A control method and system for a lower limb rehabilitation robot driven by gait data.

Abstract

This application relates to the field of rehabilitation robot control technology, and discloses a lower limb rehabilitation robot control method and system driven by gait data. The method includes: acquiring gait angle sequences of healthy individuals and patients; calculating angle differences and determining rehabilitation stage labels based on temporal change characteristics; inputting the rehabilitation stage labels, angle differences, and patient vital signs parameters into a fully connected neural network to output interpolation coefficients, thereby generating a transitional gait sequence between the patient's and healthy gait; converting this sequence into motor control commands for execution, and adjusting the interpolation coefficients according to positional errors. This application improves the personalized adaptation capability and rehabilitation effect of rehabilitation training.

Description

Technical Field

[0001] This application relates to the field of rehabilitation robot control technology, and in particular to a lower limb rehabilitation robot control method and system driven by gait data. Background Technology

[0002] Lower limb dysfunction is a common sequela of neurological diseases such as stroke and spinal cord injury, severely impacting patients' walking ability and quality of life. Traditional rehabilitation training mainly relies on therapists manually assisting patients in gait training, but this approach suffers from problems such as difficulty in quantifying training intensity, high workload for therapists, and inconsistent training effects. With the development of robotics and artificial intelligence, lower limb rehabilitation robots are gradually becoming important assistive tools in the field of neurorehabilitation. Existing lower limb rehabilitation robot systems typically use preset gait trajectories for passive training or impedance control for active assisted training. These methods can help patients recover lower limb motor function to some extent, but they still have shortcomings in generating personalized rehabilitation trajectories and adaptive adjustments during rehabilitation stages.

[0003] The main problems with existing technologies are as follows: First, most rehabilitation robot systems use fixed healthy gait trajectories as training targets, directly driving the patient's lower limbs to move according to the healthy gait. This "one-size-fits-all" training method ignores the patient's current motor ability level, easily leading to excessive training intensity causing patient resistance, or insufficient training intensity resulting in slow rehabilitation progress. Second, existing systems lack intelligent recognition mechanisms for rehabilitation stages and cannot automatically adjust training strategies according to the dynamic changes in the patient's rehabilitation progress. This requires frequent intervention from therapists for manual assessment and parameter adjustment, increasing the complexity of clinical use. In addition, existing control methods often use simple linear interpolation or fixed-ratio trajectory mixing when generating transitional gait trajectories, failing to fully consider the impact of individual patient differences such as height, weight, and limb length on rehabilitation trajectories, resulting in deviations between the generated transitional gait and the patient's actual needs. Summary of the Invention

[0004] This application provides a lower limb rehabilitation robot control method and system driven by gait data, which solves the technical problems of existing lower limb rehabilitation robot systems, such as the inability to automatically identify rehabilitation stages, the lack of a personalized transition gait generation mechanism, and the inability to dynamically adjust training parameters based on execution feedback. This improves the personalized adaptation capability and rehabilitation effect of rehabilitation training.

[0005] In a first aspect, this application provides a lower limb rehabilitation robot control method driven by gait data, the lower limb rehabilitation robot control method driven by gait data comprising: Step S1: Obtain the healthy gait angle sequence and the patient gait angle sequence; Step S2: Calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; Step S3: Input the rehabilitation stage label, the current angle difference, and the patient's vital signs parameters into a fully connected neural network. The fully connected neural network outputs stage interpolation coefficients. The angle difference is weighted using the stage interpolation coefficients to obtain the compensation angle increment. The patient's gait angle sequence and the compensation angle increment are superimposed to generate a transition gait angle sequence. The transition gait angle sequence is between the patient's gait angle sequence and the healthy gait angle sequence. Step S4: Convert the transition gait angle sequence into motor control commands and drive the motor to execute them. Collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficient according to the position error.

[0006] Secondly, this application provides a lower limb rehabilitation robot control system driven by gait data, the lower limb rehabilitation robot control system driven by gait data includes: The acquisition module is used to acquire healthy gait angle sequences and patient gait angle sequences; The determination module is used to calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; The input module is used to input the rehabilitation stage label, the current angle difference, and the patient's vital signs into a fully connected neural network. The fully connected neural network outputs stage interpolation coefficients. The angle difference is weighted using the stage interpolation coefficients to obtain a compensation angle increment. The patient's gait angle sequence and the compensation angle increment are superimposed to generate a transition gait angle sequence, which is between the patient's gait angle sequence and the healthy gait angle sequence. The drive module is used to convert the transition gait angle sequence into motor control commands and drive the motor to execute them, collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficient according to the position error.

