Exoskeleton closed-loop step control system based on electromyographic signal prediction

Abstract

This application discloses a closed-loop stepping control system for an exoskeleton based on electromyography (EMG) signal prediction. A multimodal sensing module collects real-time shoulder and elbow joint angles and EMG / force signals from the user. A physical information neural network inference engine, deployed in the microcontroller's on-chip memory, receives state vectors constructed from sensor signals and predicts in real-time the gravity and friction compensation torques required for each joint. The compliant assist control unit fuses the predicted compensation torque with the processed user-intended torque to generate a total driving torque command. The motor drive module then drives the joint motors accordingly, forming a high-response closed-loop control. Accurate and real-time fitting of the dynamic model is achieved through embedded lightweight PINN, effectively solving the problems of complex modeling, lag, and dragging sensation in traditional exoskeleton systems. This results in a low-latency, high-transparency compliant assist experience on a single chip.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an exoskeleton closed-loop stepping control system based on electromyographic signal prediction. Background Technology

[0002] Existing upper limb assistive exoskeleton systems mostly employ traditional PID control or computational torque control based on simplified physical models. However, these systems are inherently highly nonlinear multi-joint robotic arm systems, with complex frictional forces, Coriolis forces, and time-varying load characteristics in their dynamics. Traditional mathematical modeling methods struggle to accurately describe these nonlinear coupling factors, leading to significant errors in the established dynamic models. Consequently, insufficient precision often occurs in actual control, causing users to experience noticeable mechanical resistance or dragging during operation, hindering a truly zero-gravity transparent interactive experience, and impacting wearing comfort and the naturalness of operation.

[0003] To improve the accuracy of dynamic modeling, some advanced control schemes attempt to introduce data-driven methods such as neural networks to fit complex dynamics. However, these methods typically rely on high-performance industrial PCs or GPU accelerators for model training and inference, resulting in bulky systems with high power consumption, making them difficult to integrate into portable wearable devices. On the other hand, existing embedded solutions are limited by the limited computing power and storage resources of microcontrollers, often only able to run extremely simple control algorithms and unable to deploy high-precision, high-complexity deep learning models, leading to long system response times and failing to meet the low-latency control requirements for real-time human motion tracking.

[0004] Furthermore, purely data-driven neural network training methods require the collection of massive amounts of actual motion data, resulting in high data acquisition costs and long cycles. Simultaneously, these methods lack the guidance of physical laws, making them prone to overfitting during training. This leads to poor model generalization ability in unseen motion patterns or among different users, resulting in insufficient control robustness. Therefore, current technology urgently needs a lightweight intelligent predictive control scheme that can effectively integrate prior knowledge of physical dynamics and can be implemented on a low-power monolithic microcontroller. This would improve the compliance, accuracy, and physical interpretability of exoskeleton control while ensuring system real-time performance and deployability, thereby promoting the practical application of low-cost, highly portable exoskeletons. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide a closed-loop stepping control system for exoskeleton based on electromyography signal prediction. By deploying a lightweight physical information neural network that integrates prior knowledge of rigid body dynamics on a high-performance microcontroller, the system can predict and compensate for the gravity and friction torque of the exoskeleton in real time. This solves the technical problems of complex dynamic modeling, slow response, and reliance on high-performance computing platforms in traditional upper limb exoskeleton control, and achieves low-cost, high-real-time zero-gravity compliant assist control.

[0006] The objective of this application is achieved through the following technical solution:

[0007] In a first aspect, this application proposes an exoskeleton closed-loop stepping control system based on electromyographic signal prediction, comprising:

[0008] Microcontroller unit;

[0009] The multimodal sensing module, connected to the microcontroller unit, is used to collect the user's shoulder and elbow joint angle signals as well as human-computer interaction force signals.

[0010] The physical information neural network inference engine, deployed in the microcontroller unit, is used to receive the instantaneous state vector constructed from joint angle signals and interaction force signals, and output the predicted compensation torque corresponding to each joint under the current posture. The predicted compensation torque includes gravity compensation torque and friction torque.

[0011] The compliant power assist control unit is connected to the physical information neural network inference engine to fuse the predicted compensation torque with the user intention torque obtained based on the interactive force signal to generate the total driving torque command.

[0012] The motor drive module, connected to the compliant assist control unit, is used to drive the joint motors of the exoskeleton according to the total driving torque command.

[0013] In one possible implementation, the microcontroller unit establishes full-duplex asynchronous communication links with the left and right exoskeleton execution modules via the USART2 and USART3 interfaces, respectively.

[0014] The USART2 interface connects to the left exoskeleton execution module and is used to transmit multi-dimensional state vectors, electromyographic data and torque commands for the left shoulder and left elbow joints.

[0015] The USART3 interface connects to the right exoskeleton execution module and is used to transmit symmetrical data structures of the right shoulder and right elbow joints.

[0016] Both communication links support DMA transfer mode.

[0017] In one possible implementation, the multimodal sensing module includes:

[0018] Angle sensors, installed at the shoulder and elbow joints of the exoskeleton, are used to collect absolute angle information of the joints.

[0019] Force sensors or electromyography (EMG) sensors are installed on the human-machine interface to collect active interactive forces or EMG signals applied by the user.

[0020] The microcontroller unit combines the collected shoulder angle, elbow angle, and corresponding interaction force signals to construct a multidimensional instantaneous state vector;

[0021] The electromyography (EMG) sensor circuit includes:

[0022] The preamplifier unit uses an AD8226 chip to form a differential amplifier circuit, with an external gain setting resistor.

[0023] The signal conditioning link uses a TL084 operational amplifier to construct a bandpass filter unit, a precision rectifier unit, and an envelope detector unit to convert bidirectional AC electromyography signals into unidirectional DC envelope signals.

[0024] The microcontroller unit synchronously acquires conditioned analog signals from the electromyography sensor.

[0025] In one possible implementation, the system further includes a multimodal data storage array, the storage array comprising:

[0026] SDRAM memory, connected to the microcontroller unit via the FMC bus, provides off-chip RAM space for use as a dynamic activation buffer and sliding window data cache during neural network inference.

[0027] The QSPI Flash memory, connected to the microcontroller unit, provides static storage space for storing the physical information quantized by INT8, the model weights of the neural network inference engine, and the system firmware.

[0028] In one possible implementation, the physical information neural network inference engine employs a fully connected neural network architecture, including:

[0029] The input layer contains 6 nodes, corresponding to a multi-dimensional instantaneous state vector composed of shoulder angle, elbow angle, angular velocity, angular acceleration and corresponding interaction force signals;

[0030] At least one hidden layer containing 64 neurons, using the ReLU activation function;

[0031] The output layer contains two nodes, which output the predicted compensation torques corresponding to the shoulder joint and elbow joint, respectively.

