Novel lower limb rehabilitation training system based on electroencephalogram acquisition device and application

Abstract

A new lower limb rehabilitation training system based on electroencephalogram acquisition device and application, including host computer, embedded controller, portable electroencephalogram acquisition device, pressure sensor connected with host computer respectively, and lower limb rehabilitation training module wirelessly connected with embedded controller, the pressure sensor is also connected with lower limb rehabilitation training module, and power module for supplying power to the whole system is also arranged, the host computer is used for processing the received electroencephalogram signal collected by the portable electroencephalogram acquisition device and the pressure signal collected by the pressure sensor of the lower limb rehabilitation training module respectively, and generating control signal, which is wirelessly transmitted to the lower limb rehabilitation training module through the embedded controller, the embedded controller also feeds back the position information of the lower limb of the user output by the lower limb rehabilitation training module to the host computer. The application can realize man-machine interaction well, can dynamically aggregate electroencephalogram signal channel information and extract time domain features at the same time.

Description

Technical Field

[0001] This invention relates to a lower limb rehabilitation training device. In particular, it relates to a novel lower limb rehabilitation training system and its application based on an electroencephalogram (EEG) acquisition device. Background Technology

[0002] my country currently has a large number of stroke patients, with diverse and widespread causes. After a stroke, damage occurs to the brain regions supplied by the affected blood vessel, leading to weakness in one or both limbs. Patients need active motor rehabilitation training to have any hope of recovering impaired hand motor function. In addition, patients with hand fractures or who have undergone limb amputation surgery also require rehabilitation training to restore hand function.

[0003] Traditional rehabilitation assistive robots stimulate and promote the reconstruction of the nervous system that controls limb movement by having the machine drive the limbs to perform repetitive movements. Rehabilitation robots include two types: upper limb and lower limb rehabilitation robots. Most of them require the assistance of others to use, and patients cannot actively participate in them. They lack accurate feedback mechanisms, which is not conducive to the active participation of patients and accurate feedback on rehabilitation effects.

[0004] Compared to traditional EEG acquisition instruments used in medical institutions, portable EEG acquisition devices offer significantly improved portability while maintaining required accuracy and speed. Their smaller size and lighter weight, coupled with lower operating conditions, greatly enhance portability, leading to their widespread application in brain-computer interface (BCI) devices. Particularly beneficial for patients who cannot be hospitalized long-term, portable EEG acquisition systems provide home-based EEG monitoring, facilitating continuous monitoring of EEG changes for conditions requiring close observation. Furthermore, by processing the EEG signals acquired and extracted by the portable EEG acquisition device along with signals from other sensors in the rehabilitation system, interaction and feedback between the patient and the rehabilitation system can be achieved.

[0005] In recent years, numerous researchers have combined deep learning techniques with EEG signals, achieving results superior to traditional methods and demonstrating strong generalization capabilities across different subject domains. However, EEG signals based on deep learning often inherit the difficulties of designing deep learning model structures and the reliance on lengthy manual adjustments. Therefore, to address these issues, genetic evolution algorithms, reinforcement learning, and Bayesian optimization algorithms have been applied to the network structure search process. These methods have, to some extent, solved the challenges of network structure design, making it possible for models to automatically acquire appropriate network structures based on data information. Currently used lower limb rehabilitation training systems only provide position and pressure signals, without considering the user's active intentions. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a novel lower limb rehabilitation training system and application based on EEG acquisition equipment that can effectively identify the user's lower limb movement intentions, complete interaction with the user, and realize lower limb rehabilitation training functions.

[0007] The technical solution adopted in this invention is: a novel lower limb rehabilitation training system based on an EEG acquisition device, comprising a host computer, an embedded controller, a portable EEG acquisition device, a pressure sensor, and a lower limb rehabilitation training module wirelessly connected to the embedded controller. The pressure sensor is also connected to the lower limb rehabilitation training module. A power supply module is also provided to power the entire system. The host computer processes the EEG signals acquired by the portable EEG acquisition device and the pressure signals acquired by the pressure sensor from the lower limb rehabilitation training module, and generates control signals, which are wirelessly transmitted to the lower limb rehabilitation training module through the embedded controller. The embedded controller also feeds back the user's lower limb position information output by the lower limb rehabilitation training module to the host computer.

[0008] The host computer includes a control module, an internal module, a human-computer interaction interface, and an EEG signal processing module. The internal module includes a read module and a write module. The read module receives angle signals of the user's lower limb joints from the lower limb rehabilitation training module, transmitted by the embedded controller, and pressure signals from the lower limb rehabilitation training module, transmitted by the pressure sensor. The read module reads the angle and status signals of the user's lower limb joints transmitted by the embedded controller at a rate of once every 1 ms, using a 1 ms time base. The write module sends the signals generated by the control module to the embedded controller. The embedded controller generates and outputs enable signals, direction control signals, and pulse control signals to drive the power output module in the lower limb rehabilitation training module. The human-computer interaction interface executes the command lines of the control module to display the user's training status. The user can also set the exercise intensity through the human-computer interaction interface. The EEG signal processing module processes the EEG signals using a proposed adaptive spatiotemporal graph convolutional network.

[0009] The lower limb rehabilitation training module includes, in sequence, a communication module, a voltage conversion module, a power output module, and a lower limb rehabilitation training device. The communication module is wirelessly connected to the embedded controller. The output of the power output module drives the lower limb rehabilitation training device. The output signal of the angle sensor installed within the lower limb rehabilitation training device is transmitted to the embedded controller via the communication module. The pressure sensor installed on the lower limb rehabilitation training device transmits the pressure signal from the device to the communication module within the lower limb rehabilitation module, and the communication module then transmits it to the reading module in the host computer.

