Method and system for controlling lower limb exoskeleton under brain-computer interface intention confidence

Abstract

This application discloses a method and system for controlling a lower limb exoskeleton using a brain-computer interface based on intent confidence, relating to the field of medical rehabilitation technology. First, the system identifies the original electroencephalogram (EEG) signal to obtain intent confidence, and maps an impedance parameter set including joint stiffness and damping coefficients by combining gait phase information. Second, it extracts motion position and interaction force errors, compares the intent confidence with the actual execution state using a sliding time window to obtain neural matching errors, and merges these three into a total composite error. Simultaneously, it assesses human-machine coupling compliance based on changes in human joint angles and exoskeleton torque, and generates a safety gain coefficient. Finally, it performs gain calculations on the total composite error based on the impedance parameter set, and corrects it with the safety gain coefficient to obtain the target control torque for output to the actuator. This constructs a neural-force-motor three-loop architecture, achieving compliant and adaptive on-demand assistance and highly safe human-machine collaborative rehabilitation with defensive protection.

Description

Technical Field

[0001] This application relates to the field of medical rehabilitation technology, and in particular to a method and system for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control. Background Technology

[0002] In recent years, lower limb exoskeletons have played a crucial role in the rehabilitation training of patients with neurological disorders. To stimulate patients' active participation, brain-computer interface (BCI) technology has been widely incorporated into exoskeleton control. However, existing control schemes typically only interpret EEG signals into simple start-stop trigger commands, failing to leverage the continuous confidence level of motor intentions in dynamic control system design. This singular control approach prevents the system from adaptively adjusting impedance parameters such as joint stiffness and damping based on gait phase, often resulting in overly rigid auxiliary torques. Furthermore, traditional low-level controls largely rely on physical positional errors or human-computer interaction force errors for closed-loop feedback, neglecting the neural matching error between central neural commands and actual execution states. This control architecture, which separates brain intention from physical execution, struggles to achieve deep synergy between "neural-force-motor" processes, thus hindering rehabilitation outcomes.

[0003] On the other hand, there are significant deficiencies in the compliant interaction and safety protection mechanisms during rehabilitation training. Due to nerve damage, patients' lower limb muscle tone is often unstable, making them prone to sudden muscle spasms or joint stiffness during training. Existing exoskeleton systems lack a quantitative assessment mechanism for instantaneous human-machine coupling compliance (i.e., the coupling relationship between changes in human joint angles and changes in exoskeleton torque). When human-machine movement conflicts occur and compliance deteriorates sharply, due to the lack of advance compliance warnings and real-time adaptive safety gain correction strategies, the underlying actuators may continue to output huge forced compensation torques. This rigid control method not only disrupts the smoothness of gait assistance but also easily violates the physiological stress limits of the patient's joints, leading to a serious risk of secondary injury. Therefore, a control method that balances intention confidence-driven operation and compliance safety monitoring is urgently needed. Summary of the Invention

[0004] This application provides a brain-computer interface intention confidence-driven control method for lower limb exoskeleton, which breaks through the limitations of traditional fragmented control based on a single physical quantity, and constructs a compliant closed-loop control link that deeply integrates the patient's continuous movement intention, gait time-varying characteristics and real-time physical compliance state, thereby achieving human-machine collaborative rehabilitation training that takes into account both natural adaptation and high safety protection.

[0005] According to one aspect of this application, a method for controlling a lower limb exoskeleton via a brain-computer interface intention confidence-driven control is provided, comprising: S1, performing motion intention recognition on the acquired raw EEG signals to obtain an intention confidence level characterizing the intensity or credibility of the user's motion intention; S2, acquiring gait phase information and performing neural-driven variable impedance parameter mapping on the gait phase information and intention confidence level to obtain an impedance parameter set composed of joint stiffness coefficients and joint damping coefficients; S3, acquiring motion position error and interaction force error, and using a sliding time window to compare the intention confidence level with the actual execution state to obtain neural matching error, and then adjusting the motion position error and interaction force error accordingly. At least two of the mutual force error and neural matching error are fused in a closed loop to obtain the total composite error; S4, the human-machine coupling compliance is evaluated on the obtained changes in human joint angle and exoskeleton torque to obtain the human-machine coupling compliance index, and the human-machine coupling compliance index is compared with a preset safety threshold to obtain the safety gain coefficient; S5, the gain is calculated on the total composite error based on the impedance parameter set to obtain the initial control torque, and the initial control torque is corrected for safety based on the safety gain coefficient to obtain the target control torque, wherein the target control torque is output to the drive actuator of the lower limb exoskeleton to drive the patient's lower limb to complete the auxiliary movement.

[0006] According to another aspect of this application, a brain-computer interface system for controlling a lower limb exoskeleton based on intention confidence is provided, comprising: an intention confidence recognition module for recognizing the movement intention of acquired raw EEG signals to obtain an intention confidence level characterizing the intensity or credibility of the user's movement intention; an impedance parameter mapping module for acquiring gait phase information and performing neurally driven variable impedance parameter mapping on the gait phase information and intention confidence level to obtain an impedance parameter set composed of joint stiffness coefficients and joint damping coefficients; and an error fusion module for acquiring movement position error and interaction force error, and comparing the intention confidence level with the actual execution state using a sliding time window to obtain neural matching error, thereby correcting the movement position error. At least two of the following errors—error, interaction force error, and neural matching error—are fused in a closed loop to obtain the total composite error: a compliance safety assessment module is used to assess the human-machine coupling compliance of the acquired changes in human joint angles and exoskeleton torque to obtain a human-machine coupling compliance index, and compares the human-machine coupling compliance index with a preset safety threshold to obtain a safety gain coefficient; a torque gain correction module is used to perform gain calculation on the total composite error based on the impedance parameter set to obtain the initial control torque, and to perform safety correction on the initial control torque based on the safety gain coefficient to obtain the target control torque, wherein the target control torque is output to the drive actuator of the lower limb exoskeleton to drive the patient's lower limb to complete the assisted movement.

[0007] Compared to existing technologies, this application provides a brain-computer interface method for controlling a lower limb exoskeleton based on intent confidence. This method continuously incorporates intent confidence into the control gain allocation, enabling the exoskeleton to accurately perceive the patient's slight or strong motor intentions. This completely eliminates the stiff and rigid dragging sensation of traditional systems, achieving true human-machine integration and smooth follow-up, and helping to more profoundly stimulate the plasticity reconstruction of the patient's central nervous system. Simultaneously, the establishment of a neural-motor-force three-loop architecture fills the technological gap in previous systems that could not actively correct deviations when faced with patient muscle fatigue or delayed neural signals, significantly improving the compliance of trajectory tracking under complex gait conditions. More importantly, relying on innovative compliance assessment and underlying torque correction mechanisms, the system possesses adaptive safety boundary monitoring capabilities. In the instant of sudden abnormal muscle tone, spasticity, or severe human-machine aggression, it can smoothly reduce the control torque in milliseconds, substantially avoiding secondary mechanical injury to the patient caused by the device, greatly improving the reliability and clinical application value of medical rehabilitation devices.

[0008] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0009] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, wherein:

[0010] Figure 1 This is a schematic flowchart illustrating a method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control, as described in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the data flow of a brain-computer interface intent confidence-driven control method for lower limb exoskeleton according to an embodiment of this application.

[0012] Figure 3 This is a schematic flowchart illustrating step S2 of the method for controlling a lower limb exoskeleton using a brain-computer interface intent confidence-driven control, as described in an embodiment of this application.

[0013] Figure 4 This is a schematic flowchart illustrating step S3 of the method for controlling a lower limb exoskeleton using a brain-computer interface intent confidence-driven control, as described in an embodiment of this application.