[0007] The technical solution provided in this application, by acquiring the angle sequences of healthy gait and patient gait and calculating the angle difference between the two, can intuitively reflect the degree of deviation between the patient's current gait and healthy gait, providing a quantitative data foundation for subsequent rehabilitation stage determination and trajectory generation. Furthermore, by determining the rehabilitation stage label based on the temporal change characteristics of the angle difference, this application breaks through the limitation of traditional rehabilitation robots relying on manual assessment by therapists, achieving automatic intelligent identification of rehabilitation stages. This allows the system to automatically adjust training strategies according to the dynamic evolution of the patient's rehabilitation process. More importantly, this application inputs the rehabilitation stage label, the current angle difference, and the patient's vital signs into a fully connected neural network to output stage interpolation coefficients. This design fully demonstrates the advantages of neural networks in multi-source heterogeneous information fusion and nonlinear mapping. Unlike traditional methods that use fixed interpolation coefficients or simple rules, this application learns the complex correlation between rehabilitation stage, angle deviation, individual physiological characteristics, and optimal interpolation coefficients through a neural network. This allows the generated interpolation coefficients to simultaneously consider the training intensity requirements of the patient's rehabilitation stage and the current gait. The transition gait is designed by considering the difference between the target gait and the patient's biomechanical characteristics. This ensures that the transition gait is neither too close to the healthy gait, causing resistance beyond the patient's capabilities, nor too close to the patient's current gait, lacking sufficient guidance. Subsequently, a weighted calculation of the angle difference using stage interpolation coefficients is performed to obtain the compensation angle increment. The patient's gait angle sequence is then superimposed with the compensation angle increment to generate the transition gait angle sequence. This gradual trajectory generation mechanism conforms to the "gradual" training principle in rehabilitation medicine, ensuring that the transition gait forms a smooth rehabilitation path between the patient's current ability and the healthy gait. Finally, by converting the transition gait angle sequence into motor control commands to drive the motor, and collecting the positional error between the actual motor execution angle and the transition gait angle sequence, the stage interpolation coefficients are adjusted based on the positional error. This application constructs a complete closed-loop feedback optimization mechanism, enabling the system to dynamically correct the interpolation coefficients output by the neural network according to the patient's actual performance ability. This ensures that the generated transition gait always closely matches the patient's true rehabilitation state, avoiding training parameter mismatch problems caused by individual differences or fluctuations in the rehabilitation process. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of an embodiment of the lower limb rehabilitation robot control method driven by gait data in this application. Figure 2 This is a schematic diagram of the fully connected neural network training process in an embodiment of this application; Figure 3 This is a schematic diagram illustrating how the position error changes with the number of training days in an embodiment of this application. Detailed Implementation

[0010] This application provides a lower limb rehabilitation robot control method and system driven by gait data. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0011] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the lower limb rehabilitation robot control method driven by gait data in this application includes: Step S1: Obtain the healthy gait angle sequence and the patient gait angle sequence; Specifically, the three-dimensional spatial coordinates are converted into joint angle data that can be used for motor control. The infrared optical motion capture system uses multiple cameras to spatially locate reflective markers and obtain the continuous position coordinates of the hip, knee, and ankle joints during the gait cycle. The Y-axis and Z-axis coordinates are extracted from the collected coordinate data, where the Y-axis represents the direction of human forward movement and the Z-axis represents the vertical height direction. The hip joint angle is obtained by calculating the linkage angle formed by the knee and ankle joints based on the arctangent function. The knee joint angle is obtained by calculating the linkage angle formed by the hip and knee joints and subtracting it from the aforementioned angle. The angle values ​​at continuous sampling times are arranged in chronological order to form an angle sequence.

[0012] Step S2: Calculate the angle difference between the healthy gait angle sequence and the patient's gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; Specifically, determining the rehabilitation stage label requires extracting temporal change features from the angle difference. The angle difference reflects the degree of deviation between the patient's current gait and healthy gait. By setting a time window, angle differences over several consecutive days are extracted to form a subsequence. After receiving the angle difference subsequence, the Long Short-Term Memory Network controls the acceptance of new information through the input gate, the retention of historical information through the forget gate, and the final output through the output gate. The hidden layer state of the network captures the rate and trend of change of the angle difference with the rehabilitation process. The classification layer maps the hidden layer state into four rehabilitation stage labels: completely passive period, assisted active period, resistance training period, and near-normal period.

[0013] Step S3: Input the rehabilitation stage label, the current angle difference, and the patient's vital signs parameters into the fully connected neural network. The fully connected neural network outputs stage interpolation coefficients. The stage interpolation coefficients are used to perform weighted calculations on the angle difference to obtain the compensation angle increment. The patient's gait angle sequence and the compensation angle increment are superimposed to generate a transition gait angle sequence. The transition gait angle sequence is between the patient's gait angle sequence and the healthy gait angle sequence. Specifically, the fully connected neural network receives rehabilitation stage labels, current angle difference, and vital signs such as patient height and weight as input. The first hidden layer performs a linear transformation on the input features and then processes them through a nonlinear activation function. The second hidden layer further extracts the correlation between features. The output layer generates stage interpolation coefficients. The interpolation coefficients are multiplied element-wise with the angle difference to obtain the compensation angle increment. The patient's gait angle sequence is added to the compensation angle increment to obtain the transition gait angle sequence. The value of the transition gait angle sequence is between the patient's current gait and healthy gait, forming a progressive rehabilitation trajectory.