[0032] In one possible implementation, the physical information neural network in the physical information neural network inference engine is trained using a physical information-driven strategy, including:

[0033] Construct the rigid body dynamics equations of the robotic arm that include gravity terms, Coriolis force terms, centrifugal force terms, and friction force terms;

[0034] By using dynamic equations to perform random sampling in the joint motion space, a physical simulation dataset containing state inputs and theoretical torque labels is generated.

[0035] Supervised learning of neural networks is performed using physical simulation datasets to make the network approximate a real nonlinear dynamic model.

[0036] In one possible implementation, the compliant power assist control unit is used to execute torque fusion logic, which includes:

[0037] Dead zone filtering is performed on the collected raw interaction force signals;

[0038] The filtered interactive force signal is multiplied by a preset boost gain coefficient to obtain the user intention torque;

[0039] Multiply the predicted compensation torque output by the physical information neural network inference engine by the gravity compensation coefficient;

[0040] The processed user-intended torque is added to the predicted compensation torque multiplied by a coefficient to obtain the total driving torque command.

[0041] In one possible implementation, the system further includes a power management module, which includes:

[0042] A dual-channel BUCK circuit based on the TMI7003B chip generates the core voltage and IO voltage respectively.

[0043] The core voltage is filtered by inductors and capacitors and then supplied to the core and digital circuits of the microcontroller unit. The IO voltage is isolated by ferrite beads and then divided into digital power supply and analog power supply. The analog power supply powers the ADC module.

[0044] In one possible implementation, the system also includes a human-computer interaction interface and a debugging interface:

[0045] The human-machine interface includes a parallel LCD interface and a resistive touch screen interface, which transmits video data through data lines and synchronization signals to display joint angles, predict compensation torques and system status in real time.

[0046] The debugging interface includes a one-button download debugging circuit, which works in conjunction with the reset circuit and the BOOT circuit. The reset signal and the BOOT0 pin level are controlled by a single button operation, and it supports USB DFU or serial ISP program download.

[0047] In one possible implementation, the drive module employs field-oriented control technology, using a current loop as the inner loop, to directly respond to torque commands issued by the microcontroller unit.

[0048] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0049] This application discloses a closed-loop stepping control system for an exoskeleton based on electromyography (EMG) signal prediction, belonging to the field of artificial intelligence technology. A multimodal perception module collects real-time shoulder and elbow joint angles and EMG / force signals from the user; a physical information neural network inference engine, deployed in the microcontroller's on-chip memory, receives state vectors constructed from sensor signals and predicts in real-time the gravity and friction compensation torques required for each joint; the compliant assist control unit fuses the predicted compensation torque with the processed user intention torque to generate a total driving torque command, which the motor drive module uses to drive the joint motors, forming a high-response closed-loop control. Accurate and real-time fitting of the dynamic model is achieved through embedded lightweight PINN, effectively solving the problems of complex modeling, lag response, and dragging sensation in traditional exoskeleton systems, achieving a low-latency, high-transparency compliant assist experience on a single chip. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A schematic diagram of the structure of an exoskeleton closed-loop stepping control system based on electromyographic signal prediction, as proposed in an embodiment of this application, is shown.

[0052] Figure 2 The diagram shows the overall hardware architecture of the upper limb exoskeleton prediction control system proposed in this application embodiment.

[0053] Figure 3 This diagram illustrates the convergence of the physical information neural network model used in the embodiments of this application during the training process.

[0054] Figure 4 The signal flow diagram of the compliant assist control strategy is shown.

[0055] Figure 5 This is a comparison curve of the gravity compensation effect of the embodiments of the present invention in a rehabilitation training scenario.

[0056] Figure 6a The figure shown is a verification result of the shoulder joint trajectory tracking accuracy in an embodiment of the present invention.

[0057] Figure 6b The figure shown is a verification result of the elbow joint trajectory tracking accuracy in an embodiment of the present invention.

[0058] Figure 7 This diagram illustrates the core predictive performance of the Physical Information Neural Network (PINN) model of this system.

[0059] Figure 8 A schematic diagram showing the attachment position of the electromyography sensor on the human arm in an embodiment of this application is shown.

[0060] Figure 9 A typical peripheral circuit diagram of a QSPI Flash memory chip with model number W25Q256JV is shown.

[0061] Figure 10 A schematic diagram of a TF card storage interface circuit based on an SDMMC controller is shown.

[0062] Figure 11 A schematic diagram of a dual-channel synchronous buck converter circuit based on the TMI7003B power management chip is shown. Detailed Implementation

[0063] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0064] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] Please refer to Figure 1 , Figure 1 This paper presents a schematic diagram of a closed-loop stepping control system for an exoskeleton based on electromyography signal prediction, as proposed in an embodiment of this application, including:

[0066] Microcontroller unit;

[0067] The multimodal sensing module, connected to the microcontroller unit, is used to collect the user's shoulder and elbow joint angle signals as well as human-computer interaction force signals.

[0068] The physical information neural network inference engine, deployed in the microcontroller unit, is used to receive the instantaneous state vector constructed from joint angle signals and interaction force signals, and output the predicted compensation torque corresponding to each joint under the current posture. The predicted compensation torque includes gravity compensation torque and friction torque.

[0069] The compliant power assist control unit is connected to the physical information neural network inference engine to fuse the predicted compensation torque with the user intention torque obtained based on the interactive force signal to generate the total driving torque command.

[0070] The motor drive module, connected to the compliant assist control unit, is used to drive the joint motors of the exoskeleton according to the total driving torque command.

[0071] This system achieves low-cost, high-real-time intelligent upper limb assistance through embedded artificial intelligence. The core of the system lies in its highly integrated hardware architecture and intelligent software algorithms, with all functions handled by a single microcontroller unit. This microcontroller unit, serving as the system's computation and control hub, preferably employs a high-performance microcontroller based on the ARM Cortex-M7 core.

[0072] The multimodal sensing module is connected to the microcontroller unit and is responsible for acquiring the user's motion state and intentions in real time. It includes angle sensors mounted on the shoulder and elbow joints of the exoskeleton to acquire the absolute angles of the joints; and electromyography (EMG) sensors or force sensors mounted on the human-machine interface to acquire the interactive forces actively applied by the user or EMG signals characterizing muscle activity. The microcontroller synchronously acquires these signals at a frequency of at least 1 kHz and combines the shoulder angle, elbow angle, and processed interactive force signals to construct a 6-dimensional state vector that comprehensively reflects the instantaneous dynamics of the current system.

[0073] The physical information neural network inference engine is deployed directly in the on-chip Flash memory of the microcontroller unit in the form of a lightweight model. The engine receives a 6-dimensional state vector as input in real time and, based on the rigid body dynamics prior knowledge learned through the physical information-driven strategy, accurately infers and outputs the predicted compensation torque corresponding to the shoulder and elbow joints in a very short time. This torque is mainly used to counteract the robot arm's own gravity and joint friction.

