (DS) is as follows Figure 1 As shown in (f), the upper slope (UR) is as follows Figure 1 As shown in (g), the downslope (DR) is as follows: Figure 1 As shown in (h).

[0063] In practical applications, the test subject can be set to complete various typical movements continuously for a period of 180 seconds. Furthermore, in practical applications, other movement scenarios may be added based on other specific needs.

[0064] Completing any typical movement means that the test subject randomly and non-continuously completes one or more of the following: sitting (ST), standing (SS), walking on flat ground (LW), running on flat ground (LR), climbing stairs (US), descending stairs (DS), ascending an incline (UR), and descending an incline (DR). Collecting non-continuous gait data from real-life situations allows the subsequent prediction model to adapt to various movement scenarios encountered by users, avoiding the limitation of prediction models to continuous, fixed movement sequences, and thus improving the applicability of the prediction model.

[0065] like Figure 2As shown, the method for acquiring joint angle time series is as follows: inertial sensors for real-time acquisition of joint angles are deployed on both lower limbs of the test subject. A total of six inertial sensors are used: two are symmetrically deployed on the outer thighs of both lower limbs, between the hip and knee joints; two are symmetrically deployed on the mid-shin of both lower limbs, between the knee and ankle joints; and two are symmetrically deployed on the ankle surfaces of both lower limbs. Preferably, the sampling frequency of the inertial sensors is set to 240Hz, which is beneficial for accurately capturing the test subject's gait details. In practical applications, the acquired first and second raw joint angle time series can be saved in Excel spreadsheet format. Furthermore, in actual use, the sensors can be mounted on a motion capture suit for easy wear by the test subject.

[0066] The vector expressions for both the first and second original joint angle time series are ,and ,in The angle of the right hip joint as measured at time t. The angle of the left hip joint collected at time t. The angle of the right knee joint collected at time t. The angle of the left knee joint collected at time t. The angle of the right ankle joint as measured at time t. This refers to the angle of the left ankle joint captured at time t. For example... Figure 3 As shown, Indicates the angle of the right hip joint, Indicates the angle of the left hip joint. The angle of the right knee joint Indicates the angle of the left knee joint. Refers to the angle of the right ankle joint, Refers to the angle of the left ankle joint.

[0067] In this embodiment, time series of the first original joint angles were collected from 10 healthy test subjects during continuous completion of various typical movements. The time for the test subjects to continuously complete various typical movements was set to 180 seconds. The joint angle curve of one of the first original joint angle time series obtained in this collection is shown in the figure below. Figure 6 As shown; in this embodiment, the sampling frequency of the inertial sensor is 240Hz, the test subject continuously completes multiple typical movements for 180s, and each first original joint angle time series has more than 40,000 angle time series data points. In this embodiment, a total of 5 healthy test subjects completed any type of typical movement, and the second original joint angle time series were collected.

[0068] S2: Standardize the first original joint angle time series to obtain standard samples, some of which constitute the training set, and the remaining standard samples serve as the first samples; standardize the second original joint angle time series to obtain the second samples; both the first and second samples are test samples, and all test samples constitute the test set.

[0069] In this embodiment, the ratio of the standard sample to the first sample is: The training set contains 5 standard samples and 5 first samples; the test set contains 10 test samples.

[0070] Before standardizing the time series of the first and second original joint angles, the time series of the first and second original joint angles are first preprocessed by filtering, denoising and baseline correction to suppress signal noise in the joint angle time series and eliminate zero drift of inertial sensors.

[0071] Both the first and second original joint angle time series were standardized using Z-score standardization. Z-score standardization can reduce joint angle biases caused by individual first or second original joint angle time series, such as those caused by the tester's leg length or walking speed, thereby improving the convergence and generalization ability of subsequent base model training.

[0072] The formulas for standardizing the time series of the first and second original joint angles using Z-scores are as follows:

[0073] ,

[0074] ,

[0075] ;

[0076] in Hadamard division of a matrix Represents the Hadamard product of the matrix; when standardizing the first original joint angle time series, Refers to the standardized standard sample after the time series of the first original joint angles. This refers to the number of time series of the first original joint angles. This refers to the mean of all first-original joint angle time series. This refers to the standard deviation of all first-origin joint angle time series; when standardizing the second-origin joint angle time series... The second sample refers to the standardized version of the second original joint angle time series. This refers to the number of time series of the second original joint angles. This refers to the mean of all second-original joint angle time series. This refers to the standard deviation of all second original joint angle time series.

