A joint compensation balanced on-demand auxiliary force lower limb exoskeleton robot flexibility design method

By decoupling individual wearing errors, matching temporal receptive fields and gait cycles in a lower limb exoskeleton robot, and combining two-stage learning and three-channel control, the problem of cross-user gait recognition fluctuations was solved, achieving high-precision gait recognition and joint compensation balance, thus improving the adaptability and efficiency of the exoskeleton robot.

CN123103585APending Publication Date: 2026-11-17ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611145193.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing lower limb exoskeleton robots experience fluctuations in gait recognition in cross-user applications and complex gait switching scenarios due to individual wear differences and mismatches in temporal feature extraction. This leads to deviations in the system's auxiliary output and increases individual joint compensation.

Method used

By calculating the equalization normalization parameters during the calibration phase to decouple individual wear errors, matching the temporal receptive field with the gait cycle, and employing a two-stage equalization learning mechanism and a three-channel phase-separated hybrid control, a robust gait recognition and compliant control mechanism is constructed to achieve high-precision gait recognition and auxiliary output.

Benefits of technology

It improves the accuracy of cross-user gait recognition, reduces system latency, enhances joint compensation balance, and achieves deep adaptation and efficient assistance of highly compliant lower limb exoskeleton robots in multi-task, multi-body environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103585A_ABST
    Figure CN123103585A_ABST
Patent Text Reader

Abstract

This invention discloses a compliant design method for a lower limb exoskeleton robot with joint compensation and balance, enabling on-demand assisted force application. The method includes: first, acquiring a baseline gait recognition model using a local edge device and a cloud server; then, the user wears the lower limb exoskeleton robot for system calibration, establishing a high-precision gait recognition model on the local edge device through a two-stage balanced learning process of "phase-locked adaptation—global collaborative fine-tuning"; during real-time operation, acquiring human-machine interaction state data and outputting gait recognition results; and constructing a three-channel phase-separated hybrid control technology (gait phase channel, interaction force channel, and motion trajectory channel) to control the lower limb exoskeleton robot, achieving compliant assisted force application. This invention effectively alleviates excessive work on non-target joints, achieving a compliant assisted effect through joint compensation and balance, and enabling deep adaptation and efficient assistance of a highly compliant lower limb exoskeleton robot in multi-task, multi-body environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of wearable robot perception and human-machine collaborative control, specifically involving a compliant design method for a lower limb exoskeleton robot with joint compensation and balance for on-demand assisted force exertion. Background Technology

[0002] The common control logic of lower limb exoskeleton robots is as follows: multi-dimensional time series signals are collected by a distributed sensor array arranged on the soles of the feet, waist, legs and mechanical structure to identify the wearer's current gait intention, and then the exoskeleton actuator generates an auxiliary output that matches the current terrain and motion phase.

[0003] From the perspectives of biomechanics and ergonomics, exoskeleton assistance typically follows the principle of "on-demand assistance." When the auxiliary output provided by the exoskeleton is out of phase with the wearer's actual movement intention (e.g., gait transition recognition lag), the system may generate resistance inconsistent with the individual's movement direction. To maintain balance or drive the exoskeleton system, the wearer often uses adjacent joints (such as back muscles or the unexoskeletonized contralateral lower limb) to generate compensatory force or movement. Long-term asymmetrical joint compensation can easily increase muscle fatigue. Therefore, achieving joint compensation balance and compliant human-machine collaboration is an important optimization direction for lower limb exoskeleton design, which requires the control system to have low latency and high-precision motion pattern recognition capabilities.

[0004] However, in practical engineering applications, deep learning-based exoskeleton gait recognition models face the following technical challenges:

[0005] First, the model's generalization ability is reduced due to differences in physical wear among users (covariate bias). Different users have objective differences in physiological morphology such as height, leg length, and body shape, and the initial physical installation position of sensors caused by flexible straps is usually difficult to keep absolutely consistent. This physical bias introduces baseline zero bias. Some existing exoskeleton systems habitually use the global mean and variance of a pre-trained general dataset for data normalization during inference. This fixed approach is difficult to effectively eliminate static physical zero bias for new users, causing fluctuations in the model's recognition accuracy when facing new users, and easily leading to bias in auxiliary output.

[0006] Second, temporal networks lack temporal receptive field matching for joint biomechanical cycles, which can easily lead to control delays. Gait intention recognition in exoskeletons has high real-time requirements. Conventional temporal convolutional networks or one-dimensional convolutional networks often increase network depth to obtain longer historical information during design, but rarely perform mathematical mapping matching between the convolutional receptive field and the actual individual's gait cycle. If the receptive field is too large, it will introduce too much redundant historical information, increasing the system's computational load and response latency; if the receptive field is too small, it will be difficult to capture complete gait cycle features (such as the complete support and swing phases of walking on flat ground), and recognition fluctuations are likely to occur in terrain transition areas, affecting the execution effect of "on-demand assisted force application".

[0007] Third, model fine-tuning in edge environments can easily lead to the forgetting of existing feature knowledge. To adapt to the wearing characteristics of new users, the conventional approach is to collect new user data for model retraining. However, when performing global retraining on the local microprocessor of the exoskeleton with limited computing power, and when the sample size of new user calibration data is small, the model is prone to overfitting local features, thereby covering or weakening the general macroscopic gait features learned from the source domain data, lacking a lightweight and robust feature update mechanism.

[0008] Fourth, existing underlying compliant control strategies lack multi-channel coordinated adjustment capabilities for gait phase, interactive force response, and motion trajectory. Traditional methods typically handle gait recognition, interactive force response, and trajectory correction separately, making it difficult to dynamically allocate prior knowledge of force control and position control based on different gait phases. This can easily lead to problems such as abrupt changes in auxiliary output, overly rigid position tracking, or delayed interactive response, which in turn induces compensatory force exertion at non-target joints. Summary of the Invention

[0009] To address the issue that existing lower limb exoskeleton robots suffer from gait recognition fluctuations due to individual wearer differences and temporal feature extraction mismatches in cross-user applications and complex gait switching scenarios, leading to system auxiliary output deviations and increased individual joint compensation, this invention provides a compliant design method for on-demand assisted force application in lower limb exoskeleton robots with balanced joint compensation. This method rapidly reduces the physical bias of new users at the local edge, matches the network receptive field to the individual's physical gait cycle on a temporal scale, and achieves stable recognition through a robust adaptive learning mechanism. Furthermore, based on this accurate perception, a compliant control mechanism is constructed that dynamically allocates force control prior knowledge and position control prior knowledge according to gait phase, interactive forces, and motion trajectory. This provides a state basis for the exoskeleton actuator to output accurate and compliant assistance, ultimately promoting balanced joint compensation throughout the wearer's body and achieving deep adaptation and efficient assistance of a highly compliant lower limb exoskeleton robot in multi-task, multi-person environments.

[0010] The technical solution of the present invention is as follows:

[0011] In a first aspect, this invention proposes a compliance design method for an on-demand assisted force-generating lower limb exoskeleton robot based on Joint Compensation Equilibrium (JCE), the method comprising the following steps:

[0012] Step 1: Obtain the baseline gait recognition model pre-installed on the control end of the lower limb exoskeleton robot;

[0013] Step 2: The user wears the lower limb exoskeleton robot and performs system calibration. The distributed exoskeleton sensor data of the user during the calibration phase is obtained, and the baseline gait recognition model is trained based on the distributed exoskeleton sensor data to obtain a high-precision gait recognition model, which is then deployed on the lower limb exoskeleton robot control terminal.

[0014] Step 3: Acquire distributed exoskeleton sensor data and human-computer interaction status data during real-time operation. The high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data. Then, based on the gait recognition results and human-computer interaction status data, auxiliary control parameters are generated, and the lower limb exoskeleton robot is controlled based on the auxiliary control parameters.

[0015] Step 4: Repeat step 3 to continuously acquire distributed exoskeleton sensor data and human-computer interaction status data and generate corresponding auxiliary control parameters to complete the continuous control of the lower limb exoskeleton robot.