[0010] The communication module is equipped with a Wi-Fi wireless communication protocol stack. In operation, it has a load current of 59mA and a maximum Wi-Fi transmission rate of 54Mbps. It receives and parses commands from the embedded controller, and through unpacking, extracts the corresponding action commands from the data packets sent by the embedded controller, sending them to the voltage conversion module. The voltage conversion module uses a PWM control output mode, converting the PWM wave sent by the communication module into different voltages via internal power electronic switching devices, which are then applied to the power output module. The power output module consists of a stepper motor, an encoder, a driver, and a reducer connected to the stepper motor. The driver is powered by the power supply module. The embedded controller's multiple PWM signals control the voltage values ​​of the driver and reducer to change the operating state of the stepper motor. The encoder installed at the end of the stepper motor detects the current position information and feeds it back to the communication module.

[0011] The portable EEG acquisition device includes, in sequence, electrode pads and connecting devices for acquiring the user's EEG signals, a bioelectric signal acquisition module for amplifying and converting EEG signals, an STM32 processor for controlling the operation of the bioelectric signal acquisition module and receiving the EEG signals output by the bioelectric signal acquisition module, and a WIFI wireless transmission circuit for transmitting EEG signals, as well as a power supply circuit connecting the bioelectric signal acquisition module and the STM32 processor respectively; wherein, the electrode pads and connecting devices acquire EEG signals from the user's prefrontal cortex region and are connected to the bioelectric signal acquisition module via a flexible flat cable for the acquisition and transmission of bioelectric signals.

[0012] The bioelectric signal acquisition module is composed of a bioelectric signal acquisition chip, which integrates a high common-mode rejection ratio analog input module for receiving EEG signals acquired by electrode pads and connecting devices, a low-noise programmable gain amplifier for amplifying EEG signals, and a high-resolution synchronous sampling analog-to-digital converter for converting analog signals into digital signals.

[0013] The STM32 processor is used to control the acquisition mode and parameters of the bioelectric signal acquisition module, as well as the transmission mode and transmission speed of the WIFI wireless transmission circuit.

[0014] The input to the WIFI wireless transmission circuit is the EEG brain signal output by the bioelectric signal acquisition module, and the EEG brain signal is transmitted to the host computer.

[0015] The power supply circuit has an input voltage of 5V, is powered by a lithium battery, and outputs a voltage of 3.3V through a voltage conversion module to provide the operating voltage required by the system.

[0016] The embedded controller is an auxiliary controller that receives signals processed by the control module through a write module and generates output enable signals, direction control signals, and pulse control signals. The embedded controller collects position signals fed back by the encoder and pressure signals fed back by the pressure sensor, providing raw input signals for the control module in the host computer to calculate control signals and motion intensity. The embedded controller and the host computer are connected via a serial port for real-time data exchange.

[0017] After receiving the EEG signals acquired by the portable EEG acquisition device, the EEG signal processing module in the host computer performs bandpass filtering on the EEG signals and converts the processed data into a directed graph. Each graph represents a label for different actions. Each node in the graph and its corresponding node value represent the electrode channel and the EEG signal, respectively. The edges between nodes represent the connections between electrode channels. The EEG signal processing module uses an adaptive spatiotemporal graph convolutional network to process the EEG signals.

[0018] The adaptive spatiotemporal graph convolutional network consists of five layers, which are connected in sequence:

[0019] The first convolutional layer is used to receive EEG signals. The kernel size is (1,64) and the stride is (1,1). The dimension of the input data is set to (N,T) and the dimension of the adjacency matrix is ​​(N,N), where N is the number of nodes in the input convolutional layer and T is the number of nodes in the output convolutional layer. The dimension of the feature map obtained by the first convolution operation remains unchanged.

[0020] The first spatiotemporal block layer is used to receive the feature map processed by the first convolutional layer. The convolutional kernel size of the convolutional layer in the spatiotemporal block is (1,8), the stride is (1,1), the average pooling size is (1,2), and the dimension of the feature map obtained by the average pooling size is (N,T / 2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operation, aggregating the information of the center node and the neighboring nodes of the center node, without changing the dimension of the feature map.

[0021] The second spatiotemporal block layer is used to receive the feature map processed by the first spatiotemporal block layer. The convolutional kernel size of the convolutional layer in the second spatiotemporal block layer is (1,8), the stride is (1,1), and the average pooling size is (1,2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operations. The dimension of the feature map after processing by the two spatiotemporal block layers is (N,T / 4).

[0022] The second convolutional layer is used to receive the feature map obtained after processing by two spatiotemporal block layers. The convolutional kernel size is (N, 1) and the stride is (1, 1). The second convolutional layer combines the information of all nodes and reduces the dimension of the feature map to (1, T / 4). After an average pooling operation of size (1, 4), a feature map with a dimension of (1, T / 16) is obtained.

[0023] The fully connected layer is used to receive the feature map processed by the second convolutional layer and realize the classification of EEG signals. The fully connected layer is a linear connected layer with an input-output dimension of (T / 16,2).

[0024] In the aforementioned adaptive spatiotemporal graph convolutional network, all layers except the fully connected layers use the Softmax activation function, while the other layers use the ELU activation function.

[0025] The adaptive spatiotemporal graph convolutional network described above employs a binary classification cross-entropy loss function with a batch size of 20; a robust normalization strategy is used to accelerate the convergence speed of the algorithm; dropout is used to avoid overfitting; and the Adam optimizer with a learning rate of 0.001 is used for network training.

