Knee joint exoskeleton self-adaptive control method and system based on LSTM (Long Short Term Memory) multi-motion mode recognition and joint angle prediction

Abstract

The invention discloses a knee joint exoskeleton adaptive control method and system based on LSTM multi-motion mode recognition and joint angle prediction, and belongs to the field of wearable control. According to the method, X-axis Euler angle signals and Y-axis Euler angle signals of thighs and shanks are collected and preprocessed to form time sequence data, and the time sequence data are input into a two-stage LSTM model: the first-stage LSTM classification model identifies a current motion mode in real time; the second-stage LSTM regression model predicts a future knee joint angle sequence in a look-ahead mode; and finally, generating a rigidity adjusting instruction according to the identified motion mode so as to adjust the equivalent rigidity of the driver. Meanwhile, a trajectory tracking instruction is generated based on the predicted angle sequence, and a driver is controlled to output synchronous auxiliary torque. The system comprises a controller, an IMU (Inertial Measurement Unit) and a driver with a stiffness regulating mechanism. According to the invention, real-time identification and look-ahead compensation of the motion intention by the exoskeleton are realized, and the adaptive ability, man-machine cooperation efficiency and wearing comfort in multiple scenes are improved.

Description

Technical Field

[0001] This invention belongs to the field of wearable control technology, specifically relating to an adaptive control method and system for a knee exoskeleton based on LSTM multi-motion pattern recognition and joint angle prediction. Background Technology

[0002] With technological advancements and an aging population, powered knee exoskeletons, as an important assistive device for human movement, have shown great potential in areas such as functional rehabilitation, mobility assistance, and enhanced work performance. Their goal is to achieve efficient and natural human-machine coordinated movement by providing assistive torque consistent with the user's intended movement. However, existing knee exoskeletons still face the following challenges in terms of human-machine coordination efficiency and wearing comfort when dealing with complex and varied real-world usage scenarios:

[0003] First, there is a contradiction between the static preset of the control strategy and the dynamic motion requirements: Traditional knee exoskeletons generally rely on preset fixed reference trajectories or single closed-loop control to achieve assistance, with the core assumption that human-machine motion characteristics are basically constant. However, in real-world scenarios, the switching of motion modes is significantly time-varying. Fixed reference trajectories and fixed gains can easily lead to delayed responses or involuntary actions in the device, and even human-machine aggression, reducing the overall compatibility and motion efficiency of the human-machine system.

[0004] Secondly, joint state perception is lagging and lacks forward-looking predictive capabilities: Existing technologies mainly acquire real-time joint angle and velocity information through sensors, but there is an inherent delay from signal acquisition to torque output. Control commands based on current or historical states are difficult to match real-time motion requirements, especially during gait phase transitions, where the lag effect disrupts the smoothness of movement. More importantly, the systems generally lack the ability to predict future joint movement trends, failing to provide the necessary lead time to compensate for the delay in control, thus limiting the timeliness and smoothness of assistive forces.

[0005] Third, the mechanical characteristics of the actuator do not match the biomechanical characteristics of the human body: the human knee joint is not a rigid structure, and its equivalent stiffness is dynamically adjusted with the exercise task and gait stage. Traditional actuators focus on output torque and power, and their stiffness is often fixed, has a limited adjustment range, or is slow to adjust. It is difficult to adapt to the natural rapid stiffness requirements of the human body, resulting in it being too soft when high support is needed and too hard when low resistance is needed, affecting the naturalness of movement and long-term comfort.

[0006] In summary, existing technologies have shortcomings in adaptive motion pattern recognition, forward prediction of joint motion trends, and variable stiffness matching of actuators, which limit the adaptive assistance capabilities of knee exoskeletons in dynamic environments.