[0016] Further, in step 2, the baseline gait recognition model is trained based on the distributed exoskeleton sensor data to obtain a high-precision gait recognition model, including:

[0017] The distributed exoskeleton sensor data collected by the user during the calibration phase is preprocessed to obtain a temporal feature dataset. Based on this temporal feature dataset, the baseline gait recognition model is trained to obtain a high-precision gait recognition model.

[0018] Furthermore, the equalization preprocessing of the distributed exoskeleton sensor data of the user during the calibration phase to obtain a time-series feature dataset includes:

[0019] Based on the distributed exoskeleton sensor data of the user during the calibration phase, the equalization normalization parameter corresponding to the user is calculated. Then, the distributed exoskeleton sensor data of the user during the calibration phase is preprocessed using the current equalization normalization parameter to obtain preprocessed distributed exoskeleton sensor data. Then, a sliding time window sequence is extracted from the preprocessed distributed exoskeleton sensor data using a sliding time window to obtain a time-series feature dataset.

[0020] Furthermore, the step of training the baseline gait recognition model based on the time-series feature dataset to obtain a high-precision gait recognition model includes:

[0021] The maximum temporal receptive field length of the temporal feature extraction network in the baseline gait recognition model is configured to match the length of each sliding time window sequence in the temporal feature dataset. Based on this temporal feature dataset, the parameter-adjusted baseline gait recognition model is trained using a two-stage equalization learning method to obtain a high-precision gait recognition model.

[0022] Furthermore, the training of the parameter-adjusted baseline gait recognition model using a two-stage equilibrium learning method includes:

[0023] First training phase: Freeze all prior knowledge parameters of the temporal feature extraction network in the baseline gait recognition model after parameter adjustment, and iteratively update the prior knowledge parameters of the classification decision module in the baseline gait recognition model according to the first preset learning strategy, and finally obtain the gait recognition model after completing the first training phase.

[0024] Second training phase: Release the prior knowledge parameters of the temporal feature extraction network in the gait recognition model after the completion of the first training phase, and iteratively fine-tune the gait recognition model after the completion of the first training phase according to the second preset learning strategy to finally obtain a high-precision gait recognition model; the learning rate in the first preset learning strategy is greater than the learning rate in the second preset learning strategy.

[0025] Furthermore, in step 3, the high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data, including:

[0026] After equalization preprocessing of the distributed exoskeleton sensor data during real-time operation, temporal feature data is obtained; then, the temporal feature data is input into a high-precision gait recognition model, and the model outputs gait recognition results.

[0027] Furthermore, in step 3, auxiliary control parameters are generated based on the gait recognition results and human-computer interaction state data, including:

[0028] First, current interaction force information and gait phase are generated based on human-computer interaction state data. Then, trajectory correction amount and expected compliant trajectory are generated based on the current interaction force information and gait phase. Next, the trajectory correction amount is optimized based on the expected compliant trajectory to obtain the optimal trajectory correction amount. Finally, auxiliary control parameters are generated based on gait recognition results and optimal trajectory correction amount.

[0029] Furthermore, the step of generating trajectory correction amounts based on current interaction force information and gait phase includes:

[0030] Based on the current interaction force information and gait phase, a fuzzy controller is used to generate virtual damping correction and virtual stiffness correction. Then, combined with a time-varying impedance model, a trajectory correction and a desired compliant trajectory are generated.

[0031] Furthermore, the optimization of the trajectory correction amount based on the desired compliant trajectory to obtain the optimal trajectory correction amount includes:

[0032] Based on the desired smooth trajectory, the trajectory correction amount is optimized through multiple rolling steps until the optimal trajectory correction amount is obtained.

[0033] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0034] This invention decouples individual features and matches receptive field scales of distributed exoskeleton sensor data collected by a distributed sensor array, and introduces a two-stage equilibrium learning mechanism at the edge to obtain stable gait pattern determination results applicable to different users. Then, it schedules the auxiliary output of the matching phase to provide the underlying state basis for realizing joint compensation equilibrium of the exoskeleton, thereby realizing adaptive gait phase control of the lower limb exoskeleton robot and deeply adapting to different tasks and individuals.

[0035] Furthermore, this invention introduces a force output and compliant control fusion mechanism based on gait recognition results. The control system determines the current assistance stage based on the gait phase channel, judges the wearer's active force application, passive following, or support bearing state based on the interaction force channel, and generates a position correction amount based on the motion trajectory channel. The three-channel fusion result is used to dynamically allocate prior knowledge of force control and position control, enabling the exoskeleton to have compliant following ability in the active force application stage, stable support ability in the support bearing stage, and reduce abrupt changes in auxiliary output during gait switching.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] (1) Decoupling of individual wearing errors based on balanced normalized parameters. This invention calculates and solidifies balanced normalized parameters for each user during the calibration phase, thus removing the static baseline bias introduced by different users due to differences in height, leg length, body fat thickness, and physical offset of sensor straps before the features enter the neural network. This parameter no longer participates in online statistical updates during the real-time inference phase, cutting off the interference of statistical fluctuations caused by abnormal movements on control stability. Experiments show that by introducing this feature decoupling preprocessing without updating any prior network knowledge parameters, the cross-user gait recognition accuracy can be significantly improved, demonstrating the suppression effect on covariate bias.

[0038] (2) Evolutionary capability of joint compensation equilibrium model based on cloud-edge collaborative deployment. The baseline gait recognition model of this invention supports acquisition and updating from a remote cloud server via wireless communication networks such as Wi-Fi and 4G / 5G. The cloud server aggregates anonymized motion feature statistics uploaded by multiple exoskeleton devices with user authorization, continuously optimizes the prior knowledge parameters of the model through privacy-preserving methods such as federated learning or incremental aggregation, and deploys the optimized model to each exoskeleton device. This collaborative closed loop of "edge acquisition - cloud aggregation - edge deployment" enables the exoskeleton system to continuously obtain model evolution without the need for users to perform complex manual upgrades; in an offline environment, the system automatically switches to local loading mode, reads the deployed model files from the device's internal storage, ensures that the gait recognition program can still run normally under network conditions, outputs gait pattern categories and confidence probabilities, and continuously improves the robustness of cross-user gait recognition and the accuracy of on-demand assistance.

[0039] (3) Low-latency feature extraction based on temporal receptive field-gait cycle scale matching. This invention uses a mathematically constrained strategy to increase the convolutional kernel size, number of layers, and dilation factor of the temporal feature extraction network, aligning its maximum temporal receptive field length with the individual's single standard physical gait cycle on the time scale. This design ensures that the features captured by the network during a single forward inference are neither fragmented nor redundant, covering the complete gait support phase, swing phase, and terrain transition phase; at the same time, by combining causal convolution and unidirectional sequence pruning, the mathematical architecture blocks the leakage of future time step information, reduces the phase delay of the system's real-time response, and provides low-latency, high-confidence gait state input for the underlying compliant control.

[0040] (4) Lightweight edge-side model adaptation based on two-stage balanced learning. This invention proposes a two-stage balanced learning mechanism of "phase-locked adaptation - global collaborative fine-tuning": In the first stage, the prior knowledge parameters of the network are extracted by freezing the temporal features with a high learning rate, and only the classification decision module is updated to quickly anchor the user's local decision boundary; in the second stage, the prior knowledge parameters of the network are released with an extremely low learning rate for collaborative fine-tuning, and micro-feature characterization is performed on the gait manifold of the target individual. Under the condition of extremely small calibration samples on the edge side, this mechanism avoids both overfitting of the model to local features and catastrophic forgetting of general macro-gait knowledge in the source domain. Experiments show that, under different calibration data ratios, the median accuracy improvement brought by this mechanism to various test groups is in a stable positive gain range, thereby achieving deep adaptation and efficient assistance of highly compliant lower limb exoskeleton robots in multi-task multi-person environments.