[0026] Both the first and second spatiotemporal block layers are augmented with a trainable adaptive matrix W, which dynamically adjusts the normalized adjacency matrix in the graph convolution operation, resulting in an adaptive graph convolution layer. The formula for the adaptive graph convolution layer is as follows:

[0027]

[0028] Among them, H t It is the output of graph convolution, a diagonal matrix. It is the metric matrix of the graph. Let X be the normalized adjacency matrix, ⊙ be the Hadamard product, and X be the normalized adjacency matrix. t Let θ be the EEG signal at time t. t b is the coefficient of the Chebyshev polynomial at time t. t Let be the bias of the adaptive graph convolution at time t. It is an adaptive adjacency matrix.

[0029] The control module in the host computer uses a fuzzy logic actuator with self-learning capability to generate control signals. The position signal of the lower limb rehabilitation training device collected by the encoder, the EEG signal collected by the portable EEG acquisition device processed by the EEG signal processing module, and the pressure signal of the lower limb rehabilitation training device collected by the pressure sensor are normalized and then input into the fuzzy logic evaluator. The reinforcement signal generated in the fuzzy logic evaluator is input into the fuzzy logic actuator, thereby updating the control parameters of the fuzzy logic actuator. Combined with the current position signal and the desired trajectory signal, a control signal is generated and sent to the embedded controller through the write module in the internal module. The embedded controller ultimately realizes the control of the lower limb rehabilitation training device in the lower limb rehabilitation training module.

[0030] The fuzzy logic actuator formulates a fuzzy rule and passes it to the fuzzy logic evaluator. If the control signal generated by the fuzzy rule can ultimately force the lower limb rehabilitation training device to reach the optimal control target and complete the desired training and treatment task, then the fuzzy logic evaluator rewards the fuzzy logic actuator with a reinforcement signal; otherwise, it punishes the fuzzy logic actuator with a reinforcement signal. Through this iterative "reward-punishment" mechanism, the fuzzy logic actuator is forced to repeatedly modify its fuzzy rules, ultimately enabling the lower limb rehabilitation training device to assist the user in completing the desired optimal trajectory tracking training and treatment task.

[0031] The design method for the fuzzy logic executor and fuzzy logic evaluator is as follows:

[0032] The ideal control rate of the control module is set as follows:

[0033] u F =h+d+K r r

[0034] Where h represents the uncertain part, d represents the disturbance, and K r The positive definite gain matrix r is the filtering error;

[0035] The fuzzy logic actuator is used to adaptively approximate the uncertain part h in the ideal control law, as shown in the formula:

[0036]

[0037] Among them, W F Ψ is the weight vector of the fuzzy logic executor. F Let ε be the fuzzy basis function. F χ represents the reconstruction error, and χ is the input to the fuzzy logic executor.

[0038] The adaptive law for the weights of a fuzzy logic executor is:

[0039]

[0040] Among them, Υ F M is a positive definite invertible constant matrix. R The positive definite inertia matrix is ​​symmetrical at the distal end of the lower limb rehabilitation training equipment. and Here, R represents the estimated weights of the fuzzy logic evaluator, and R is the reward signal. It is the derivative of the weight estimate of the fuzzy logic executor. It is the derivative of the fuzzy basis function;

[0041] The fuzzy logic evaluator is used to approximate the reward signal R, and the formula is:

[0042]

[0043] in, Let e ​​be the transpose of the center and width matrix containing Gaussian functions, and let e be the input vector of the evaluation unit. These are the weight estimates of the fuzzy logic evaluator. Representing the fuzzy basis function, ρ is a solution to the differential equation, which is:

[0044]

[0045] Where, k c β c sgn(·) is the positive definite control gain, which is the robust compensation term used to overcome uncertainties in the system evaluation process. M is the system parameter, μ0 is the robust integral error feedback control term, and ∧2 is the positive definite control gain. The differential equation contains not only the system's tracking error but also the current control strategy, which means that the reward function can evaluate the current system's tracking accuracy and control strategy.

[0046] The application of a novel lower limb rehabilitation training system based on electroencephalogram (EEG) acquisition equipment includes the following steps:

[0047] Step 1) Open the rehabilitation system, initialize each interface of the control module, display the current interface status through the human-computer interaction interface, and wait for command input after initialization is completed;

[0048] Step 2) After the user completes all preparations and selects the corresponding exercise mode and rehabilitation exercise trajectory through the human-computer interaction interface, the control module starts to execute the program. First, it listens to the user's keyboard input to prevent emergency interruption or reselection of exercise mode, and then starts to run the control program.

[0049] Step 3) For each user, the EEG signal collected by the portable EEG acquisition device is first processed by the EEG signal processing module in the host computer. The signal is then combined with the position signal collected by the encoder and the pressure signal collected by the pressure sensor. After data standardization processing, the signal is input into the fuzzy logic evaluator to generate a reinforcement signal. This signal is then input into the fuzzy logic actuator to generate a control signal, thereby realizing the trajectory tracking control of the lower limb rehabilitation training module.

[0050] Step 4) During the cycle of outputting the motion trajectory, the motion intensity generated after processing by the control module is transmitted to the human-computer interaction interface for display. The user can adjust the motion intensity according to the training situation to achieve a better rehabilitation effect.

[0051] This invention relates to a novel lower limb rehabilitation training system and its application based on an EEG acquisition device. The system obtains EEG signals from the human body using a portable EEG acquisition device, and then processes these signals along with position and pressure signals through a fuzzy logic evaluator to generate reinforcement signals. These reinforcement signals are then input into a fuzzy logic actuator to produce control commands, enabling effective human-computer interaction. Furthermore, for the classification of EEG signals related to motor imagery, an Adaptive Spatiotemporal Graph Convolutional Network (ASTGCN) is proposed to process the EEG signals. This network dynamically aggregates EEG signal channel information and simultaneously extracts temporal features, exhibiting high accuracy, high efficiency, and strong robustness to cross-testing and cross-subject variations in EEG signal classification. Attached Figure Description

[0052] Figure 1 This is a block diagram of the novel lower limb rehabilitation training system based on electroencephalogram (EEG) acquisition equipment of the present invention;

[0053] Figure 2 This is a block diagram of the portable EEG acquisition device of the present invention;

[0054] Figure 3 This is a control structure diagram of the present invention. Detailed Implementation

[0055] The novel lower limb rehabilitation training system based on electroencephalogram (EEG) acquisition equipment and its application according to the present invention will be described in detail below with reference to embodiments and accompanying drawings.