[0014] Step S4: Convert the transition gait angle sequence into motor control commands and drive the motor to execute them. Collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficients according to the position error.

[0015] Specifically, the transition gait angle sequence is transmitted from the host computer to the intermediate control layer via a data distribution service protocol. The intermediate control layer splits the floating-point angle data into byte format and transmits it to the bottom execution layer via a serial communication protocol. The bottom execution layer reassembles the byte data into floating-point angle instructions to drive the brushless DC motor. The encoder collects the actual rotation angle of the motor shaft in real time and feeds it back to the control layer. The difference between the actual execution angle and the transition gait angle sequence is calculated to obtain the position error. When the statistical characteristics of the position error exceed a preset threshold, the online update of the weights of the neural network output layer is triggered. The updated stage interpolation coefficients make the generated transition gait angle sequence more closely match the patient's current actual execution ability.

[0016] In one specific embodiment, step S1 includes: The three-dimensional position coordinate data of the hip, knee, and ankle joints of the lower limbs of healthy volunteers and patients in the spatial coordinate system were collected by an infrared optical motion capture system to obtain the set of healthy joint coordinates and the set of patient joint coordinates. The Y-axis coordinates of the ankle joint, Z-axis coordinates of the ankle joint, Y-axis coordinates of the knee joint, Z-axis coordinates of the knee joint, Y-axis coordinates of the hip joint, and Z-axis coordinates of the hip joint were extracted from the healthy joint coordinate set and the patient joint coordinate set, respectively. The initial hip joint angle is obtained by using arctangent calculation based on the differences between the Y-axis coordinates of the knee joint and the Y-axis coordinates of the ankle joint, and the differences between the Z-axis coordinates of the knee joint and the Z-axis coordinates of the ankle joint. The intermediate transition angle is obtained by using arctangent calculation based on the differences between the Y-axis coordinates of the hip joint and the Y-axis coordinates of the knee joint, and the differences between the Z-axis coordinates of the hip joint and the Z-axis coordinates of the knee joint. The knee joint angle is obtained by calculating the difference between the initial hip joint angle and the intermediate transition angle. The initial hip joint angle and knee joint angle are arranged in time series to generate healthy gait angle sequences and patient gait angle sequences, respectively.

[0017] Specifically, the Y-axis and Z-axis coordinates extracted from the healthy joint coordinate set and the patient joint coordinate set correspond to the spatial position information of the human gait in the forward direction and vertical height direction, respectively. The Y-axis coordinate reflects the displacement change of the joint along the forward direction during walking, and the Z-axis coordinate reflects the height change of the joint in the swing phase and the support phase. By extracting the coordinate values ​​in these two directions, the motion trajectory of the lower limb joints can be reconstructed in a two-dimensional plane. The difference between the Y-axis coordinate of the knee joint and the Y-axis coordinate of the ankle joint represents the projected length of the lower leg link in the forward direction, and the difference between the Z-axis coordinate of the knee joint and the Z-axis coordinate of the ankle joint represents the projected length of the lower leg link in the vertical direction. The angle between the lower leg link and the vertical axis obtained by the arctangent operation is the initial angle of the hip joint. Similarly, the difference between the Y-axis coordinate of the hip joint and the Y-axis coordinate of the knee joint, and the difference between the Z-axis coordinate of the hip joint and the Z-axis coordinate of the knee joint, are obtained by the arctangent operation and the angle between the thigh link and the vertical axis, which is the intermediate transition angle.

[0018] The difference calculation between the initial hip joint angle and the intermediate transition angle actually calculates the relative angle between the thigh link and the lower leg link. This angle is the bending angle of the knee joint during the gait cycle. The initial hip joint angle and knee joint angle at consecutive sampling moments are arranged in chronological order to form a time series data structure. The angle time series of healthy volunteers constitutes the healthy gait angle sequence, and the angle time series of patients constitutes the patient gait angle sequence. The angle values ​​at corresponding moments in the two sequences can be directly differenced. The length of the sequence depends on the sampling frequency and the number of gait cycles. When the sampling frequency is 120Hz, a single gait cycle contains approximately 120 to 150 sampling points.