[0041] (5) Joint compensation balance and compliant human-machine collaboration based on three-channel phase-separated hybrid control. This invention constructs a three-channel phase-separated hybrid control framework consisting of gait phase channel, interaction force channel, and motion trajectory channel. The control system determines the current assistance stage based on the gait phase, evaluates the wearer's active force exertion, passive following, or support bearing state in real time based on the interaction force and its rate of change, and dynamically allocates force control prior knowledge and position control prior knowledge in combination with motion trajectory constraints—reducing position constraints to enhance force control prior knowledge to achieve compliant following when actively exerting force, and increasing position constraints to enhance virtual stiffness and provide stable support when supporting bearing. Furthermore, the phase-separated hybrid control results are smoothly converted into the optimal trajectory correction amount through multi-step rolling optimization, reducing the abrupt changes in auxiliary output and human-machine aggression during gait switching, so that the exoskeleton's auxiliary effect is accurately matched with the individual's motion phase, thereby reducing the wearer's tendency to call non-target joints (such as the lower back, contralateral healthy lower limb) for compensatory force exertion, achieving coordination and balance of lower limb joint biomechanical work, and thus realizing adaptive gait phase control of the lower limb exoskeleton robot, deeply adapting to different tasks and individuals. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention.

[0043] Figure 2 This is a system framework diagram of the present invention.

[0044] Figure 3 It is the confusion matrix of the pre-set baseline gait recognition model in the source domain subject data in the embodiments of the present invention; it is used to quantify the basic classification performance of the model before cross-user application, as well as the confusion limitations at the edge of similar terrain, and to provide a benchmark for subsequent adaptation experiments.

[0045] Figure 4 This is the gait confidence curve of the baseline gait recognition model in a continuous walking task in an embodiment of the present invention.

[0046] Figure 5 This is a bar chart comparing the impact of different processing strategies on the accuracy of cross-user gait recognition in embodiments of the present invention.

[0047] Figure 6 This is a box plot showing the distribution of recognition accuracy gain among different individual subjects generated by the two-stage balanced learning mechanism in this embodiment of the invention.

[0048] Figure 7 This is a trend line graph showing the impact of different two-stage balanced learning strategy configurations and calibration data ratios on recognition accuracy in embodiments of the present invention.

[0049] Figure 8 It is the confusion matrix of the high-precision gait recognition model after two-stage equalization learning optimization in the embodiments of the present invention.

[0050] Figure 9 This is the gait confidence curve of the high-precision gait recognition model after two-stage equalization learning optimization in the embodiments of the present invention in a continuous walking task.

[0051] Figure 10 This embodiment of the invention describes the human-computer interaction push-off propulsion force characteristics based on sensor arrays for lower limb exoskeleton joint compensation and balance.

[0052] Figure 11 This invention relates to a human-computer interaction landing buffer force feature based on a sensor array for lower limb exoskeleton joint compensation and balance.

[0053] Figure 12 This is a graph showing the dynamic adjustment response of human-computer interaction input, virtual compliance parameters, and position correction amount of a lower limb exoskeleton robot designed for on-demand assisted force application compliance in two degrees of freedom.

[0054] Figure 13 This is a diagram showing the auxiliary output following results and error curves of the lower limb exoskeleton robot with on-demand assisted force generation for joint compensation and balance in an embodiment of the present invention during the gait cycle. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to illustrate the technical mechanism and engineering effects of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make various equivalent substitutions or modifications to the technical solutions of the present invention.

[0056] This invention proposes a compliant design method for on-demand assisted force-generating lower limb exoskeleton robots with joint compensation and equalization. Its core control logic is: through physical feature decoupling of distributed sensor arrays, temporal scale matching of sliding time windows, and lightweight two-stage equalization learning, it provides low-latency, high-confidence state input to the underlying actuators of the exoskeleton, and further supports three-channel phase-separated hybrid force-position compliant control.

[0057] like Figure 1 As shown, the method in this embodiment specifically includes the following steps:

[0058] Step 1: Obtain the baseline gait recognition model pre-installed on the control terminal of the lower limb exoskeleton robot. The baseline gait recognition model includes a connected temporal feature extraction network (whose prior knowledge parameters are denoted as...). ), and the classification decision module (whose prior knowledge parameters are denoted as The temporal feature extraction network is used to extract spatiotemporal physiological features, and the classification decision module is used to output gait recognition results.

[0059] In one implementation, the baseline gait recognition model is deployed from a remote cloud server and distributed to the control terminal of the lower limb exoskeleton robot via a wireless communication network (such as Wi-Fi, 4G / 5G mobile communication, etc.). The cloud server is responsible for the version management, storage, and deployment scheduling of the baseline gait recognition model. When the exoskeleton device is powered on for the first time or when an updated model version is detected in the cloud, the control system automatically or with user confirmation downloads the latest baseline gait recognition model from the cloud, completing local deployment. In addition, with user authorization, each exoskeleton device can upload anonymized gait statistics collected during the calibration phase to the cloud. The cloud continuously optimizes the prior knowledge parameters of the baseline gait recognition model through federated learning or incremental aggregation, forming a co-evolutionary closed loop of "edge acquisition - cloud aggregation - edge deployment". When the exoskeleton device is operating in an offline environment (such as underground mines, remote mountainous areas, or other scenarios where wireless communication is unavailable), the control system automatically switches to local loading mode, reads the deployed baseline gait recognition model file from the device's non-volatile storage, and ensures that the gait recognition program can still start and run normally under network conditions.

[0060] In one feasible implementation, the baseline gait recognition model is trained based on historical sensor data. For example, it is trained offline based on large-scale distributed exoskeleton sensor data from multiple source domain subjects in complex terrain, aiming to capture common macroscopic gait characteristics of humans. The typical gait pattern set defined by the system is denoted as M, which mainly covers the terrains with the greatest differences in dynamic requirements in daily exoskeleton assistance, including horizontal walking (LG), uphill (RA), downhill (RD), upstairs (SA), and downstairs (SD).

[0061] The confusion matrix of the baseline gait recognition model is as follows Figure 3 As shown, the model's basic classification performance before cross-user applications and its limitations in handling confusion at similar terrain edges are used to quantify its characteristics, providing a benchmark for subsequent adaptation experiments. The model exhibits good basic classification performance within the source domain data distribution, demonstrating high recognition ability for terrains requiring higher auxiliary output to overcome gravity (e.g., uphill recognition rate of 93.51%). However, due to objective differences in individual physiological morphology (e.g., leg length ratio, gait frequency habits) and slight zero bias during initial sensor installation, similar gait manifolds (e.g., the moment of heel strike during horizontal walking and extremely gentle downhill walking) overlap to some extent in the local feature space, leading to cross-classification errors.

[0062] also, Figure 4This is the gait confidence curve of the baseline gait recognition model in a continuous walking task. From top to bottom, it displays the recognition confidence, relative amplitude of resultant acceleration, relative amplitude of resultant angular velocity, cadence trend, and true label for five gait modes (LG, RA, RD, SA, SD). Without individual equalization parameter adaptation, the model output exhibits the following problems: high-frequency fluctuations in confidence and cadence sensitivity. The probability stream of the model output shows significant transient abrupt changes in non-target classes. Comparing the cadence curve, it can be seen that the cadence of healthy adults fluctuates naturally within a range of approximately ±3~5 steps / minute, while the baseline gait recognition model is overly sensitive to this—when the cadence slightly increases, the false confidence of the uphill (RA) category rises accordingly; when the cadence slightly decreases, confusion arises between the downhill (RD) and flat walking (LG) categories. This indicates that the baseline gait recognition model has not yet decoupled "normal cadence fluctuations" from "gait mode switching precursors." False triggering of acceleration fluctuations also exists. Local peaks in the combined acceleration waveform (originating from stride fine-tuning or torso swaying) also trigger pulse-like jumps in non-target confidence, which, together with stride frequency drift, exacerbate recognition jitter. If such intent tags carrying high-frequency noise are directly sent to the underlying controller, it can easily cause unexpected fluctuations in the exoskeleton-assisted output, reducing the smoothness of human-machine collaboration.

[0063] Step 2: To address the covariate offset issue in cross-user applications, the system enters a brief calibration phase when a user first wears the lower limb exoskeleton or after readjusting the strap tightness during daily use. During this phase, the user, guided by the control system, performs a short guided movement including a basic gait (such as walking on flat ground), and the system acquires distributed exoskeleton sensor data during the calibration phase.