[0056] like Figure 1As shown, the novel lower limb rehabilitation training system based on an EEG acquisition device of the present invention includes a host computer 1, an embedded controller 2, a portable EEG acquisition device 3, a pressure sensor 4, and a lower limb rehabilitation training module 5 wirelessly connected to the embedded controller 2, wherein the pressure sensor 4 is also connected to the lower limb rehabilitation training module 5, and a power supply module 6 is provided to supply power to the entire system. The host computer 1 processes the EEG signals acquired by the portable EEG acquisition device 3 and the pressure signals acquired by the pressure sensor 4 from the lower limb rehabilitation training module 5, and generates control signals, which are wirelessly transmitted to the lower limb rehabilitation training module 5 through the embedded controller 2. The embedded controller 2 also feeds back the position information of the user's lower limbs output by the lower limb rehabilitation training module 5 to the host computer 1.

[0057] like Figure 1 As shown, the host computer 1 includes a control module 11, an internal module 12, a human-computer interaction interface 13, and an EEG signal processing module 14. The internal module 12 includes a read module 121 and a write module 122. The read module 121 receives angle signals of the user's lower limb joints output by the lower limb rehabilitation training module 5 from the embedded controller 2, and pressure signals from the lower limb rehabilitation training module 5 output by the pressure sensor 4. The read module 121 reads the angle and status signals of the user's lower limb joints transmitted by the embedded controller 2 at a rate of once every 1 ms. The rate is used as the time base; the write module 122 sends the signal generated by the control module 11 to the embedded controller 2, which generates and outputs enable signal, direction control signal and pulse control signal to drive the power output module in the lower limb rehabilitation training module 5; the human-computer interaction interface 13 executes the command line of the control module 11 to display the user's training status, and the user can also set the exercise intensity through the human-computer interaction interface 13; the EEG signal processing module 14 processes the EEG signal through a proposed adaptive spatiotemporal graph convolutional network (ASTGCN).

[0058] like Figure 1 As shown, the lower limb rehabilitation training module 5 includes, in sequence, a communication module 51, a voltage conversion module 52, a power output module 53, and a lower limb rehabilitation training device 54. The communication module 51 is wirelessly connected to the embedded controller 2. The output of the power output module 53 drives the lower limb rehabilitation training device 54. The output signal of the angle sensor installed in the lower limb rehabilitation training device 54 is transmitted to the embedded controller 2 through the communication module 51. The pressure sensor 4 is installed on the lower limb rehabilitation training device 54 and is used to transmit the pressure signal of the lower limb rehabilitation training device 54 to the communication module 51 in the lower limb rehabilitation module 5, and then the communication module 51 transmits it to the reading module 121 in the host computer 1.

[0059] The communication module 51 uses a chip of model ESP8266, model W600, or model BK7231, equipped with a Wi-Fi wireless communication protocol stack. In operation, the load current is 59mA, and the maximum Wi-Fi transmission rate is 54Mbps. It receives and parses instructions sent by the embedded controller 2, and through unpacking, extracts the corresponding action instructions from the data packets sent by the embedded controller 2, sending them to the voltage conversion module 52. The voltage conversion module 52 uses a PWM control output mode, converting the PWM wave sent by the communication module 51 into different voltages through internal power electronic switching devices, which are then applied to the power output module 53. The power output module 53 consists of a stepper motor, an encoder, a driver, and a reducer connected to the stepper motor. The driver is powered by the power module 6. The multiple PWM signals from the embedded controller 2 control the voltage values ​​of the driver and reducer to change the operating state of the stepper motor. The encoder installed at the end of the stepper motor detects the current position information and feeds it back to the communication module 51.

[0060] like Figure 2 As shown, the portable EEG acquisition device 3 includes, in sequence, electrode pads and a connecting device 31 for acquiring EEG signals from the user, a bioelectric signal acquisition module 32 for amplifying and converting EEG signals, an STM32 processor 33 for controlling the operation of the bioelectric signal acquisition module 32 and receiving the EEG signals output by the bioelectric signal acquisition module 32, and a WIFI wireless transmission circuit 34 for transmitting EEG signals, as well as a power supply circuit 35 connecting the bioelectric signal acquisition module 32 and the STM32 processor 33 respectively; wherein, the electrode pads and the connecting device 31 acquire EEG signals from the user's prefrontal cortex region and are connected to the bioelectric signal acquisition module 32 via a flexible flat cable for the acquisition and transmission of bioelectric signals.

[0061] The bioelectric signal acquisition module 32 is composed of an ADS129x series bioelectric signal acquisition chip. The bioelectric signal acquisition chip integrates a high common-mode rejection ratio analog input module for receiving EEG brain signals acquired by the electrode pads and the connection device 31, a low-noise programmable gain amplifier (PGA) for amplifying EEG brain signals, and a high-resolution synchronous sampling analog-to-digital converter (ADC) for converting analog signals into digital signals.

[0062] The STM32 processor 33 is used to control the acquisition mode and parameters of the bioelectric signal acquisition module 32, and to control the transmission mode and transmission speed of the WIFI wireless transmission circuit 34.

[0063] The input of the WIFI wireless transmission circuit 34 is the EEG brain signal output by the bioelectric signal acquisition module 32, and the EEG brain signal is transmitted to the host computer 1.