[0074] The compliant power steering control unit connects to the PINN inference engine and sensing module, executing crucial torque fusion decisions. This unit first processes the raw interactive force or electromyographic signals and maps them to the user's intended torque using a preset gain coefficient. Simultaneously, it multiplies the predicted compensation torque output by PINN by a gravity compensation coefficient preferably set between 0.8 and 0.95 to achieve partial compensation, thereby reducing the load while preserving necessary proprioception. Finally, the processed user-intention torque and the adjusted compensation torque are vector-added to generate the final total driving torque command.

[0075] The motor drive module receives the total drive torque command from the compliant assist control unit and converts it into control signals for the joint motor. The motor is preferably a closed-loop stepper motor, and the drive module directly drives the motor driver using pulse width modulation (PWM) pulses with adjustable frequency generated by the microcontroller. This direct control method—torque command → pulse frequency / number—combines the inherent open-loop position accuracy of the stepper motor with closed-loop torque control, achieving high-precision, high-frequency response output of the exoskeleton joint assist torque, thus completing fully closed-loop intelligent control from user intent recognition to power execution.

[0076] The microcontroller unit establishes full-duplex asynchronous communication links with the left and right exoskeleton execution modules via the USART2 and USART3 interfaces, respectively.

[0077] The USART2 interface is connected to the left module via pins PA2 and PA3, and is used to transmit multidimensional state vectors, electromyographic data and torque commands for the left shoulder and left elbow joints.

[0078] The USART3 interface is connected to the right module via pins PB10 and PB11 and is used to transmit symmetrical data structures of the right shoulder and right elbow joints.

[0079] Both communication links support DMA transfer mode.

[0080] The USART2 interface connects to the shoulder and elbow joint execution modules of the left exoskeleton via the PA2 (transmitter) and PA3 (receiver) pins of the microcontroller, forming the first full-duplex asynchronous communication link. This link is responsible for uploading the multi-dimensional state vector of the left limb and the raw data from the electromyography sensor in real time, while simultaneously sending down the left joint gravity compensation torque and assist commands calculated by the core control algorithm.

[0081] Symmetrically, the USART3 interface establishes a second full-duplex asynchronous communication link with the right exoskeleton execution module through the PB10 (transmitter) and PB11 (receiver) pins of the microcontroller, which is used to transmit sensor data and control commands of the symmetrical right limb.

[0082] Both USART2 and USART3 interfaces are configured with high-speed baud rate, 8 data bits, 1 stop bit, no parity bit, and support DMA transfer mode. USART2 receives the 6-dimensional state vector and raw electromyography data from the left exoskeleton, while USART3 receives the symmetrical data structure from the right exoskeleton, enabling synchronous data acquisition at a dual-channel 1kHz control frequency.

[0083] The RGB LCD interface uses a 40P FPC 0.5mm flip-top connector, supporting 5V backlight power supply and 3.3V logic level. Video data is transmitted via 24 data lines (LCD_R0 to R7, G0 to G7, B0 to B7) and synchronization signals (LCD_HS, LCD_VS, LCD_CLK, LCD_DE), and a reserved LCD_BL backlight PWM control pin is provided for brightness adjustment.

[0084] The multimodal sensing module includes:

[0085] Angle sensors, installed at the shoulder and elbow joints of the exoskeleton, are used to collect absolute angle information of the joints.

[0086] Force sensors or electromyography (EMG) sensors are installed on the human-machine interface to collect active interactive forces or EMG signals applied by the user.

[0087] The microcontroller unit combines the collected shoulder angle, elbow angle, and corresponding interaction force signals to construct a multidimensional instantaneous state vector;

[0088] The electromyography (EMG) sensor circuit includes:

[0089] The preamplifier unit uses an AD8226 chip to form a differential amplifier circuit, with an external gain setting resistor.

[0090] The signal conditioning link uses a TL084 operational amplifier to construct a bandpass filter unit, a precision rectifier unit, and an envelope detector unit to convert bidirectional AC electromyography signals into unidirectional DC envelope signals.

[0091] The microcontroller unit synchronously acquires conditioned analog signals from the electromyography sensor.

[0092] Angle sensors are precisely mounted at the rotation centers of the shoulder and elbow joints of the exoskeleton to measure the absolute angle information of the joints in real time at a sampling frequency of no less than 1 kHz, thereby accurately acquiring the position and movement status of the limb. Simultaneously, electromyography (EMG) sensors or force sensors are integrated into the wearable contact surface of the arm. The EMG sensor captures the user's muscle activity intentions by detecting electrophysiological signals on the skin surface; the force sensor directly measures the pressure or torque at the human-machine interface. These signals collectively reflect the user's active control intentions.

[0093] The microcontroller unit synchronously receives all the above sensor signals. It fuses the acquired shoulder angle, elbow angle, calculated angular velocity and angular acceleration, and the corresponding shoulder and elbow interaction torques to construct a 6-dimensional instantaneous state vector.

[0094] The electromyography (EMG) sensor employs a high-performance, low-noise analog signal conditioning circuit to ensure the quality of acquisition of weak bioelectrical signals. Its core circuitry is a two-stage amplification and processing architecture.

[0095] Preamplifier unit: Utilizing an AD8226 chip to construct a differential amplifier circuit, its positive and negative input terminals are connected to the measuring and reference electrodes respectively via matching resistors, forming a high common-mode rejection ratio input stage. By connecting an external 240-ohm gain setting resistor, the amplification factor of this stage is set to approximately 200 times, achieving initial amplification of the original microvolt-level electromyographic signal.

[0096] Signal conditioning circuit: Built on the TL084 quad operational amplifier chip. The amplified signal first passes through a bandpass filter unit composed of operational amplifiers and RC networks, effectively filtering out baseline drift and high-frequency noise. Subsequently, the signal enters a precision full-wave rectifier unit composed of precision diodes and operational amplifiers, converting the bidirectional AC electromyography signal into a unidirectional signal. This signal is then smoothed by an envelope detector unit composed of RC integrator circuits to extract the DC envelope signal reflecting muscle contraction intensity. The final stage uses a potentiometer-adjustable in-phase proportional amplifier unit for output buffering and gain fine-tuning to accommodate signal strength differences among users. Finally, the circuit outputs a standard analog voltage signal with an amplitude suitable for analog-to-digital conversion.

[0097] The system also includes a multimodal data storage array, which includes:

[0098] SDRAM memory, connected to the microcontroller unit via the FMC bus, provides off-chip RAM space for use as a dynamic activation buffer and sliding window data cache during neural network inference.

[0099] The QSPI Flash memory, connected to the microcontroller unit, provides static storage space for storing the physical information quantized by INT8, the model weights of the neural network inference engine, and the system firmware.