[0019] In one specific embodiment, step S2 includes: The angle difference time series is obtained by performing point-by-point difference calculation between the healthy gait angle sequence and the patient gait angle sequence at corresponding time points. A sliding time window is used to extract the angle difference time series to obtain a seven-day continuous angle difference subsequence. The angle difference subsequence is input into the long short-term memory network, and the long short-term memory network extracts the temporal variation features of the angle difference subsequence. Based on the temporal variation characteristics, the rehabilitation stage labels are output through the classification layer. The rehabilitation stage labels include the completely passive stage, the assisted active stage, the resistance training stage, and the near-normal stage.

[0020] Specifically, sliding time window extraction is an operation that extracts angle difference time series in segments according to a fixed time length. The time window length is set to seven days because rehabilitation training is usually assessed in stages on a weekly basis. Each day of rehabilitation training will generate angle difference data for several gait cycles. The angle difference data of seven consecutive days are spliced ​​in chronological order to form an angle difference subsequence. This subsequence contains complete dynamic information on the patient's gait improvement within a week. The step length of the sliding window is set to one day, which means that a seven-day subsequence is extracted every other day. During the window sliding process, the data of the first six days will overlap between adjacent windows. The new window only updates the data of the last day compared to the old window. The trend of the numerical change in the angle difference subsequence reflects the speed and stability of the patient's rehabilitation process.

[0021] The Long Short-Term Memory (LSTM) network extracts temporal features from the angle difference subsequence through three gating structures: input gate, forget gate, and output gate. The input gate determines which information from the current angle difference needs to be memorized into the cell state, the forget gate determines which parts of the angle difference information from previous moments need to be forgotten, and the output gate determines which information from the current cell state needs to be output to the hidden layer state. The hidden layer state transmits and accumulates the angle difference change patterns from previous moments in the time dimension. The temporal change features extracted by the network include statistical characteristics such as the mean change rate of the angle difference, the variance change trend, and the frequency of extreme values. The classification layer maps the hidden layer state into a four-dimensional vector through a fully connected layer. After softmax normalization of the four-dimensional vector, the probability distribution of the four rehabilitation stage labels is obtained. The label with the highest probability is selected as the final output rehabilitation stage label.

[0022] In one specific embodiment, step S3 involves inputting the rehabilitation stage label, the current angle difference, and the patient's vital signs parameters into a fully connected neural network, which then outputs stage interpolation coefficients, including: The rehabilitation stage labels are converted into one-hot encoded vectors. The mean and standard deviation of the current angle difference are used as statistical features of the difference. The patient's height, weight, thigh length, calf length and hip width are used as patient vital signs parameters. The one-hot encoded vector, the difference statistical features, and the patient's vital signs parameters are concatenated to obtain the input feature vector of the fully connected neural network; The input feature vector of the fully connected neural network is fed into a fully connected neural network containing two hidden layers with 64 and 32 neurons respectively. After processing by a non-linear activation function, the interpolation coefficients are generated by the output layer.

[0023] Specifically, one-hot encoding vectors are an encoding method that converts rehabilitation stage labels from discrete categories into numerical vectors. The four rehabilitation stage labels correspond to four-dimensional vectors: the completely passive stage is encoded as [1, 0, 0, 0], the assisted active stage as [0, 1, 0, 0], the resistance training stage as [0, 0, 1, 0], and the near-normal stage as [0, 0, 0, 1]. One-hot encoding ensures the equidistant nature of different rehabilitation stages in the numerical space and avoids introducing sequential bias between stages. The mean of the current angle difference is obtained by calculating the arithmetic mean of the most recent sampling points in the angle difference time series. The standard deviation is obtained by calculating the square root of the sum of squares of the deviations of the angle difference from the mean. The mean reflects the overall deviation of the patient's current gait from the healthy gait, while the standard deviation reflects the stability and fluctuation of the patient's gait. The five vital signs parameters of the patient—height, weight, thigh length, calf length, and hip width—are directly obtained from the patient's physiological measurement data. These parameters determine the biomechanical characteristics and range of motion of the patient's lower limbs.

[0024] The first hidden layer of the fully connected neural network contains 64 neurons. Each neuron receives all dimensions of the input feature vector and performs a linear transformation through the weight matrix. The result of the linear transformation is processed by the ReLU nonlinear activation function and outputs a 64-dimensional feature vector. The second hidden layer contains 32 neurons that receive the 64-dimensional output of the first layer. After linear transformation through the weight matrix and processing by the ReLU activation function, it outputs a 32-dimensional feature vector. The output layer contains 1 neuron that receives the 32-dimensional output of the second layer. After linear transformation, the output is processed by the sigmoid activation function to map the values ​​to between 0 and 1 to obtain the stage interpolation coefficients. The magnitude of these coefficients determines the degree of transition from the patient's current gait to a healthy gait. The closer the coefficient is to 0, the closer the transition gait is to the patient's current gait. The closer the coefficient is to 1, the closer the transition gait is to the healthy gait.