[0014] Figure 5 This is a schematic flowchart illustrating step S4 of the method for controlling a lower limb exoskeleton using a brain-computer interface intent confidence-driven control, as described in an embodiment of this application.

[0015] Figure 6 This is a schematic block diagram of a brain-computer interface intent confidence-driven control system for lower limb exoskeleton, as described in an embodiment of this application. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by this application in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of this application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0017] like Figure 1 and Figure 2 As shown, this application provides a method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control, comprising: S1, performing motion intent recognition on the acquired raw EEG signals to obtain an intent confidence score characterizing the intensity or credibility of the user's motion intent. It should be understood that traditional brain-computer interface exoskeleton control schemes typically treat the parsed motion intent as a simple start / stop trigger signal, ignoring the continuously dynamic changes in the strength of the patient's active intent due to neurological damage. By extracting continuous intent confidence scores, the strength and certainty of the patient's brain's expectation of the current movement can be quantitatively assessed in real time. This avoids being limited to a single, rigid triggering mode, instead using intent confidence scores as a deep control driving signal for continuously adjusting control gain and performing variable impedance mapping, thereby providing the underlying physical actuator with a smooth control benchmark that highly matches the patient's instantaneous physiological expectations.

[0018] In one embodiment of this application, the acquisition of raw EEG signals is used to identify motor intent in order to obtain an intent confidence score that characterizes the intensity or credibility of the user's motor intent. This includes: firstly, continuously acquiring raw EEG signals from the patient's scalp using multi-channel EEG detection hardware; then, directly applying a digital filter to bandpass filter the signal, focusing on extracting specific frequency band features highly correlated with motor imagery or execution, such as calculating the power spectral density within a specific frequency band using Fast Fourier Transform, or extracting event-related desynchronization energy features. Next, the extracted frequency band energy features are input into a traditional linear classifier (e.g., a linear discriminant analysis classifier or support vector machine) for classification. After obtaining the decision distance or feature vector amplitude output by the classifier, the system uses a linear normalization function based on a preset decision threshold to directly map it to a continuous value between 0 and 1. For example, by conducting extensive basic tests, a lower threshold for the baseline energy in a normal resting state and an upper threshold for the saturation energy of a strong intention are predetermined. When the calculated energy characteristic is lower than the lower threshold, the intention confidence output is directly set to 0; when the energy characteristic is higher than the upper threshold, the intention confidence output is set to 1; when the characteristic value is between the two, a decimal between 0 and 1 is calculated using a linear interpolation algorithm as the current intention confidence parameter.

[0019] In another embodiment of this application, step S1 includes: First, acquiring raw EEG signals with a sampling frequency set between 256 and 1000 Hz using a multi-channel EEG acquisition device with 8-channel electrodes. The predetermined values ​​of the sampling frequency and the number of channels are chosen to balance the preservation of high-frequency details in the multimotor cortex region with the system's computational response time, while satisfying the Nyquist sampling theorem. After acquiring the raw EEG signals, the system imports them into a preset bandpass digital filter and performs discrete-time signal filtering calculations strictly according to a frequency band cutoff range of 8 to 30 Hz to filter out low-frequency baseline drift and high-frequency electromyography artifacts, thereby accurately extracting the μ rhythm (8 to 12 Hz) and β rhythm (13 to 30 Hz) feature components containing the patient's clear motor intentions, and generating a frequency band filtered signal.

[0020] Subsequently, the noise-filtered frequency band signal is reconstructed into a two-dimensional spatiotemporal data matrix suitable for network input, and then fed into a pre-trained convolutional neural network layer. This network layer uses a set convolutional kernel with specific weights to perform sliding discrete convolution operations in both the spatial electrode dimension and the temporal sampling point dimension. After adding a bias term to the result of the discrete convolution operation, a feature map reflecting the activation pattern of the cerebral cortex is extracted through a nonlinear activation function mapping. This feature map data constitutes deep spatiotemporal features. This operation process can be expressed by the following formula:

[0021]

[0022] in, For depth-space-time feature tensors; For non-linear activation functions used in network layers, the rectified linear function (ReLU) is usually intended to improve the non-linear representation ability of the model and avoid computational failure due to gradient vanishing. represents the feature extraction convolution kernel weight matrix obtained through large-scale offline pre-training in a convolutional neural network; * represents the mathematical symbol for discrete two-dimensional convolution operation. The input is the frequency band filtered signal matrix data; This is a predetermined bias term vector constant in the convolutional neural network layer.

[0023] Finally, the deep spatiotemporal features are flattened along the time step and input into the temporal neuron nodes within the Long Short-Term Memory network to capture the temporal evolution dependency of deep EEG signals, thereby calculating and outputting the hidden layer temporal vector. Subsequently, the hidden layer temporal vector is directly input into the fully connected layer and the Softmax classifier at the end of the network topology. The Softmax classifier performs exponential normalization on the linear output of the fully connected layer, reducing the dimensionality of complex abstract features to a probability distribution value for each set motion intention category. The system extracts the decimal value corresponding to the target category with the highest predicted probability and outputs it directly as the final intention confidence. In the preset mechanism of this invention, the value range of the intention confidence output probability is explicitly predetermined to be [0,1], and the numerical range of 0.3 to 0.7 is specifically defined as the transition interval of the confidence threshold. This transition interval is dedicated to the subsequent smoothing of the continuous nonlinear adjustment of the impedance function. The operation logic of this built-in output can be expressed by the formula:

[0024]

[0025] in, The confidence level of the final output that belongs to the closed interval [0,1]. This is a specific weight vector in the fully connected layer structure that corresponds to the target prediction category; The hidden layer temporal state vector is the output of the deep spatiotemporal feature signal after it has undergone processing by a long short-term memory network. is the bias constant corresponding to the target prediction category in the fully connected layer structure; The total number of categories for classifying predefined exoskeleton movement intentions is a constant; j is the number of categories that cycle from 1 to... An accumulator index for categorized traversal and stacking; This is the weight vector corresponding to the j-th subclass in the fully connected layer structure; is the bias constant corresponding to the j-th subclass in the fully connected layer structure.

[0026] As those skilled in the art will know, the predetermined coefficients such as convolutional kernel weights, fully connected layer neuron weights, and network bias constants involved in the above embodiments are all hyperparameter values ​​that were directly stored in the microcontroller's flash memory after the loss function was iteratively optimized using the backpropagation algorithm during the offline training phase of the model before the system was put into actual rehabilitation applications, by collecting the patient's previous historical motor imagery EEG dataset.

[0027] S2 acquires gait phase information and performs neurally driven variable impedance parameter mapping on the gait phase information and intention confidence to obtain an impedance parameter set composed of joint stiffness coefficient and joint damping coefficient. It should be understood that traditional lower limb exoskeleton rehabilitation systems typically employ rigid segmented impedance control or fixed impedance model logic. This control method results in the mechanical assistance characteristics provided by the exoskeleton being rigid throughout the gait cycle. It cannot integrate with the varying strengths and weaknesses of the patient's active movement intentions caused by central nervous system injury, nor can it meet the heterogeneous force interaction requirements of the lower limbs under different gait phases, such as absorbing impact during the support phase and free forward thrust during the swing phase. By using the confidence level of intent extracted from continuous quantification of the brain as a factor for continuous adjustment and control gain allocation, and combining it with physical gait, the impedance function is designed to be a continuous dynamic mapping relationship that can be directly related to neural signals and gait timing. This allows the exoskeleton controller to spontaneously output just the right compliant assistive features to the patient. On the basis of eliminating human-machine rigidity and ensuring fall prevention support safety, it maximizes the patient's potential for autonomous exertion and the reconstruction of neural pathways.