[0100] This storage array mainly consists of two parts:

[0101] SDRAM memory: As the system's off-chip dynamic random access memory, it is connected at high speed to the microcontroller unit's FMC (Flexible Memory Controller) bus. It interfaces via a 32-bit wide data bus (FMC_D0 to FMC_D31), address bus (FMC_A0 to FMC_A11), and control signal lines such as clock (FMC_SDCLK) and clock enable (FMC_SDCKE), providing up to 32MB of cache space. This SDRAM core is used to store dynamically activated buffer data generated during the forward propagation of the physical information neural network inference engine, as well as sliding window timing data from sensors, providing the necessary high-speed data exchange space for complex real-time calculations. Figure 9 This diagram illustrates a typical peripheral circuit for the W25Q256JV QSPI Flash memory chip. The W25Q256JV QSPI Flash memory chip communicates with the host controller via a four-wire SPI (QSPI) interface. CS is the chip select signal, CLK is the clock signal, and DI (IO0), DO (IO1), WP (IO2), and HOLD (IO3) form a four-wire bidirectional data bus, supporting high-speed data transmission. The chip power supply (VCC) is connected to 3.3V and is equipped with a decoupling capacitor C43 (0.1µF) to filter power supply noise. GND is the ground terminal. This circuit is used to store non-volatile data such as firmware and neural network models in embedded systems and supports XIP (Execute in Local) mode, facilitating direct program execution from the Flash memory.

[0102] QSPI Flash Memory: Serving as the system's non-volatile storage medium, it connects to the microcontroller unit via a four-wire SPI (QSPI) interface. Specifically, it communicates using the QSPI_BK1_NCS (chip select), QSPI_CLK (clock), and IO0 to IO3 (data input / output) signal lines, with a storage capacity of 32MB. This Flash memory is primarily used to store the lightweight physical information neural network model weights processed by INT8 quantization, as well as the firmware program for the entire system. Furthermore, it supports XIP (Execute In Place) mode, allowing the microcontroller to directly read and execute code from it, further improving system startup and operation efficiency.

[0103] In the embedded system constructed in this application, its startup, operation, and data management processes have been carefully designed to ensure high performance and high reliability. When the system powers on, the microcontroller unit first reads the lightweight physical information neural network model weights quantized by INT8 from the Flash memory connected to the QSPI interface, and loads them into the SDRAM connected via the FMC bus, completing the initialization of the core algorithm model. Subsequently, the system displays the ready status on the screen through the RGB LCD interface, completing the startup process.

[0104] Once in the main operating loop, the microcontroller unit fully utilizes its built-in DSP instruction set to accelerate forward inference computation of the neural network deployed in SDRAM. After optimization, the complete inference time from the state vector input to the compensation torque output is controlled within 1 millisecond, thus meeting the real-time requirements of control loops up to 1kHz and achieving imperceptible delay tracking of human movement.

[0105] After the system powers on, the TMI7003B dual-channel buck converter in the multi-level power management module starts working, efficiently converting the input voltage into a 1.8V core voltage and a 3.3V interface voltage. The 1.8V voltage, after passing through a filter network composed of inductors and capacitors, powers the microcontroller core and digital logic; the 3.3V voltage, after being isolated by a ferrite bead, is separated into a dedicated analog power supply (VDDA) for the ADC module. This minimizes the interference of digital noise on the acquisition accuracy of weak analog signals such as electromyography signals.

[0106] In terms of data storage architecture, SDRAM, as high-speed dynamic memory, provides an activation buffer for real-time inference of neural networks through the FMC bus; QSPI Flash, as non-volatile storage, stores system firmware and neural network weights; while the TF card connected through the SDMMC interface serves as a large-capacity extended storage, used to record long-term raw electromyography data and system logs, facilitating offline analysis and evaluation.

[0107] The physical information neural network inference engine adopts a fully connected neural network architecture, including:

[0108] The input layer contains 6 nodes, corresponding to a multi-dimensional instantaneous state vector composed of shoulder angle, elbow angle, angular velocity, angular acceleration and corresponding interaction force signals;

[0109] At least one hidden layer containing 64 neurons, using the ReLU activation function;

[0110] The output layer contains two nodes, which output the predicted compensation torques corresponding to the shoulder joint and elbow joint, respectively.

[0111] The physics-information neural network inference engine can be deployed in the on-chip memory of a microcontroller to process 6-dimensional state vectors constructed by the sensing module in real time and output accurate joint compensation torque predictions. This engine employs a specially designed and optimized fully connected neural network (MLP) architecture. This network architecture includes: an input layer with 6 nodes, the number of which corresponds to the 6-dimensional instantaneous state vector, used to receive real-time acquired sensor data; at least one hidden layer for high-dimensional nonlinear mapping to learn dynamic relationships; the network contains two hidden layers, each with 64 neurons and using the ReLU activation function to achieve a balance between model complexity and computational efficiency; and an output layer with 2 nodes, whose output values ​​correspond to the predicted compensation torques required for the shoulder and elbow joints, respectively. The network is trained using a physics-information driven strategy, utilizing simulation data generated based on rigid body dynamics equations as supervision signals, enabling the network to learn embedded physical laws, thereby achieving excellent generalization ability and prediction reliability. After quantization and embedded optimization, the parameter weights of the complete network model occupy only about 19 kB of storage space, and the runtime dynamic memory (RAM) occupies less than 1 kB. Combined with the hardware acceleration capability of the microcontroller, it can achieve extreme performance with a single inference time of less than 1 millisecond, thus meeting the real-time requirements of the exoskeleton system with a control frequency of up to 1 kHz.

[0112] The physical information neural network in the physical information neural network inference engine is trained using a physical information-driven strategy, including:

[0113] Construct the rigid body dynamics equations of the robotic arm that include gravity terms, Coriolis force terms, centrifugal force terms, and friction force terms;

[0114] By using dynamic equations to perform random sampling in the joint motion space, a physical simulation dataset containing state inputs and theoretical torque labels is generated.

[0115] Supervised learning of neural networks is performed using physical simulation datasets to make the network approximate a real nonlinear dynamic model.

[0116] The Physical Information Neural Network (PINN) inference engine employs a physically-driven strategy for training, incorporating physical laws as prior knowledge into the neural network's learning process. This includes the following steps: First, constructing the rigid body dynamics equations for the robotic arm, including gravity, Coriolis force, centrifugal force, and friction terms. For example, the equations established for a two-bar robotic arm would be... ,in For gravity, The equation, consisting of Coulomb friction and viscous friction terms, fully describes the nonlinear dynamic characteristics of the system. Secondly, large-scale random sampling is performed within the joint motion space using the dynamic equations to generate a high-quality physical simulation dataset containing state inputs and theoretical torque labels. This dataset generation does not rely on expensive and time-consuming real data acquisition. Finally, the physical simulation dataset is used for supervised learning of the neural network. By minimizing the error between the network's predicted torque and the theoretical labels, the network can approximate the real nonlinear dynamic model. This training strategy ensures that the network not only learns the data distribution but also strictly follows physical laws. Therefore, even when faced with motion states not covered by the training data after embedded deployment, it can maintain stable prediction accuracy and excellent generalization ability due to its physical consistency.

