Lower limb exoskeleton control system and method based on multi-modal intention fusion and gait parameter dynamic adjustment

Abstract

The application discloses a lower limb exoskeleton control system and method based on multi-modal intention fusion and gait parameter dynamic adjustment, the system comprises a multi-modal physiological signal acquisition module configured to synchronously acquire motor imagery electroencephalogram signals, frontal lobe attention representation signals and limb proprioceptive feedback signals of a user; an intention fusion and decoding module configured to perform fusion processing to identify the motor intention of the user, and evaluate the certainty of the motor intention based on the frontal lobe attention representation signals; a gait parameter mapping module configured to determine target gait parameters through a nonlinear mapping relationship according to the motor intention intensity corresponding to the motor control instruction; an environment perception and safety control module configured to acquire real-time state information of the exoskeleton according to the limb proprioceptive feedback signals, and generate safety control instructions; and an exoskeleton execution module configured to drive the exoskeleton to perform actions according to the target gait parameters and the safety control instructions, thereby improving the refinement of control.

Description

Technical Field

[0001] This invention relates to the field of lower limb exoskeleton control technology, and in particular to a lower limb exoskeleton control system and method based on multimodal intent fusion and dynamic adjustment of gait parameters. Background Technology

[0002] Lower limb exoskeletons have shown great potential in assisting walking and rehabilitation training. Motor imagery brain-computer interfaces, as a non-invasive control method, allow users to control external devices through imagined movements, providing a new interactive pathway for patients with limb dysfunction.

[0003] However, existing lower limb exoskeleton systems controlled by brain-computer interfaces for motor imagery still face many challenges: First, most systems use "on / off" control, meaning they can only achieve simple start, stop, or preset gait switching, making it difficult to achieve real-time, continuous adjustment of fine gait parameters such as stride length and speed. This makes the exoskeleton's walking mode stiff and unnatural, making it difficult to adapt to complex and changing environments and users' personalized needs.

[0004] Secondly, decoding a single EEG signal is susceptible to noise, artifacts, and fluctuations in user attention, leading to insufficient accuracy and robustness of control commands. This may cause false triggering or control delays, affecting user experience and security. While existing technologies include bimodal fusion solutions for EEG and EMG (such as CN110495893A), they do not introduce frontal lobe attention signals as an independent evaluation dimension for intent certainty, nor do they establish a signal quality-driven dynamic weight adjustment mechanism.

[0005] Furthermore, in practical applications, the stability control of exoskeletons under different terrains or sudden situations is also a problem that urgently needs to be solved. Most existing systems lack effective perception of the walking environment, cannot adaptively adjust gait parameters according to terrain changes (such as slopes, uneven ground), and lack a sound emergency avoidance mechanism, posing a risk of falls.

[0006] Therefore, improving the precision, accuracy, and environmental adaptability of exoskeleton control is a pressing technical problem that needs to be solved. Summary of the Invention

[0007] This invention aims to address the problems of insufficient gait fine control, poor accuracy and robustness of control commands, and unsatisfactory environmental adaptability and safety in existing motor imagery brain-computer interface lower limb exoskeleton control systems. It provides a lower limb exoskeleton control system and method based on multimodal intention fusion and dynamic adjustment of gait parameters.

[0008] The first aspect of this invention provides a lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters, comprising: The multimodal physiological signal acquisition module is configured to simultaneously acquire the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals. The intention fusion and decoding module is connected to the multimodal physiological signal acquisition module and is configured to preprocess, extract features and fuse the motor imagery EEG signal, frontal lobe attention representation signal and limb proprioceptive feedback signal to identify the user's motor intention, and evaluate the certainty of the motor intention based on the frontal lobe attention representation signal, intention confidence and signal quality evaluation index, and selectively output motor control commands according to the evaluation results. The gait parameter mapping module, connected to the intent fusion and decoding module, is configured to determine intermediate gait parameters based on the intensity of the motion intent corresponding to the motion control command through a nonlinear mapping relationship, and to constrain the intermediate gait parameters based on at least one of the constraints of gait parameter amplitude, joint range of motion, gait phase change rate and environmental state to obtain target gait parameters. The environmental perception and safety control module is connected to the intent fusion and decoding module and is configured to obtain real-time status information of the exoskeleton based on the limb proprioceptive feedback signal and generate safety control commands based on the real-time status information. The exoskeleton execution module is connected to the gait parameter mapping module and the environment perception and safety control module, and is configured to drive the exoskeleton to perform actions according to the target gait parameters and the safety control commands. In this context, the environmental perception and safety control module has a higher control priority over the exoskeleton execution module than the intent fusion and decoding module.

[0009] Preferably, the multimodal physiological signal acquisition module is configured to execute the following logic: The user's scalp EEG signals are collected, and the motor imagery EEG signals and the frontal lobe attention representation signals are obtained from the scalp EEG signals based on electrode location or frequency band division. Inertial measurement data of the lower limb joints of the exoskeleton and contact mechanics data of the sole of the foot are collected, and the inertial measurement data and contact mechanics data are combined to form the limb proprioceptive feedback signal. The motor imagery EEG signal, frontal lobe attention representation signal and limb proprioceptive feedback signal are time-synchronized and aligned by hardware trigger signal or unified clock source. Signal quality analysis is performed on each modal signal to generate signal quality evaluation indicators that characterize the effectiveness of the signal, and the signal quality evaluation indicators are sent to the intent fusion and decoding module.

[0010] Preferably, the intent fusion and decoding module is configured as follows: Receives each modal signal sent by the multimodal physiological signal acquisition module; Motor imagery features are extracted from the motor imagery EEG signals, attention level features are extracted from the frontal lobe attention representation signals, and proprioceptive motor features are extracted from the limb proprioceptive feedback signals, respectively. The contribution of each modal feature in the fusion is determined based on the signal quality of each modal signal, and the motion imagination feature, attention level feature and propriomotion feature are fused according to the contribution to generate fused features; The fused features are input into a classifier to identify the user's motion intent category and the corresponding intent confidence level; The intent certainty metric is calculated based on the attention level features and the intent confidence, and then compared with a preset threshold. Motion control commands are selectively output based on the comparison results.

[0011] Preferably, the intent fusion and decoding module is configured as follows: The fusion weights of motor imagery EEG features, frontal lobe attention features, and propriomotor features are determined, and the sum of each fusion weight is a fixed value. The fusion weights are adjusted in real time based on the signal quality evaluation index of each modal signal. When the signal quality evaluation index of any modal signal decreases, the fusion weight of that modal feature is reduced, and the fusion weight of at least one other modal feature is increased accordingly. The modal features are weighted according to the modified fusion weights, and the weighted motor imagery EEG features, frontal lobe attention features, and proprioceptive features are spliced ​​or weighted summed to generate the fusion features.

[0012] Preferably, the intent fusion and decoding module determines the motion intent determinism according to the following logic: Obtain the intent confidence score, the normalized frontal attention level index, and the comprehensive signal quality evaluation value; The intention confidence score, the normalized frontal attention level index, and the comprehensive signal quality evaluation value are weighted and summed to obtain the intention certainty measure. When the intent determinism metric is not lower than the first determinism threshold for multiple consecutive control cycles, the motion control command is output. When the intent certainty metric is lower than the second certainty threshold, output no action command or maintain the current gait state; Wherein, the second deterministic threshold is less than the first deterministic threshold, so as to form an intent triggering logic with hysteresis characteristics.