[0064] Distributed exoskeleton sensor data is obtained by a distributed sensor array deployed on the soles of the feet, waist, legs, and / or the exoskeleton body. This data can include one or more of the following: foot support force distribution, trunk posture changes, lower limb movement states, exoskeleton body movement states, and human-machine contact states. The combined distributed exoskeleton sensor data provides a state basis for gait phase recognition, interaction intent determination, and compliant control parameter adjustment. In specific implementations, the sensor array can be implemented using sensing elements capable of obtaining posture, force, displacement, angle, contact state, or deformation state; this invention does not limit the specific sensor type.

[0065] Specifically:

[0066] During calibration, the system uses sensor arrays deployed on the soles of the feet, waist, legs, and / or the exoskeleton to simultaneously collect distributed time-series data on support forces, trunk posture changes, lower limb movement states, and human-machine contact states, thereby constructing a state feature vector. At any given time... Sensor input vector , Let be the total number of sensing channels. Each channel can correspond to pressure distribution, posture changes, displacement changes, joint motion states, contact states, or their derived features. By temporally fusing foot support information, trunk motion trends, and leg motion information, a state feature space can be constructed for gait phase recognition and interaction intent determination. Let the total number of sampling points in the calibration phase be . Then the first The calibration data set of each channel is represented as .

[0067] Then, based on the acquired distributed exoskeleton sensor data, the baseline gait recognition model is trained to obtain a high-precision gait recognition model suitable for users and deployed on the control end of the lower limb exoskeleton robot;

[0068] In one feasible implementation, a baseline gait recognition model is trained based on the distributed exoskeleton sensor data to obtain a high-precision gait recognition model, which is then deployed on the control end of the lower limb exoskeleton robot, including:

[0069] The distributed exoskeleton sensor data collected by the user during the calibration phase is preprocessed to obtain a time-series feature dataset. Based on this time-series feature dataset, the baseline gait recognition model is trained to obtain a high-precision gait recognition model. Its prior knowledge parameters are solidified and encapsulated into a gait recognition program and deployed on the control end of the lower limb exoskeleton robot.

[0070] In one feasible implementation, the distributed exoskeleton sensor data of the user during the calibration phase undergoes equalization preprocessing to obtain a time-series feature dataset, including:

[0071] Based on the distributed exoskeleton sensor data of the user during the calibration phase, the equalization normalization parameter corresponding to the user is calculated. Then, the distributed exoskeleton sensor data of the user during the calibration phase is preprocessed using the current equalization normalization parameter to obtain preprocessed distributed exoskeleton sensor data. This preprocessed distributed exoskeleton sensor data is time-series feature data after individual adaptation mapping. Then, a sliding time window sequence is extracted from the preprocessed distributed exoskeleton sensor data using a sliding time window to obtain the time-series feature dataset.

[0072] Optionally, the sliding step size between two adjacent sliding time windows is a fixed number of frames; the fixed number of frames is less than the length of the sliding time window, so that data overlap occurs between adjacent windows, thereby enhancing the continuous coverage of features in the gait pattern transition interval. In this way, the model can extract continuous temporal features that match the complete gait cycle.

[0073] In a preferred embodiment of the present invention, the sliding time window sequence is constructed using a fixed-length window and a fixed frame step size, and a high proportion of data overlap is set between adjacent windows (e.g., an overlap ratio greater than 90%). The high overlap sliding strategy is beneficial for dividing the non-smooth transition intervals between different terrains into multiple continuous sub-windows, enhancing the system's ability to capture state transition boundary features.

[0074] Due to differences in waist circumference, fat thickness, and relative slippage of the exoskeleton straps among different subjects, the raw physical signals acquired by the sensors often carry individual static biases. To isolate this physical wear error from the signal source, statistical parameters were calculated for each sensing channel in the distributed exoskeleton sensor data during the calibration phase. These statistical parameters include at least the mean parameter characterizing the user's static bias during the calibration phase. and scale parameters characterizing dynamic amplitude features The formula is as follows:

[0075]

[0076]

[0077] in, This represents the number of sampling points during the calibration phase. Number the sensor channel; Number the sampling points; For the first The first sensing channel in the first The original physical signal values ​​at each sampling point; For the first The mean parameter of each sensing channel is used to characterize the user's static bias on that channel; For the first The scale parameter of each sensing channel is used to characterize the dynamic amplitude characteristics of the user on that channel.

[0078] Traditional online sliding average normalization is highly susceptible to statistical abrupt changes caused by local abnormal actions (such as sudden tripping), leading to system phase drift and lag on the time axis. The proposed equalization normalization parameter in this invention is used to compensate for deviations and unify scales in the user's real-time sensor data, achieving feature decoupling of individual physical wear errors. The equalization normalization parameter is persistently fixed after calculation. During the real-time identification phase, it is not updated again using real-time data from the online sliding window. Instead, it is used as a fixed mapping benchmark for the user during the real-time identification phase to avoid interference from real-time statistical fluctuations on system control compliance. This embodiment mathematically eliminates the interference of online statistical fluctuations on system stability by locking the equalization parameter, ensuring the benchmark consistency of subsequent control commands.

[0079] Through this step, the present invention decouples and isolates static wear errors introduced by user differences in body shape, sensor physical installation position offset, etc., and blocks control fluctuations that may be caused by real-time inference data participating in the dynamic update of statistical quantities.

[0080] In one feasible implementation, based on the time-series feature dataset, a baseline gait recognition model is trained to obtain a high-precision gait recognition model suitable for the user, which is then deployed on the control end of the lower limb exoskeleton robot, including:

[0081] The maximum temporal receptive field length of the temporal feature extraction network in the baseline gait recognition model is set to the length of each sliding time window sequence in the temporal feature dataset; that is, the maximum temporal receptive field length is equal to the size of the sliding time window. This gives the temporal feature extraction network a temporal receptive field alignment mechanism, specifically: the number of layers and the expansion rate of the temporal feature extraction network are controlled by biomechanical prior knowledge, ensuring that its maximum temporal receptive field length is mathematically constrained and scale-matched to the length of a single standard physical gait cycle of the lower limbs. Furthermore, a unidirectional sequence pruning operation prevents the leakage of future time step information. Based on the temporal feature dataset, a two-stage equilibrium learning method is used to train the parameter-adjusted baseline gait recognition model, resulting in a high-precision gait recognition model.

[0082] Subsequently, the control system constructs a sliding time window sequence and uses it as the input temporal feature extraction backbone network. This network employs a dilated convolutional architecture with strict causal properties. For a one-dimensional input sequence... It has a size of convolution kernel and the Layer expansion factor Its causal dilation convolution operation Defined as: To ensure gradient propagation in deep networks, residual mapping is introduced internally: ,in This represents the feature transformation terms after sequence pruning and convolution operations. The convolution kernel size is strictly constrained. Network layers and the strategy of increasing the expansion factor for each layer This makes the maximum temporal receptive field length of the backbone network... The following geometric series constraint equations must be satisfied: The system will adjust the length of the sliding time window. Configured to match the maximum receptive field length Matching, or corresponding to the maximum receptive field length within a preset range. Thus, the input feature tensor is constructed. .

[0083] During the project implementation, the system comprehensively considered the gait rhythm of healthy adults walking on flat ground without load, as well as the objective fact that gait frequency decreases under exoskeleton load and assistance conditions. Therefore, the system pre-configured various network architecture parameters so that the length calculated by the above formula... The corresponding physical time can cover or largely cover the single standard physical gait cycle of an individual under complex working conditions. This matching mechanism ensures that the features extracted by the top-level neurons during each forward inference are neither fragmented nor contain too much redundant history, but rather cover a complete gait support phase, swing phase, and transition phase. Simultaneously, due to the unidirectional sequence pruning and causal convolution properties at the algorithm level, the model outputs the current intent only based on historical and current frame features, mathematically avoiding or suppressing the leakage of future information and ensuring the real-time response capability of the control system.

[0084] Thus, a scale-matching relationship is formed between the physical length of the sliding time window, the maximum temporal receptive field length of the network, and the target gait cycle, enabling the model to cover or essentially cover the complete gait cycle and the transition features before and after terrain switching during a single forward inference. Simultaneously, combining causal convolution and unidirectional sequence pruning operations reduces the leakage of information from future time steps and lowers the phase delay of the system's real-time response.