[0025] Figure 2 This is a schematic diagram illustrating the training process of a fully connected neural network in an embodiment of this application. Figure 2As shown in the figure, this figure illustrates the performance of the fully connected neural network during the training process. Subfigure (a) shows the convergence curve of the loss function, with the horizontal axis representing the training epochs and the vertical axis representing the loss value. The black solid line represents the training loss, and the gray solid line represents the validation loss. Both curves show a rapid decrease and gradual convergence with increasing training epochs. The training loss decreases from approximately 1.2 to approximately 0.1, and the validation loss decreases from approximately 1.3 to approximately 0.15, indicating that the network effectively learns the mapping relationship between the rehabilitation stage and the interpolation coefficients. Subfigure (b) shows the improvement curve of classification accuracy, with the horizontal axis representing the training epochs and the vertical axis representing the percentage of accuracy. The black solid line represents the training accuracy, and the gray solid line represents the validation accuracy. Both curves show a rapid increase and then tend to stabilize. Ultimately, both the training accuracy and validation accuracy reach approximately 100%, verifying the network's excellent performance in the rehabilitation stage recognition task. Furthermore, the closeness between the training and validation curves indicates that the network has not overfitted and possesses good generalization ability.

[0026] In one specific embodiment, step S3 involves using stage interpolation coefficients to weight the angle difference to obtain the compensation angle increment, including: The stage interpolation coefficients and angle differences are multiplied element by element to obtain the weighted angle difference. The weighted angle difference is subjected to time smoothing filtering to eliminate high-frequency jitter components in the weighted angle difference, thereby obtaining the compensation angle increment.

[0027] Specifically, the element-wise multiplication operation multiplies the stage interpolation coefficients with the angle difference at each moment in the angle difference time series. Since the stage interpolation coefficients are scalar values ​​while the angle difference is a time series, the multiplication operation is actually multiplying each angle difference in the series by the same interpolation coefficient. The weighted angle difference obtained after multiplication retains the time variation pattern of the original angle difference, but the amplitude is scaled by the interpolation coefficient. The time smoothing filtering process uses a moving average filter to smooth the weighted angle difference series. The filter slides on the series with a fixed-length time window, and the arithmetic mean of multiple sampling points within the window is used as the filtered output at the center of the window. The filtering operation eliminates the high-frequency jitter components in the weighted angle difference caused by sampling noise or unstable patient movement. The filtered series retains the overall trend of angle compensation, and this series is the compensation angle increment.

[0028] In one specific embodiment, step S3 involves superimposing the patient's gait angle sequence with the compensation angle increment to generate a transitional gait angle sequence, including: The hip joint angle in the patient's gait angle sequence is added to the corresponding hip joint compensation component in the compensation angle increment to obtain the transition hip joint angle sequence. The knee joint angle in the patient's gait angle sequence is added to the corresponding knee joint compensation component in the compensation angle increment to obtain the transition knee joint angle sequence. The transitional hip joint angle sequence and the transitional knee joint angle sequence are combined to generate the transitional gait angle sequence.

[0029] Specifically, the patient gait angle sequence includes two independent time series: a hip joint angle subsequence and a knee joint angle subsequence. Similarly, the compensation angle increment also includes two independent time series: a hip joint compensation component and a knee joint compensation component. The hip joint compensation component is obtained by weighting and filtering the difference in hip joint angles between the healthy gait and the patient gait using interpolation coefficients. The knee joint compensation component is obtained by processing the difference in knee joint angles using the same method. The addition operation is performed at the corresponding time, that is, the hip joint angle value at the nth time of the patient gait angle sequence is added to the hip joint compensation component at the nth time of the compensation angle increment to obtain the angle value at the nth time of the transition hip joint angle sequence. The knee joint addition operation follows the same time correspondence. After the addition operation, the value of the transition hip joint angle sequence is between the patient's original hip joint angle and the healthy hip joint angle, and the value of the transition knee joint angle sequence is between the patient's original knee joint angle and the healthy knee joint angle.

[0030] Combining the transitional hip joint angle sequence and the transitional knee joint angle sequence refers to integrating two independent time series according to the joint type in terms of data structure. The combined transitional gait angle sequence contains two angle values ​​at each moment, corresponding to the hip joint and the knee joint respectively. The time length of the sequence is consistent with the original patient gait angle sequence. The combination operation is achieved by array concatenation or matrix construction. The hip joint angle sequence is used as the first dimension data and the knee joint angle sequence is used as the second dimension data to form a two-dimensional array structure. The combination of angle values ​​at each moment in the transitional gait angle sequence describes the complete movement posture of the lower limb at that moment. This sequence can be directly converted into motor control commands to drive the hip and knee joint motors of the rehabilitation robot to move according to the transitional gait trajectory.