[0028] like Figure 3 As shown, in one embodiment of this application, step S2 includes: S21, acquiring gait phase information; S22, retrieving preset minimum stiffness boundary values ​​and maximum stiffness boundary values ​​from the storage register of the underlying controller, and performing stiffness parameter mapping on the intention confidence based on the minimum stiffness boundary values ​​and maximum stiffness boundary values ​​to obtain the joint stiffness coefficient; S23, retrieving preset minimum damping boundary values ​​and maximum damping boundary values ​​in the control system for the current patient's vital signs, and performing damping parameter mapping on the gait phase information based on the minimum damping boundary values ​​and maximum damping boundary values ​​to obtain the joint damping coefficient; S24, dynamically constructing an impedance parameter group based on the joint damping coefficient and the joint stiffness coefficient.

[0029] Specifically, firstly, the microprocessor of the main control system concurrently receives the intent confidence score relayed from the feature parsing channel in the communication thread, and simultaneously acquires the gait phase information provided by the external gait system within the current interruption cycle. After acquiring the above multi-dimensional parameters, the stiffness parameter mapping process based on the intent confidence score is initiated. To constrain the calculation threshold and prevent stiffness from escalating and becoming uncontrollable, the control system directly retrieves the pre-set minimum stiffness boundary value and maximum stiffness boundary value safety parameters from the non-volatile memory register of the lower limb exoskeleton's underlying controller via the internal I / O bus.

[0030] Next, the linear interpolation mapping function engine is invoked, using the input intent confidence level as a continuously adjusted weighting factor for stiffness fluctuations. Rigorous numerical calculations are then performed to accurately derive the joint stiffness coefficient that meets the patient's neuro-assisted needs in the current execution cycle. The formula for calculating the joint stiffness coefficient is:

[0031]

[0032] in, It refers to the joint stiffness coefficient determined at the current discrete-time sampling time t, and its general unit is set as Newton-meter per radian (Nm / rad). Refers to the minimum stiffness boundary value of the fall-prevention exoskeleton joint that has been specified. The maximum stiffness boundary value of the exoskeleton joints refers to the maximum safe power supply of the system. It refers to the confidence level of an intention to follow the dynamic changes and fluctuations of the human body's neurophysiological responses.

[0033] Those skilled in the art will understand that the extreme value pre-determined coefficients involved in the above formulas must be pre-determined and fixed in the calibration experiment based on the stress requirements of different patients' weights and conditions. For example, the minimum stiffness boundary value... The value range is calibrated and constrained to be within 5 to 15 Nm / rad. This setting ensures that even when the patient experiences muscle fatigue or extremely low confidence of intent, the exoskeleton joint retains a certain degree of folding stiffness at the physical level to guarantee basic standing and impact resistance safety. Conversely, the maximum stiffness boundary value... The range of values ​​used is set to be between 25 and 60 Nm / rad to ensure that the joint has sufficient auxiliary stiffness to drive the limb to initiate and propel itself when the system senses and captures a strong and stable subjective intention of the patient to initiate or push the nerve.

[0034] Then, the minimum and maximum damping boundary values ​​specifically set and matched to the current patient's specific vital signs are retrieved from the firmware's preset control form. Using a pre-loaded damping dynamic adjustment function curve, the input gait phase information is treated as a time-varying independent variable within the corresponding gait evolution cycle and imported into the function for calculation and mapping. This outputs a precise joint damping coefficient that can match in real time whether the user's heel is just landing or in a transitional state. The formula for adaptive adjustment of damping characteristics is:

[0035]

[0036] in, The joint damping coefficient calculated at time sampling time t is typically obtained using a unit system of Newton-meter-second per radian (Nm·s / rad). Refers to the minimum damping boundary value used to reduce the drag sensation of heavily loaded systems; This refers to the maximum damping boundary value used to absorb impact force waveforms; and This refers to the dynamic gait phase information that is in the current transition state and whose value is between [0,1].

[0037] As those skilled in the art will know, the predetermined minimum damping boundary value The value is typically taken to be between 0.5 and 2 Nm·s / rad, so that the back electromotive force and mechanical hysteresis resistance of the servo motor gears can be reduced during the swinging period of the leg as it leaves the ground; the predetermined maximum damping boundary value The value of is limited to 3 to 8 Nm·s / rad, so as to significantly increase the energy absorption damping level during the support period after heel strike, and smoothly dissipate the instantaneous contact interaction impact peak torque of the echo from the ground rigid frame contact point.

[0038] After the microprocessor completes the floating-point arithmetic derivation of the two core impedance elements—joint stiffness and damping—the system performs forced type conversion of the storage bit width and reordering and realigning of the register cache addresses on the two scalar coefficient data according to the communication protocol specifications of the lower-level hardware driver. The stiffness and damping data, after being aligned with the bit width and address specifications, are further constructed and assembled into a unified-addressable two-dimensional control parameter matrix or control vector block. Finally, it undergoes rigorous redundancy checks and encapsulation to form the final impedance parameter set, which is then passed to the next-level closed-loop error adjustment pipeline layer for low-level comprehensive synthesis of control torque.

[0039] S3 acquires the motion position error and interaction force error, and uses a sliding time window to compare the intention confidence with the actual execution state to obtain the neural matching error. Then, it performs closed-loop fusion processing on at least two of the motion position error, interaction force error, and neural matching error to obtain the total composite error. It is understandable that traditional lower limb exoskeleton control largely relies on closed-loop feedback of physical-level motion position errors or human-computer interaction force errors for adjustment. This control architecture separates the upper-level brain intention from the lower-level physical execution, completely ignoring the misalignment and delay between the central neural command issuance and the actual electromechanical execution state. If this neural-level matching error cannot be quantitatively assessed, the system cannot achieve true deep "neuro-force-motor" synergy, resulting in the generated auxiliary control torque failing to adapt to the patient's actual neural expectations, thus producing a sense of stickiness and resistance at the human-computer interaction level.

[0040] like Figure 4 As shown, in one embodiment of this application, step S3 includes: S31, extracting neural matching error based on a sliding time window from the acquired intent confidence and actual execution state to obtain neural matching error; S32, performing multi-dimensional error fusion weight allocation on the received neural matching error, motion position error and interaction force error to obtain a weight factor group; S33, based on the weight factor group, performing three-closed-loop composite error weighted fusion on the neural matching error, motion position error and interaction force error to obtain the total composite error.

[0041] Specifically, the microprocessor receives motion position error and interaction force error calculated from the lower-level sensors. The microprocessor temporarily stores these two sets of physical error data in a cache without modification for subsequent transmission. Simultaneously, it receives the intent confidence score, representing the real-time strength of the brain's command, and the actual execution state, mapped and normalized to a value range of 0 to 1 by the joint encoder signal. A sliding time window based on a first-in-first-out queue is constructed in memory, and continuously sampled frames within this time window are aligned and matched. The mean absolute deviation between the intent confidence score and the actual execution result is calculated, and this deviation result is output as the neural matching error. The formula for calculating the neural matching error is:

[0042]

[0043] in, The neural matching error calculated at the current moment; This represents the total number of discrete sampling points contained within the sliding time window. This is the latest discrete-time sampling time. The index for the accumulator of historical sampling moments within the sliding time window is traversed from t-N+1 to t; The confidence level of the intent is defined as the value collected at the k-th sampling time that is between 0 and 1. This represents the actual execution state after recording and normalization at the k-th sampling time.

[0044] As those skilled in the art will understand, the actual duration of the predetermined sliding time window is set between 100 and 300 milliseconds, meaning the total number of sampling points N equals the set time divided by the system's sampling period. If the current rehabilitation patient has high spasticity sensitivity and requires extremely high real-time neural response, the time window is predetermined to be fixed at 100 milliseconds; if the patient is in a chronic recovery phase and the intention to move is relatively slow and prolonged, a value of 300 milliseconds is predetermined to obtain a smoother and more stable mean of matching error.