[0117] The torque fusion logic includes:

[0118] Dead zone filtering is performed on the collected raw interaction force signals;

[0119] The filtered interactive force signal is multiplied by a preset boost gain coefficient to obtain the user intention torque;

[0120] Multiply the predicted compensation torque output by the physical information neural network inference engine by the gravity compensation coefficient;

[0121] The processed user-intended torque is added to the predicted compensation torque multiplied by a coefficient to obtain the total driving torque command.

[0122] The compliant assist control unit is a key decision-making module connecting user intent recognition and exoskeleton power execution. It receives raw interactive force (or electromyographic) signals from the multimodal perception module and predicted compensation torque from the physical information neural network inference engine, and executes the following torque fusion logic to generate the final total driving torque command:

[0123] First, the raw interactive force signal is filtered for dead zone. Then, the filtered clean interactive force signal is multiplied by a preset boost gain coefficient, which maps the physical signal linearly or nonlinearly into a control quantity, thereby obtaining the torque component that characterizes the user's motion intention.

[0124] Simultaneously, the predicted gravity and friction compensation torque output in real time by the PINN inference engine for the current posture is multiplied by a specific gravity compensation coefficient. This undercompensation strategy allows the exoskeleton motors to bear most of the robotic arm's weight, while the user still perceives a small amount of residual gravity. This significantly reduces the user's workload and achieves a zero-gravity tracking experience while retaining necessary physical gravity feedback, thus significantly enhancing the user's proprioception and the naturalness of human-computer interaction. It avoids the discomfort and safety hazards caused by weightlessness resulting from complete compensation. Finally, the control unit vector-adds the processed user-intention torque with the coefficient-adjusted predicted compensation torque to synthesize the final total drive torque command.

[0125] The system also includes a power management module, which includes:

[0126] A dual-channel BUCK circuit based on the TMI7003B chip generates the core voltage and IO voltage respectively.

[0127] The core voltage is filtered by inductors and capacitors and then supplied to the core and digital circuits of the microcontroller unit. The IO voltage is isolated by ferrite beads and then divided into digital power supply and analog power supply. The analog power supply powers the ADC module.

[0128] Figure 11 This diagram illustrates a dual-channel synchronous buck converter circuit built upon the TMI7003B power management chip. The power management module utilizes a dual-channel synchronous buck (BUCK) circuit based on the TMI7003B chip as its core, efficiently and stably generating the two critical operating voltages required by the system from the input power supply: one is a 1.8V core voltage (VDDCORE), and the other is a 3.3V input / output interface voltage (VDDIO). The 1.8V core voltage, after passing through a filter network composed of inductors and capacitors, is dedicated to powering the microcontroller unit and other high-speed digital logic circuits, ensuring stable core operations. The 3.3V interface voltage is further divided into digital and analog power supplies after isolation via a ferrite bead. The analog power supply (VDDA) is independently and cleanly supplied to the microcontroller's analog-to-digital converter (ADC) module. This design effectively isolates switching noise generated by the digital circuits, thereby significantly improving the accuracy and signal-to-noise ratio of weak analog signal acquisition, such as electromyography (EMG) signals. In addition, this power management architecture supports seamless switching between AC adapters (AC / DC) and backup batteries, enhancing the system's power supply flexibility and continuity in different application scenarios.

[0129] The system also includes a human-computer interaction interface and a debugging interface:

[0130] The human-machine interface includes a parallel LCD interface and a resistive touch screen interface, which transmits video data through data lines and synchronization signals to display joint angles, predict compensation torques and system status in real time.

[0131] The debugging interface includes a one-button download debugging circuit, which works in conjunction with the reset circuit and the BOOT circuit. The reset signal and the BOOT0 pin level are controlled by a single button operation, and it supports USB DFU or serial ISP program download.

[0132] The core of the debugging interface is a cleverly designed one-button download and debugging circuit. This circuit consists of a reset circuit and a BOOT (boot) circuit working together. Its key innovation lies in using a transistor to construct a single-button trigger logic. With a single operation of a physical button (SW1), the hardware reset signal (RST) of the microcontroller and the level state of the system boot configuration pin (BOOT0) can be controlled simultaneously. The user only needs to perform a "press and release" operation once, and the system can automatically enter the boot mode from the system memory. This allows for the download and updating of applications via USB device firmware upgrade (DFU) or serial port in-system programming (ISP), greatly simplifying the development, debugging, and subsequent firmware maintenance process.

[0133] In addition, the system provides a TF card storage interface as an important data expansion port. This interface connects to a standard MicroSD card slot via the microcontroller's Secure Digital Input / Output / Multimedia Card (SDMMC) controller, specifically using the SDMMC1_D0 to SDMMC1_D3 data signal lines, the SDMMC1_CMD command line, and the SDMMC1_CLK clock line. This interface allows the system to offline store large amounts of information, such as raw multi-channel electromyography signal data, joint motion trajectories, and system event logs acquired over long periods of monitoring, on a high-capacity MicroSD memory card.

[0134] Figure 10 A schematic diagram of a TF card storage interface circuit based on an SDMMC controller is shown. This circuit multiplexes specific pins of the microcontroller as SDMMC1 interface signals: SDMMC1_D0 to D3 (data lines) are connected to PC8-PC11 pins, SDMMC1_CMD (command line) is connected to PD2 pin, and SDMMC1_CLK (clock line) is connected to PC12 pin. Both the data and command lines are connected to a 10kΩ pull-up resistor (R17-R21) to a 3.3V power supply to ensure bus signal stability and compliance with SD card electrical specifications. The circuit includes a TF card slot, whose power supply pin (VDD) is filtered through a 0.1µF decoupling capacitor (C48) before being connected to 3.3V.

[0135] The microcontroller unit converts the trained neural network model into optimized C code using an embedded neural network deployment tool, and accelerates matrix operations using the microcontroller's internal DSP instruction set.

[0136] The motor drive module adopts field-oriented control technology, with the current loop as the inner loop, to directly respond to the torque command issued by the microcontroller unit.

[0137] The motor drive module is the key actuator for achieving high-precision torque output, employing Field Oriented Control (FOC) technology. This technology decouples motor control into independent control of the magnetic field and torque components, using a current loop as the innermost control circuit. Through this control architecture, the drive module can directly and accurately respond to torque commands issued by the microcontroller unit, quickly converting them into the corresponding q-axis current setpoint. Combined with high-frequency PWM modulation and real-time current sampling feedback, this module achieves precise and smooth control of motor torque, effectively suppressing torque ripple.

[0138] The microcontroller unit synchronously acquires multiple sensor signals through the ADC module in a timer-triggered manner, and directly transfers the data to the circular buffer in memory via DMA.