[0077] S3: A basic model is constructed based on an adaptive radial basis function neural network model and an enhanced proportional differential feedback module. The basic model is trained using standard joint time angle sequences from the training set to obtain a prediction model for predicting lower limb joint angles.

[0078] like Figure 4 As shown, the basic model consists of an Adaptive Radial Basis Function Neural Network (ARBFNN) model and an Enhanced Proportional Differential Feedback (EPD) module. The standard samples of the training set are used as input to the ARBFNN model, which outputs a feedforward prediction result. The error between the standard samples of the training set and the output prediction result of the basic model is used as input to the EPD module, which outputs an error correction. The sum of the feedforward prediction result and the error correction is the final prediction result of the basic model.

[0079] The adaptive radial basis function neural network model consists of an input layer, a hidden layer, and an output layer; the standard samples in the training set are used as the input vector. Entering the input layer, i.e. The hidden layer includes One radial basis function neuron (RBF); the output layer outputs the predicted joint angle. .

[0080] The activation function of the radial basis function neurons is a Gaussian function, and the activation function of the i-th radial basis function neuron in the hidden layer is... ,in , The center vector of the i-th radial basis function neuron, and ; The width of the Gaussian function of the i-th radial basis function neuron. Input vector With the center vector The Euclidean distance.

[0081] The output layer outputs the predicted joint angles. And joint angle The calculation formula is ,in The weight matrix of the adaptive radial basis function neural network model, and , The number is consistent with that of inertial sensors; The bias vector of the adaptive radial basis function neural network model, and .

[0082] The method of training the base model using standard samples from the training set is as follows:

[0083] S31: The K-means clustering algorithm is used to cluster the angular time series data points of the standard samples contained in the training set to obtain the initial center vector of each radial basis function neuron in the hidden layer. and the width of the Gaussian function .

[0084] S311: Set the total number of radial basis function neurons to the number of clusters, that is, set the number of clusters to... In this embodiment, the number of radial basis function neurons is 50, that is... .

[0085] S312: Calculate the Euclidean distance from each standard sample in the training set to the center of each cluster, and assign the corresponding standard sample to the cluster with the closest Euclidean distance.

[0086] S313: Calculate the mean of the standard samples contained in each cluster, and use the mean as the cluster center of the corresponding cluster.

[0087] S314: Return to step S312 until the difference between the cluster centers calculated in two adjacent iterations for all clusters is less than the preset value, or the number of iterations reaches the predicted number, then proceed to the next step.

[0088] S315: Use the obtained cluster centers as the initial center vectors of the corresponding radial basis function neurons. And the initial center vector It is a six-dimensional vector.

[0089] S316: For each cluster, calculate the mean Euclidean distance from all the standard samples it contains to its cluster center. ,Will The mean is used as the initial Gaussian function width for the corresponding radial basis function neuron. .Will The initial Gaussian function width of the corresponding radial basis function neuron This ensures that the standard samples within a cluster respond effectively. In practical applications, if the number of standard samples in a cluster is less than or equal to a preset threshold, preferably 5, then the average distance between all cluster centers is used as the mean of that cluster. To avoid parameter anomalies. Initial center vector. and the initial width of the Gaussian function The settings ensure that the activation function effectively captures the nonlinear characteristics of the joint angles, providing a basis for the subsequent weight matrix. and bias vector Iterative optimization provides well-fitting initial values.

[0090] S32: Weight matrix of adaptive radial basis function neural network model combining mini-batch gradient descent and Adam optimization algorithms and bias vector Perform iterative optimization.

[0091] S321: Obtain partial time series data points from the training set to construct the optimization dataset. ; ,in Data set The number of time series data points at the angle.

[0092] In this implementation, 48 time-series data points from the training set were acquired. The 48 data points are compatible with a 240Hz sampling frequency, which can fully capture the gait features of a single movement, while also meeting the computational requirements of a lightweight model. In practical applications… The specific quantity can be set according to actual needs.

[0093] S322: Define the loss function for an adaptive radial basis function neural network model. ,and ,in This refers to optimizing the number of standard samples included in the dataset. Optimize the first in the dataset A standard sample, The output of the adaptive radial basis function neural network model is related to... The corresponding feedforward prediction results.