[0045] Subsequently, the system receives the generated neural matching error and the temporarily stored propagated motion position and interaction force errors. Based on the current patient's different rehabilitation stages, it retrieves and extracts the preset three types of control weight constants from the parameter matrix memory of the main control unit. These are then encapsulated into a unified addressable weight factor group for direct use by subsequent stages. To meet the needs of targeted and differentiated control support throughout the entire rehabilitation cycle, the selection of weights has corresponding normalized extreme value boundaries to prevent any error from expanding indefinitely and causing control instability.

[0046] In one specific embodiment, the motion error weight responsible for the deviation in exoskeleton motion trajectory tracking is constrained to the range of 0.3 to 0.6. For example, in the completely passive training stage during absolute bed rest or the very early stage of rehabilitation, when the patient has no intention of exerting force voluntarily, this weight value will be set to the upper limit of 0.6 to perform strict position trajectory tracking. The interaction force error weight is constrained to the range of 0.2 to 0.4 to adjust the active tolerance of human-machine confrontation. The neural matching error weight is set to the range of 0.1 to 0.3. For example, for patients in the late stage of active rehabilitation and advanced voluntary movement training, this weight coefficient is predetermined to be set to the upper limit of 0.3 to maximize the influence dimension of neural closure in the overall error feedback, thereby strengthening the reconstruction effect of neural pathways.

[0047] Finally, the input weight factor set is parsed, and the released independent weight coefficients are multivariately linearly weighted and summed with their corresponding error objects of the same level. The motion position error, interaction force error, and neural matching error are each multiplied by their respective weight coefficients, and then the product of these three factors is accumulated and summed to output the total composite error representing the overall multimodal deviation of the system. The multidimensional fusion calculation formula is expressed as:

[0048]

[0049] in, This represents the total composite error output within this calculation cycle; For motion error weights; The error in motion position between the preset trajectory acquired in real time and the actual joint motion; For interaction force error weights; The error of the interaction force is derived from the torque sensor. The neural matching error weights; The neural matching error is inferred through cumulative comparison over a time window. Using the multidimensional fusion calculation formula described above, the three separate single physical and physiological scales are integrated into a holistic bias data stream with complex multimodal information content, and then steadily transferred to the external gain calculation component in the next calculation cycle for accurate synthesis of the initial control torque.

[0050] It should be noted that the closed-loop fusion processing of at least two of the motion position error, interaction force error, and neural matching error is achieved through dynamic sparse allocation of the weight factor group. This application can preset one or more weight factors to 0 according to the patient's different rehabilitation stages or preset training modes, thereby achieving flexible selection or combination of error terms.

[0051] Specifically, in one embodiment, if the patient is in the early stages of passive rehabilitation, the system can weight the neural matching error. Setting the value to 0 enables high-intensity trajectory tracking control through the fusion of motion position error and interaction force error. In another embodiment, if the patient has some active movement ability but abnormal muscle tone fluctuations, the system can adjust the weight of the motion position error. Set to 0, and implement compliant rehabilitation assistance by fusing the dual closed-loop fusion of interaction force error and neural matching error. When , and When all values ​​are non-zero, a three-loop composite error weighted fusion is formed, achieving full-dimensional, deep, and coordinated control of the neural-motor-force systems. Through this dynamic weight allocation mechanism, the system can smoothly switch between different error combination modes according to actual working conditions, balancing the flexibility of control with the targeted nature of rehabilitation effects.

[0052] S4 assesses the human-machine coupling compliance of the acquired changes in human joint angles and exoskeleton torque to obtain a human-machine coupling compliance index. This index is then compared with a preset safety threshold to obtain a safety gain coefficient. It is understandable that during rehabilitation training, due to nerve damage, the patient's lower limb muscle tone is often extremely unstable, making them highly susceptible to sudden muscle spasms, local joint stiffness, or involuntary abnormal resistance during rehabilitation walking. Without a quantitative assessment and early warning mechanism for the instantaneous human-machine coupling physical compliance state, when human-machine movement conflicts occur, the underlying mechanical actuators will continue to output huge forced compensatory pulling torques according to the trajectory and error requirements of the upper layers. This not only severely disrupts the smooth progression of the gait cycle but also easily exceeds or even violates the physiological force limits of the patient's currently damaged joints, leading to a serious risk of secondary injury.

[0053] like Figure 5 As shown, in one embodiment of this application, step S4 includes: S41, acquiring the change in human joint angle and the change in exoskeleton torque, and determining a human-machine coupling compliance index characterizing the patient's limb following ability under a unit auxiliary torque based on the change in human joint angle and the change in exoskeleton torque; S42, retrieving a preset lower limit and upper limit of the safety threshold from the system's non-volatile memory, comparing and calculating the human-machine coupling compliance index with the safety threshold range, determining that the system is in a safe following state if the index is within the range, and determining that the system is in an abnormal risk state if the index exceeds the safety threshold range, and encapsulating the determined logical state into a safety determination result; S43, generating a safety gain coefficient based on the safety determination result.

[0054] Specifically, the change in human joint angle is first obtained by differential calculation of the exoskeleton joint encoder, and the change in exoskeleton torque is obtained by feedback from the torque sensor and differential processing.

[0055] In the first embodiment of this application, a basic calculation in the form of a derivative ratio is performed, that is, the dynamic division and comparison of the angular increment of human motion and the torque increment generated by the exoskeleton is directly performed. The result of this calculation directly characterizes the patient's limb's follow-up ability under a unit auxiliary torque, and this scalar value is defined as the human-machine coupling compliance index. Specifically, the human-machine coupling compliance index is determined by the following formula:

[0056]

[0057] in, This represents the change in the angle of a human joint. This represents the change in exoskeleton torque. It is an indicator of human-machine coupling compliance.

[0058] The first embodiment described above provides a compliance assessment method with low computational overhead and fast response, suitable for basic training scenarios with extremely low sensor noise or extremely high real-time requirements. However, in the actual application scenarios of complex lower limb exoskeleton rehabilitation training, the first embodiment reveals significant technical defects. The physiological state of the patient's limbs differs fundamentally at different stages of the gait cycle: during the support phase, the foot bears the weight load, and the lower limb joints need to maintain body stability. The muscles of the hip, knee, and ankle joints are in isometric or concentric contraction, and joint stiffness naturally increases to resist gravity and ground reaction forces. The angular change produced under the same exoskeleton auxiliary torque is significantly reduced, resulting in a lower calculated compliance index value. During the swing phase, the patient's leg is off the ground in a free forward swing motion state, the limb does not bear the weight load, and the muscles around the joint switch to eccentric contraction or relaxation, greatly enhancing joint flexibility. The same torque can drive a larger angular change, resulting in a correspondingly higher compliance index value.

[0059] The first embodiment uses a unified evaluation standard to quantify these two distinct physiological states, which cannot distinguish between normal gait phase differences and real human-machine conflict events. When the system detects that the compliance index is lower than the safety threshold during the support phase, it may misjudge that the patient is rigid or resistant, triggering an unnecessary safety downgrade mechanism, which may interfere with the normal rehabilitation training rhythm. Conversely, if the patient does have an abnormally high muscle tone during the swing phase, the abnormal state may be masked and the protective response may not be triggered in time because the baseline value of the compliance index is already high at this stage.