[0007] In view of this, the present invention is hereby proposed. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive control method and system for a knee exoskeleton based on LSTM multi-motion mode recognition and joint angle prediction. This addresses the technical problems of existing knee exoskeletons, such as poor adaptability of control strategies, lagging joint motion perception, and mismatch of actuator mechanical characteristics, when facing dynamic switching of multiple motion modes. The invention enables the exoskeleton to quickly and accurately perceive and proactively adapt to the wearer's motion intentions, thereby improving its assistive efficiency, human-machine compatibility, and wearing comfort in complex dynamic environments.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] In a first aspect, the present invention provides an adaptive control method for a knee exoskeleton based on LSTM multi-motion pattern recognition and joint angle prediction, comprising the following steps:

[0011] Step 1, Data Acquisition and Preprocessing: Motion signals of the ipsilateral thigh and calf segments of the knee exoskeleton wearer are acquired by two inertial measurement units, and the motion signals are screened, filtered and normalized to form multi-channel time-series input data;

[0012] Step 2, Motion Pattern Recognition: The multi-channel temporal input data obtained in Step 1 is input into the pre-trained first-level LSTM neural network classification model. The first-level LSTM neural network classification model identifies the current motion pattern category of the wearer in real time based on the temporal characteristics of the input data.

[0013] Step 3, Knee joint angle prediction: Input the multi-channel time series input data obtained in Step 1 into the pre-trained second-level LSTM neural network regression model. The second-level LSTM neural network regression model outputs the knee joint angle prediction sequence within a set future time window from the current moment in a look-ahead manner.

[0014] Step 4, Adaptive Control Execution: Based on the current motion mode identified in Step 2, and based on the preset mapping relationship between the motion mode and the equivalent stiffness level of the actuator, an adaptive stiffness adjustment command is generated to adjust the equivalent stiffness of the actuator in the knee exoskeleton; simultaneously, the knee joint angle prediction sequence output in Step 3 is used as the desired trajectory to generate a trajectory tracking command; combining the stiffness adjustment command and the trajectory tracking command, the actuator is controlled to output a corresponding auxiliary torque to achieve adaptive assistance synchronized with the wearer's movement intention.

[0015] Further, in step 1, the motion signal specifically refers to Euler angle data characterizing the sagittal flexion-extension and coronal oscillation postures measured by two six-axis IMU sensors respectively;

[0016] The Euler angle data includes the Euler angles of the X-axis and Y-axis of the thigh segment IMU sensor and the Euler angles of the X-axis and Y-axis of the calf segment IMU sensor. The four Euler angle data together constitute four-channel timing input data.

[0017] Furthermore, in step 2, the first-level LSTM neural network classification model is a dual LSTM layer network structure, which includes a classification input layer for receiving multi-channel time-series data, a first LSTM layer, a second LSTM layer, a first ReLU activation function layer, a first fully connected layer, and a Softmax classification output layer connected in sequence.

[0018] The first LSTM layer has 64 hidden units and its output mode is sequence output; the second LSTM layer has 32 hidden units and its output mode is end state output; the output dimension of the first fully connected layer is equal to the number of preset motion pattern categories; the Softmax classification output layer is used to output motion pattern category labels.

[0019] Furthermore, the preset exercise mode categories include four types: walking on flat ground, running on flat ground, walking uphill, and climbing stairs.

[0020] Furthermore, in step 3, the second-level LSTM neural network regression model is a dual LSTM layer network structure, which includes, in sequence: a regression input layer for receiving multi-channel time series data, a third LSTM layer, a fourth LSTM layer, a second ReLU activation function layer, a second fully connected layer, and a regression output layer;

[0021] The third LSTM layer has 32 hidden units and its output mode is sequence output; the fourth LSTM layer has 16 hidden units and its output mode is end state output; the second fully connected layer has an output dimension of 1, and the regression output layer is used to output the knee joint angle prediction sequence.