[0064] The power supply circuit 35 has an input voltage of 5V, is powered by a lithium battery, and outputs a voltage of 3.3V through a voltage conversion module to provide different operating voltages required by the system.

[0065] The embedded controller 2 is an auxiliary controller. It receives the signals processed by the control module 11 through the write module 122 and generates output enable signals, direction control signals and pulse control signals. The power output module 53 collects the position signals fed back by the encoder and the pressure signals fed back by the pressure sensor 4, providing the original input signals for the control module 11 in the host computer 1 to calculate the motion intensity. The embedded controller 2 is connected to the host computer 1 through a serial port for real-time data exchange.

[0066] like Figure 3 As shown, the control module 11 in the host computer 1 uses a fuzzy logic actuator with self-learning capability to generate control signals. The position signal of the lower limb rehabilitation training device 54 collected by the encoder, the EEG signal collected by the portable EEG acquisition device 3 processed by the EEG signal processing module 14, and the pressure signal of the lower limb rehabilitation training device 54 collected by the pressure sensor 4 are processed by data normalization and then input into the fuzzy logic evaluator. The reinforcement signal generated in the fuzzy logic evaluator is input into the fuzzy logic actuator, thereby updating the control parameters of the fuzzy logic actuator. Combined with the current position signal and the desired trajectory signal, a control signal is generated and sent to the embedded controller 2 through the write module 122 in the internal module 12. The embedded controller 2 ultimately realizes the control of the lower limb rehabilitation training device 54 in the lower limb rehabilitation training module 5.

[0067] The fuzzy logic actuator formulates a fuzzy rule and passes it to the fuzzy logic evaluator. If the control signal generated by the fuzzy rule can ultimately force the lower limb rehabilitation training device 54 to reach the optimal control target and complete the desired training and treatment task, then the fuzzy logic evaluator rewards the fuzzy logic actuator with a reinforcement signal; otherwise, it punishes the fuzzy logic actuator with a reinforcement signal. Through this iterative "reward-punishment" mechanism, the fuzzy logic actuator is forced to repeatedly modify its fuzzy rules, ultimately enabling the lower limb rehabilitation training device 54 to assist the user in completing the desired optimal trajectory tracking training and treatment task.

[0068] The design method for the fuzzy logic executor and fuzzy logic evaluator is as follows:

[0069] The ideal control rate of control module 11 is set as follows:

[0070] u F =h+d+K r r

[0071] Where h represents the uncertain part, d represents the disturbance, and K r The positive definite gain matrix r is the filtering error;

[0072] The fuzzy logic actuator is used to adaptively approximate the uncertain part h in the ideal control law, as shown in the formula:

[0073]

[0074] Among them, W F Ψ is the weight vector of the fuzzy logic executor. F Let ε be the fuzzy basis function. F χ represents the reconstruction error, and χ is the input to the fuzzy logic executor.

[0075] The adaptive law for the weights of a fuzzy logic executor is:

[0076]

[0077] Among them, Υ F M is a positive definite invertible constant matrix. R The positive definite inertia matrix is ​​symmetrical at the distal end of the lower limb rehabilitation training equipment. and Here, R represents the estimated weights of the fuzzy logic evaluator, and R is the reward signal. It is the derivative of the weight estimate of the fuzzy logic executor. It is the derivative of the fuzzy basis function;

[0078] The fuzzy logic evaluator is used to approximate the reward signal R, and the formula is:

[0079]

[0080] in, Let e ​​be the transpose of the center and width matrix containing Gaussian functions, and let e be the input vector of the evaluation unit. These are the weight estimates of the fuzzy logic evaluator. Representing the fuzzy basis function, ρ is a solution to the differential equation, which is:

[0081]

[0082] Where, k c β csgn(·) is the positive definite control gain, which is the robust compensation term used to overcome uncertainties in the system evaluation process. M is the system parameter, μ0 is the robust integral error feedback control term, and ∧2 is the positive definite control gain. The differential equation contains not only the system's tracking error but also the current control strategy, which means that the reward function can evaluate the current system's tracking accuracy and control strategy.

[0083] After receiving the EEG signals acquired by the portable EEG acquisition device, the EEG signal processing module 14 in the host computer 1 performs bandpass filtering on the EEG signals and converts the processed data into a directed graph. Each graph represents a label for different actions. Each node in the graph and its corresponding node value represent the electrode channel and the EEG signal, respectively. The edges between nodes represent the connections between electrode channels. The EEG signal processing module 14 uses an adaptive spatiotemporal graph convolutional network (ASTGCN) to process the EEG signals. ASTGCN consists of multiple convolutional layers and the proposed spatiotemporal block (STB). The STB is composed of a proposed adaptive graph convolutional layer (AGCL) and convolutional layers, which can dynamically aggregate EEG signal channel information and simultaneously extract temporal features.

[0084] The aforementioned Adaptive Spatiotemporal Graph Convolutional Network (ASTGCN) consists of five layers, which are connected in sequence:

[0085] The first convolutional layer is used to receive EEG signals. The kernel size is (1,64) and the stride is (1,1). The dimension of the input data is set to (N,T) and the dimension of the adjacency matrix is ​​(N,N), where N is the number of nodes in the input convolutional layer and T is the number of nodes in the output convolutional layer. The dimension of the feature map obtained by the first convolution operation remains unchanged.

[0086] The first spatiotemporal block layer is used to receive the feature map processed by the first convolutional layer. The convolutional kernel size of the convolutional layer in the spatiotemporal block is (1,8), the stride is (1,1), the average pooling size is (1,2), and the dimension of the feature map obtained by the average pooling size is (N,T / 2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operation, aggregating the information of the center node and the neighboring nodes of the center node, without changing the dimension of the feature map.