[0013] Preferably, the gait parameter mapping module is configured as follows: Obtain the motion intent intensity output by the intent fusion and decoding module; The intensity of the motion intention is input into a nonlinear mapping model, which is used to characterize the nonlinear functional relationship between the intensity of the motion intention and the gait parameters, and outputs intermediate gait parameters. The intermediate gait parameters are constrained according to preset constraints to generate the final target gait parameters, which include at least the target stride length and the target gait speed.

[0014] The constraints include at least one of the following: gait parameter amplitude limits, joint range of motion limits, gait phase change rate limits, and environmental adaptive correction based on environmental state information.

[0015] Preferably, the environmental perception and safety control module is configured as follows: Acquire plantar pressure distribution information from the limb proprioceptive feedback signal, and identify the current support state based on the plantar pressure distribution information. The support state includes at least one of two-foot support, one-foot support, heel contact, and toe contact. Acquire inertial measurement data from the limb proprioceptive feedback signal, and calculate the real-time posture information and posture change rate of the exoskeleton based on the inertial measurement data. The posture information includes at least pitch angle and roll angle. Based on the support status, the real-time attitude information, and the attitude change rate, determine the terrain type and stability status of the current walking environment; When the current walking environment is determined to be a slope based on the terrain type, the slope direction and slope magnitude are determined based on the pitch angle, and the driving torque and range of motion of the knee and ankle joints are adjusted accordingly. When the current walking environment is determined to be uneven ground based on the plantar pressure distribution information and the real-time posture information, the target walking speed is reduced, the rate of change of gait parameters is limited, and / or the foot placement and joint range of motion are adjusted. When it is determined that the current state is abnormal or unstable based on the support status, the real-time posture information and the posture change rate, the maximum output torque of the exoskeleton joint is reduced, the target walking speed is reduced or the swing range is limited.

[0016] A second aspect of the present invention also provides a lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters, comprising: Simultaneously collect the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals; The motor imagery EEG signal, frontal lobe attention representation signal, and limb proprioceptive feedback signal are preprocessed, feature extracted, and fused to identify the user's motor intention. The determinism of the motor intention is evaluated based on the frontal lobe attention representation signal, intention confidence, and signal quality evaluation index. Motor control commands are selectively output according to the evaluation results. Based on the intensity of the motion intention corresponding to the motion control command, intermediate gait parameters are determined through a nonlinear mapping relationship. The intermediate gait parameters are then constrained according to at least one of the following constraints: gait parameter amplitude, joint range of motion, gait phase change rate, and environmental state, to obtain the target gait parameters. The real-time status information of the exoskeleton is obtained based on the limb proprioceptive feedback signal, and safety control commands are generated based on the real-time status information. The exoskeleton is driven to perform actions according to the target gait parameters and the safety control commands; In this context, when driving the exoskeleton to perform actions, the safety control commands have a higher execution priority than the motion control commands.

[0017] A third aspect of the present invention also provides an electronic device, characterized in that it comprises: The memory is used to store the processing program; The processor, when executing the processing program, implements the lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters as described above.

[0018] A fourth aspect of the present invention also provides a computer-readable storage medium, characterized in that the storage medium contains a computer program for implementing the lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters as described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1) Improve the precision of control: Through the nonlinear mapping between the intensity of motion imagination features and gait parameters, real-time and continuous adjustment of parameters such as stride length and speed of the exoskeleton is realized, making the exoskeleton walking more natural and flexible.

[0020] 2) Enhanced control accuracy and robustness: The introduction of frontal lobe attention level signals as an intention deterministic evaluation index effectively reduced the false trigger rate; multimodal signal fusion decoding improved the accuracy and anti-interference ability of motion intention recognition.

[0021] 3) Enhanced environmental adaptability and safety: Combined with the environmental perception module, the exoskeleton can adaptively adjust its gait according to changes in terrain and has an emergency stop mechanism, which significantly improves the user's walking safety in complex environments and the system's robustness.

[0022] 4) Promote the effectiveness of rehabilitation training: More natural and precise control methods help patients participate in rehabilitation training better, and improve training compliance and effectiveness. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall architecture of the lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of the fusion and decoding module in this embodiment of the invention. Figure 3 This is a schematic diagram illustrating the principle of the dynamic mapping model for gait parameters in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the process of the environmental perception and safety control module in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment provides a lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters, which mainly includes a multimodal physiological signal acquisition module, an intent fusion and decoding module, a gait parameter mapping module, an environmental perception and safety control module, and an exoskeleton execution module.

[0026] The multimodal physiological signal acquisition module is configured to simultaneously acquire the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals. Specifically, this module acquires the user's scalp EEG signals and extracts motor imagery EEG signals and frontal lobe attention representation signals from the scalp EEG signals based on electrode location or frequency band division; it also acquires inertial measurement data of the exoskeleton's lower limb joints and contact mechanics data of the sole of the foot, which together constitute the limb proprioceptive feedback signal. This module also performs time synchronization alignment of each modality signal through a hardware trigger signal or a unified clock source, and performs signal quality analysis on each modality signal to generate signal quality evaluation indicators to characterize signal effectiveness. High-precision EEG acquisition equipment (such as 64-lead or 128-lead electrode caps), as well as a high-sampling-rate IMU (such as the MPU-9250) and a plantar pressure sensor array (such as an FSR sensor) are used. All sensors are connected to the main control unit through a customized hardware interface board to achieve synchronous data acquisition.

[0027] The intent fusion and decoding module, connected to the multimodal physiological signal acquisition module, is configured to preprocess, extract features from, and fuse the motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals to identify the user's motor intent. It then evaluates the determinism of the motor intent based on the frontal lobe attention representation signals, intent confidence, and signal quality evaluation metrics, selectively outputting motor control commands according to the evaluation results. A pre-trained deep learning model is run on an embedded high-performance processor (such as an NVIDIA Jetson series or an ARM Cortex-A72). The model is deployed on an inference engine such as TensorRT to accelerate decoding speed. This module also includes a real-time intent determinism evaluation submodule, which determines whether to activate exoskeleton movement by comparing the attention level index and classifier confidence with a preset threshold.

[0028] The gait parameter mapping module, connected to the intent fusion and decoding module, is configured to determine the target gait parameters based on the intensity of the motion intent corresponding to the motion control command through a nonlinear mapping relationship. It also runs on an embedded processor, enabling real-time calculation of the nonlinear mapping model. The model parameters are stored in non-volatile memory and can be adjusted online or offline according to user-specific needs. The output of this module is the target angle, angular velocity, and torque commands for each joint of the exoskeleton.

[0029] The environmental perception and safety control module, connected to the intent fusion and decoding module, is configured to acquire real-time status information of the exoskeleton based on the limb proprioceptive feedback signals and generate safety control commands based on this real-time status information. An independent microcontroller (such as an STM32 series) processes IMU and plantar pressure sensor data for attitude calculation and terrain recognition. This module exchanges data with the main control unit via high-speed serial communication (such as a CAN bus) and independently runs a safety monitoring algorithm. Upon detecting an anomaly, it can directly send an emergency stop command to the exoskeleton execution module.