[0022] Furthermore, in step 3, the second-level LSTM neural network regression model outputs the knee joint angle prediction sequence in a look-ahead manner, which is achieved through data preprocessing during the model training phase, specifically:

[0023] The true knee joint angle sequence is shifted forward by a fixed look-ahead time offset relative to the multi-channel time-series input data to form training labels.

[0024] During model training, zero-value frames are padded to the front end of the multi-channel time-series input data to align the input data sequence during training with the label sequence having look-ahead time offset in length and index.

[0025] Furthermore, the fixed look-ahead time offset is 100 milliseconds, corresponding to 10 sampling points at a sampling frequency of 100 Hz.

[0026] Secondly, the present invention also provides a knee joint exoskeleton system, controlled by the aforementioned adaptive control method, comprising:

[0027] The controller is used to perform motion pattern recognition, knee joint angle prediction, and adaptive control.

[0028] Two inertial measurement units are fixedly installed on the thigh and lower leg segments of the knee exoskeleton, respectively, to collect motion signals and transmit them to the controller;

[0029] The actuator, controlled by the controller, has a stiffness adjustment function for adjusting the equivalent stiffness to match the current motion mode according to the instructions of the controller and outputting an auxiliary torque.

[0030] Furthermore, the driver includes:

[0031] The motor is fixed to the housing;

[0032] A rope reel is connected to the output shaft of the motor;

[0033] Multiple flexible fiber ropes are arranged in parallel, with one end of each flexible fiber rope wound around the rope winding wheel;

[0034] A slider is connected to the other end of all the aforementioned flexible fiber ropes;

[0035] A guide rail, laid on the housing, is used to guide the linear movement of the slider;

[0036] A reset spring acts on the slider to provide a reset force;

[0037] The stiffness adjustment mechanism includes a cam driven by a servo motor via a clutch shaft. The servo motor drives the cam to rotate to different preset angle positions according to the stiffness adjustment command. The outer contour of the cam selectively engages different combinations of the flexible fiber ropes, so that the engaged flexible fiber ropes are tensioned and connected to the transmission chain, while the unengaged flexible fiber ropes remain slack, thereby achieving graded adjustment of the equivalent stiffness of the actuator.

[0038] Furthermore, the controller adopts a hierarchical architecture, including:

[0039] The embedded single-board computer, serving as the host computer, is used to deploy and run the first-level LSTM neural network classification model and the second-level LSTM neural network regression model to perform motion pattern recognition and knee joint angle prediction accordingly.

[0040] As a lower-level STM32 control board, it receives instructions from the upper-level computer via serial communication and executes the servo control of the motor in the driver and the drive of the servo motor in the stiffness adjustment mechanism.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. This invention uses the Euler angles of the thigh and calf along the X and Y axes as signal sources, eliminating the noise and irregularity of acceleration signals and the significant drift of the Z-axis attitude angle, thus fundamentally improving the stability and temporal correlation of the input signal. Based on this, by building and training a first-level LSTM neural network classification model, it can accurately and in real-time identify the current motion pattern (such as walking on flat ground, running on flat ground, walking uphill, and climbing stairs) from the signal. This solves the contradiction that traditional exoskeletons, relying on fixed reference trajectories, cannot adapt to dynamic motion scenarios, providing crucial prior information for subsequent adaptive stiffness adjustment and significantly improving the exoskeleton's adaptability and human-machine collaboration efficiency in multiple scenarios.

[0043] 2. This invention shifts the training labels (true knee joint angle values) forward by 100 milliseconds relative to the input sequence (IMU signal) on the time axis, and inserts zero-value frames at the beginning of the input sequence to maintain length alignment. The resulting second-level LSTM neural network regression model can directly output a predicted knee joint angle sequence for the next 100 milliseconds based solely on current and historical sensor data. This prediction result serves as the desired trajectory of the actuator, providing advance capability for the control system and effectively compensating for the inherent sensing, computation, and execution delays of the system. This solves the problem of asynchronous auxiliary torque output and impaired motion smoothness caused by perception lag in existing technologies, achieving smoother and more synchronized human-computer interaction.