[0087] The second spatiotemporal block layer is used to receive the feature map processed by the first spatiotemporal block layer. The convolutional kernel size of the convolutional layer in the second spatiotemporal block layer is (1,8), the stride is (1,1), and the average pooling size is (1,2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operations. The dimension of the feature map after processing by the two spatiotemporal block layers is (N,T / 4).

[0088] The second convolutional layer is used to receive the feature map obtained after processing by two spatiotemporal block layers. The convolutional kernel size is (N, 1) and the stride is (1, 1). The second convolutional layer combines the information of all nodes and reduces the dimension of the feature map to (1, T / 4). After an average pooling operation of size (1, 4), a feature map with a dimension of (1, T / 16) is obtained.

[0089] The fully connected layer is used to receive the feature map processed by the second convolutional layer and realize the classification of EEG signals. The fully connected layer is a linear connected layer with an input-output dimension of (T / 16,2).

[0090] In the aforementioned adaptive spatiotemporal graph convolutional network, except for the fully connected layers which use the Softmax activation function, all other layers use the ELU activation function;

[0091] The adaptive spatiotemporal graph convolutional network described above employs a binary classification cross-entropy loss function with a batch size of 20; a robust normalization strategy is used to accelerate the convergence speed of the algorithm; dropout is used to avoid overfitting; and the Adam optimizer with a learning rate of 0.001 is used for network training.

[0092] Both the first and second spatiotemporal block layers are augmented with a trainable adaptive matrix W, which dynamically adjusts the normalized adjacency matrix in the graph convolution operation, resulting in an adaptive graph convolutional layer (AGCL). The formula for the adaptive graph convolutional layer is as follows:

[0093]

[0094] Among them, H t It is the output of graph convolution, a diagonal matrix. It is the metric matrix of the graph. Let X be the normalized adjacency matrix, ⊙ be the Hadamard product, and X be the normalized adjacency matrix. t Let θ be the EEG signal at time t. t b is the coefficient of the Chebyshev polynomial at time t. t Let be the bias of the adaptive graph convolution at time t. It is an adaptive adjacency matrix.

[0095] The application of the novel lower limb rehabilitation training system based on electroencephalogram (EEG) acquisition equipment of the present invention includes the following steps:

[0096] Step 1) Open the rehabilitation system, initialize each interface of the control module, display the current interface status through the human-computer interaction interface, and wait for command input after initialization is completed;

[0097] Step 2) After the user completes all preparations and selects the corresponding exercise mode and rehabilitation exercise trajectory through the human-computer interaction interface, the control module starts to execute the program. First, it listens to the user's keyboard input to prevent emergency interruption or reselection of exercise mode, and then starts to run the control program.

[0098] Step 3) For each user, EEG signals are first collected using a portable EEG acquisition device. These signals are then combined with position signals collected by the encoder and pressure signals collected by the pressure sensor. After data standardization processing, the data is input into the fuzzy logic evaluator to generate reinforcement signals. These signals are then input into the fuzzy logic actuator to generate control signals, thereby achieving trajectory tracking control of the lower limb rehabilitation training module.

[0099] Step 4) During the cycle of outputting the motion trajectory, the motion intensity generated after processing by the control module is transmitted to the human-computer interaction interface for display. The user can adjust the motion intensity according to the training situation to achieve a better rehabilitation effect.