[0060] Further analysis revealed that the lower limb exoskeleton system involves three main drive joints: the hip, knee, and ankle. Each joint exhibits significant heterogeneity in anatomical structure, physiological function, torque tolerance, and range of motion. The hip joint, as the largest ball-and-socket joint in the human body, is surrounded by powerful muscle groups such as the gluteus maximus and iliopsoas. Under normal gait, it can generate a peak torque of 80 to 120 Nm, with a flexion-extension range of approximately 120 degrees. The knee joint, a typical hinge joint, is controlled by the quadriceps femoris and hamstring muscles, with a peak torque range of 60 to 90 Nm and a flexion-extension range of up to 140 degrees. The ankle joint has a relatively delicate structure, primarily driven by the gastrocnemius and tibialis anterior muscles, with a peak torque of only 40 to 60 Nm and a combined plantar flexion and dorsiflexion range of approximately 50 degrees. The first embodiment uses the same ratio calculation method for all joints, ignoring the essential differences in the torque-angle response characteristics of each joint. This results in a lack of comparability of compliance indices across joints: the hip joint, due to its large torque base and relatively small range of motion, naturally has a lower compliance index; while the ankle joint exhibits a higher compliance value due to its small torque base. When a uniform safety threshold is set for discrimination, it may be overly sensitive to the hip joint and sluggish in response to the ankle joint, failing to achieve accurate joint-specific safety monitoring.

[0061] The first embodiment directly uses the ratio of angle change to torque change within a single control cycle for calculation. This method is extremely sensitive to sensor noise and transient disturbances. In actual rehabilitation training scenarios, joint encoders are affected by factors such as mechanical vibration and electromagnetic interference, and the output signal inevitably has high-frequency noise components superimposed on it. When measuring human-machine interaction forces, torque sensors capture various transient disturbances such as muscle tremors of the patient's limbs, elastic deformation of the exoskeleton structure, and ground impacts. When the instantaneous value of the torque change approaches zero or sensor glitch occurs, the denominator of the ratio calculation approaches zero, causing the compliance index to explode or fluctuate violently. The system may repeatedly trigger the safety judgment logic in a short period of time, causing frequent switching of control gain. This not only affects the smoothness of exoskeleton assistance but may also exacerbate the patient's discomfort due to control oscillations, reducing the compliance and effectiveness of rehabilitation training.

[0062] To address the aforementioned technical deficiencies, this application proposes a second embodiment to achieve accurate quantitative assessment of human-machine coupling state. Specifically, in this second embodiment, the changes in human joint angles and exoskeleton torque are acquired, and a human-machine coupling compliance index characterizing the patient's limb follow-up ability under a unit auxiliary torque is determined based on these changes. This includes: acquiring joint type identifiers; determining gait phase weighting factors and joint normalization coefficients based on joint type identifiers and gait phase information; performing time-series smoothing on the changes in human joint angles and exoskeleton torque to obtain filtered angle changes and filtered torque changes; performing a ratio calculation on the filtered angle changes and filtered torque changes and introducing a joint normalization coefficient for normalization to obtain a basic compliance index; and using the gait phase weighting factor as a dynamic modulation coefficient to adaptively weight and fuse the basic compliance index to obtain the human-machine coupling compliance index.

[0063] First, joint type identifiers are acquired, and based on these identifiers and gait phase information, gait phase weighting factors and joint normalization coefficients are determined. It should be understood that the first embodiment failed to distinguish the physiological characteristics differences between different stages of the gait cycle, resulting in a phase blind spot in compliance assessment. Specifically, gait phase information provided by the gait detection system is first received. This information is typically represented as a normalized phase angle within the gait cycle, ranging from 0 to 1, where 0 to 0.6 corresponds to the support phase and 0.6 to 1.0 corresponds to the swing phase. Based on the biomechanical characteristics of the human lower limb, a gait phase adaptive weighting function is constructed. This function uses piecewise cosine modulation, assigning a higher weight value between 1.2 and 1.5 during the support phase to compensate for the decrease in compliance caused by increased joint stiffness, and assigning a lower weight value between 0.6 and 1.0 during the swing phase to suppress excessively high indices caused by excessive compliance.

[0064] The weighting function is expressed as follows:

[0065]

[0066] in Represents the gait phase weighting factor. This represents the normalized gait phase information. The introduction of the cosine function ensures a smooth change in weights during the phase transition region, avoiding abrupt control changes caused by step switching.

[0067] Simultaneously, it receives joint type identifiers and calculates joint normalization coefficients based on the physiological torque range and range of motion of each joint. The calculation process for these joint normalization coefficients is as follows:

[0068]

[0069] in Represents the joint normalization coefficient. This indicates the reference torque range for this joint type (80 Nm for hip joint, 60 Nm for knee joint, and 40 Nm for ankle joint). This indicates the reference range of motion for this joint type (120 degrees for the hip, 140 degrees for the knee, and 50 degrees for the ankle).

[0070] The above steps achieve differentiated compensation for the support and swing phases through cosine modulation, enabling compliance assessment to adapt to the dynamic changes in the gait cycle. Simultaneously, a unified assessment benchmark across joints is established, resolving the joint heterogeneity issue. The purpose of this execution is to provide phase-adaptive modulation parameters and joint normalization benchmarks for subsequent compliance calculations. The execution effect is reflected in the accurate differentiation between normal gait phase differences and real human-machine conflict, avoiding misjudgment of the support phase and omission of the swing phase.

[0071] Then, the changes in human joint angles and exoskeleton torques are time-series smoothed to obtain filtered angle and torque changes. The ratios of these filtered angle and torque changes are then calculated, and a joint normalization coefficient is introduced for normalization to obtain the basic compliance index. It should be noted that the first embodiment directly uses single-cycle raw sensor data for ratio calculation, lacking the ability to suppress noise and transient disturbances.

[0072] Specifically, the system first receives the raw changes in human joint angles and exoskeleton torque. To suppress sensor noise and transient disturbances, a first-order low-pass filter is introduced to smooth the timing of the two signals. The filter's cutoff frequency is set to 5-10 Hz. This frequency band preserves the main dynamic response characteristics of human movement (normal walking frequency is approximately 0.8-1.2 Hz, with the corresponding motion spectrum mainly distributed below 5 Hz) while effectively filtering out high-frequency sensor noise and mechanical vibration interference. The filtering algorithm uses a recursive form, and the filtering process for the angle changes is expressed as follows:

[0073]

[0074] in This represents the change in angle after filtering. This represents the change in the angle of a human joint at the current moment. This represents the filter smoothing coefficient (values ​​range from 0.3 to 0.5). This represents the filtered output from the previous moment. The filtering formula for the torque change is:

[0075]

[0076] in This represents the change in torque after filtering. This represents the change in exoskeleton torque at the current moment.

[0077] After filtering, the basic compliance index is calculated using the filtered data, and a joint normalization coefficient is introduced for normalization. The calculation process is as follows:

[0078]

[0079] in Indicators of basic compliance Represents the division-by-zero protection constant (values) (To avoid numerical anomalies when the denominator is zero) This represents the joint normalization coefficient.

[0080] In this step, the low-pass filter effectively suppresses sensor noise and transient disturbances, improving the robustness of compliance calculation and avoiding numerical explosion and control oscillations caused by noise in the first embodiment. Normalization processing allows compliance indices for different joints to be safely distinguished on a unified scale, solving the problem of lack of comparability across joints. The goal is to generate basic compliance indices that have undergone noise suppression and joint normalization. The execution effect is reflected in a significant improvement in the stability of the compliance values, providing high-quality input data for subsequent phase-weighted fusion.

[0081] Finally, the gait phase weighting factor is used as a dynamic modulation coefficient to adaptively weight and fuse the basic compliance index to achieve accurate compensation for the dynamic characteristics of the gait phase, thus obtaining the human-machine coupling compliance index. In other words, the dynamic characteristics of the gait phase need to be incorporated into the compliance assessment to achieve accurate compensation for different gait phases.

[0082] Specifically, the basic compliance index and gait phase weighting factor are received, and an adaptive weighted fusion operation is performed. The gait phase weighting factor is used as a dynamic modulation coefficient to perform phase compensation on the basic compliance index. The fusion process is expressed as follows:

[0083]

[0084] in Indicators representing human-machine coupling compliance Indicators of basic compliance This represents the gait phase weighting factor at the current moment. This indicates the gait phase information at the current moment.