[0044] 3. The invention adaptively converts the identified motion patterns into adjustment commands for the equivalent stiffness of the actuator through a preset mapping relationship. The stiffness control mechanism in the actuator selectively engages different rope (4 aramid fiber ropes) groups via a servo-driven cam outer contour, enabling rapid and graded adjustment of its output stiffness (e.g., switching between 7 stiffness levels within 3198 to 24469 N / m within 24 milliseconds). This mechanism allows the exoskeleton's mechanical output to dynamically match the biomechanical needs of the human knee joint in different motion phases (e.g., high stiffness during the support phase and low stiffness during the swing phase) and different modes, fundamentally solving the problem of human-machine mechanical mismatch caused by fixed stiffness or limited adjustment capabilities in traditional actuators, significantly improving the flexibility of movement and long-term wear comfort.

[0045] 4. The controller of this invention adopts a hierarchical architecture. This architecture deploys the computationally intensive LSTM model (pattern recognition, angle prediction) on a host computer (such as a Raspberry Pi), while the real-time-critical low-level servo control (motor and servo drive) is handled by a lower-level computer (such as an STM32). The two communicate via an internal bus (such as a serial port). This design ensures both the high-performance operation of the complex neural network model and the accuracy and determinism of actuator control, while avoiding the risk of performance degradation or unstable response due to task overload of a single processor. This architecture enables the system to operate independently and reliably under lightweight and low-power conditions, providing an effective solution for the practical application of intelligent sensing and control technology in the exoskeleton field.

[0046] In summary, this invention systematically solves the problems of existing knee exoskeletons in terms of environmental adaptability, real-time response, human-machine compatibility, and engineering feasibility through signal filtering, high-precision real-time motion pattern recognition, forward-looking knee joint angle prediction, adaptive stiffness control, and efficient hierarchical system architecture design. It significantly improves the overall performance and practical value of knee exoskeleton devices in dynamic and complex environments. Attached Figure Description

[0047] The accompanying drawings are incorporated in and form part of this specification, and together with the description serve to explain the principles of the invention.

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the adaptive control method for the knee exoskeleton of the present invention;

[0050] Figure 2 The diagram shows the IMU acceleration signals under four motion modes of the present invention; (a) is the thigh X-axis acceleration signal; (b) is the calf X-axis acceleration signal; (c) is the thigh Y-axis acceleration signal; (d) is the calf Y-axis acceleration signal; (e) is the thigh Z-axis acceleration signal; (f) is the calf Z-axis acceleration signal.

[0051] Figure 3 The diagram shows the Euler angle signals of the IMU under four motion modes of the present invention; (a) is the Euler angle signal of the thigh X-axis; (b) is the Euler angle signal of the lower leg X-axis; (c) is the Euler angle signal of the thigh Y-axis; (d) is the Euler angle signal of the lower leg Y-axis; (e) is the Euler angle signal of the thigh Z-axis; (f) is the Euler angle signal of the lower leg Z-axis.

[0052] Figure 4 The measured triaxial Euler angle signals of the dual IMU under four motion modes of the present invention are as follows: (a) flat ground walking mode; (b) flat ground running mode; (c) stair climbing mode; (d) hill climbing walking mode.

[0053] Figure 5 This is a schematic diagram of the neuron structure of the LSTM neural network of the present invention;

[0054] Figure 6 This is a schematic diagram of the first-level LSTM neural network classification model architecture of the present invention;

[0055] Figure 7 This is a schematic diagram of the second-level LSTM neural network regression model architecture of the present invention;

[0056] Figure 8 These are the experimental results of the second-level LSTM neural network regression model for the flat ground walking mode of this invention; (a) is the loss function curve; (b) is the knee joint angle prediction result.