[0030] The exoskeleton execution module, connected to the gait parameter mapping module and the environmental perception and safety control module, is configured to drive the exoskeleton to perform actions based on the target gait parameters and the safety control commands. It consists of a high-power brushless DC motor, a reducer, a joint encoder, and a motor driver. The motor driver receives torque or position commands from the main control unit and precisely controls the motor output through PID or other advanced control algorithms to achieve the movement of the exoskeleton joints. The encoder provides joint angle feedback for closed-loop control.

[0031] In this context, the environmental perception and safety control module has a higher control priority over the exoskeleton execution module than the intent fusion and decoding module.

[0032] In this embodiment, the core of the system lies in converting user intent (acquired through multimodal physiological signals) into precise exoskeleton movements. Its working principle is a closed-loop control of perception-decision-execution-feedback. First, the multimodal physiological signal acquisition module acquires real-time information from the user's brain (EEG), attention (EEG), and limbs (IMU / EMG). This raw data is sent to the intent fusion and decoding module, which uses complex algorithms (such as deep learning classifiers) to identify the user's movement intent and assess its certainty. If the intent is clear, its intensity is passed to the gait parameter mapping module, which converts the intent intensity into specific gait parameters such as stride length and speed based on a preset nonlinear model. Simultaneously, the environmental perception and safety control module continuously monitors the exoskeleton's posture, plantar pressure, and environmental terrain, providing real-time environmental information for gait adjustment and safety assurance. Finally, the exoskeleton execution module receives instructions from the gait parameter mapping and safety control modules, driving the exoskeleton motors to achieve precise gait movements. Throughout the process, the exoskeleton's motion status and environmental information are continuously fed back to the environmental perception and safety control module, forming a dynamically adjusted closed loop to ensure the safety and stability of walking.

[0033] In one embodiment, the multimodal physiological signal acquisition module is configured to execute the following logic: The system collects the user's scalp EEG signals and extracts the motor imagery EEG signals and frontal lobe attention representation signals from these signals based on electrode location or frequency band division. The multimodal physiological signal acquisition module employs high-precision EEG acquisition equipment (such as 64- or 128-lead electrode caps), a high-sampling-rate inertial measurement unit (such as an MPU-9250), and a plantar pressure sensor array (such as a force-sensitive resistor sensor). All sensors are connected to the main control unit via a customized hardware interface board for synchronous data acquisition. The raw EEG signals are acquired using dry or wet electrode caps. Bandpass filtering (e.g., 0.5-40Hz) is applied to the raw EEG signals to remove DC drift and high-frequency noise, and a notch filter (e.g., 50 / 60Hz) is used to remove power frequency interference. Based on electrode location or frequency band division, motor imagery EEG signals and frontal lobe attention representation signals are obtained from the scalp EEG signals. Motor imagery EEG signals are EEG signals in motor imagery-related frequency bands (e.g., mu rhythm 8-12Hz, beta rhythm 18-30Hz). Frontal lobe attention representation signal acquisition involves performing power spectral density analysis on the data from the frontal lobe region electrodes in the original EEG signals, calculating the power spectral density ratio of a specific frequency band (e.g., theta / alpha ratio), and normalizing the ratio to a predetermined range (e.g., 0-1) to obtain the frontal lobe attention representation signal. A higher attention level index indicates a more focused user attention.

[0034] Inertial measurement data of the lower limb joints and contact mechanics data of the foot are collected from the exoskeleton. These inertial measurement data and contact mechanics data are combined to form the limb proprioceptive feedback signal. The limb proprioceptive feedback signal is acquired by collecting joint angle and angular velocity data through an inertial measurement unit installed at the lower limb joints of the exoskeleton, and by collecting ground reaction force data through a pressure sensor array integrated into the foot of the exoskeleton. The inertial measurement unit can be installed at the hip, knee, or ankle joints of the exoskeleton, providing triaxial acceleration, angular velocity, and magnetic field data. The foot pressure sensor uses a high-density force-sensitive resistor sensor array integrated into the insole or exoskeleton foot to provide pressure distribution data for different areas of the foot.

[0035] The motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals are synchronized in time using a hardware trigger signal or a unified clock source. Employing a hardware trigger synchronization mechanism, all sensors are clock-synchronized through the main control unit to ensure consistent timestamps for data acquisition and avoid data misalignment caused by asynchronous sampling.

[0036] Signal quality analysis is performed on each modality signal to generate signal quality evaluation indicators characterizing signal effectiveness. These indicators are then sent to the intent fusion and decoding module. The signal quality evaluation indicators include at least one or more of the following: artifact ratio, signal-to-noise ratio, and effective channel ratio of the EEG signal; saturation state of the inertial measurement unit signal; and effective contact ratio of the plantar pressure signal. The signal quality evaluation indicators are sent to the intent fusion and decoding module for weighted fusion calculation.

[0037] like Figure 2 As shown, the intent fusion and decoding module receives signals from various modalities sent by the multimodal physiological signal acquisition module, including motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals, as well as signal quality evaluation indicators corresponding to each modality. In practical applications, the motor imagery EEG signal is a microvolt-level raw EEG signal containing a large amount of noise and artifacts; the frontal lobe attention representation signal is an attention level index calculated based on the frontal lobe region EEG signals; and the limb proprioceptive feedback signal includes triaxial acceleration, angular velocity, and magnetic field data from the inertial measurement unit, as well as pressure distribution data from the plantar pressure sensor. This embodiment does not rely on a single EEG signal but constructs a fusion and decoding mechanism for three heterogeneous signals. Figure 2 As shown, the intention fusion and decoding module simultaneously receives motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals. These three signals describe the user's motor intention from three dimensions: cortical motor intention (ERD / ERS features), prefrontal cognitive state (attention level index), and limb physical response (joint angle / angular velocity / plantar pressure).

[0038] The intent fusion and decoding module is configured to execute the following processing logic: 1) Signal Preprocessing: This step aims to ensure signal quality and lay the foundation for subsequent feature extraction. For EEG signals, the raw EEG signal is typically in the microvolt range and contains a large amount of noise and artifacts. To address this characteristic, a bandpass filter (e.g., 0.5-40Hz) is used to remove DC drift and high-frequency noise, a notch filter (e.g., 50 / 60Hz) is used to remove power frequency interference, and further independent component analysis (ICA) or wavelet transform is used to remove artifacts such as electrooculography (EOG) and electromyography (EMG). For attention level signals, it can be the theta / alpha power ratio derived from EEG. To reduce signal abrupt changes, the attention index derived from EEG is smoothed. For proprioceptive feedback signals, it includes triaxial acceleration, angular velocity, and magnetic field data provided by the IMU, and muscle electrical activity intensity provided by the EMG. To address this characteristic, Kalman filtering or complementary filtering is applied to the IMU data for attitude calculation to obtain accurate kinematic parameters; rectification, low-pass filtering, and root mean square (RMS) calculation are performed on the EMG data (if available) to obtain muscle activation intensity.

[0039] 2) Feature Extraction: Motor imagery features are extracted from the motor imagery EEG signals, attention level features are extracted from the frontal lobe attention representation signals, and proprioceptive motor features are extracted from the limb proprioceptive feedback signals. Specifically, for the motor imagery EEG signals, time-frequency analysis is performed using short-time Fourier transform or wavelet transform to extract the power spectral density changes in the mu rhythm (8-12Hz) and beta rhythm (18-30Hz) frequency bands, calculate the event-related desynchronization or event-related synchronization amplitude, and further employ a co-spatial mode algorithm to extract discriminative spatial features that distinguish different motor imagery tasks as motor imagery features. The intensity of motor imagery features can be quantified by the percentage change in energy of event-related desynchronization / synchronization, with a range set from 0-100%.