Claims

1. A novel lower limb rehabilitation training system based on an electroencephalogram (EEG) acquisition device, comprising a host computer (1), an embedded controller (2) connected to the host computer (1), a portable EEG acquisition device (3), a pressure sensor (4), and a lower limb rehabilitation training module (5) wirelessly connected to the embedded controller (2), wherein the pressure sensor (4) is also connected to the lower limb rehabilitation training module (5), and a power supply module (6) is provided to supply power to the entire system, characterized in that, The host computer (1) processes the EEG signals collected by the portable EEG acquisition device (3) and the pressure signals collected by the pressure sensor (4) from the lower limb rehabilitation training module (5), and generates control signals, which are wirelessly transmitted to the lower limb rehabilitation training module (5) through the embedded controller (2). The embedded controller (2) also feeds back the position information of the user's lower limbs output by the lower limb rehabilitation training module (5) to the host computer (1). The host computer (1) includes a control module (11), an internal module (12), a human-computer interaction interface (13), and an EEG signal processing module (14); wherein, The EEG signal processing module (14) uses an adaptive spatiotemporal graph convolutional network to process the EEG signals to classify the EEG signals of motor imagery; the adaptive spatiotemporal convolutional network includes an adaptive graph convolutional layer obtained through an adaptive matrix W; and The control module (11) uses a fuzzy logic actuator with self-learning capability to generate control signals. The position signal of the lower limb rehabilitation training device (54) collected by the encoder, the EEG signal collected by the portable EEG acquisition device (3) processed by the EEG signal processing module (14), and the pressure signal of the lower limb rehabilitation training device (54) collected by the pressure sensor (4) are processed by data normalization and then input into the fuzzy logic evaluator. The adaptive spatiotemporal graph convolutional network consists of five layers, which are connected in sequence: The first convolutional layer is used to receive EEG signals. The kernel size is (1,64) and the stride is (1,1). The input data dimension is set to (N, T) and the adjacency matrix dimension is set to (N, N), where N is the number of nodes in the input convolutional layer and T is the number of nodes in the output convolutional layer. The dimension of the feature map obtained by the first convolution operation remains unchanged. The first spatiotemporal block layer is used to receive the feature map processed by the first convolutional layer. The convolutional kernel size of the convolutional layer in the spatiotemporal block is (1,8), the stride is (1,1), the average pooling size is (1,2), and the dimension of the feature map obtained by the average pooling size is (N, T / 2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operation, aggregating the information of the center node and the neighboring nodes of the center node, without changing the dimension of the feature map. The second spatiotemporal block layer is used to receive the feature map processed by the first spatiotemporal block layer. The convolutional kernel size of the convolutional layer in the second spatiotemporal block layer is (1,8), the stride is (1,1), and the average pooling size is (1,2). The adaptive graph convolutional layer in the spatiotemporal block performs adaptive graph convolution operations. The dimension of the feature map after processing by the two spatiotemporal block layers is (N, T / 4). The second convolutional layer is used to receive the feature map obtained after processing by two spatiotemporal block layers. The convolutional kernel size is (N, 1) and the stride is (1, 1). The second convolutional layer combines the information of all nodes and reduces the dimension of the feature map to (1, T / 4). After an average pooling operation of size (1, 4), a feature map with a dimension of (1, T / 16) is obtained. The fully connected layer is used to receive the feature map processed by the second convolutional layer and realize the classification of EEG signals. The fully connected layer is a linear connected layer with an input-output dimension of (T / 16,2). In the aforementioned adaptive spatiotemporal graph convolutional network, all layers except the fully connected layers use the Softmax activation function, while the other layers use the ELU activation function. The adaptive spatiotemporal graph convolutional network described above uses a binary classification cross-entropy loss function with a batch size of 20; it employs a robust normalization strategy to accelerate the convergence speed of the algorithm; it uses dropout to avoid overfitting; and it uses an Adam optimizer with a learning rate of 0.001 for network training. The first and second spatiotemporal block layers both incorporate a trainable adaptive matrix W, dynamically adjusting the normalized adjacency matrix in the graph convolution operation to obtain the adaptive graph convolution layer. The formula for the adaptive graph convolution layer is: ； in, It is the output of graph convolution, a diagonal matrix. It is the metric matrix of the graph. For the normalized adjacency matrix, For Hadama accumulation, The EEG signal at time t These are the coefficients of the Chebyshev polynomial at time t. Let be the bias of the adaptive graph convolution at time t. An adaptive adjacency matrix; The design method for the fuzzy logic executor and fuzzy logic evaluator is as follows: The ideal control rate of the control module (11) is set as follows: ； in, The uncertain part, For interference, For filtering error Positive definite gain matrix; The fuzzy logic actuator is used to handle the uncertain part of the ideal control law. To achieve adaptive approximation, the formula is: ； in, For the weight vector of the fuzzy logic executor, For fuzzy basis functions, For reconstruction error, For the input of the fuzzy logic executor; The adaptive law for the weights of a fuzzy logic executor is: ； in, It is a positive definite invertible constant matrix. The positive definite inertia matrix is ​​symmetrical at the distal end of the lower limb rehabilitation training equipment. and These are the weight estimates for the fuzzy logic evaluator. As a reward signal, It is the derivative of the weight estimate of the fuzzy logic executor. It is the derivative of the fuzzy basis function; The fuzzy logic evaluator is used to approximate the reward signal. The formula is: ； in, , The transpose of the center and width matrix containing the Gaussian function. It is the input vector of the evaluation unit. These are the weight estimates of the fuzzy logic evaluator. Represents fuzzy basis functions. It is a solution to the differential equation, which is: ； in, , For positive definite control gain, This is a sign function, representing a robust compensation term used to overcome uncertainties during the system evaluation process. For system parameters, For robust integral error feedback control term, Assuming a positive definite control gain, the differential equation contains not only the system's tracking error but also the current control strategy. This means that the reward function can evaluate the current system's tracking accuracy and control strategy.

2. The novel lower limb rehabilitation training system based on EEG acquisition equipment according to claim 1, characterized in that, The internal module (12) includes a read module (121) and a write module (122). The read module (121) is used to receive the angle signal of the user's lower limb joints output by the lower limb rehabilitation training module (5) transmitted by the embedded controller (2), and the pressure signal of the lower limb rehabilitation training module (5) output by the pressure sensor (4). The read module (121) reads the angle and status signal of the user's lower limb joints transmitted by the embedded controller (2) at a rate of once every 1ms, and uses a rate of once every 1ms as the time base. The write module (122) sends the signal generated by the control module (11) to the embedded controller (2). The embedded controller (2) generates and outputs the enable signal, the direction control signal and the pulse control signal to drive the power output module in the lower limb rehabilitation training module (5). The human-machine interface (13) executes the command line of the control module (11) to display the user's training status. The user can also set the exercise intensity through the human-machine interface (13).

3. The novel lower limb rehabilitation training system based on EEG acquisition equipment according to claim 1, characterized in that, The lower limb rehabilitation training module (5) includes, in sequence: a communication module (51), a voltage conversion module (52), a power output module (53), and a lower limb rehabilitation training device (54). The communication module (51) is wirelessly connected to the embedded controller (2). The output of the power output module (53) drives the lower limb rehabilitation training device (54). The output signal of the angle sensor installed in the lower limb rehabilitation training device (54) is transmitted to the embedded controller (2) through the communication module (51). The pressure sensor (4) is installed on the lower limb rehabilitation training device (54) and is used to transmit the pressure signal of the lower limb rehabilitation training device (54) to the communication module (51) in the lower limb rehabilitation module (5), and the communication module (51) transmits it to the reading module (121) in the host computer (1).

4. The novel lower limb rehabilitation training system based on EEG acquisition equipment according to claim 3, characterized in that, The communication module (51) is equipped with a Wi-Fi wireless communication protocol stack. The load current is 59mA and the maximum Wi-Fi transmission rate is 54Mbps. It receives and parses the instructions sent by the embedded controller (2). Through unpacking, it parses the data packets sent by the embedded controller (2) to extract the corresponding action instructions and sends them to the voltage conversion module (52). The voltage conversion module (52) adopts the PWM control output mode and converts the PWM wave sent by the communication module (51) into different voltages through internal power electronic switching devices and loads it onto the power output module (53). The power output module (53) consists of a stepper motor, an encoder, a driver, and a reducer connected to the stepper motor. The driver is powered by the power module (6). The multiple PWM signals of the embedded controller (2) change the operating state of the stepper motor by controlling the voltage values ​​of the driver and the reducer. The encoder installed at the end of the stepper motor detects the current position information and feeds it back to the communication module (51).