[0085] It is understandable that the adaptive weighted fusion mechanism achieves precise compensation for the dynamic characteristics of gait phase, enabling compliance assessment to adapt to the physiological differences between the stance and swing phases: during the stance phase, the base compliance index is low due to increased joint stiffness, which is compensated by multiplying by a weight factor greater than 1 to bring it back to the normal assessment range; during the swing phase, the base compliance index is high due to increased joint flexibility, which is suppressed by multiplying by a weight factor less than 1 to prevent artificially high values ​​from masking real anomalies. The goal is to generate a human-machine coupling compliance index that comprehensively considers gait phase characteristics, joint physiological differences, and temporal smoothing characteristics. The execution effect is reflected in its ability to accurately distinguish between normal phase differences and real human-machine conflicts, significantly improving the accuracy and reliability of human-machine conflict detection, and providing high-quality input for subsequent safety judgment.

[0086] The second embodiment of this application solves the defects of phase blind zone, joint heterogeneity and noise sensitivity in the first embodiment through three-dimensional collaborative optimization of gait phase weight function, joint normalization coefficient and temporal filtering operator. It realizes accurate quantitative assessment of human-machine coupling state, enabling the exoskeleton system to accurately identify the patient's true movement intention and limb state, and avoids misjudgment and omission due to assessment bias.

[0087] Subsequently, the preset lower and upper limits of the safety threshold are read from the system's non-volatile memory. The currently obtained human-machine coupling compliance index is directly compared and calculated with the safety threshold range, and classification is performed for different types of abnormal states: if the index value is within the set range, the exoskeleton system is determined to be normal and marked as a safe follow-up state; if the compliance index value is lower than the lower threshold, it indicates that the unit torque cannot drive sufficient angular displacement, that is, the patient has experienced abnormal muscle rigidity and spasm or severe human-machine physical aggression, at which point the system forcibly determines and marks it as a first-type abnormal risk state (low compliance rigidity state); correspondingly, if the compliance index value is greater than the upper threshold, it indicates that the human limb may have abnormal high-frequency tremors that are not visible to the naked eye, sudden loss of muscle tone (complete flaccidity) during the support phase, or a sudden severe drift of the system sensors causing the compliance value to be abnormally high. At this time, the system loses effective physical support and is very likely to cause loss of control or patient fall, so it is determined as a second-type abnormal risk state (high compliance instability state).

[0088] After the judgment is completed, the above logical state is encapsulated into a safety judgment result and an adaptive safety gain coefficient decision is executed. For different safety judgment results, the system runs differentiated gain correction algorithms and defense mechanisms: if the received safety judgment result is a safe follow-up state, the full-size gain output is maintained, and the safety gain coefficient is set to a constant of 1.0, without any intervention in the initial control torque; if the received safety judgment result is a first-type abnormal risk state, the control compliant degradation mechanism is activated. The system sets the safety gain coefficient to a preset attenuation ratio constant (set to a value range of 0.4 to 0.7), actively yielding to the patient's resistance by reducing a certain proportion of the control gain, thereby effectively preventing the forced pulling torque of the underlying mechanical drive from exceeding the physiological force limit of the patient's damaged joint; if the received safety judgment result is a second-type abnormal risk state, considering that simply removing a certain proportion of the auxiliary torque cannot solve or may even exacerbate the risk of the patient (especially those in the support phase) falling due to lower limb weakness, the system will forcibly activate the emergency parking defense mechanism. At this point, the system immediately cuts off the safety gain coefficient and sets it to a minimum maintenance value (e.g., 0 or 0.05) to instantly deprive the active traction force source that might amplify high-frequency tremors. Simultaneously, an emergency stop command is issued in the underlying communication handshake bus, triggering each joint actuator to enter a high-damping biomechanical self-locking mode (or mechanical locking state), providing the patient with instantaneous, absolute, rigid fall-prevention support until manual intervention for repositioning. The decision logic formula is expressed as follows:

[0089]

[0090] in, The safety gain coefficient for the output; The human-machine coupling compliance index is calculated and obtained in real time for the above execution steps; The lower limit of the security threshold set for the system. The upper limit of the security threshold set for the system; This is a control gain attenuation ratio constant that is forcibly implemented when the system confirms entering a first-type abnormal risk state. As an example of a predetermined coefficient setting in this embodiment, to prevent false alarms while ensuring safety, the lower limit of the safety threshold is predetermined to be 0.02 rad / Nm, and the upper limit of the safety threshold is predetermined to be 0.15 rad / Nm. Gain attenuation ratio constant The value range is preset to any constant within the interval of 0.4 to 0.7. For example, the preset coefficient can be configured to be 0.5, which means that when the compliance index falls below the threshold extreme point, the rigid compensation torque of the underlying mechanical drive is instantly forcibly stripped and the control gain energy is reduced by 50%.

[0091] S5 calculates the initial control torque based on the gain of the total composite error using impedance parameters. Then, it performs a safety correction on the initial control torque using a safety gain coefficient to obtain the target control torque. This target control torque is output to the actuator of the lower limb exoskeleton to drive the patient's lower limbs to complete assisted movements. It can be understood that the brain's intention, multi-dimensional composite deviations, and human-machine compliance assessment calculated within the control system's computational bus must be projected and converted into actual mechanical rotational torque at the physical level to truly apply a smooth, neurally controlled, "on-demand assistance" support thrust to the patient. Simultaneously, because the theoretical compensation torque calculated directly from the model can be highly destructive under extreme conditions, it must be subject to the mandatory boundary management of a safety gain defense strategy before being sent to the underlying hardware. This prevents the risk of actuator overload or even secondary injury caused by instantaneous large pose deviations.

[0092] In one embodiment of this application, step S5 includes: dividing the difference between the total composite error at the current sampling time and the total composite error value between the previous control cycle by the system sampling cycle to obtain the error change rate; extracting the joint stiffness coefficient and joint damping coefficient from the impedance parameter set, and determining the initial control torque based on the joint stiffness coefficient, joint damping coefficient, total composite error and error change rate; and performing a safety correction on the initial control torque based on the safety gain coefficient to obtain the target control torque.

[0093] Specifically, the main control unit receives the total composite error, impedance parameter set, and safety gain coefficient from the upstream computing pipeline. Using a built-in first-order fixed-step backward differential operator, the main control unit calculates the difference between the latest total composite error entering the register at the current sampling moment and the old error value stored in the previous control cycle. This extracted difference is then directly divided by the fixed sampling update cycle set at the system's underlying layer, thereby deriving the error change rate, which characterizes the dynamic extension speed of the system's overall deviation.

[0094] In one specific embodiment, considering the balance between the high-frequency response requirements of the exoskeleton system's geared motor and the load-bearing limit of the microprocessor's floating-point operations, the predetermined coefficient for the fixed error sampling period is typically configured to a fixed constant range of 1 millisecond to 4 milliseconds (ms). The derivative formula for the error rate of change is expressed as:

[0095]

[0096] in, This represents the rate of change of error over time. This is the total composite error input value for sampling in this processing cycle; This is the total composite error value left over from the previous control cycle; The sampling time interval is a pre-fixed value.