[0031] In one specific embodiment, step S4 includes: The transition gait angle sequence is transmitted to the intermediate control layer via the data distribution service protocol, and the intermediate control layer converts the transition gait angle sequence into byte data format; The byte data format is transmitted to the underlying execution layer through a serial communication protocol, and the underlying execution layer reassembles the byte data format into floating-point angle instructions; The brushless DC motor is driven by a magnetic field orientation control algorithm based on floating-point angle commands, and the actual execution angle of the brushless DC motor is collected in real time by an encoder. The positional error between the actual execution angle and the transition gait angle sequence is calculated. When the average positional error over three consecutive days is greater than 1.5 times the historical average error, the weights of the output layer of the fully connected neural network are fine-tuned online to update the stage interpolation coefficients.

[0032] Specifically, the data distribution service protocol is a real-time communication mechanism based on a publish-subscribe model. The host computer, as the publisher, publishes the transition gait angle sequence to a designated topic. The intermediate control layer, as the subscriber, receives the angle sequence data from the topic. The received floating-point angle data needs to be converted into byte data format to adapt to the transmission requirements of the serial communication protocol. In the conversion process, each floating-point number is split into four bytes according to the IEEE 754 standard. Multiple angle values ​​are arranged in the order of hip joint angle in front and knee joint angle behind to form a byte array. The byte data format is transmitted byte by byte to the underlying execution layer through the transmit buffer of the serial communication protocol. The underlying execution layer reads the byte data from the receive buffer and reassembles it into a floating-point value according to the rule of four bytes per group. The reassembly process combines the four bytes into a 32-bit floating-point number through bit operations. The restored floating-point angle command is completely consistent with the original angle value sent by the host computer in terms of value.

[0033] The field-oriented control algorithm converts the three-phase current of the brushless DC motor into two-phase stationary coordinate system current through Clarke transformation, and then into rotating coordinate system current through Park transformation. In the rotating coordinate system, the direct-axis current controls the magnetic flux, and the quadrature-axis current controls the torque. The floating-point angle command is used as the target value of the position loop, and the difference between it and the actual execution angle fed back by the encoder is calculated to obtain the position error. The position error is input to the PI controller to generate the speed command. The difference between the speed command and the actual speed of the motor is input to the current loop PI controller to generate the quadrature-axis current command. The quadrature-axis current command is converted into a three-phase voltage to drive the motor after inverse Park transformation and SVPWM modulation. The average position error for three consecutive days is obtained by calculating the arithmetic mean of the position error sequence recorded after each day's training, and then calculating the arithmetic mean of the three averages. The historical average error is the arithmetic mean of the average position error of all training days from the start of rehabilitation training to the current moment. When the average of three consecutive days exceeds 1.5 times the historical average error, it indicates that the interpolation coefficients are no longer suitable for the patient's rehabilitation progress at the current stage, triggering the gradient descent update of the weights of the output layer of the fully connected neural network. The update process fixes the weights of the first two hidden layers and only adjusts the weights of the output layer to reduce the position error.

[0034] Figure 3 This is a schematic diagram illustrating the change in position error with the number of training days in an embodiment of this application. Figure 3As shown in the figure, this figure illustrates the trend of positional error between the actual motor execution angle and the transitional gait angle sequence as a function of training days. The horizontal axis represents the number of training days, the vertical axis represents the degree of positional error, the solid black line represents the real-time change curve of the positional error, and the horizontal dashed line represents the threshold for triggering online fine-tuning. The curve shows that the positional error was approximately 7 degrees in the early stages of training. As training progressed, the positional error exhibited a significant decreasing trend, rapidly decreasing to approximately 4 degrees in the first 30 days, continuing to decrease to below approximately 2 degrees between days 30 and 60, and stabilizing within 1 degree after 60 days with minimal fluctuations, eventually converging to approximately 0.5-0.6 degrees. The trigger threshold was set at 1.5 times the historical average error, approximately 6.9 degrees. When the average positional error exceeds this threshold for three consecutive days, the system automatically triggers the online fine-tuning mechanism of the fully connected neural network output layer weights, thereby achieving dynamic optimization of the stage interpolation coefficients and ensuring that the transitional gait trajectory continuously matches the patient's current actual execution ability. This result verifies the effectiveness of the adaptive feedback optimization mechanism proposed in this application.