[0085] In another preferred embodiment of the present invention, in applications targeting lower limb exoskeletons, considering the typical gait cycle of healthy adults with exoskeleton assistance and the extended transitional features such as turning from flat ground to a slope, the maximum receptive field length of the temporal feature extraction network is configured to correspond to a certain physical time range (e.g., the number of sampling frames required to include a complete stride). The value of its maximum receptive field is equal to... ,in The kernel size is [size]. This corresponds to the inflation rate of the residual layer. By adjusting these architectural parameter combinations, the model can achieve precise alignment between the time extraction scale and the physical gait cycle without increasing the number of parameters excessively.

[0086] To further eliminate high-dimensional nonlinear feature offsets that are difficult to resolve through linear preprocessing, the system constructs a cross-entropy loss function based on the user's calibration data. Used to measure the probability of model prediction Compared with real gait labels Differences:

[0087]

[0088] in, To fine-tune the batch size for updates; The sample number in the current batch; A set of preset gait patterns; The category index in the gait pattern set; For the first The sample at the th The true label on the gait pattern, when the sample belongs to the first... The value is 1 if the condition is met, and 0 otherwise. For the model to predict the first The sample belongs to the first The probability of gait-like patterns; This is the cross-entropy loss function, used to measure the difference between the model's predicted probability distribution and the true gait label distribution.

[0089] Considering the computational limitations of the exoskeleton edge processor and the need to prevent catastrophic forgetting, a two-stage equalization learning method is used to train the parameter-adjusted baseline gait recognition model, including:

[0090] The first training phase (phase-locked adaptation phase) involves freezing all prior knowledge parameters of the temporal feature extraction network in the baseline gait recognition model after parameter adjustment, and iteratively updating the prior knowledge parameters of the classification decision module in the baseline gait recognition model individually according to the first preset learning strategy. Specifically, it is gradient descent iteration ( This process aims to quickly anchor the local feature boundaries of the user, ultimately obtaining the gait recognition model after the first training stage is completed. This stage aims to utilize the general features of the source domain to quickly anchor the local decision boundaries of the user after mapping.

[0091] Second training phase (global collaborative fine-tuning phase): Releasing the prior knowledge parameters of the temporal feature extraction network in the gait recognition model after the completion of the first training phase. Based on the second preset learning strategy, the gait recognition model after the first training phase is collaboratively iteratively fine-tuned to ultimately obtain a high-precision gait recognition model. The iterative fine-tuning specifically involves... , Represents the gradient, with an extremely low learning rate. This is equivalent to performing micro-level feature characterization on the user's individual gait manifold, enabling the convolutional kernel to capture the user's unique subtle force exertion habits and gait rhythm. Simultaneously, this constraint prevents the model from overfitting to the limited sample data during the calibration phase, ensuring that the macro-gait features of the source domain are not compromised. The second training phase aims to comprehensively adapt the model's low-level features with a relatively small adjustment range, refining the micro-gait rhythms of the target individual while avoiding the forgetting of general macro-gait knowledge.

[0092] Learning rate in the first pre-set learning strategy Compared to the learning rate in the second preset learning strategy Large, meaning the update step size of the second preset learning strategy is strictly smaller than that of the first preset learning strategy.

[0093] Through a series of cross-validation experiments, the progressive gain effect of the above processing steps was objectively demonstrated: Figure 5 This is a bar chart comparing the impact of different processing strategies on cross-user gait recognition accuracy in embodiments of the present invention. Combined with... Figure 5 The accuracy comparison histogram shows that when the unadapted baseline gait recognition model is directly applied, the average accuracy is only 13.82% due to covariate bias, which is difficult to meet the requirements of stable control. After introducing individual feature decoupling (using the equilibrium normalization formula), the accuracy jumps to 92.31%~95.81% even without updating the network's prior knowledge parameters, proving the role of physical zero-bias stripping in recovering the model's feature perception. On the basis of feature alignment, after superimposing two-stage equilibrium learning, the accuracy further climbs to over 95%~98%. Figure 6 This is a box plot showing the distribution of recognition accuracy gain among different individual subjects generated by the two-stage balanced learning mechanism in this embodiment of the invention. Figure 6 As can be seen from the gain distribution box plot, under different proportions of data division, the median improvement in accuracy brought by the two-stage balanced learning to the test group covering different heights and body types is robustly in the positive gain range (average absolute improvement of 2.45% to 4.15%), which reflects the good applicability of the learning mechanism. Figure 7 This is a line graph showing the trend of the impact of different two-stage balanced learning strategy configurations and calibration data ratios on recognition accuracy in embodiments of the present invention. Figure 7 As can be seen from the line trend chart, there is a diminishing marginal effect in the selection of the number of iterations in the first stage. Using a limited number of iterations in the first stage often yields better overall test accuracy, which experimentally verifies the engineering scientific basis of the "fast phase-locked fitting to prevent classifier overfitting" strategy.

[0094] Step 3: Acquire distributed exoskeleton sensor data and human-machine interaction status data during real-time operation. The high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data. Then, based on the gait recognition results and human-machine interaction status data, generate auxiliary control parameters and control the lower limb exoskeleton robot based on the auxiliary control parameters.

[0095] In one feasible implementation, the high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data, including:

[0096] After equalization preprocessing of the distributed exoskeleton sensor data during real-time operation, temporal feature data is obtained, which is feature data adapted to the individual. This temporal feature data is then input into a high-precision gait recognition model, and the model outputs gait recognition results.

[0097] This invention utilizes balanced normalization parameters to perform channel-by-channel bias removal and scale unification processing on continuously acquired real-time distributed exoskeleton sensor data. The formula is as follows:

[0098]

[0099] in, To prevent tiny constants with a denominator of zero; For the real-time operation phase Each sensing channel at time The acquired raw physical signal values; and The calculations and storage of the first calibration stage are respectively Mean and scale parameters of each sensing channel; These are the real-time normalized feature values ​​after bias removal and scale unification.

[0100] After this step, the original sensor data with physical units and individual biases are transformed into dimensionless feature values ​​distributed near the standard normal space. This process forces the feature space to align, making the temporal feature data after individual adaptation mapping statistically close to the source domain data distribution during the training of the baseline gait recognition model. This reduces the impact of covariate bias on recognition accuracy and reduces the recognition pressure on the model when facing new data.

[0101] In one feasible implementation, the gait recognition result includes gait patterns. Gait patterns include, but are not limited to, one or more state switching intentions among walking on flat ground, walking uphill, walking downhill, walking up stairs, and walking down stairs.

[0102] In one feasible implementation, the gait recognition result includes gait pattern and motion state category.

[0103] In real-time operation, the tensor output by the model is mapped to the confidence probability distribution of various target gait patterns through the Softmax function:

[0104]

[0105] in, For the network to the first The feature bias score of the gait-like output.

[0106] The confusion matrix of the high-precision gait recognition model after two-stage equalization learning optimization is as follows: Figure 8 As shown, boundary states that were easily confused in the baseline gait recognition model (such as horizontal walking and slow downhill walking) became extremely clear in their classification boundaries after individual adaptation. Meanwhile, as... Figure 9 As shown in the confidence curves and physical waveform panel, the model optimized through two-stage equalization learning achieves improvements on the same task: confidence depth smoothing and step frequency decoupling are completed. Confidence scores for each dominant category are also presented. During continuous walking for hundreds of seconds, the gait score remained consistently high, approaching 1.0, with spurious activations in non-target classes completely suppressed. Crucially, compared to the cadence curve—which also exhibits natural fluctuations of ±3-5 steps / minute—the model confidence score no longer drifted in response to this. For example, during flat walking, when the cadence increased from approximately 55 steps / minute to approximately 68 steps / minute, the LG class confidence score remained stable. This indicates that the second-stage global collaborative fine-tuning has encoded the user's individual cadence characteristics and their natural variability into the prior knowledge parameters of the temporal feature extraction network, achieving decoupling between cadence fluctuations and gait classification. Acceleration perturbations are immune, and terrain transitions exhibit low-latency response. Transient fluctuations in the combined acceleration and synthetic angular velocity waveforms no longer trigger model misjudgments. At the transition points to real terrain (marked by real labels), the confidence of the old category rapidly decays while the confidence of the new category rapidly increases. The short and stable mixed probability during the transition period reflects the model's reasonable perception of the physical characteristics of the transition state (adaptive adjustment of gait frequency and changes in acceleration amplitude) rather than confusion in recognition, providing a precise and timely gait phase switching trigger signal for the three-channel phase-separated hybrid control.