[0097] After obtaining the dynamic rate of change, the system analyzes and decomposes the impedance parameter set, extracting the joint stiffness coefficient and joint damping coefficient of the internal load to construct a neurally driven virtual impedance calculation model control law. The hardware multiplier of the computation pipeline performs a dot product operation on the total composite error and the joint stiffness coefficient to generate a composite elastic feedback force parameter to simulate elastic self-correction. Simultaneously, it further performs a dot product operation on the previously calculated error rate of change and the joint damping coefficient to generate a dynamic damping force parameter to simulate viscous kinetic resistance. The system's accumulation control unit directly performs algebraic summation and merging of the elastic self-correction parameter and the damping force term, autonomously and in real time synthesizing the initial control torque. The formula for calculating the initial control torque is expressed as:

[0098]

[0099] in, Characterizes the initial control torque, measured in the basic unit Newton-meter (Nm); Characterizes the instantaneous joint stiffness coefficients separated analytically; Characterizes the total composite error amplitude transmitted upstream; To analyze and separate the joint damping coefficients; This represents the rate of change of error over time.

[0100] Subsequently, a scaling multiplier is used to multiply the initial control torque amplitude extracted from the calculation with the received safety gain coefficient to reshape the system boundary. In a specific fixed-point implementation example, if the peripheral module determines in real time that it is in a risk-compliant state of abnormal locking or human-machine interference, the forced input safety gain coefficient will be a pre-configured attenuation lower limit coefficient constant. The predetermined value range of this constant is limited to the range of 0.4 to 0.7 (e.g., set to 0.5). The main control system can reduce the amplitude of the original initial torque by a certain proportion in real time (e.g., reduce it by 50%) through this product operation. If the peripheral confirms that the current state is an absolutely interference-free safety state, the gain is set to a predetermined constant of 1.0, and the control law is then allowed to proceed as is. The execution instruction with the absolute system safety lower limit requirement formed after dynamic safety scaling through the above strong constraint means is the target control torque, and its arithmetic formula is determined as follows:

[0101]

[0102] in, The ultimate goal of controlled contraction or normal release is the control torque; The security gain coefficient value called for the current control frame; This is the source of the initial control torque generated by the impedance model.

[0103] Finally, the system master controller transmits the target control torque in scalar form in full-duplex mode to the micro-servo main power supply of the peripheral physical actuator via the lower-level communication bus. The latter, based on predetermined configuration parameters such as the inherent static torque constant of the exoskeleton frame motor and the reduction ratio of the planetary assembly of the mechanical body, performs an electromechanical inverse mapping operation from the torque scalar space to the three-phase current vector space. The physical drive current configured for electromagnetic conversion is numerically strictly equal to the target control torque directly divided by the scalar product of the motor torque constant and the reduction ratio. The control formula is:

[0104]

[0105] in, The required scalar input current command is given to the drive actuator to actually inject the stator coils; To calculate the target control torque issued; The rated torque constant of the micro motor driving the exoskeleton is a predetermined value. In this embodiment, the predetermined coefficient is, for example, 0.12 Nm / A. The multiplier is a predetermined parameter for the reduction ratio of a transmission device (such as a harmonic reducer integrated into the middle of a joint). In engineering implementation, this ratio is usually preset to be on the order of a constant, such as 50 or 120.

[0106] The underlying AC controller ultimately generates a high-frequency pulse width modulated three-phase electrical vector based on the decoupled current command, driving the mechanical joints worn by the patient's lower limbs to output corresponding compliant torques to provide substitute auxiliary support movements such as starting or stopping.

[0107] During the forward cycle of the human-machine interface operation with assisted force intervention, the high-precision magnetic encoder microelectronic module embedded in the joint peripheral device and the thin-film strain torque sensor array are activated in parallel to capture and extract the latest actual forced joint physical displacement and interactive echo force parameter stream generated by the patient. This dynamic data is integrated into the execution feedback data of the current pulse beat and then backflowed upwards, transmitted back to step S3 via a high-speed industrial data bus to trigger the error value refresh and overwrite for the next control cycle. The above closed-loop feedback link realizes continuous data updates, thereby achieving a deep closed-loop rehabilitation exoskeleton drive regulation mechanism that takes into account the compliance with deep commands from the central brain, dynamic adjustment of lower limb gait stiffness and damping, and self-resistance and defense against sudden muscle rigidity and other physical overloads.

[0108] To verify the effectiveness of the brain-computer interface intention confidence-driven control method for lower limb exoskeleton provided in this application in actual medical rehabilitation scenarios, a related clinical controlled evaluation experiment was conducted. Ten stroke rehabilitation patients were recruited as subjects, and the entire training period lasted four weeks. The experiment included a control group using a traditional fixed impedance control method and an experimental group using the brain-computer interface intention confidence-driven method (adaptive compliance control) provided in this application. After four weeks of comparative training and testing, multi-dimensional clinical and physical data were collected and analyzed. The results showed that: First, in terms of stimulating patients' active rehabilitation awareness, the mean square value of electromyography (EMG) of active lower limb muscle involvement in the experimental group was 28.6% higher than that in the control group. Second, in terms of gait quality and safety, the gait symmetry index of patients wearing exoskeleton-assisted walking improved by 19.3%, and due to the introduction of the system's human-machine coupling compliance safety defense mechanism and dynamic damping adjustment, the peak value of interactive impact caused by abnormal interference was significantly reduced by 35.2%. Finally, in terms of subjective experience of human-machine interaction, the subjective comfort score of patients provided by the system assistance significantly improved by 32% as collected through dynamic scale assessment. The above quantitative data fully demonstrate that the method provided in this application not only effectively stimulates patients' active movement potential but also ensures the safety and stability of rehabilitation training.

[0109] In summary, the brain-computer interface intention confidence-driven control method for lower limb exoskeleton according to the embodiments of this application has been elucidated. First, the raw EEG signal is deeply analyzed to obtain continuously changing intention confidence, which is then combined with the patient's current gait phase information to perform neurally driven variable impedance parameter mapping, thereby dynamically generating joint stiffness and damping coefficients that highly match the current motion state. Based on this, a sliding time window is used to compare the intention confidence with the actual physical execution state, extracting the neural matching error characterizing the delay or deviation between the brain's intention and the mechanical response. This neural dimension error is further fused with traditional mechanical motion position error and interaction force error using a three-closed-loop feature fusion to calculate a total composite error that more comprehensively reflects human-machine coordination deviation. To ensure absolute rehabilitation safety, changes in human joint angles and exoskeleton torque are monitored in real time, and a human-machine coupling compliance assessment model is established. Once the compliance index deviates from the safety threshold, a damping or degraded safety gain coefficient is immediately generated. Finally, the impedance parameter set is used to calculate the gain of the total composite error, and a safety gain coefficient is superimposed to dynamically reduce or correct the output torque, generating a target control torque that both responds to the brain's intention and takes into account physical collision avoidance to drive the actuator.

[0110] Figure 6 This is a schematic block diagram of a system for controlling a lower limb exoskeleton based on the intent confidence drive of a brain-computer interface according to an embodiment of this application. Figure 6As shown, this application also provides a brain-computer interface intention confidence-driven control system 600 for lower limb exoskeleton, including: an intention confidence recognition module 610, used to perform motion intention recognition on the acquired raw EEG signals to obtain an intention confidence level characterizing the intensity or credibility of the user's motion intention; an impedance parameter mapping module 620, used to acquire gait phase information and perform neurally driven variable impedance parameter mapping on the gait phase information and intention confidence level to obtain an impedance parameter set composed of joint stiffness coefficient and joint damping coefficient; and an error fusion module 630, used to acquire motion position error and interaction force error, and use a sliding time window to compare the intention confidence level with the actual execution state to obtain neural matching error, thereby controlling the motion position. At least two of the errors, interaction force errors, and neural matching errors are subjected to closed-loop fusion processing to obtain the total composite error; the compliance safety assessment module 640 is used to evaluate the human-machine coupling compliance of the acquired changes in human joint angles and exoskeleton torque to obtain a human-machine coupling compliance index, and compare the human-machine coupling compliance index with a preset safety threshold to obtain a safety gain coefficient; the torque gain correction module 650 is used to perform gain calculation on the total composite error based on the impedance parameter set to obtain the initial control torque, and to perform safety correction on the initial control torque based on the safety gain coefficient to obtain the target control torque, wherein the target control torque is output to the drive actuator of the lower limb exoskeleton to drive the patient's lower limb to complete the assisted movement.