[0139] The microcontroller unit utilizes the on-chip ADC1 module for dual-channel synchronous data acquisition. This acquisition process is triggered by timer TIM1 at a frequency of 1 kHz and undergoes 16-bit oversampling processing to output data with a 12-bit effective resolution. The acquired data is directly transferred to a preset circular buffer in memory via DMA, without CPU intervention, significantly reducing system overhead. Simultaneously, the encoder interface reads the magnetic encoder values ​​of the left and right joints via SPI2 / 3 interrupts at a frequency of 1 kHz, employing a DMA double-buffering mechanism to ensure the real-time performance and reliability of the angle data. Furthermore, the motor control section uses timer TIM3 to generate PWM pulses with a dynamically adjustable frequency within the range of 0-50 kHz, directly driving the STEP pin of the 42 mm three-phase closed-loop stepper driver to achieve precise torque output.

[0140] Example 1: Applying the electromyography signal prediction-based closed-loop stepping control system for exoskeleton proposed in this application to a lightweight upper limb exoskeleton for hospital rehabilitation:

[0141] Mechanical Structure and Wearable Design: Figure 8This diagram illustrates the placement of the electromyography (EMG) sensor on the human arm in an embodiment of this application. The entire design is extremely lightweight, with a total weight of only 185 grams per arm. This includes a 25-gram control board, a 120-gram 24V / 2Ah battery, and the remainder consisting of wires and electrodes. The joint support is made of 3D-printed PLA material, with one degree of freedom each at the shoulder and elbow. The reducer uses a 10:1 reduction ratio, achieving a peak output torque of 0.35 N·m. The range of motion for each joint is limited to -90° to +90°, ensuring safe and reasonable movement. The wearable component uses elastic and breathable shoulder straps with 8mm wide Velcro rails on both shoulder straps for easy adjustment to different user body types. The EMG / force sensor in the sensing module employs a differential electrode layout: the central electrode is attached to the brachioradialis muscle belly, while the reference electrode is placed behind the acromion, with a distance of approximately 2 cm between them. This design allows for rapid adjustment based on arm circumference, ensuring stable and reliable signal acquisition.

[0142] Hardware control core configuration: Figure 2 This diagram illustrates the overall hardware architecture of the upper limb exoskeleton predictive control system proposed in this application. The top layer of the architecture is the input acquisition module, which includes angle sensors for measuring joint positions and electromyography (EMG) / force sensors for sensing the user's active intentions. The multimodal signals acquired by these sensors are fed into the core STM32H7 main control unit. Inside the main control unit, a state vector is first constructed, combining the original signals into a 6-dimensional instantaneous state vector. Subsequently, the system is divided into two parallel computing paths: one path is intention recognition and gain adjustment, converting the EMG / force signals into user-assisted torques; the other path's core is a physical information neural network inference engine, which, based on prior knowledge of dynamics, performs real-time calculations on the input 6-dimensional vector and outputs gravity / friction compensation torques. These two torques are merged into a total control torque within the main control unit and issued to the underlying actuators. The actuators consist of motor drivers and exoskeleton joint motors, responsible for outputting physical forces. The movement of the motors generates new joint angles, forming a physical motion feedback loop, making the entire system a complete real-time control system based on sensor feedback and model prediction.

[0143] The hardware control core of this embodiment uses the high-performance STM32H750XBH6 microcontroller based on the ARM Cortex-M7 core. This controller is configured with a main frequency of 480 MHz, and its computational potential is fully utilized by enabling instruction and data cache (I / D-Cache) and setting the Flash wait cycle to 5. For data acquisition, the system uses the on-chip ADC1 module for dual-channel synchronous acquisition, triggered by timer TIM1 at a frequency of 1 kHz (TRGO). The sampled data undergoes 16-bit oversampling processing to obtain a 12-bit effective resolution. This data is automatically transferred to a 256 ms × 2 circular buffer via direct memory access (DMA), without CPU intervention, ensuring real-time acquisition and low latency. Joint angle information is read from the magnetic encoder (14-bit resolution, range 0-16383) via the SPI2 / 3 interface, also interrupted at a frequency of 1 kHz, and efficiently transmitted using a DMA double-buffering mechanism. For motor drive, the system uses timer TIM3 to generate pulse width modulation (PWM) pulses with a frequency that is dynamically adjustable in the range of 0 to 50 kHz, which directly drive the STEP pin of the 42 mm three-phase closed-loop stepper driver to achieve precise control of the output torque.

[0144] In terms of storage resource configuration, the system has made fine-grained allocations: 512 kB of AXI-SRAM is used to store the activation buffer during the neural network inference process; 320 kB of SRAM1 is used to store sliding window data and feature vectors; and 128 kB of SRAM2 is used to run the FreeRTOS operating system and inference task stack. The first 512 kB of Flash memory is used to store the bootloader and application programs, and the subsequent 1536 kB is used to store the PINN model weights (approximately 19 kB) quantized by INT8 and related lookup tables. The remaining space is reserved to support the over-the-air (OTA) upgrade function with dual backup.

[0145] Multimodal sensing and feature construction: The multimodal sensing and feature construction stage is responsible for converting raw physiological and motion signals into high-quality inputs that can be processed by intelligent algorithms. At the sensor front end, electromyography (EMG) signal acquisition uses a dedicated chip AD8232 for conditioning, with its gain set to 100. A 50 Hz dual-T notch filter is configured to suppress power frequency interference, and a third-order Sallen-Key low-pass filter with a cutoff frequency of 200 Hz is used to ensure that the noise of the output signal is below 5 µV RMS, thereby obtaining stable and clean EMG data. In the state vector construction stage, the microcontroller acquires the shoulder joint angle in real time (…). ), elbow joint angle ( ) and human-computer interaction torque ( The system will calculate the joint angular velocity in real time to further capture the dynamic characteristics of the system, including raw signals such as angular velocity. ) and angular acceleration ( Finally, these signals are fused to construct a 6-dimensional instantaneous state vector, represented as follows: This vector is directly used as input to the subsequent Physical Information Neural Network (PINN), providing it with digital features that comprehensively describe the current system attitude, motion state, and interaction intent.

[0146] Construction and Training of Physical Information Neural Network (PINN): The core of this embodiment lies in the construction of the Physical Information Neural Network (PINN) model. Unlike purely data-driven "black box" models, this design innovatively integrates prior knowledge of rigid body dynamics into the learning process of the neural network. First, the dynamic equations describing the motion of the two-bar linkage robotic arm are established: ,in Represents the gravity term. The effects of Coulomb friction and viscous friction are combined. Based on this equation, large-scale random sampling is performed in the joint motion space to automatically generate a physical simulation dataset containing 50,000 sets of "[state vector, theoretical torque]" correspondences. The model adopts a fully connected network (MLP) architecture, specifically: Input(6) → Dense(64, ReLU) → Dense(64, ReLU) → Output(2). Supervised learning of this network is performed using the TensorFlow framework, and the loss function integrates mean squared error (MSE) and physical constraint terms. After training, the model is further quantized to INT8 format, compressing its weight size to only about 19 kB. Figure 7 The diagram shows the core prediction performance of the Physical Information Neural Network (PINN) model of this system. The predicted torque of the trained PINN model has a very high correlation with the theoretical dynamic torque (R²>0.99), which fully proves that the model has successfully learned and internalized the physical laws of the system.