[0040] For frontal lobe attention representation signals, the power spectral density ratio of a specific frequency band (such as the power ratio of the theta band to the alpha band) is calculated based on preprocessed frontal lobe EEG data. This ratio is then normalized to the range of 0-1 and used as an attention level feature. A higher value indicates a more focused user attention. For example, the normalized value of Attention_Index = (Theta_Power / Alpha_Power) is used, where Attention_Index is the attention level index, Theta_Power represents the power spectral density of the theta band, which refers to the energy level of the theta band (usually 4-8 Hz) in the EEG signal, and Alpha_Power represents the power spectral density of the alpha band, which refers to the energy level of the alpha band (usually 8-13 Hz) in the EEG signal.

[0041] For limb proprioceptive feedback signals, kinematic parameters such as joint angles, joint angular velocities, and limb end-effector velocities are extracted from inertial measurement unit data, and the center of plantar pressure (COP) trajectory and pressure distribution characteristics are extracted from plantar pressure data, which together constitute proprioceptive motion characteristics.

[0042] 3) Adaptive Weighted Fusion: The contribution of each modality feature in the fusion is determined based on the signal quality of each modality signal. The motor imagery feature, attention level feature, and proprioceptive feature are then fused according to their respective contributions to generate a fused feature. Specifically, this module determines the fusion weights for the motor imagery EEG feature, frontal lobe attention feature, and proprioceptive feature, and the sum of these weights is normalized to a fixed value. Initial weights can be set to 0.6 for motor imagery EEG features, 0.2 for frontal lobe attention features, and 0.2 for proprioceptive features. Subsequently, the fusion weights are adjusted in real-time based on the signal quality evaluation index of each modality signal. When the signal quality evaluation index of any modality signal decreases (e.g., increased EEG noise leading to a lower signal-to-noise ratio, or poor electrode contact leading to a lower effective channel ratio), the fusion weight of that modality feature is reduced, and the fusion weight of at least one other modality feature is correspondingly increased. Conversely, when the quality of a modality signal improves, the fusion weight of that modality feature is correspondingly increased. Through this adaptive weight adjustment mechanism, modal signals with higher signal quality contribute more to the fusion, while modal signals with lower signal quality contribute less. The modal features are weighted according to the revised fusion weights, and the weighted motor imagery EEG features, frontal lobe attention features, and proprioceptive features are spliced ​​or weighted summed to generate fusion features.

[0043] 4) The fused features are input into a classifier to identify the user's motion intention category and corresponding intention confidence. This classifier employs a hybrid model combining a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. The CNN layer extracts the spatiotemporal patterns of motion imagery features; its structure consists of 3-5 convolutional layers, each with 32 filters and a 3×3 kernel, using the ReLU activation function. The LSTM layer processes the temporal series information of the body's motion features; its structure consists of 2 LSTM layers, each with 64 units. Attention level features are used as additional input features. After the outputs of the CNN and LSTM layers are fused, they are processed by two fully connected layers using the ReLU activation function. The final output layer selects the Softmax activation function based on the intention type to output the probability distribution of each intention category, with the highest probability value serving as the intention confidence. The motion intention categories output by the classifier can include stepping with the left foot, stepping with the right foot, and stopping.

[0044] This classifier is trained using supervised learning with a labeled multimodal dataset. Cross-entropy loss is used as the loss function, and Adam is employed as the optimizer. Backpropagation is used to adjust network weights. The training dataset contains multimodal physiological signal data from at least 20 healthy subjects and 10 patients with limb dysfunction, covering various states including motor imagery, relaxation, concentration, and distraction. Each state lasts 5-10 minutes and is repeated 10-20 times. K-fold cross-validation is used for model training and validation. Evaluation metrics include accuracy, F1 score, false trigger rate, and control latency.

[0045] 5) Intent certainty assessment and decision output: The intent fusion and decoding module is first configured to execute feature fusion logic: calculate an intent certainty metric based on the attention level features and the intent confidence, and compare the intent certainty metric with a preset threshold. The intent fusion and decoding module is configured to: determine the fusion weights of motor imagery EEG features, frontal lobe attention features, and proprioceptive features, with the sum of each fusion weight being a fixed value; adjust the fusion weights in real time based on the signal quality evaluation index of each modality signal; when the signal quality evaluation index of any modality signal decreases, reduce the fusion weight of that modality feature and correspondingly increase the fusion weight of at least one other modality feature; weight each modality feature according to the adjusted fusion weights, and perform feature concatenation or weighted summation on the weighted motor imagery EEG features, frontal lobe attention features, and proprioceptive features to generate the fused feature.

[0046] After generating fused features and identifying the intent confidence score through a classifier, the intent fusion and decoding module determines the determinism of the motion intent according to the following logic: The intent confidence score (i.e., the highest probability value output by the classifier), the normalized frontal attention level index, and the comprehensive signal quality evaluation value are obtained. The comprehensive signal quality evaluation value is calculated based on the weighted sum or weighted average of the signal quality evaluation indicators corresponding to each modality signal. These signal quality evaluation indicators include at least one or more of the following: artifact ratio of EEG signal, signal-to-noise ratio, effective channel ratio, saturation state of inertial measurement unit signal, and effective contact ratio of plantar pressure signal.

[0047] The intent certainty metric is obtained by weighted summation of the intent confidence, the normalized frontal attention level index, and the comprehensive signal quality evaluation value. In a preferred implementation, the intent certainty metric is calculated as follows: Intent_Certainty = Attention_level * Classifier_Confidence, where Intent_Certainty is the intent certainty metric; Attention_level is the normalized frontal attention level index, reflecting the user's current level of concentration; and Classifier_Confidence is the intent confidence, i.e., the highest probability value output by the classifier, reflecting the classifier's confidence in the recognition result. Through this product calculation, intent certainty effectively reflects the joint constraint relationship between attention level and classifier confidence: when user attention is distracted, causing Attention_level to decrease, even if the classifier outputs a high Classifier_Confidence, the intent certainty metric will decrease significantly. In another implementation, the intent certainty metric can be expressed as C = α·P + β·A + γ·Q, where C is the intent certainty metric, P is the intent confidence level, A is the normalized frontal attention level index, Q is the comprehensive signal quality evaluation value, and α, β, and γ are preset fusion coefficients. Through this calculation method, intent certainty simultaneously reflects the classifier's grasp of the recognition result, the user's current level of concentration, and the reliability of signal acquisition.

[0048] Based on the comparison results, motion control commands are selectively output. A first deterministic threshold (e.g., THRESHOLD_CERTAINTY is set to 0.6) and a second deterministic threshold less than the first deterministic threshold are set to form an intention triggering logic with hysteresis characteristics. When the intention deterministic metric is not lower than the first deterministic threshold for multiple consecutive control cycles, the user's intention is determined to be clear, and the system responds to the motion intention. At this time, the decoded target intention takes the maximum value in the classifier's output probability distribution, i.e., Decoded_Intent = ArgMax(Classifier_Output), and the motion control command and the corresponding motion intention intensity are output. When the intention deterministic metric is lower than the second deterministic threshold, the user's intention is determined to be unclear, the system does not respond to the motion intention, and outputs a no-action command (i.e., Decoded_Intent = NO_ACTION) or maintains the current gait state.