5. The novel lower limb rehabilitation training system based on an EEG acquisition device according to claim 1, characterized in that, The portable EEG acquisition device (3) includes, in sequence, electrode pads and connecting devices (31) for acquiring EEG signals from the user, a bioelectric signal acquisition module (32) for amplifying and converting EEG signals, an STM32 processor (33) for controlling the operation of the bioelectric signal acquisition module (32) and receiving EEG signals output by the bioelectric signal acquisition module (32), and a WIFI wireless transmission circuit (34) for transmitting EEG signals, as well as a power supply circuit (35) connecting the bioelectric signal acquisition module (32) and the STM32 processor (33) respectively; wherein, the electrode pads and connecting devices (31) acquire EEG signals from the user's prefrontal cortex region and are connected to the bioelectric signal acquisition module (32) via a flexible flat cable for acquiring and transmitting bioelectric signals.

6. The novel lower limb rehabilitation training system based on an EEG acquisition device according to claim 5, characterized in that, The bioelectric signal acquisition module (32) is composed of a bioelectric signal acquisition chip. The bioelectric signal acquisition chip integrates a high common-mode rejection ratio analog input module for receiving EEG brain signals acquired by the electrode pads and the connecting device (31), a low-noise programmable gain amplifier for amplifying EEG brain signals, and a high-resolution synchronous sampling analog-to-digital converter for converting analog signals into digital signals. The STM32 processor (33) is used to control the acquisition mode and parameters of the bioelectric signal acquisition module (32), and to control the transmission mode and transmission speed of the WIFI wireless transmission circuit (34). The input of the WIFI wireless transmission circuit (34) is the EEG brain signal output by the bioelectric signal acquisition module (32), and the EEG brain signal is transmitted to the host computer (1). The power supply circuit (35) has an input voltage of 5V, is powered by a lithium battery, and outputs a voltage of 3.3V through a voltage conversion module to provide the operating voltage required by the system.

7. The novel lower limb rehabilitation training system based on an EEG acquisition device according to claim 1, characterized in that, The embedded controller (2) is an auxiliary controller. It receives the signal processed by the control module (11) through the write module (122) and generates the output enable signal, direction control signal and pulse control signal. The embedded controller (2) collects the position signal fed back by the encoder and the pressure signal fed back by the pressure sensor (4) to provide the original input signal for the control module (11) in the host computer (1) to calculate the control signal and motion intensity. The embedded controller (2) is connected to the host computer (1) through a serial port to exchange data in real time.

8. The novel lower limb rehabilitation training system based on EEG acquisition equipment according to claim 2, characterized in that, After receiving the EEG signal collected by the portable EEG acquisition device, the EEG signal processing module (14) in the host computer (1) performs bandpass filtering on the EEG signal and converts the processed data into a directed graph. Each graph represents a label of different actions. Each node in the graph and its corresponding node value represent the electrode channel and the EEG signal, respectively. The edges between nodes represent the connection between the electrode channels.

9. The novel lower limb rehabilitation training system based on an EEG acquisition device according to claim 1, characterized in that, The reinforcement signal generated in the fuzzy logic evaluator is input into the fuzzy logic actuator, thereby updating the control parameters of the fuzzy logic actuator and generating a control signal by combining the current position signal and the desired trajectory signal. The signal is then sent to the embedded controller (2) via the write module (122) in the internal module (12). The embedded controller (2) ultimately controls the lower limb rehabilitation training equipment (54) in the lower limb rehabilitation training module (5). The fuzzy logic actuator formulates a fuzzy rule and passes it to the fuzzy logic evaluator. If the control signal generated by the fuzzy rule can eventually force the lower limb rehabilitation training device (54) to reach the optimal control target and complete the expected training and treatment task, then the fuzzy logic evaluator will reward the fuzzy logic actuator with the generated reinforcement signal; otherwise, it will punish the fuzzy logic actuator with the reinforcement signal. Through this repeated "reward-punishment" mechanism, the fuzzy logic actuator is forced to repeatedly modify its fuzzy rule, so that the lower limb rehabilitation training device (54) can assist the user in completing the expected optimal trajectory tracking training and treatment task.

10. An application of the novel lower limb rehabilitation training system based on an EEG acquisition device as described in claim 1, characterized in that, Includes the following steps: Step 1) Open the rehabilitation system, initialize each interface of the control module, display the current interface status through the human-computer interaction interface, and wait for command input after initialization is completed; Step 2) After the user completes all preparations and selects the corresponding exercise mode and rehabilitation exercise trajectory through the human-computer interaction interface, the control module starts to execute the program. First, it listens to the user's keyboard input to prevent emergency interruption or reselection of exercise mode, and then starts to run the control program. Step 3) For each user, the EEG signal collected by the portable EEG acquisition device is first processed by the EEG signal processing module (14) in the host computer (1). The EEG signal is then processed by the position signal collected by the encoder and the pressure signal collected by the pressure sensor. After data standardization processing, the signal is input into the fuzzy logic evaluator to generate a reinforcement signal. The signal is then input into the fuzzy logic actuator to generate a control signal, thereby realizing the trajectory tracking control of the lower limb rehabilitation training module. Step 4) During the cycle of outputting the motion trajectory, the motion intensity generated after processing by the control module is transmitted to the human-computer interaction interface for display. The user can adjust the motion intensity according to the training situation to achieve a better rehabilitation effect.

Images