[0094] S323: Calculate the weight matrix based on the mini-batch gradient descent algorithm. and bias vector gradient, , ;in Weight matrix No. Line number The elements of the column, and ; For prediction error The 1 element, and , ; For bias vector The Each element.

[0095] S324: Setting the weight matrix using the Adam optimization method and bias vector The update rules are used to obtain the weight matrices respectively. Adaptive law and bias vector Adaptive law ;in refer to Time weight matrix No. Line number Column elements, refer to Time weight matrix No. Line number Column elements, Learning rate Refers to small, stable positive constants. refer to Moment The first moment estimate, refer to Moment The second moment estimate.

[0096] Calculate the weight matrix The update rules are implemented as follows:

[0097] Calculate time t First moment estimation and second-order moment estimation , , ,in The exponential decay rate of the first moment. The exponential decay rate of the second-order moment coefficients, and , Weight matrix The gradient, and .

[0098] To eliminate the first-moment estimate and second-order moment estimation The bias introduced by an initial value of 0 affects the estimation of the first moment. and second-order moment estimation Bias correction is performed, and the corrected first-order moment estimate is obtained. for Corrected second-order moment estimate for , , .

[0099] Calculate the bias vector The update rules are implemented as follows:

[0100] Calculate time t First moment estimation and second-order moment estimation , , ,in The exponential decay rate of the first moment. The exponential decay rate of the second-order moment coefficients, and , Bias vector The gradient, and .

[0101] To eliminate the first-moment estimate and second-order moment estimation The bias introduced by an initial value of 0 affects the estimation of the first moment. and second-order moment estimation Bias correction is performed, and the corrected first-order moment estimate is obtained. for Corrected second-order moment estimate for , , In this embodiment It is 0.01. for , It is 0.9. It is 0.999.

[0102] S325: Set the preset value of the loss function and adopt an adaptive law. and For the weight matrix respectively and bias vector Perform parameter optimization until the loss function is less than the preset value, then stop iterative optimization to obtain the prediction model.

[0103] Enhance the input of the proportional-derivative feedback module Before inputting into the enhanced proportional-derivative feedback module, for The wave error is obtained by performing moving average filtering. Wave error Differentiation yields .

[0104] , Pointing to The corresponding prediction results.

[0105] ,in The size of the sliding window used in moving average filtering.

[0106] .

[0107] The error correction amount of the enhanced proportional-derivative feedback module output is: ,and ,in This refers to proportional control gain. Refers to the differential control gain. Indicates gain.

[0108] Robust control terms are used to control gain. Adaptive update, and gain The adaptive update rule is ,in Refers to the linear growth rate. Refers to the error threshold, and Gain The adaptive update rule states: if the error norm exceeds the threshold... Gain According to linear growth rate The gain is increased linearly to quickly correct and reduce the error; otherwise... By rate Reduce to avoid over-adjustment. Simultaneously, increase the gain. Constrained Within a certain range, to ensure control stability. In this embodiment... It is 24. It is 0.1. The gain is 0.4. It is constrained to the range [0.01, 0.1].

[0109] Standard samples as input vectors Input adaptive radial basis function neural network model, feedforward prediction results of the adaptive radial basis function neural network model With input vector error Adaptively adjust its parameters; the prediction results of the prediction model. With input vector error The input is processed by the enhanced proportional-derivative feedback module, which corrects the error and outputs the error correction amount. Feedforward prediction results of the adaptive radial basis function neural network model Error correction amount of the enhanced proportional-derivative feedback module The sum is the predicted output of the prediction model. , .

[0110] This invention utilizes the feedforward prediction results of an adaptive radial basis function neural network model. Compared with the true value Error in obtaining the difference Based on this error The weights and biases of the adaptive radial basis function neural network model are adaptively adjusted to regulate the output; simultaneously, based on the error... An enhanced proportional-derivative feedback module is incorporated, and a robust term is added to improve error handling. The combination of rapid correction capability and long-term stability forms a closed-loop control, ultimately yielding a predictive model. Furthermore, the adaptive radial basis function neural network model has a simple structure and relatively few parameters. In practical applications, it can be coupled with a lightweight deployment strategy to adapt to the resource constraints of exoskeleton embedded systems, thus solving the problem of low inference efficiency in existing deep learning models.

[0111] S4: Input the test set into the prediction model, and then use the prediction results from the prediction model... and its corresponding test samples Calculate the evaluation metrics used to evaluate the predictive model.