[0049] In practical applications, when the user's attention is focused and the EEG signal is clear, the intent certainty remains high, and the system responds normally to the motor intent. However, when the user's attention is distracted, causing a decrease in the frontal lobe attention level index, even if the classifier outputs a high intent confidence score, the intent certainty may still be below the first certainty threshold. Therefore, the system does not trigger exoskeleton movements, thus avoiding false triggers caused by attention fluctuations. For example, when a user attempts to visualize movement with their left or right foot, the EEG signal generates a specific event-related desynchronization / synchronization (ERD / ERS) pattern, which is extracted as a motor imagery feature. However, EEG signals are easily interfered with by environmental noise, poor electrode contact, and other factors. In such cases, the motor imagery feature may still be recognized as a valid intent by the classifier, resulting in false triggers. To overcome this problem, this embodiment introduces the frontal lobe attention level signal as an independent evaluation dimension for intent certainty. When the user's attention level is high, the certainty of their motor intent is also higher; when the user's attention is distracted, even if the classifier outputs a high intent confidence score, the system will refuse to trigger the action due to insufficient intent certainty. Meanwhile, the limb proprioceptive feedback signal serves as a third information source, further verifying or correcting the accuracy of EEG signal decoding through the subtle limb movements or posture changes detected by the IMU.

[0050] The three signals are fused using an adaptive weighted fusion strategy. Initial weights can be set to 0.6 for the EEG signal, 0.2 for the attention level signal, and 0.2 for the proprioceptive feedback signal. The system dynamically adjusts the fusion weights based on the real-time signal quality evaluation metrics of each modality. When the quality of a signal deteriorates (e.g., increased noise in the EEG signal leads to a lower signal-to-noise ratio), the fusion weight of that signal is automatically reduced, while the weights of the remaining signal branches are increased accordingly. This ensures that the fusion feature is always dominated by the signal with the highest current quality.

[0051] After the fusion feature input classifier identifies the motion intention category and intention confidence, the system further performs secondary verification through an intention determinism evaluation mechanism. The system only outputs motion control commands when the intention determinism is not lower than the first determinism threshold for multiple consecutive control cycles; if it is lower than the second determinism threshold, the system outputs no action command or maintains the current state. Simultaneously, the second determinism threshold is lower than the first determinism threshold, creating a hysteresis characteristic to prevent the exoskeleton's movements from frequently starting and stopping at the threshold boundaries. Using this module, even if a certain modal signal is interfered with, the system can still maintain reliable intention recognition through the complementarity of other modal signals and adaptive weight adjustment; even if the classifier outputs a high-confidence result, the lack of attention level can still prevent the system from responding to "unintentional but misidentified" motion intentions, thereby effectively reducing the false trigger rate.

[0052] Through the multi-dimensional deterministic evaluation mechanism of intent confidence, attention level, and signal quality, the exoskeleton movement is activated only when the user's intent is clear enough and the signal quality is reliable, effectively reducing the false trigger rate. At the same time, the hysteresis feature design prevents the exoskeleton movement from frequently starting and stopping at the threshold boundary due to small signal fluctuations.

[0053] like Figure 3 As shown, the gait parameter mapping module is configured to: acquire the motion intention intensity output by the intention fusion and decoding module; input the motion intention intensity into a nonlinear mapping model, which characterizes the nonlinear functional relationship between the motion intention intensity and gait parameters, and outputs intermediate gait parameters; constrain the intermediate gait parameters according to preset constraints to generate the final target gait parameters. This module transforms the abstract motion imagination intensity into specific, continuous exoskeleton gait parameters. The motion imagination feature intensity is input, which is one of the outputs of the intention fusion and decoding module and represents the intensity of the user's motion intention. This intensity value is fed into the nonlinear mapping model. This model is pre-trained using experimental data and can capture the complex nonlinear relationship between motion imagination intensity and desired gait parameters. Internally, the model uses mathematical functions (such as polynomials and neural networks) to convert the input intensity value into the corresponding target stride length and target gait speed. For example, when the user's motion imagination intensity is low, the model outputs a smaller stride length and gait speed to simulate slow walking; when the intensity increases, the model outputs a larger stride length and gait speed to simulate fast walking. These target gait parameters are then passed to the exoskeleton execution module to guide the actual movement of the exoskeleton.

[0054] Specifically, the intensity of motor intention is quantified by the percentage change in ERD / ERS energy of the motor imagery EEG signal, with a quantization range of 0-100%. This intensity value represents the strength of the user's motor intention. The nonlinear mapping model aims to transform the abstract intensity of motor imagery into specific, continuous exoskeleton gait parameters.

[0055] In one specific implementation, the nonlinear mapping model can use a polynomial fitting function to establish the mapping relationship between the intensity of movement intention and stride speed and stride length. The specific formula is as follows: ; Wherein, MI_Intensity represents the intensity of the motor intention, ranging from 0 to 100%; Speed ​​represents the target gait speed, ranging from 0.1 to 0.8 m / s; Length represents the target stride length, ranging from 0.3 to 0.8 m; and A1, B1, C1, A2, B2, and C2 are coefficients determined through training. This quadratic polynomial can characterize the nonlinear characteristics of the relationship between motor intention intensity and gait parameters, characterized by "gradual changes in the low-intensity region, sensitive changes in the medium-intensity region, and saturation in the high-intensity region," which is more consistent with the physiological laws of human motor control compared to linear mapping.

[0056] As another preferred implementation, to more accurately capture the complex nonlinear characteristics of human gait, the nonlinear mapping model employs a small feedforward neural network (FNN). Through hidden layers, it automatically discovers higher-order correlations between the intensity of motor intention and gait parameters, achieving more refined and natural gait adjustment. This network structure specifically includes: an input layer with one node receiving the normalized intensity of motor intention; one or two hidden layers, each with 10-20 nodes, using the ReLU activation function to introduce nonlinear features; and an output layer with two nodes, outputting the target stride length and target stride speed respectively, using a linear activation function. Compared to traditional linear mapping, this feedforward neural network model can learn and approximate arbitrarily complex nonlinear functions. By automatically discovering higher-order correlations between the intensity of motor intention and gait parameters through hidden layers, the exoskeleton's response more closely aligns with human biomechanical characteristics and the user's subjective experience, achieving more refined and natural gait adjustment. Furthermore, this network structure allows for rapid personalized adjustments using a small amount of training data to adapt to the gait habits of different users.

[0057] The parameters of the nonlinear mapping model are determined by collecting stride length and gait speed data from multiple subjects under different intensities of motor imagery, representing their desired exoskeleton performance. Training data is derived from gait parameters precisely measured by external sensors (such as photogates and high-speed cameras) when subjects wear the exoskeleton and perform motor imagery tasks in a laboratory environment. For polynomial models, the least squares method is used for parameter fitting; for neural network models, a neural network training algorithm (such as backpropagation) is used for parameter optimization. After model training, the parameters are stored in non-volatile memory and can be adjusted online or offline according to the user's personalized needs.