[0111] It should be noted that the brain-computer interface intent confidence-driven control system for lower limb exoskeleton in this application embodiment is similar in principle to the aforementioned brain-computer interface intent confidence-driven control method for lower limb exoskeleton. Therefore, the implementation process, implementation principle, and beneficial effects of the brain-computer interface intent confidence-driven control system for lower limb exoskeleton can be found in the description of the implementation process, implementation principle, and beneficial effects of the aforementioned method, and will not be repeated.

[0112] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0113] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for controlling a lower limb exoskeleton using a brain-computer interface based on intent confidence, characterized in that... include: S1, perform motion intention recognition on the acquired raw EEG signals to obtain the intention confidence level, which characterizes the intensity or credibility of the user's motion intention; S2, acquire gait phase information, and perform neural-driven variable impedance parameter mapping on gait phase information and intention confidence to obtain an impedance parameter set composed of joint stiffness coefficient and joint damping coefficient; S3, obtain the motion position error and interaction force error, and use a sliding time window to compare the intention confidence with the actual execution state to obtain the neural matching error. Then, perform closed-loop fusion processing on at least two of the motion position error, interaction force error and neural matching error to obtain the total composite error. S4. The human-machine coupling compliance is evaluated by the obtained changes in human joint angles and exoskeleton torque to obtain the human-machine coupling compliance index. The human-machine coupling compliance index is then compared with a preset safety threshold to obtain the safety gain coefficient. S5, based on the impedance parameter set, the gain calculation of the total composite error is used to obtain the initial control torque, and based on the safety gain coefficient, the initial control torque is corrected for safety to obtain the target control torque, wherein the target control torque is output to the drive actuator of the lower limb exoskeleton to drive the patient's lower limb to complete the auxiliary movement.

2. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 1, characterized in that, Step S2 includes: Obtain gait phase information; The minimum stiffness boundary value and the maximum stiffness boundary value are retrieved from the storage register of the underlying controller, and the joint stiffness coefficient is obtained by mapping the intention confidence based on the minimum stiffness boundary value and the maximum stiffness boundary value. The minimum and maximum damping boundary values ​​preset in the control system for the current patient's vital signs are retrieved, and the gait phase information is mapped to damping parameters based on the minimum and maximum damping boundary values ​​to obtain the joint damping coefficient. Impedance parameter sets are dynamically constructed based on joint damping coefficient and joint stiffness coefficient.

3. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 1, characterized in that, Step S3 includes: The neural matching error is obtained by extracting the acquired intent confidence and actual execution status based on a sliding time window. The received neural matching error, motion position error and interaction force error are multidimensionally fused and weighted to obtain a weight factor group; Based on the weight factor group, the neural matching error, motion position error and interaction force error are weighted and fused using a three-closed-loop composite error to obtain the total composite error.

4. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 1, characterized in that, Step S4 includes: The changes in human joint angles and exoskeleton torques were obtained, and human-machine coupling compliance indices characterizing the patient's limb follow-up ability under a unit auxiliary torque were determined based on these changes. The system retrieves the preset lower and upper limits of the safety threshold from the system's non-volatile memory, compares and calculates the human-machine coupling compliance index with the safety threshold range, and determines that the system is in a safe follow-up state if the index is within the range, and determines that the system is in an abnormal risk state if the index exceeds the safety threshold range. The determined logical state is then encapsulated as a safety determination result. Based on the security determination results, a security gain coefficient is generated.

5. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 4, characterized in that, The changes in human joint angles and exoskeleton torque are obtained, and a human-machine coupling compliance index characterizing the patient's limb follow-up ability under a unit auxiliary torque is determined based on these changes. This includes determining the human-machine coupling compliance index using the following formula: in, This represents the change in the angle of a human joint. This represents the change in exoskeleton torque. It is an indicator of human-machine coupling compliance.

6. The method for controlling a lower limb exoskeleton using a brain-computer interface intent confidence-driven control according to claim 4, characterized in that, The changes in human joint angles and exoskeleton torque are acquired, and human-machine coupling compliance indices characterizing the patient's limb follow-up ability under a unit auxiliary torque are determined based on these changes. These indices include: Get the joint type identifier; Based on joint type identification and gait phase information, determine the gait phase weighting factor and joint normalization coefficient; The changes in human joint angles and exoskeleton torques are processed by time-series smoothing to obtain filtered changes in angles and torques. The ratio of the filtered changes in angles and torques is calculated and a joint normalization coefficient is introduced for normalization to obtain the basic compliance index. The gait phase weighting factor is used as a dynamic modulation coefficient to adaptively weight and fuse the basic compliance index to obtain the human-machine coupling compliance index.

7. The method for controlling a lower limb exoskeleton using a brain-computer interface intent confidence-driven control according to claim 4, characterized in that, Based on the security determination results, a security gain coefficient is generated, including: If the safety determination result is a safe follow-up state, set the safety gain coefficient to 1.0; If the safety assessment result is a Class I abnormal risk state, that is, the human-machine coupling compliance index is lower than the lower limit of the safety threshold, the safety gain coefficient is determined as a preset attenuation ratio constant to proportionally reduce the initial control torque. If the safety assessment result is a Class II abnormal risk state, that is, the human-machine coupling compliance index is higher than the upper limit of the safety threshold, the safety gain coefficient is set to 0 or 0.05, and an emergency parking command is issued to make the drive actuator enter the high-damping self-locking mode.

8. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 1, characterized in that, Step S5 includes: Based on the total composite error at the current sampling time and the total composite error value between the previous control cycle, the error change rate is obtained by dividing the difference by the system sampling cycle. The joint stiffness coefficient and joint damping coefficient are extracted from the impedance parameter set, and the initial control torque is determined based on the joint stiffness coefficient, joint damping coefficient, total composite error and error change rate. The initial control torque is corrected for safety based on the safety gain coefficient to obtain the target control torque.

9. The method for controlling a lower limb exoskeleton using a brain-computer interface with intent confidence-driven control according to claim 1, characterized in that, Also includes: The joint encoder and force sensor collect joint angle data and interaction force data in real time as execution feedback data. The execution feedback data is sent back to step S3 to update the motion position error and interaction force error, thereby realizing closed-loop adaptive control of the neural-motor-force system.

10. A system for controlling a lower limb exoskeleton via a brain-computer interface intent confidence drive, characterized in that, include: The intent confidence recognition module is used to identify the motor intent of the acquired raw EEG signals in order to obtain the intent confidence level, which characterizes the intensity or credibility of the user's motor intent. The impedance parameter mapping module is used to acquire gait phase information and perform neural-driven variable impedance parameter mapping on the gait phase information and intention confidence to obtain an impedance parameter set composed of joint stiffness coefficient and joint damping coefficient. The error fusion module is used to obtain motion position error and interaction force error, and to compare the intention confidence with the actual execution state using a sliding time window to obtain neural matching error. Then, at least two of the motion position error, interaction force error and neural matching error are fused in a closed loop to obtain the total composite error. The compliance safety assessment module is used to evaluate the human-machine coupling compliance of the acquired changes in human joint angles and exoskeleton torque to obtain a human-machine coupling compliance index, and compare the human-machine coupling compliance index with a preset safety threshold to obtain a safety gain coefficient. The torque gain correction module is used to perform gain calculation on the total composite error based on the impedance parameter set to obtain the initial control torque, and to perform safety correction on the initial control torque based on the safety gain coefficient to obtain the target control torque. The target control torque is output to the drive actuator of the lower limb exoskeleton to drive the patient's lower limb to complete the assisted movement.

Images