[0057] Figure 9 These are the experimental results of the second-level LSTM neural network regression model for the flat running mode of this invention; (a) is the loss function curve; (b) is the knee joint angle prediction result.

[0058] Figure 10 These are the experimental results of the second-level LSTM neural network regression model of the hill-climbing walking mode of this invention; (a) is the loss function curve; (b) is the knee joint angle prediction result.

[0059] Figure 11 These are the experimental results of the second-level LSTM neural network regression model of the stair climbing mode of this invention; (a) is the loss function curve; (b) is the knee joint angle prediction result.

[0060] Figure 12 This is a schematic diagram of the control principle structure of the knee joint exoskeleton of the present invention;

[0061] Figure 13 This is a schematic diagram of the overall structure of the knee exoskeleton system of the present invention;

[0062] Figure 14 This is a schematic diagram of the internal structure of the actuator in the knee exoskeleton system of the present invention;

[0063] Figure 15 The diagrams are as follows: (a) is a schematic diagram of the stiffness control mechanism of the present invention; (b) is a schematic diagram of the working principle;

[0064] Figure 16 This is the experimental result of knee joint angle during flat walking according to the present invention;

[0065] Figure 17 This is the experimental result of knee joint angle during flat running according to the present invention;

[0066] Figure 18 These are the experimental results of the knee joint angle during stair climbing according to the present invention;

[0067] Figure 19 These are the experimental results of the knee joint angle during hill climbing walking according to the present invention;

[0068] Figure 20 This is the result of an experimental study on knee joint torque during flat walking in this invention;

[0069] Figure 21 The results of the electromyography (EMG) experiment on flat ground walking according to the present invention are shown in (a) for the control group and (b) for the experimental group.

[0070] Figure 22 The results of the electromyography (EMG) test on flat ground running are shown in the present invention; (a) is the control group; (b) is the experimental group.

[0071] Figure 23 The results of the electromyographic signal experiment for stair climbing in this invention are shown in (a) and (b) are as follows:

[0072] Figure 24 The results of the electromyographic signal experiment of climbing walking in this invention are shown in (a) and (b) are the control group and experimental group, respectively.

[0073] in:

[0074] 1 is the motor; 2 is the rope reel; 3 is the flexible fiber rope; 4 is the slider; 5 is the guide rail; 6 is the return spring; 7 is the stiffness adjustment mechanism; 71 is the servo motor; 72 is the clutch shaft; 73 is the cam. Detailed Implementation

[0075] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses consistent with some aspects of the invention as detailed in the appended claims.

[0076] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0077] Example 1

[0078] Please see Figure 1This invention provides an adaptive control method for a knee exoskeleton based on LSTM multi-motion pattern recognition and joint angle prediction. The method first relies on motion signals collected by dual inertial measurement units (six-axis IMU sensors); then, it uses a two-stage LSTM neural network to achieve motion pattern recognition and knee joint angle prospective prediction respectively; finally, it combines preset stiffness mapping rules and trajectory tracking control to drive the knee exoskeleton to achieve adaptive assistance synchronized with the wearer's movement intentions. Specifically, it includes the following steps:

[0079] Step 1: Data Acquisition and Preprocessing

[0080] 1.1) Signal Acquisition: A six-axis IMU sensor (model IMU-600) is fixed directly above the thigh and calf clamps on the same side of the knee exoskeleton wearer. The sensor installation posture is strictly calibrated: the X-axis is kept horizontal, the Y-axis is parallel to the axial direction of the thigh and calf, and the Z-axis is aligned with the normal direction of the upper surface of the clamp. This installation method ensures that the acquired motion signals accurately reflect the flexion and extension movements of the lower limb in the sagittal plane and the swing posture in the coronal plane, avoiding signal distortion caused by installation posture deviations. After the sensor is started, static zero-point calibration is required. The sensor is placed still for 2 minutes. The static data acquisition during this period is used to estimate the three-axis zero bias and gravity direction reference, laying the foundation for accurate acquisition of subsequent motion signals. After calibration, the two IMU sensors synchronously acquire motion signals at a sampling frequency of 100Hz to ensure the consistency and integrity of the timing data.