[0147] Embedded Deployment and Inference: To enable efficient operation of the trained physical information neural network on resource-constrained microcontrollers, a dedicated embedded neural network deployment tool was first used to automatically convert the trained model, saved in TFLite format, into C language code highly optimized for the target hardware. After deployment, the inference performance was tested on a selected STM32H750 microcontroller, showing a single forward inference time of only 0.82 milliseconds and a peak RAM usage of 60kB. The network's final output is the gravity compensation torque required for the shoulder joint and elbow joint under the current human posture, calculated by the system. These two torque values ​​will be directly used for subsequent compliant control.

[0148] Compliant Assist Control Strategy: Figure 4 The signal flow diagram of the compliant assist control strategy is shown. First, by fusing prior knowledge of rigid body dynamics through a physical information neural network, it can accurately fit the nonlinear friction and variable load characteristics that are difficult to describe by traditional mathematical models, significantly reducing mechanical resistance in the human-computer interaction process and achieving a near-realistic "zero gravity" high transparency experience. Second, the system achieves extreme lightweight and low cost, running entirely on a single MCU chip, with network model weights of only about 19kB and RAM usage of less than 1kB, eliminating the need for expensive external GPUs or industrial control computers, greatly reducing system size and power consumption. Third, thanks to the lightweight network design and embedded optimization, the AI ​​inference time is less than 1ms, and the overall system control loop frequency can reach over 1kHz, achieving ultra-low latency real-time response and completely solving the problem of response lag in traditional solutions. Finally, this solution has strong physical interpretability. Compared to purely data-driven black-box models, the PINN of this invention is trained based on physical equations and can maintain stable prediction output even in unseen motion states, demonstrating excellent generalization ability and robustness.

[0149] This embodiment employs a composite compliant assist control strategy combining "feedforward gravity compensation" and "impedance control." In the gravity compensation branch, the predicted compensation torque output in real time by the Physical Information Neural Network (PINN) is used. Multiplying by a compensation coefficient of 0.9 (kg), this means the exoskeleton motors will actively bear approximately 90% of the robotic arm's weight, leaving the user to bear only about 10% of the remaining weight. This "partial compensation" design significantly reduces user load and achieves a "lightweight" experience while deliberately retaining necessary physical gravity feedback to maintain the user's proprioception and naturalness. In the intent recognition branch, the system performs dead-time processing on the acquired raw interaction force signal Tint to filter out noise, then multiplies it by an adjustable assist gain Ka, converting it into a torque component reflecting the user's movement intent. Finally, the control unit fuses the two processed torque components to generate a total drive torque command. This instruction is further converted into a corresponding current or pulse command, which is then sent to the motor driver to drive the joint motor to perform precise auxiliary movements.

[0150] Simulation verification and effect analysis: In the simulation verification and effect analysis phase, the overall performance of the system was fully verified by constructing a high-fidelity physical simulation environment. Figure 5This is a comparison curve of the gravity compensation effect of the embodiment of the present invention in a rehabilitation training scenario. In the simulated gravity compensation experiment, when the assist function based on physical information neural network (PINN) is enabled, the amplitude of the simulated electromyographic signal required to maintain the same body posture is significantly reduced. This directly proves that the system can effectively share the human body load applied to the user. Figure 6a The figure shown is a verification result of the shoulder joint trajectory tracking accuracy in an embodiment of the present invention. Figure 6b The diagram shows the verification results of the elbow joint trajectory tracking accuracy in an embodiment of the present invention. The angles of each joint of the exoskeleton can closely follow the preset human movement trajectory, with a maximum hysteresis angle error of less than 2°, fully demonstrating the excellent dynamic response and real-time tracking performance of the system. In summary, because PINN achieves precise compensation for the nonlinear frictional force of the system, this control scheme can theoretically eliminate more than 90% of the mechanical resistance, thereby fundamentally improving the compliance and transparency of human-computer interaction.

[0151] In another possible embodiment, Embodiment 2 provides a power exoskeleton design for industrial material handling scenarios. Compared to the rehabilitation assistance application of Embodiment 1, this solution features comprehensive upgrades in mechanical structure, load capacity, and system configuration to meet the demands of high-load, long-duration industrial operations.

[0152] In terms of mechanics and load capacity, the structural materials have been upgraded to 6061 aluminum alloy articulated arms and fiberglass belts, ensuring strength while controlling weight. The entire machine weighs approximately 1.8 kg, and the rated load capacity has been increased to 15 kg. The joint degrees of freedom have been expanded to 5 (2 at the shoulder, 1 at the elbow, and 2 at the wrist), and a variable reduction ratio design (range 15:1-30:1) has been adopted, with a maximum output torque of 35 N·m to support diverse gripping and handling movements.

[0153] The control system is optimized while keeping the core microcontroller (STM32H750) unchanged: the main frequency is moderately reduced to 400 MHz to reduce power consumption, while ensuring that the single inference time of the neural network is still less than 1 ms; the ADC sampling rate of the sensing module is increased to 2 kHz, and an electromyography signal acquisition channel for the biceps brachii and posterior deltoid is added to more comprehensively capture the upper limb movement intention; the drive module is upgraded to a high-performance field-oriented control (FOC) driver with 48 V power supply, supports 2 Mbps RS-485 bus communication, and its current loop closed-loop bandwidth exceeds 2 kHz.

[0154] To match the multi-degree-of-freedom model, this embodiment extends and retrains the Physical Information Neural Network (PINN). The network's input vector is expanded to 10 dimensions, and the output is correspondingly expanded to 5 dimensions, representing the compensation torques of the five joints. The expanded network size is moderately increased, with a weight size of approximately 95 kB. On the optimized hardware, the measured inference time is approximately 0.96 ms, which still meets the real-time requirements of the system's 1 kHz control loop.

[0155] Figure 3 This diagram illustrates the convergence of the physical information neural network model used in the embodiments of this application during the training process. The training loss and validation loss curves are plotted with the number of training epochs on the horizontal axis and the mean squared error (MSE) on the vertical axis. After training begins, both loss curves rapidly decrease from their initial values, converging quickly to extremely low levels close to zero within approximately 10 epochs and remaining stable in subsequent training. This indicates that supervised learning of the neural network using the simulation dataset generated by the physical information-driven strategy is highly efficient and successful, enabling the model to quickly and accurately learn and approximate the real nonlinear dynamics model with a small number of training epochs. Furthermore, the training loss and validation loss curves highly overlap and decrease synchronously, meaning that the model performs consistently on both the training and validation datasets, without overfitting.