[0107] In one feasible implementation, auxiliary control parameters are generated based on gait recognition results and human-computer interaction state data, including:

[0108] First, current interaction force information and gait phase are generated based on human-computer interaction state data. Then, trajectory correction amount and expected compliant trajectory are generated based on the current interaction force information and gait phase. Next, the trajectory correction amount is optimized based on the expected compliant trajectory to obtain the optimal trajectory correction amount. Finally, auxiliary control parameters are generated based on gait recognition results and optimal trajectory correction amount.

[0109] In one feasible implementation, a trajectory correction amount is generated based on the current interaction force information and gait phase, including:

[0110] Based on the current interaction force information and gait phase, a fuzzy controller is used to generate virtual damping correction and virtual stiffness correction. Then, combined with a time-varying impedance model, a trajectory correction and a desired compliant trajectory are generated.

[0111] Specifically, it includes:

[0112] While obtaining low-latency, high-confidence gait pattern recognition results, the system simultaneously acquires real-time human-machine interaction state data between the wearer and the exoskeleton robot. Then, based on this data, it generates current interaction force information and gait phase. The interaction force information includes the magnitude, direction, and rate of change of the interaction force. The gait phase refers to the specific step a person is taking, such as lifting the leg, stepping forward, or touching the ground. This invention constructs a three-channel phase-separated hybrid control mechanism: a gait phase channel, an interaction force channel, and a motion trajectory channel. Within a control cycle, it synchronously coordinates gait phase information, human-machine interaction force information, and motion trajectory constraint information, dynamically allocating prior knowledge of force control and position control to form a compliant force-position hybrid control output.

[0113] like Figure 10 and Figure 11 As shown, in one embodiment, an array of pressure sensors is arranged on the sole of the wearer's foot or on the exoskeleton's foot pedal, serving as a human-computer interaction force acquisition unit in a distributed sensor array. Based on human anatomical characteristics, the sole is divided into four core force monitoring areas: the phalangeal region (PH), the metatarsal region (ME), the arch / midfoot region (MF), and the heel region (HE). In actual walking conditions, when the wearer is walking uphill or in the push-off phase of a flat-ground gait, the center of force shifts significantly forward, with the maximum interactive pressure concentrated in the metatarsal region (ME), reaching a measured peak of 529.1 N, while the local minimum pressure is only 44.4 N. When the wearer is walking downhill or in the heel strike cushioning phase of a gait, the force support surface shifts significantly backward to absorb ground impact, with the maximum interactive pressure appearing in the heel region (HE), surging to a peak of 690.2 N, corresponding to a minimum pressure of 57.5 N. The plantar pressure distribution characteristics reveal the migration pattern of the force center in different gait modes, providing a biomechanical basis for the phased adjustment of the force-position mixing ratio in compliant control.

[0114] In the interactive force channel, the control system extracts the human-computer interaction force in the above-mentioned areas in real time. and its rate of change As input variables, the fuzzy logic inference engine adjusts the prior knowledge of force control, prior knowledge of position control, and virtual compliance parameters in real time to output time-varying impedance parameters. These time-varying impedance parameters include virtual damping corrections and virtual stiffness corrections. The basic adjustment principle of the fuzzy logic inference engine is: when the direction of the interaction force is consistent with the desired motion direction in the current gait phase, and... Larger When the absolute value is large, it is determined that the wearer has an intention to actively exert force. The system outputs instructions to reduce virtual stiffness and virtual damping, increasing the weight of the force control prior knowledge, so that the exoskeleton exhibits low-resistance compliant motion characteristics; when the interaction force is concentrated in the support area and When the impact discrimination condition is met, the wearer is determined to be in a supported bearing state. The system outputs instructions to increase virtual stiffness and improve virtual damping, and increases the allocation weight of position control prior knowledge to provide stable support. When the interaction force amplitude is small and changes smoothly, the preset force-position mixing ratio remains unchanged.

[0115] In the motion trajectory channel, the time-varying impedance parameters, adjusted by the fuzzy logic inference engine, are substituted into the time-varying impedance model to represent the human-computer interaction force. Real-time conversion to trajectory correction amount .

[0116] To overcome the mechanical stiffness and response lag caused by the use of fixed compliance parameters in traditional exoskeletons, the time-varying impedance model of this invention is as follows:

[0117]

[0118] in, For virtual mass constant matrix, and These are the time-varying damping matrix and time-varying stiffness matrix calculated in real time by the system, respectively. This represents the correction amount for the desired trajectory of the exoskeleton.

[0119] The aforementioned time-varying damping matrix and time-varying stiffness matrix are constructed by superimposing the reference value and the dynamic compensation amount output in real time by the fuzzy logic inference engine, as shown in the following formulas:

[0120] ,

[0121] in, and The preset reference impedance parameters, and For the fuzzy logic inference engine based on the current human-computer interaction force and its rate of change The virtual damping correction and virtual stiffness correction are output in real time.

[0122] Furthermore, the trajectory correction amount is compared with the initial reference trajectory. After superposition, a corrected, smooth trajectory reflecting the wearer's actual force application intention and exercise needs is obtained. ,satisfy This process achieves a hybrid coordination of force space and position space—the greater the interaction force and the more consistent its direction with the motion, the greater the trajectory correction and the stronger the exoskeleton's compliance; when the interaction force is an impact-type reverse force, the stiffness increases to resist disturbances.

[0123] This invention determines the auxiliary stage (early / middle / late stage or swing stage) of the current moment through gait recognition results, providing phase context constraints for fuzzy inference of the interactive force channel and generation of correction amount of the motion trajectory channel, so that the phase adjustment of the force-position mixing ratio is synchronized with the gait cycle.

[0124] The three channels work together to achieve a force-position hybrid compliant control output based on gait phase, interactive force state, and motion trajectory constraints.

[0125] In one implementation, to ensure the dynamic stability of the system under time-varying impedance parameters, after correcting the impedance output by the fuzzy logic, the transient damping ratio of each degree of freedom is further calculated. : ,in, For the first The time-varying damping coefficient for each degree of freedom is determined by the preset reference damping for that degree of freedom. Damping correction amount in real-time output of fuzzy logic Superimposed to form, For the first The virtual mass coefficients for each degree of freedom are preset constant parameters. For the first The time-varying stiffness coefficients for each degree of freedom are determined by the preset reference stiffness for that degree of freedom. Stiffness correction amount in real-time output of fuzzy logic Superimposed structure, with a defined damping ratio safety boundary. , This is the lower limit threshold for the damping ratio, used to prevent underdamped oscillations in the system caused by excessive stiffness or insufficient damping. This is the upper limit threshold for the damping ratio, used to prevent response hysteresis and loss of compliance due to excessive damping or insufficient stiffness. When When the impedance parameters exceed this range, a safety check and correction should be performed to prevent system oscillation or response lag caused by excessive parameter adjustment.

[0126] like Figure 12 As shown, in time-varying interaction forces ( Figure 12 Under the excitation of (red dotted line), the virtual stiffness of the system ( Figure 12 The blue solid line indicates that the system stiffness can be smoothly and adaptively adjusted according to the load level. Taking the first rotational degree of freedom, representing the flexion and extension motion of the lower limbs, as an example, the system stiffness can be smoothly adjusted within the range of 85.7 N / m to 150.5 N / m depending on the force applied, correspondingly generating a compliance trajectory correction amount from -0.1004 rad to 0.1004 rad. Figure 12 (Green dashed line), thus aligning with the power generation requirements of different time-phase states. From Figure 12It can be seen that fuzzy inference logic has the ability to smoothly and adaptively adjust the virtual compliant parameters and trajectory of the system when faced with time-varying interactive input.