[0035] The control method for a lower limb rehabilitation robot driven by gait data in the embodiments of this application has been described above. The control system for a lower limb rehabilitation robot driven by gait data in the embodiments of this application is described below. One embodiment of the control system for a lower limb rehabilitation robot driven by gait data in the embodiments of this application includes: The acquisition module is used to acquire healthy gait angle sequences and patient gait angle sequences; The determination module is used to calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; The input module is used to input the rehabilitation stage label, the current angle difference, and the patient's vital signs into a fully connected neural network. The fully connected neural network outputs stage interpolation coefficients. The angle difference is weighted using the stage interpolation coefficients to obtain a compensation angle increment. The patient's gait angle sequence and the compensation angle increment are superimposed to generate a transition gait angle sequence, which is between the patient's gait angle sequence and the healthy gait angle sequence. The drive module is used to convert the transition gait angle sequence into motor control commands and drive the motor to execute them, collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficient according to the position error.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for a lower limb rehabilitation robot driven by gait data, characterized in that, The method includes: Step S1: Obtain the healthy gait angle sequence and the patient gait angle sequence; Step S2: Calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; Step S3: Input the rehabilitation stage label, the current angle difference, and the patient's vital signs parameters into a fully connected neural network. The fully connected neural network outputs stage interpolation coefficients. The angle difference is weighted using these stage interpolation coefficients to obtain a compensation angle increment. The patient's gait angle sequence is superimposed with the compensation angle increment to generate a transitional gait angle sequence. This transitional gait angle sequence lies between the patient's gait angle sequence and the healthy gait angle sequence. This includes: converting the rehabilitation stage label into a one-hot encoded vector; using the mean and standard deviation of the current angle difference as statistical features; and using the patient's height, weight, thigh length, calf length, and hip width as the patient's vital signs parameters; concatenating the one-hot encoded vector, the statistical features, and the patient's vital signs parameters to obtain a fully connected neural network input feature vector; and then inputting the fully connected neural network into the feature vector. The eigenvector is input into a fully connected neural network containing two hidden layers with 64 and 32 neurons respectively. After processing by a nonlinear activation function, the output layer generates stage interpolation coefficients. The stage interpolation coefficients are then multiplied element-wise with the angle difference to obtain a weighted angle difference. The weighted angle difference is then subjected to time smoothing filtering to eliminate high-frequency jitter components, resulting in a compensation angle increment. The hip joint angle in the patient's gait angle sequence is added to the corresponding hip joint compensation component in the compensation angle increment to obtain a transitional hip joint angle sequence. The knee joint angle in the patient's gait angle sequence is added to the corresponding knee joint compensation component in the compensation angle increment to obtain a transitional knee joint angle sequence. The transitional hip joint angle sequence and the transitional knee joint angle sequence are combined to generate the transitional gait angle sequence. Step S4: Convert the transition gait angle sequence into motor control commands and drive the motor to execute them. Collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficient according to the position error.

2. The lower limb rehabilitation robot control method driven by gait data according to claim 1, characterized in that, Step S1 includes: The three-dimensional position coordinate data of the hip, knee, and ankle joints of the lower limbs of healthy volunteers and patients in the spatial coordinate system were collected by an infrared optical motion capture system to obtain the set of healthy joint coordinates and the set of patient joint coordinates. The Y-axis coordinates of the ankle joint, Z-axis coordinates of the ankle joint, Y-axis coordinates of the knee joint, Z-axis coordinates of the knee joint, Y-axis coordinates of the hip joint, and Z-axis coordinates of the hip joint are extracted from the healthy joint coordinate set and the patient joint coordinate set, respectively. The initial hip joint angle is obtained by performing an arctangent operation based on the differences between the Y-axis coordinates of the knee joint and the Y-axis coordinates of the ankle joint, and the differences between the Z-axis coordinates of the knee joint and the Z-axis coordinates of the ankle joint. An intermediate transition angle is obtained by performing an arctangent operation based on the differences between the Y-axis coordinates of the hip joint and the Y-axis coordinates of the knee joint, and the differences between the Z-axis coordinates of the hip joint and the Z-axis coordinates of the knee joint. The knee joint angle is obtained by performing a difference operation between the initial hip joint angle and the intermediate transition angle. The initial hip joint angle and the knee joint angle are arranged in time sequence to generate healthy gait angle sequence and patient gait angle sequence, respectively.

3. The lower limb rehabilitation robot control method driven by gait data according to claim 1, characterized in that, Step S2 includes: The healthy gait angle sequence and the patient gait angle sequence are subjected to point-by-point difference calculation at corresponding time points to obtain the angle difference time series; A sliding time window is used to extract the angle difference time series to obtain a seven-day angle difference subsequence. The angle difference subsequence is input into a long short-term memory network, and the long short-term memory network extracts the temporal variation features of the angle difference subsequence. Based on the temporal change characteristics, a rehabilitation stage label is output through a classification layer. The rehabilitation stage label includes the completely passive stage, the assisted active stage, the resistance training stage, and the near-normal stage.

4. The lower limb rehabilitation robot control method driven by gait data according to claim 1, characterized in that, Step S4 includes: The transition gait angle sequence is transmitted to the intermediate control layer via a data distribution service protocol, and the intermediate control layer converts the transition gait angle sequence into byte data format. The byte data format is transmitted to the underlying execution layer via a serial communication protocol, and the underlying execution layer reassembles the byte data format into a floating-point angle instruction. Based on the floating-point angle command, the brushless DC motor is driven to execute through a magnetic field orientation control algorithm, and the actual execution angle of the brushless DC motor is collected in real time by the encoder. Calculate the position error between the actual execution angle and the transition gait angle sequence. When the average position error over three consecutive days is greater than 1.5 times the historical average error, fine-tune the output layer weights of the fully connected neural network online to update the stage interpolation coefficients.