[0156] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0157] First, by fusing prior knowledge of rigid body dynamics with a physical information neural network, it is possible to accurately fit complex dynamic characteristics, including nonlinear friction. This significantly reduces mechanical resistance in human-computer interaction, providing users with a near-realistic "zero-gravity" responsive experience.

[0158] Secondly, the entire system runs entirely on a single microcontroller chip, with neural network model weights of only about 19kB and RAM usage of less than 1kB during operation. It eliminates the need for an external GPU or industrial PC, significantly reducing system cost, size, and power consumption, making it ideal for portable wearable devices.

[0159] Third, thanks to the lightweight network design and embedded optimization, the AI ​​inference time is less than 1 millisecond, and the system control loop frequency can reach over 1 kHz. This completely solves the problem of lag in traditional solutions, enabling real-time and natural tracking of human movement.

[0160] Fourth, the PINN proposed in this application is trained based on physical equations, rather than a purely data-driven black-box model, thus possessing stronger physical interpretability. It can maintain stable prediction output even under unseen motion states, demonstrating good generalization ability and robustness.

[0161] This invention also provides a predictive control method for an upper limb exoskeleton based on any of the aforementioned system embodiments. The method mainly includes the following steps: First, a rigid body dynamics model of the exoskeleton arm is established, and a physical simulation dataset is generated based on this model to train a physical information neural network. Then, the trained and optimized lightweight PINN model is deployed to the on-chip memory of the microcontroller unit. During real-time operation, the system simultaneously performs data acquisition and intelligent inference: On one hand, it collects the user's shoulder and elbow joint angles and human-computer interaction force (or electromyography) signals in real time and constructs a 6-dimensional state vector; on the other hand, it inputs this state vector into the deployed PINN model in real time to infer the gravity and friction compensation torque required to maintain the current posture. Simultaneously, the system calculates the user's intention torque based on the interaction force signals and fuses it with the aforementioned compensation torque according to a preset strategy to generate the final total driving torque command, thereby driving the joint motors to output precise auxiliary torque.

[0162] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A closed-loop stepping control system for exoskeletons based on electromyographic signal prediction, characterized in that, include: Microcontroller unit; The multimodal sensing module, connected to the microcontroller unit, is used to collect the user's shoulder and elbow joint angle signals as well as human-computer interaction force signals. The physical information neural network inference engine, deployed in the microcontroller unit, is used to receive the instantaneous state vector constructed from joint angle signals and interaction force signals, and output the predicted compensation torque corresponding to each joint under the current posture. The predicted compensation torque includes gravity compensation torque and friction torque. The physical information neural network in the physical information neural network inference engine is trained using a physical information-driven strategy, including: Construct the rigid body dynamics equations of the robotic arm that include gravity terms, Coriolis force terms, centrifugal force terms, and friction force terms; By using dynamic equations to perform random sampling in the joint motion space, a physical simulation dataset containing state inputs and theoretical torque labels is generated. Supervised learning of neural networks is performed using physical simulation datasets to make the network approximate a real nonlinear dynamic model. The compliant power assist control unit is connected to the physical information neural network inference engine to fuse the predicted compensation torque with the user intention torque obtained based on the interactive force signal to generate the total driving torque command. The compliant power steering control unit is used to execute torque fusion logic, which includes: Dead zone filtering is performed on the collected raw interaction force signals; The filtered interactive force signal is multiplied by a preset boost gain coefficient to obtain the user intention torque; Multiply the predicted compensation torque output by the physical information neural network inference engine by the gravity compensation coefficient; The processed user-intended torque is added to the predicted compensation torque multiplied by a coefficient to obtain the total driving torque command. The motor drive module, connected to the compliant assist control unit, is used to drive the joint motors of the exoskeleton according to the total driving torque command.

2. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The microcontroller unit establishes full-duplex asynchronous communication links with the left and right exoskeleton execution modules via the USART2 and USART3 interfaces, respectively. The USART2 interface connects to the left exoskeleton execution module and is used to transmit multi-dimensional state vectors, electromyographic data and torque commands for the left shoulder and left elbow joints. The USART3 interface connects to the right exoskeleton execution module and is used to transmit symmetrical data structures of the right shoulder and right elbow joints. Both communication links support DMA transfer mode.

3. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The multimodal sensing module includes: Angle sensors, installed at the shoulder and elbow joints of the exoskeleton, are used to collect absolute angle information of the joints. Force sensors or electromyography (EMG) sensors are installed on the human-machine interface to collect active interactive forces or EMG signals applied by the user. The microcontroller unit combines the collected shoulder angle, elbow angle, and corresponding interaction force signals to construct a multidimensional instantaneous state vector; The electromyography (EMG) sensor circuit includes: The preamplifier unit uses an AD8226 chip to form a differential amplifier circuit, with an external gain setting resistor. The signal conditioning link uses a TL084 operational amplifier to construct a bandpass filter unit, a precision rectifier unit, and an envelope detector unit to convert bidirectional AC electromyography signals into unidirectional DC envelope signals. The microcontroller unit synchronously acquires conditioned analog signals from the electromyography sensor.

4. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The system also includes a multimodal data storage array, which comprises: SDRAM memory, connected to the microcontroller unit via the FMC bus, provides off-chip RAM space for use as a dynamic activation buffer and sliding window data cache during neural network inference. The QSPI Flash memory, connected to the microcontroller unit, provides static storage space for storing the physical information quantized by INT8, the model weights of the neural network inference engine, and the system firmware.

5. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The physical information neural network inference engine adopts a fully connected neural network architecture, including: The input layer contains 6 nodes, corresponding to a multi-dimensional instantaneous state vector composed of shoulder angle, elbow angle, angular velocity, angular acceleration and corresponding interaction force signals; At least one hidden layer containing 64 neurons, using the ReLU activation function; The output layer contains two nodes, which output the predicted compensation torques corresponding to the shoulder joint and elbow joint, respectively.

6. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The system also includes a power management module, which comprises: The dual-channel BUCK circuit based on the TMI7003B chip generates core voltage and IO voltage respectively. The core voltage is filtered by inductors and capacitors and then supplied to the core and digital circuits of the microcontroller unit. The IO voltage is isolated by ferrite beads and then divided into digital power supply and analog power supply. The analog power supply powers the ADC module.

7. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The system also includes a human-computer interaction interface and a debugging interface: The human-machine interface includes a parallel liquid crystal interface and a resistive touch screen interface, which transmits video data through data lines and synchronization signals to display joint angles, predict compensation torques and system status in real time. The debugging interface includes a one-button download debugging circuit, which works in conjunction with the reset circuit and the BOOT circuit. The reset signal and the BOOT0 pin level are controlled by a single button operation, and it supports USB DFU or serial ISP program download.

8. The exoskeleton closed-loop stepping control system as described in claim 1, characterized in that, The motor drive module adopts field-oriented control technology, with a current loop as the inner loop, to directly respond to the torque commands issued by the microcontroller unit.

Images