[0112] The evaluation indicators include those used to measure the prediction results. and its corresponding input test samples Root mean square error of the deviation between Mean absolute error, used to reflect the prediction error of the prediction model. and the coefficient of determination used to measure the accuracy of the prediction model fit. Coefficient of determination The value range is 0-1, and the coefficient of determination A larger value indicates a higher fitting accuracy.

[0113] ,

[0114] ,

[0115] ,

[0116] in This refers to the number of test samples contained in the test set. Refers to the h-th test sample. Pointing to The corresponding prediction results The mean of all test samples.

[0117] In this embodiment, after one test sample is input into the prediction model, the comparison diagram between the prediction result output by the prediction model and the joint angle curve of the actual input test sample is shown below. Figure 7 As shown, and Figure 7 (a) shows a comparison of hip joint angle curves. Figure 7 (b) shows a comparison of knee joint angle curves. Figure 7(c) represents a comparison of ankle joint angle curves. Figure 7 The blue curve represents the true value, i.e., the actual joint angle value of the test sample, and the red curve represents the predicted value, i.e., the predicted joint angle value. Figure 7 As can be clearly seen, the blue curve and the red curve highly overlap, verifying the high prediction accuracy of the prediction model of this invention.

[0118] like Figure 5 As shown, in practical applications, the prediction model is deployed on existing walking-assisted exoskeleton devices, and inertial sensors that collect joint angles in real time are placed on the user's lower limbs. The inertial sensors collect the user's joint angle data in real time, and the joint angle data is input into the prediction model after being filtered, denoised, baseline corrected, and standardized in step S2. The prediction model outputs the predicted angle (prediction result), and the prediction exoskeleton device controls the user's hip, knee, and ankle joints based on the predicted angle to assist the user in walking.

[0119] The prediction model (ARFBNN-EPD) trained in this embodiment was compared with the existing BPNN model, RBFNN model, ARBFNN model and LSTM model in offline and online comparison experiments.

[0120] The online comparison experiment is as follows: All test samples were selected and input into the prediction model (ARFBNN-EPD) and existing technologies such as BPNN, RBFNN, ARBFNN, and LSTM. Prediction results for these five models were obtained at test durations of 100ms, 150ms, and 200ms. Based on the test samples and prediction results, evaluation indices for the hip, knee, and ankle joints were calculated for each of the five models at these three test durations. The comparison chart of the evaluation indices is shown below. Figure 8 As shown. From Figure 8 It can be intuitively seen that the root mean square error of the prediction model of the present invention is lower than that of the prediction results under three different test durations. The mean absolute error of the prediction model of this invention is smaller than that of the other four existing models at three different test durations. The coefficients of determination of the prediction model of this invention are all smaller than those of the other four existing models, and the prediction results under three different test durations are all smaller. All of these results are greater than the other four existing models. This online comparative experiment verifies that the prediction model of this invention has excellent predictive performance.

[0121] The offline comparative experiment is as follows: The prediction model of this invention (ARFBNN-EPD) and the existing BPNN, RBFNN, ARBFNN, and LSTM models were deployed on five existing exoskeleton devices. Five users wore these five exoskeleton devices in turn for testing. The prediction results of the five users wearing these five exoskeleton devices at test durations of 100ms, 150ms, and 200ms were obtained. Based on the test samples and prediction results, the evaluation indicators of the hip, knee, and ankle joints were calculated for the five users wearing these five exoskeleton devices at test durations of 100ms, 150ms, and 200ms. The comparison chart of the obtained evaluation indicators is shown below. Figure 9 As shown, Figure 9 In this context, Sub1 refers to the 1st user, Sub2 refers to the 2nd user, Sub3 refers to the 3rd user, Sub4 refers to the 4th user, and Sub5 refers to the 5th user. Figure 8 and Figure 9 The comparison clearly shows that, compared with the offline comparison results, the prediction performance of the existing BPNN, RBFNN, ARBFNN, and LSTM models decreased at different durations in online comparison experiments. However, the prediction model of the present invention still maintains the best performance, further verifying that the prediction model of the present invention has excellent prediction performance.

[0122] Example 2

[0123] This invention discloses a lower limb joint angle prediction system, including a raw data acquisition unit, a raw data processing unit, a prediction model training unit, and a prediction model evaluation unit.

[0124] The raw data acquisition unit collects the first raw joint angle time series during multiple healthy test subjects continuously completing various typical movements, and collects the second raw joint angle time series during multiple healthy test subjects completing any type of typical movement. The raw data acquisition unit executes step S1 in Example 1.