5. A lower limb rehabilitation robot control system driven by gait data, characterized in that, For implementing the lower limb rehabilitation robot control method driven by gait data as described in any one of claims 1-4, the lower limb rehabilitation robot control system driven by gait data includes: The acquisition module is used to acquire healthy gait angle sequences and patient gait angle sequences; The determination module is used to calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference; The input module is used to input the rehabilitation stage label, the current angle difference, and the patient's vital signs parameters into a fully connected neural network. The fully connected neural network outputs stage interpolation coefficients, and the angle difference is weighted using the stage interpolation coefficients to obtain a compensation angle increment. The patient's gait angle sequence is superimposed with the compensation angle increment to generate a transitional gait angle sequence, which is between the patient's gait angle sequence and the healthy gait angle sequence. This includes: converting the rehabilitation stage label into a one-hot encoded vector; using the mean and standard deviation of the current angle difference as difference statistical features; and using the patient's height, weight, thigh length, calf length, and hip width as the patient's vital signs parameters; concatenating the one-hot encoded vector, the difference statistical features, and the patient's vital signs parameters to obtain an input feature vector for the fully connected neural network; and inputting the fully connected neural network... The feature vector is input into a fully connected neural network containing two hidden layers with 64 and 32 neurons respectively. After processing by a nonlinear activation function, the output layer generates stage interpolation coefficients. The stage interpolation coefficients are multiplied element-wise with the angle difference to obtain a weighted angle difference. The weighted angle difference is then subjected to time smoothing filtering to eliminate high-frequency jitter components, resulting in a compensation angle increment. The hip joint angle in the patient's gait angle sequence is added to the corresponding hip joint compensation component in the compensation angle increment to obtain a transitional hip joint angle sequence. The knee joint angle in the patient's gait angle sequence is added to the corresponding knee joint compensation component in the compensation angle increment to obtain a transitional knee joint angle sequence. The transitional hip joint angle sequence and the transitional knee joint angle sequence are combined to generate the transitional gait angle sequence. The drive module is used to convert the transition gait angle sequence into motor control commands and drive the motor to execute them, collect the position error between the actual execution angle of the motor and the transition gait angle sequence, and adjust the stage interpolation coefficient according to the position error.

6. The system according to claim 5, characterized in that, Obtain healthy gait angle sequences and patient gait angle sequences, including: The three-dimensional position coordinate data of the hip, knee, and ankle joints of the lower limbs of healthy volunteers and patients in the spatial coordinate system were collected by an infrared optical motion capture system to obtain the set of healthy joint coordinates and the set of patient joint coordinates. The Y-axis coordinates of the ankle joint, Z-axis coordinates of the ankle joint, Y-axis coordinates of the knee joint, Z-axis coordinates of the knee joint, Y-axis coordinates of the hip joint, and Z-axis coordinates of the hip joint are extracted from the healthy joint coordinate set and the patient joint coordinate set, respectively. The initial hip joint angle is obtained by performing an arctangent operation based on the differences between the Y-axis coordinates of the knee joint and the Y-axis coordinates of the ankle joint, and the differences between the Z-axis coordinates of the knee joint and the Z-axis coordinates of the ankle joint. An intermediate transition angle is obtained by performing an arctangent operation based on the differences between the Y-axis coordinates of the hip joint and the Y-axis coordinates of the knee joint, and the differences between the Z-axis coordinates of the hip joint and the Z-axis coordinates of the knee joint. The knee joint angle is obtained by performing a difference operation between the initial hip joint angle and the intermediate transition angle. The initial hip joint angle and the knee joint angle are arranged in time sequence to generate healthy gait angle sequence and patient gait angle sequence, respectively.

7. The system according to claim 5, characterized in that, Calculate the angle difference between the healthy gait angle sequence and the patient gait angle sequence, and determine the rehabilitation stage label based on the temporal change characteristics of the angle difference, including: The healthy gait angle sequence and the patient gait angle sequence are subjected to point-by-point difference calculation at corresponding time points to obtain the angle difference time series; A sliding time window is used to extract the angle difference time series to obtain a seven-day angle difference subsequence. The angle difference subsequence is input into a long short-term memory network, and the long short-term memory network extracts the temporal variation features of the angle difference subsequence. Based on the temporal change characteristics, a rehabilitation stage label is output through a classification layer. The rehabilitation stage label includes the completely passive stage, the assisted active stage, the resistance training stage, and the near-normal stage.

Images