[0058] After outputting intermediate gait parameters, the module further constrains these parameters according to preset constraints to generate final target gait parameters. The target gait parameters include at least target stride length and target gait speed. The constraints include at least one of the following: gait parameter amplitude limits (e.g., limiting gait speed to a safe threshold), joint range of motion limits (e.g., limiting hip and knee joint angles to within physiological ranges), gait phase change rate limits, and environmental adaptive correction based on environmental state information. For example, when the user's motor imagination intensity is low, the model outputs a smaller stride length and gait speed to simulate slow walking; when the intensity increases, the model outputs a larger stride length and gait speed to simulate fast walking. The final target gait parameters generated after constraint processing are converted into target angle, angular velocity, and torque commands for each joint of the exoskeleton and transmitted to the exoskeleton execution module.

[0059] This embodiment does not decode motion intention into discrete "start / stop" or "fast / slow" levels. Instead, it extracts continuous motion intention intensity values ​​(MI_Intensity, quantization range 0-100%) and converts them into continuous target gait parameters through a nonlinear mapping model. Specifically, when the user's motion intention intensity is low (e.g., 20%), the system maps it to a smaller stride (e.g., 0.35 meters) and walking speed (e.g., 0.2 meters / second) to simulate slow walking. When the intensity increases to 50%, the stride and walking speed increase accordingly. When the intensity reaches above 80%, the stride and walking speed approach the user's upper limit (stride 0.8 meters, walking speed 0.8 meters / second) to simulate fast walking. The entire process is continuous and smooth, allowing the user to naturally control the exoskeleton's walking speed and stride by adjusting the intensity of the motion intention. Unlike traditional simple linear mapping, this embodiment uses a nonlinear mapping model to fit the biomechanical characteristics of human gait. With the above modules, users no longer need to switch between limited preset levels. Instead, they can continuously and infinitely adjust the stride and speed of the exoskeleton by the intensity of their motor imagination, making the exoskeleton walking more natural, flexible, and adaptable to personalized needs.

[0060] like Figure 4 As shown, the environmental perception and safety control module is configured to execute environmental perception, safety judgment, and emergency stop logic to build a safety assurance mechanism independent of intent decoding. The overall workflow of this module begins with sensor data input, followed by data preprocessing, environmental and balance assessment, and finally, the safety policy decision module generates and outputs safety control commands. The specific implementation logic is as follows: Data Acquisition and Preprocessing: The plantar pressure distribution information from the limb proprioceptive feedback signal is acquired, and the current support state is identified based on this information. Inertial measurement data from the limb proprioceptive feedback signal is acquired, and the real-time posture information and rate of change of the exoskeleton are calculated based on this data. Specifically, a high-density force-sensitive resistor (FSR) array is integrated into the insole or exoskeleton foot to provide pressure distribution data for different areas of the foot. Multiple inertial measurement units (IMUs, such as MPU-9250) are installed at the hip, knee, and ankle joints of the exoskeleton to provide triaxial acceleration, angular velocity, and magnetic field data. After acquiring the above raw data, the data preprocessing stage begins: the plantar pressure data is calibrated, denoised, and normalized to extract the center of plantar pressure (COP) and total pressure value; Kalman filtering or complementary filtering is applied to the IMU data, and accelerometer, gyroscope, and magnetometer data are fused to achieve high-precision posture calculation. The real-time posture information of the exoskeleton is represented using quaternions or Euler angles.

[0061] Environmental Information Recognition and Balance Assessment: The current support state and terrain information are identified based on the preprocessed data. The current support state is identified based on the plantar pressure distribution information, including at least one of the following: two-foot support, one-foot support, heel contact, toe contact, left-foot support, right-foot support, and both feet off the ground. Simultaneously, the real-time attitude information of the exoskeleton is calculated based on the inertial measurement data, including at least pitch angle, roll angle, yaw angle, and attitude change rate.

[0062] Building upon this foundation, based on plantar pressure distribution patterns and IMU posture change rates, machine learning classifiers (such as SVM or neural networks) are used to identify the current terrain type (e.g., flat ground, uphill, downhill, obstacles, etc.). For example, a heel-first strike followed by toe-back and forward COP pattern may indicate walking on flat ground; a continuously backward COP and forward-leaning posture may indicate uphill walking. Simultaneously, based on posture data, angular velocity, and COP position, the zero-moment point (ZMP) or gait stability margin is calculated to assess the exoskeleton's dynamic balance.

[0063] Adaptive fine-tuning control: Based on the support status, the real-time attitude information, and the attitude change rate, determine the terrain type and stability status of the current walking environment.

[0064] Based on the above recognition results, this module executes an adaptive adjustment strategy: Level 1 - Slope Adaptive: When the current walking environment is determined to be a slope based on the terrain category, the slope direction and magnitude are determined based on the pitch angle, and the driving torque and range of motion of the knee and ankle joints are adjusted accordingly. For example, when going uphill, the ankle plantar flexion / dorsiflexion torque and knee flexion trajectory are adjusted in synergy based on the plantar pressure center, exoskeleton pitch angle, and joint motion state; the adjustment direction and amplitude are determined by the calibrated terrain state and stability control law, rather than a fixed increase in torque in a single direction.

[0065] Level 2 - Uneven Ground Adaptation: When the current walking environment is determined to be uneven ground based on the plantar pressure distribution information, the target gait parameters are adjusted to increase the support area and reduce the target walking speed. By expanding the plantar contact area, stability is improved and falls are prevented.

[0066] Level 3 - Support Anomaly Adaptation: When it is determined that the current state is in a support anomaly or posture instability state based on the support state, the real-time posture information and the posture change rate (such as a large posture disturbance in the single-leg support state), the joint torque and angle are adjusted in real time through model predictive control (MPC) or adaptive PID control algorithm, or the maximum output torque of the exoskeleton joint is reduced, the target walking speed is reduced or the swing range is limited, in order to maintain dynamic balance.

[0067] Furthermore, the environmental perception and safety control module is configured to execute the following safety judgment logic: after the motion intention command is issued, the system further performs intention-support state consistency verification: It obtains the motion intention category output by the intention fusion and decoding module, which includes at least a left step intention, a right step intention, and a stop intention; it obtains the current support state determined based on the plantar pressure distribution information, which includes at least one of left foot support, right foot support, both feet support, and both feet off the ground; it compares the motion intention category with the current support state: when the motion intention category is a left step intention and the current support state is not right foot support, it is determined to be an abnormal support state, triggering safety intervention; when the motion intention category is a right step intention and the current support state is not left foot support, it is determined to be an abnormal support state, triggering safety intervention; when the motion intention category is a stop intention and the current support state is not both feet support, it triggers safety intervention; the safety intervention includes at least one of reducing the exoskeleton joint output torque, reducing the target step speed, or limiting the joint swing range, to prevent unsafe actions caused by incorrect intention recognition or misjudgment of support state.