[0127] Because external human-machine interaction forces exhibit nonlinearity and high-frequency noise, directly using them to drive the underlying actuators can easily lead to trajectory overshoot or abrupt changes in auxiliary output using traditional feedback control. Therefore, this invention introduces model predictive control, which smoothly transforms the force-position hybrid control result into the optimal trajectory correction amount through multi-step rolling optimization.

[0128] In one feasible implementation, the trajectory correction amount is optimized based on the desired compliant trajectory to obtain the optimal trajectory correction amount, including:

[0129] Based on the desired smooth trajectory, the trajectory correction amount is optimized through multiple rolling steps until the optimal trajectory correction amount is obtained.

[0130] Specifically as follows:

[0131] First, construct the objective cost function, which includes a system state error penalty term and a control increment penalty term. The formula is as follows:

[0132]

[0133] in, and These are the prediction time domain and the control time domain, respectively. and For the state penalty prior knowledge matrix and the control increment penalty prior knowledge matrix, To predict the first time domain The system's predicted state vector for each step typically consists of state variables such as the current position and velocity of each joint of the exoskeleton. For the first The reference state vector of the step, i.e., the desired compliant trajectory Discrete sampled values ​​at corresponding times, For the first The control input vector to be optimized in the next step. For the first The ideal impedance control output vector of the step is the desired control quantity calculated from the time-varying impedance model. This is the state feedback gain matrix, used to convert the state deviation... The compensation amount is converted into feedback control quantity, which can usually be calculated offline based on a linear quadratic regulator. The state penalty prior knowledge matrix is ​​a symmetric positive definite or semi-positive definite matrix used to apply a weighted penalty to the deviation between the predicted state and the reference state. The prior knowledge matrix for controlling the incremental penalty is a symmetric positive definite matrix used to apply a weighted penalty to the degree to which the control input deviates from the ideal impedance control law.

[0134] In the objective cost function, the first term measures the deviation between the predicted state and the reference state, characterizing the trajectory tracking accuracy requirement; the second term measures the degree to which the control increment deviates from the ideal impedance control law, characterizing the compliance constraint.

[0135] In the set prediction time domain Within the system, the pose trajectory response for the next several steps is predicted based on the current system state, and this is then implemented in the control time domain. Inner rolling solution makes the objective cost function The smallest optimal control increment sequence. The first element of this optimal control increment sequence is extracted and applied to the bottom-level controller, smoothly transforming the phase-mixed force-position compliant control result into the optimal trajectory correction at the current moment. .

[0136] like Figure 12 As shown by the green dashed line, the trajectory correction amount after rolling optimization suppresses high-frequency fluctuations while preserving the system's ability to adapt to external interaction forces. To quantitatively evaluate the system's compliant response performance under dynamic interaction, this embodiment calculates the interaction forces on each degree of freedom. With trajectory correction The cross-correlation function between them is used to evaluate the average phase lag:

[0137]

[0138] in, The cross-correlation function value represents the number of lag steps. The similarity between the trajectory correction and the interaction force sequences; The total length of the sampled sequences participating in the cross-correlation calculation; The lag step (also called the translation step) represents the offset of the interaction force sequence relative to the trajectory correction sequence on the time axis, and its value ranges from [value missing]. ; For the first Trajectory correction at each sampling point; For relative to Lag The interaction force value at each sampling point.

[0139] Cross-correlation function The physical meaning is: if the trajectory correction is a "response" to the interaction force, then there must be a time difference between the two sequences; when In a certain When the cross-correlation function reaches its maximum value, it indicates that the matching degree between the two sequence waveforms is highest at that lag step number. This lag step number corresponds to the number of sampling steps corresponding to the average response delay from sensing the interaction force to generating trajectory correction. When the cross-correlation function reaches its global maximum value, the corresponding translation step number is the lag step number between the two sequences. And then through Converted to response latency Experimental data shows that the compliant response delays of the system in the two degrees of freedom are 0.23 s and 0.32 s, respectively, maintaining a smooth range of approximately 300 ms overall, without producing a noticeable lag. The multi-step prediction capability ensures the smoothness and stability of the execution trajectory even in the presence of unmodeled dynamics and external disturbances.

[0140] In one feasible implementation, auxiliary control parameters are generated based on gait recognition results and optimal trajectory correction, including:

[0141] During real-time walking, the exoskeleton's bottom-level control unit retrieves the corresponding baseline auxiliary torque curve parameters from a pre-set biomechanical auxiliary curve library based on the current gait mode category (including but not limited to one or more state switching intentions such as walking on flat ground, walking uphill, walking downhill, walking up stairs, and walking down stairs) and the current gait phase, as output by the high-precision gait recognition model; simultaneously, it adjusts the optimal trajectory correction amount. The parameters are superimposed onto a preset reference trajectory to obtain the final target trajectory. The aforementioned baseline auxiliary torque curve parameters, the final target trajectory, along with the force control prior knowledge weights and position control prior knowledge weights output by fuzzy logic, together constitute the auxiliary control parameters, serving as the basis for the exoskeleton's motion scheduling. When the high-precision gait recognition model detects a change in the wearer's motion state, the control system updates the auxiliary torque parameters and force-position hybrid control parameters that match the current gait mode based on the current gait recognition result, current gait phase, and human-machine interaction state, while simultaneously incorporating the optimal trajectory correction amount. The local control parameters are processed to smooth the transition, reducing mechanical stiffness and human-machine interaction caused by recognition lag or abrupt intervention. The above-mentioned auxiliary control parameters are superimposed by force-position hybrid weighting—the reference auxiliary torque is weighted with the prior knowledge weight of force control, the position loop torque required to track the target trajectory is weighted with the prior knowledge weight of position control, and time-varying damping torque components are superimposed—to synthesize the final torque command of each joint motor, driving the exoskeleton actuator to output auxiliary force / torque as needed.

[0142] To quantify the controller's performance, this embodiment measures, for example... Figure 13The diagram shows the follow-up results and error curves of the assisted output during the gait cycle. In this embodiment, the assisted output is represented by a joint force rectangle, and the measured torque (red dashed line) actually output by the exoskeleton actuator can smoothly follow the expected reference torque (blue solid line). As can be seen from the error curve below, even during the peak force and gait switching phases, the following error of the assisted output remains at a low level (the maximum error is controlled within 2.2 N·m in this embodiment). This objectively demonstrates that after introducing three-channel phase-separated hybrid control adjustment and MPC multi-step rolling optimization, the exoskeleton actuator can accurately and stably execute on-demand assisted output, avoiding abrupt changes and lag in the assisted output.

[0143] Step 4: Repeat step 3 to continuously acquire distributed exoskeleton sensor data and human-computer interaction status data and generate corresponding auxiliary control parameters to complete the continuous control of the lower limb exoskeleton robot.

[0144] In achieving joint compensation balance, the essence is to reduce the additional compensation demand on non-target joints by improving the consistency of human-machine movement direction and rhythm. This auxiliary output, which matches the wearer's physiological intention, helps reduce the reverse drag force or mechanical stiffness caused by the lag in state machine switching in traditional exoskeletons. When the on-demand assistance of the exoskeleton is smoothly integrated into the individual's motor system, the wearer can reduce the generation of abnormal movement compensation strategies when facing complex terrain. Objectively, this means that the wearer can reduce the need to compensate for forward momentum by leaning forward and twisting the trunk (which helps reduce unnatural compensation by the back muscles) and also reduces excessive reliance on the unexoskeletonized contralateral healthy limb (which helps reduce the risk of asymmetrical force distribution on both lower limbs). The mechanical energy output by the exoskeleton system smoothly distributes the load on the target joints (such as the hip and knee joints), making the biomechanical work distribution of the wearer's joints more coordinated and balanced. Ultimately, at the physical execution level, this achieves the core purpose of the system's compliant design, namely joint compensation balance, enabling a highly compliant lower limb exoskeleton robot with deep adaptation and efficient assistance in multi-task, multi-body environments.