[0125] The raw data processing unit standardizes the first raw joint angle time series to obtain standard samples, some of which constitute the training set, and the remaining standard samples serve as the first samples. It also standardizes the second raw joint angle time series to obtain the second samples. Both the first and second samples are test samples, and all test samples constitute the test set. The raw data processing unit then executes step S2 in Example 1.

[0126] The prediction model training unit constructs a basic model based on an adaptive radial basis function neural network model and an enhanced proportional-differential feedback module. It trains the basic model using standard joint time-angle sequences from the training set to obtain a prediction model for predicting lower limb joint angles. The prediction model training unit executes step S3 in Example 1.

[0127] The prediction model evaluation unit inputs the test set into the prediction model and calculates the evaluation index for evaluating the prediction model based on the prediction output of the prediction model and its corresponding test samples. The prediction model evaluation unit also executes step S4 in Example 1.

Claims

1. A method for predicting lower limb joint angles, characterized in that: Includes the following steps, The time series of the first original joint angles were collected from multiple healthy test subjects during the continuous completion of various typical movements, and the time series of the second original joint angles were collected from multiple healthy test subjects during the completion of any type of typical movement. The first original joint angle time series is standardized to obtain standard samples, some of which constitute the training set, and the remaining standard samples are used as the first samples; the second original joint angle time series is standardized to obtain the second samples; both the first and second samples are used as test samples, and all test samples constitute the test set. A basic model is constructed based on an adaptive radial basis function neural network model and an enhanced proportional differential feedback module. The basic model is trained using standard joint time angle sequences from the training set to obtain a prediction model for predicting lower limb joint angles. The basic model consists of an adaptive radial basis function neural network model and an enhanced proportional-differential feedback module. The standard samples of the training set are used as input to the radial basis function neural network model, which outputs a feedforward prediction result. The error between the standard samples of the training set and the output prediction result of the basic model is used as input to the enhanced proportional-differential feedback module, which outputs an error correction. The sum of the feedforward prediction result and the error correction is the final prediction result output by the basic model. The adaptive radial basis function neural network model consists of an input layer, a hidden layer, and an output layer. The method of training the basic model using standard samples from the training set is as follows: K-means clustering algorithm is used to cluster the angle time series data points of the standard samples in the training set to obtain the initial center vector and Gaussian function width of each radial basis function neuron in the hidden layer; the weight matrix and bias vector of the adaptive radial basis function neural network model are iteratively optimized by combining mini-batch gradient descent and Adam optimization algorithms. The test set is input into the prediction model, and the evaluation index used to evaluate the prediction model is calculated based on the prediction results of the prediction model and its corresponding test samples.

2. The method for predicting lower limb joint angles according to claim 1, characterized in that: The test involves continuously performing multiple typical movements: at the start of the test, the test subject remains seated, and then continuously performs static standing, walking on flat ground, running on flat ground, climbing stairs, descending stairs, climbing a slope, and descending a slope. Completion of any typical movement means that the test subject randomly and non-continuously completes one or more of the following: maintaining a seated posture, standing statically, walking on flat ground, running on flat ground, climbing stairs, descending stairs, climbing up a slope, and descending a slope. The method for acquiring the first and second original joint angles is as follows: inertial sensors for real-time acquisition of joint angles are arranged on both lower limbs of the test subject, with two inertial sensors symmetrically deployed on the outer thighs of both lower limbs, two inertial sensors symmetrically deployed on the middle of the lower legs of both lower limbs, and two symmetrically deployed on the ankle joint surfaces of both lower limbs.

3. The method for predicting lower limb joint angles according to claim 1, characterized in that: Before standardizing the time series of the first and second original joint angles, filtering, denoising, and baseline correction are performed on the time series of the first and second original joint angles. The standardization method for the time series of the first and second original joint angles is Z-score standardization.

4. The method for predicting lower limb joint angles according to claim 1, characterized in that: The steps to obtain the initial center vector and Gaussian function width of each radial basis function neuron in the hidden layer are as follows: S1: Set the total number of radial basis function neurons to the number of clusters; S2: Calculate the Euclidean distance from each standard sample in the training set to the center of each cluster, and assign the corresponding standard sample to the cluster with the closest Euclidean distance; S3: Calculate the mean of the standard samples contained in each cluster, and use the mean as the cluster center of the corresponding cluster; S4: Return to step S2 until the difference between the cluster centers calculated in two consecutive iterations of all clusters is less than the preset value, or the number of iterations reaches the predicted number, then proceed to the next step; S5: Use the obtained cluster centers as the initial center vectors of the corresponding radial basis function neurons; S6: For each cluster, calculate the mean Euclidean distance of all standard samples contained within it to its cluster center, and then... The mean is used as the initial Gaussian function width for the corresponding radial basis function neuron.