[0068] The environmental perception and safety control module is also equipped with an emergency stop mechanism. Abnormal posture detection monitors in real time whether the angles of each joint of the exoskeleton exceed preset safety ranges, such as limiting the knee flexion angle to 0-120 degrees, and whether the overall tilt angle of the exoskeleton (roll angle, pitch angle) exceeds safety thresholds, such as a roll angle greater than 15 degrees or a pitch angle greater than 20 degrees. When an abnormal posture is detected, or when an emergency stop signal is received from a physical emergency stop button, a voice recognition unit, or a specific EEG safety mode, an emergency stop signal is immediately sent to the exoskeleton execution module. The environmental perception and safety control module uses an independent microcontroller (such as the STM32 series) to process IMU and plantar pressure sensor data, perform posture calculation and terrain recognition, exchange data with the main control unit via high-speed serial communication (such as a CAN bus), and independently run safety monitoring algorithms. This module sends emergency stop control commands directly to the exoskeleton execution module through an independent safety control channel, causing the exoskeleton execution module to stop responding to control commands from the intent fusion and decoding module and the gait parameter mapping module, and to perform power output cut-off or joint braking operations. The generation and transmission of emergency stop control commands do not depend on the operation status of the intent fusion and decoding module, ensuring the safety of users in emergency situations.

[0069] In this embodiment, the environmental perception and safety control module ensures the exoskeleton's dynamic balance and user safety in complex environments. Sensor data input includes plantar pressure sensor data and IMU data. This data first enters the data preprocessing stage, where it is filtered and calibrated. Subsequently, the environmental information recognition module uses plantar pressure data to determine the ground contact state (e.g., single-leg support, double-leg support, off-ground) and combines it with IMU data for terrain recognition (e.g., flat ground, uphill, downhill, obstacles). Simultaneously, the attitude calculation and balance assessment module uses IMU data to calculate the exoskeleton's attitude (pitch angle, roll angle, yaw angle) and angular velocity in real time, and assesses its dynamic balance state. The safety strategy decision module integrates environmental information, balance assessment results, and user intent commands to generate safety control commands. This module includes abnormal attitude detection and emergency stop triggering logic. If an abnormal attitude is detected (e.g., excessive tilt angle, joint angle exceeding limits) or a specific safety command is received from the user, an emergency stop is immediately triggered. Otherwise, the exoskeleton's power output and joint angles are fine-tuned based on environmental information to adapt to terrain changes and maintain balance. Ultimately, safety control commands (including fine-tuning commands or emergency stop commands) are transmitted to the exoskeleton execution module. The environmental perception and safety control module uses a microcontroller (such as the STM32 series) independent of the main controller, exchanges data with the main control unit via a CAN bus, and runs safety monitoring algorithms independently. This module sends emergency stop control commands directly to the exoskeleton execution module through an independent safety control channel, causing the exoskeleton execution module to stop responding to control commands from the intent fusion and decoding module and the gait parameter mapping module, and to perform power output cutoff or joint braking. The generation and transmission of emergency stop control commands do not depend on the operational state of the intent fusion and decoding module; even if the main control system crashes or experiences delays, the safety module can still independently complete emergency braking.

[0070] In one embodiment, the complete control flow timing from signal acquisition to exoskeleton execution is as follows: signal acquisition phase 0-50ms (synchronous acquisition by EEG, IMU, and pressure sensors); signal preprocessing phase 50-100ms (denoising, filtering, and feature extraction); intent decoding and deterministic evaluation phase 100-200ms (multimodal fusion, classifier calculation, and threshold judgment); gait parameter mapping phase 200-250ms (nonlinear model calculation); environmental perception and safety control phase 250-300ms (posture judgment, terrain recognition, and fine-tuning calculation); and exoskeleton execution phase 300-350ms (motor drive and joint movement). The overall system response latency is controlled within 350ms. This is ensured by employing high-performance embedded processors, optimizing algorithm efficiency, hardware acceleration (such as FPGA or GPU), and a distributed computing architecture to guarantee real-time performance and user experience.

[0071] Based on the same concept, this invention provides a lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters, including: Simultaneously collect the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals; The motor imagery EEG signal, frontal lobe attention representation signal, and limb proprioceptive feedback signal are preprocessed, feature extracted, and fused to identify the user's motor intention. The determinism of the motor intention is evaluated based on the frontal lobe attention representation signal, intention confidence, and signal quality evaluation index. Motor control commands are selectively output according to the evaluation results. Based on the intensity of the motion intention corresponding to the motion control command, intermediate gait parameters are determined through a nonlinear mapping relationship. The intermediate gait parameters are then constrained according to at least one of the following constraints: gait parameter amplitude, joint range of motion, gait phase change rate, and environmental state, to obtain the target gait parameters. The real-time status information of the exoskeleton is obtained based on the limb proprioceptive feedback signal, and safety control commands are generated based on the real-time status information. The exoskeleton is driven to perform actions based on the target gait parameters and the safety control commands; wherein, when driving the exoskeleton to perform actions, the execution priority of the safety control commands is higher than that of the motor control commands. It can be understood that during normal walking, the multimodal physiological signal acquisition module continuously and synchronously acquires motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals, and performs time synchronization alignment and signal quality assessment, generating signals for each modality and corresponding signal quality evaluation indicators, which are then sent to the intention fusion and decoding module. Upon receiving the above signals, the intention fusion and decoding module sequentially performs preprocessing, feature extraction, adaptive weighted fusion, classifier recognition, and intention determinism assessment. When the intention determinism is not lower than the first determinism threshold for several consecutive control cycles, the motor control command and the corresponding motor intention intensity are output to the gait parameter mapping module; otherwise, no action command is output, and subsequent control processes are not activated. After receiving the intensity of the motion intention, the gait parameter mapping module inputs it into a nonlinear mapping model and outputs intermediate gait parameters (target stride length and target stride speed). Then, it performs constraint processing according to preset constraints (amplitude limits, joint range limits, rate of change limits, and environmental adaptive correction) to generate the final target gait parameters. These parameters are then converted into target angle, angular velocity, and torque commands for each joint of the exoskeleton and sent to the exoskeleton execution module. Simultaneously, the environmental perception and safety control module operates independently, continuously calculating support status, posture information, and terrain type based on plantar pressure distribution information and inertial measurement data. After the motion intention command is issued, this module performs intention-support status consistency verification and generates safety control commands (fine-tuning, downgrading, or emergency stop) based on terrain recognition and balance assessment results. This module has the highest control priority over the exoskeleton execution module, and its safety control commands can override or interrupt motion commands from the gait parameter mapping module. The data flow and control flow relationships between the modules have been described in the previous embodiments and will not be repeated here.

[0072] It should be noted that the division of the various modules in this device / system embodiment is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software through processing element calls; they can also be implemented entirely in hardware; or some units can be implemented by processing element calls to software, while others can be implemented in hardware.

[0073] Based on the same concept, an electronic device is also provided in some embodiments of this application. This electronic device includes a memory and a processor, wherein the memory stores a processing program, and the processor executes the processing program according to instructions. When the processor executes the processing program, the lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters described in the foregoing embodiments is realized.

[0074] In some embodiments of this application, a readable storage medium is also provided, which can be a non-volatile readable storage medium or a volatile readable storage medium. The readable storage medium stores instructions that, when executed on a computer, cause an electronic device containing such a readable storage medium to perform the aforementioned lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters.