[0145] In traditional control modes, due to sensory recognition lag, the mechanical assistance output of the exoskeleton is out of phase with the individual's actual movement intention. The wearer may use the lower back or other non-target joints for compensatory force, leading to local muscle fatigue and joint stress imbalance. However, with the method of this invention, relying on the low-latency, high-confidence intention recognition results output by the control system, the exoskeleton actuator can output assistance output that matches the current gait phase, achieving the goal of on-demand assistance. This mechanism helps reduce the feeling of mechanical drag, making the individual's biomechanical work more coordinated and distributed among the joints, ultimately achieving the goal of compliant control with balanced joint compensation at the physical execution level.

[0146] like Figure 2 As shown, the compliant design system for a lower limb exoskeleton robot with joint compensation and on-demand assisted force application proposed in this invention includes:

[0147] The distributed sensor array acquisition module is used to collect real-time distributed exoskeleton sensor data and human-computer interaction status data of the wearer during exercise.

[0148] The individual adaptation parameter generation module is used to calculate the equalization and normalization parameters based on the user's sensor data during the calibration phase, thereby achieving feature decoupling of individual wear errors.

[0149] The two-stage balanced learning and update module is used to perform backbone network parameter freezing update and global collaborative fine-tuning on the preset baseline gait recognition model based on the time-series feature dataset of the calibration stage, so as to obtain a high-precision gait recognition model.

[0150] The data preprocessing and time series recognition module is used to call the equalization normalization parameters to map real-time data, construct a sliding time window sequence, load the gait recognition program and perform real-time inference, and output the gait recognition results.

[0151] The three-channel phase-separated hybrid control module is used to construct a three-channel phase-separated hybrid control mechanism based on the gait phase channel, interaction force channel and motion trajectory channel. It uses fuzzy logic to output time-varying impedance parameters and force / position control prior knowledge weights in real time according to human-machine interaction force, and generates trajectory correction amount through time-varying impedance model.

[0152] The predicted trajectory optimization module is used to smoothly transform the trajectory correction amount into the optimal trajectory correction amount through multi-step rolling optimization.

[0153] The joint compensation and equalization execution module is used to fuse gait recognition results with the optimal trajectory correction amount, retrieve the reference auxiliary torque curve parameters, and use the force control prior knowledge weight and the position control prior knowledge weight to weight and superimpose the reference auxiliary torque and the position loop torque to synthesize the joint torque command, which drives the exoskeleton execution mechanism to complete the on-demand auxiliary force application;

[0154] The cloud-deployed communication and model management module is used to establish a connection with a remote cloud-deployed server through a wireless communication network, enabling remote downloading, version verification, and incremental updates of baseline gait recognition models and / or high-precision gait recognition models.

[0155] Furthermore, those skilled in the art should understand that the present invention can also automate and program the steps of the above methods through an electronic terminal device including a processor and a computer-readable storage medium containing a sequence of computer instructions. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for the compliant design of a lower-limb exoskeleton robot for on-demand assistance in force with joint-compensation equalization, characterized by, Includes the following steps: Step 1: Obtain the baseline gait recognition model pre-installed on the control end of the lower limb exoskeleton robot; Step 2: The user wears the lower limb exoskeleton robot and performs system calibration. The distributed exoskeleton sensor data of the user during the calibration phase is obtained, and the baseline gait recognition model is trained based on the distributed exoskeleton sensor data to obtain a high-precision gait recognition model, which is then deployed on the lower limb exoskeleton robot control terminal. Step 3: Acquire distributed exoskeleton sensor data and human-computer interaction status data during real-time operation. The high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data. Then, based on the gait recognition results and human-computer interaction status data, auxiliary control parameters are generated, and the lower limb exoskeleton robot is controlled based on the auxiliary control parameters. Step 4: Repeat step 3 to continuously acquire distributed exoskeleton sensor data and human-machine interaction status data and generate corresponding auxiliary control parameters to complete the continuous control of the lower limb exoskeleton robot.

2. The method of claim 1, wherein the joint-compensated, balanced, on-demand assisted force generating lower extremity exoskeleton robot is characterized by, In step 2, the baseline gait recognition model is trained based on the distributed exoskeleton sensor data to obtain a high-precision gait recognition model, including: The distributed exoskeleton sensor data collected by the user during the calibration phase is preprocessed to obtain a temporal feature dataset. Based on this temporal feature dataset, the baseline gait recognition model is trained to obtain a high-precision gait recognition model.

3. The compliance design method for a lower limb exoskeleton robot with joint compensation and on-demand assisted force application according to claim 2, characterized in that, The process of equalizing and preprocessing the distributed exoskeleton sensor data of the user during the calibration phase to obtain a time-series feature dataset includes: Based on the distributed exoskeleton sensor data of the user during the calibration phase, the equalization normalization parameter corresponding to the user is calculated. Then, the distributed exoskeleton sensor data of the user during the calibration phase is preprocessed using the current equalization normalization parameter to obtain preprocessed distributed exoskeleton sensor data. Then, a sliding time window sequence is extracted from the preprocessed distributed exoskeleton sensor data using a sliding time window to obtain a time-series feature dataset.

4. The method of claim 2, wherein, The step of training the baseline gait recognition model based on the time-series feature dataset to obtain a high-precision gait recognition model includes: The maximum temporal receptive field length of the temporal feature extraction network in the baseline gait recognition model is configured to match the length of each sliding time window sequence in the temporal feature dataset. Based on this temporal feature dataset, the parameter-adjusted baseline gait recognition model is trained using a two-stage equalization learning method to obtain a high-precision gait recognition model.

5. The compliant design method of the joint-compensated and balanced on-demand assisted force-exertion lower-limb exoskeleton robot according to claim 4, wherein The training of the baseline gait recognition model with adjusted parameters using a two-stage equilibrium learning method includes: First training phase: Freeze all prior knowledge parameters of the temporal feature extraction network in the baseline gait recognition model after parameter adjustment, and iteratively update the prior knowledge parameters of the classification decision module in the baseline gait recognition model according to the first preset learning strategy, and finally obtain the gait recognition model after completing the first training phase. Second training phase: Release the prior knowledge parameters of the temporal feature extraction network in the gait recognition model after the completion of the first training phase, and iteratively fine-tune the gait recognition model after the completion of the first training phase according to the second preset learning strategy to finally obtain a high-precision gait recognition model; the learning rate in the first preset learning strategy is greater than the learning rate in the second preset learning strategy.

6. The compliant design method of the joint-compensated balanced on-demand assisted force-exertion lower-limb exoskeleton robot according to claim 1, wherein In step 3, the high-precision gait recognition model at the control end outputs gait recognition results based on the distributed exoskeleton sensor data, including: After equalization preprocessing of the distributed exoskeleton sensor data during real-time operation, temporal feature data is obtained; then, the temporal feature data is input into a high-precision gait recognition model, and the model outputs gait recognition results.

7. The compliance design method for a lower limb exoskeleton robot with joint compensation and on-demand assisted force application according to claim 1, characterized in that, In step 3, auxiliary control parameters are generated based on the gait recognition results and human-computer interaction state data, including: First, current interaction force information and gait phase are generated based on human-computer interaction state data. Then, trajectory correction amount and expected compliant trajectory are generated based on the current interaction force information and gait phase. Next, the trajectory correction amount is optimized based on the expected compliant trajectory to obtain the optimal trajectory correction amount. Finally, auxiliary control parameters are generated based on gait recognition results and optimal trajectory correction amount.

8. The compliance design method for a lower limb exoskeleton robot with joint compensation and on-demand assisted force application according to claim 7, characterized in that, The generation of trajectory correction based on current interaction force information and gait phase includes: Based on the current interaction force information and gait phase, a fuzzy controller is used to generate virtual damping correction and virtual stiffness correction. Then, combined with a time-varying impedance model, a trajectory correction and a desired compliant trajectory are generated.

9. The compliance design method for a lower limb exoskeleton robot with joint compensation and on-demand assisted force application according to claim 7, characterized in that, The optimization of the trajectory correction amount based on the desired compliant trajectory to obtain the optimal trajectory correction amount includes: Based on the desired smooth trajectory, the trajectory correction amount is optimized through multiple rolling steps until the optimal trajectory correction amount is obtained.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the compliant design method for a joint compensation and balancing on-demand assisted force exertion lower limb exoskeleton robot as described in any one of claims 1 to 9.