5. The method for predicting lower limb joint angles according to claim 4, characterized in that: Hidden layer The activation function of the radial basis function neurons is: ,in Refers to the input vector, and The vector expression for the standard sample; , The number of radial basis function neurons in an adaptive radial basis function neural network model; Refers to the first The center vector of each radial basis function neuron; Refers to the first The Gaussian function width of each radial basis function neuron Input vector With the center vector The Euclidean distance; The method for iteratively optimizing the weight matrix and bias vector of the adaptive radial basis function neural network model by combining mini-batch gradient descent and Adam optimization algorithms is as follows: A subset of time-series data points from the training set are used to construct an optimized dataset. Define the loss function for an adaptive radial basis function neural network model. ,and ,in This refers to optimizing the number of standard samples included in the dataset. Optimize the first in the dataset A standard sample, The output of the adaptive radial basis function neural network model is related to... The corresponding feedforward prediction results; The weight matrix is ​​calculated based on the mini-batch gradient descent algorithm. and bias vector gradient, , ;in Weight matrix No. Line number The elements of the column, and ; For prediction error The 1 element, and , ; For bias vector The One element; The Adam optimization method is used to set the weight matrix. and bias vector The update rules are used to obtain the weight matrices respectively. Adaptive law and bias vector Adaptive law ;in refer to Time weight matrix No. Line number Column elements, refer to Time weight matrix No. Line number Column elements, The learning rate. Small, stable positive constants. refer to Moment The first moment estimate, refer to Moment The second moment estimate; and , ,in The exponential decay rate of the first moment. The exponential decay rate of the second-order moment coefficients, and ; Set a preset value for the loss function and adopt an adaptive approach. and For the weight matrix respectively and bias vector Perform parameter optimization until the loss function is less than the preset value, then stop iterative optimization to obtain the prediction model.

6. The method for predicting lower limb joint angles according to claim 1, characterized in that: Enhance the input of the proportional-derivative feedback module Before inputting into the enhanced proportional-derivative feedback module, for The wave error is obtained by performing moving average filtering. Wave error Differentiation yields ; Set the correction rules for the enhanced proportional-derivative feedback module in the base model and prediction model: The output of the enhanced proportional-derivative feedback module is ,and ,in This refers to proportional control gain. Refers to the differential control gain. Gain; Robust control terms are used to control gain. Adaptive update, and gain The adaptive update rule is ,in Refers to the linear growth rate. Refers to the error threshold, and .

7. The method for predicting lower limb joint angles according to claim 1, characterized in that: The evaluation metrics include the root mean square error, which measures the deviation between the predicted result and its corresponding input test sample. Mean absolute error, used to reflect the prediction error of the prediction model. and the coefficient of determination used to measure the accuracy of the prediction model fit. ; ， ， ， in This refers to the number of test samples contained in the test set. Refers to the h-th test sample. Pointing to The corresponding prediction results The mean of all test samples.

8. A lower limb joint angle prediction system, characterized in that: The method for predicting lower limb joint angles according to any one of claims 1-7 includes, The raw data acquisition unit collects the first raw joint angle time series during multiple healthy test subjects continuously completing various typical movements, and collects the second raw joint angle time series during multiple healthy test subjects completing any kind of typical movement. The raw data processing unit standardizes the first raw joint angle time series to obtain standard samples, some of which constitute the training set and the remaining standard samples serve as the first samples; it also standardizes the second raw joint angle time series to obtain the second samples. Both the first and second samples are test samples, and all the test samples constitute the test set. Prediction model training unit: A basic model is constructed based on an adaptive radial basis function neural network model and an enhanced proportional differential feedback module. The basic model is trained using the standard joint time angle sequence of the training set to obtain a prediction model for predicting lower limb joint angles. The prediction model evaluation unit inputs the test set into the prediction model and calculates the evaluation index for evaluating the prediction model based on the prediction output of the prediction model and its corresponding test samples.

Images