[0075] It is understood that the aforementioned lower limb exoskeleton control methods based on multimodal intent fusion and dynamic adjustment of gait parameters, if implemented as software functional modules and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0077] The program code for executing the technical solutions disclosed in this application can be written in any combination of one or more programming languages. These programming languages ​​include object-oriented programming languages ​​such as Python and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters, characterized in that, include: The multimodal physiological signal acquisition module is configured to simultaneously acquire the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals. The intention fusion and decoding module is connected to the multimodal physiological signal acquisition module and is configured to preprocess, extract features and fuse the motor imagery EEG signal, frontal lobe attention representation signal and limb proprioceptive feedback signal to identify the user's motor intention, and evaluate the certainty of the motor intention based on the frontal lobe attention representation signal, intention confidence and signal quality evaluation index, and selectively output motor control commands according to the evaluation results. The gait parameter mapping module, connected to the intent fusion and decoding module, is configured to determine intermediate gait parameters based on the intensity of the motion intent corresponding to the motion control command through a nonlinear mapping relationship, and to constrain the intermediate gait parameters based on at least one of the constraints of gait parameter amplitude, joint range of motion, gait phase change rate and environmental state to obtain target gait parameters. The environmental perception and safety control module is connected to the intent fusion and decoding module and is configured to obtain the real-time status information of the exoskeleton based on the limb proprioceptive feedback signal and generate safety control commands based on the real-time status information. The exoskeleton execution module is connected to the gait parameter mapping module and the environment perception and safety control module, and is configured to drive the exoskeleton to perform actions according to the target gait parameters and the safety control commands. In this context, the environmental perception and safety control module has a higher control priority over the exoskeleton execution module than the intent fusion and decoding module.

2. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 1, characterized in that, The multimodal physiological signal acquisition module is configured to execute the following logic: The user's scalp EEG signals are collected, and the motor imagery EEG signals and the frontal lobe attention representation signals are obtained from the scalp EEG signals based on electrode location or frequency band division. Inertial measurement data of the lower limb joints of the exoskeleton and contact mechanics data of the sole of the foot are collected, and the inertial measurement data and contact mechanics data are combined to form the limb proprioceptive feedback signal. The motor imagery EEG signal, frontal lobe attention representation signal and limb proprioceptive feedback signal are time-synchronized and aligned by hardware trigger signal or unified clock source. Signal quality analysis is performed on each modal signal to generate signal quality evaluation indicators that characterize the effectiveness of the signal, and the signal quality evaluation indicators are sent to the intent fusion and decoding module.

3. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 1, characterized in that, The intent fusion and decoding module is configured as follows: Receives each modal signal sent by the multimodal physiological signal acquisition module; Motor imagery features are extracted from the motor imagery EEG signals, attention level features are extracted from the frontal lobe attention representation signals, and proprioceptive motor features are extracted from the limb proprioceptive feedback signals, respectively. The contribution of each modal feature in the fusion is determined based on the signal quality of each modal signal, and the motion imagination feature, attention level feature and propriomotion feature are fused according to the contribution to generate fused features; The fused features are input into a classifier to identify the user's motion intent category and the corresponding intent confidence level; The intent certainty metric is calculated based on the attention level features and the intent confidence, and then compared with a preset threshold. Motion control commands are selectively output based on the comparison results.

4. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 3, characterized in that, The intent fusion and decoding module is configured as follows: The fusion weights of motor imagery EEG features, frontal lobe attention features, and propriomotor features are determined, and the sum of each fusion weight is a fixed value. The fusion weights are adjusted in real time based on the signal quality evaluation index of each modal signal. When the signal quality evaluation index of any modal signal decreases, the fusion weight of that modal feature is reduced, and the fusion weight of at least one other modal feature is increased accordingly. The modal features are weighted according to the modified fusion weights, and the weighted motor imagery EEG features, frontal lobe attention features, and proprioceptive features are spliced ​​or weighted summed to generate the fusion features.

5. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 3, characterized in that, The intent fusion and decoding module determines the motion intent determinism according to the following logic: Obtain the intent confidence score, the normalized frontal attention level index, and the comprehensive signal quality evaluation value; The intention confidence score, the normalized frontal attention level index, and the comprehensive signal quality evaluation value are weighted and summed to obtain the intention certainty measure. When the intent determinism metric is not lower than the first determinism threshold for multiple consecutive control cycles, the motion control command is output. When the intent certainty metric is lower than the second certainty threshold, output no action command or maintain the current gait state; Wherein, the second deterministic threshold is less than the first deterministic threshold, so as to form an intent triggering logic with hysteresis characteristics.

6. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 1, characterized in that, The gait parameter mapping module is configured as follows: Obtain the motion intent intensity output by the intent fusion and decoding module; The intensity of the motion intention is input into a nonlinear mapping model, which is used to characterize the nonlinear functional relationship between the intensity of the motion intention and the gait parameters, and outputs intermediate gait parameters. The intermediate gait parameters are constrained according to preset constraints to generate the final target gait parameters, which include at least the target stride length and the target gait speed. The constraints include at least one of the following: gait parameter amplitude limits, joint range of motion limits, gait phase change rate limits, and environmental adaptive correction based on environmental state information.

7. The lower limb exoskeleton control system based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 1, characterized in that, The environmental perception and safety control module is configured as follows: Acquire plantar pressure distribution information from the limb proprioceptive feedback signal, and identify the current support state based on the plantar pressure distribution information. The support state includes at least one of two-foot support, one-foot support, heel contact, and toe contact. Acquire inertial measurement data from the limb proprioceptive feedback signal, and calculate the real-time posture information and posture change rate of the exoskeleton based on the inertial measurement data. The posture information includes at least pitch angle and roll angle. Based on the support status, the real-time attitude information, and the attitude change rate, determine the terrain type and stability status of the current walking environment; When the current walking environment is determined to be a slope based on the terrain type, the slope direction and slope magnitude are determined based on the pitch angle, and the driving torque and range of motion of the knee and ankle joints are adjusted accordingly. When the current walking environment is determined to be an uneven ground based on the plantar pressure distribution information and the real-time posture information, the target walking speed is reduced, the rate of change of the target gait parameters is limited, and / or the foot placement and joint range of motion are adjusted. When it is determined that the current state is abnormal or unstable based on the support status, the real-time posture information and the posture change rate, the maximum output torque of the exoskeleton joint is reduced, the target walking speed is reduced or the swing range is limited.

8. A lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters, characterized in that, include: Simultaneously collect the user's motor imagery EEG signals, frontal lobe attention representation signals, and limb proprioceptive feedback signals; The motor imagery EEG signal, frontal lobe attention representation signal, and limb proprioceptive feedback signal are preprocessed, feature extracted, and fused to identify the user's motor intention. The determinism of the motor intention is evaluated based on the frontal lobe attention representation signal, intention confidence, and signal quality evaluation index. Motor control commands are selectively output according to the evaluation results. Based on the intensity of the motion intention corresponding to the motion control command, intermediate gait parameters are determined through a nonlinear mapping relationship. The intermediate gait parameters are then constrained according to at least one of the following constraints: gait parameter amplitude, joint range of motion, gait phase change rate, and environmental state, to obtain the target gait parameters. The real-time status information of the exoskeleton is obtained based on the limb proprioceptive feedback signal, and safety control commands are generated based on the real-time status information. The exoskeleton is driven to perform actions according to the target gait parameters and the safety control commands; In this context, when driving the exoskeleton to perform actions, the safety control commands have a higher execution priority than the motion control commands.

9. An electronic device, characterized in that, include: The memory is used to store the processing program; The processor, when executing the processing program, implements the lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 8.

10. A computer-readable storage medium, characterized in that, The readable storage medium contains a computer program for implementing the lower limb exoskeleton control method based on multimodal intent fusion and dynamic adjustment of gait parameters as described in claim 8.

Images