[0081] 1.2) Signal Filtering: The acquired raw signals include 3-axis acceleration and 3-axis Euler angles output from each IMU sensor, totaling 12-dimensional time-series data. Comparative analysis of signal characteristics under four typical exercise modes—walking on flat ground, running on flat ground, walking uphill, and climbing stairs—revealed (e.g.) Figures 2 to 4 As shown in the figure, the 3-axis acceleration signal is significantly affected by foot impact and muscle contraction vibration, resulting in weak signal regularity and obvious peak shifts and amplitude fluctuations across movement patterns and subjects. The Z-axis Euler angle data exhibits significant low-position bias and step drift in movement patterns such as climbing slopes and stairs. The Z-axis data from the lower leg also shows significant axial torsional interference in the contact and lift-off neighborhoods, leading to low correlation with knee joint angle changes. Therefore, neither of these is suitable as a core input feature. Based on signal stability and feature correlation analysis, the original signals were filtered, retaining only the X-axis and Y-axis Euler angles from the thigh segment IMU sensor and the lower leg segment IMU sensor. These four data points together constitute a four-channel temporal input data set. This data can stably characterize the periodic movement features of the lower limb along the flexion-extension axis and shows a clear phase alignment with knee joint angle changes, providing high-confidence input for subsequent movement pattern recognition and angle prediction.

[0082] 1.3) The four-channel time-series input data after screening is further filtered and normalized: Zero-phase low-pass filtering is used to effectively remove high-frequency noise and measurement interference in the signal, while avoiding phase shift during the filtering process, ensuring the accuracy of the time-series features; normalization is used to map the data of each channel to a uniform magnitude range, eliminating the feature weight imbalance problem caused by the difference in physical dimensions between different channels; at the same time, based on the combination of preset outlier judgment rules and manual review, abnormal data segments caused by signal jumps, baseline drift, sensor slippage, etc. are removed to ensure the reliability and continuity of the input data, and finally form four-channel time-series input data that meets the input requirements of LSTM neural network, providing high-quality data support for subsequent model inference.

[0083] Step 2, Motion Pattern Recognition

[0084] The four-channel time-series input data preprocessed in step 1 is input into the pre-trained first-level LSTM neural network classification model. This model is used to accurately determine the current motion mode category of the wearer, providing a priori basis for subsequent stiffness adjustment.

[0085] This invention employs an LSTM neural network to perform motion pattern recognition and knee joint angle tracking. The core reason is that the LSTM neural network can effectively handle the training and prediction of long sequence input data. Its neuron architecture is as follows: Figure 5 As shown. Compared to the neurons in a traditional recurrent neural network, the neurons in an LSTM network have added input gates, forget gates, and output gates. The gate functions are defined by the Sigmoid function (also known as...). The function is defined as follows:

[0086]

[0087] As can be seen from the function definition above, the Sigmoid function will convert the input data... Mapping within the range of [0,1] allows it to selectively "forget" (i.e., zero out) some data while effectively preserving useful data from the preceding sequence and rationally allocating the proportion of data in the loop computation, thereby avoiding the "gradient explosion" and "gradient vanishing" problems common in traditional recurrent neural networks.

[0088] The connections between neurons in an LSTM are consistent with those in a traditional recurrent neural network. Each neuron simultaneously receives state information, output information, and external input information from the input layer from its connected neurons. At time t, the th hidden layer of the LSTM neural network The forgetting gate of each neuron The calculation method is as follows:

[0089]

[0090] In the formula: This is the external input information vector passed from the network input layer to the hidden layer; The output vector is composed of the output information of all neurons in the hidden layer of the network. These are the bias parameters used in the forget gate operation; These are the input weight parameters in the forget gate operation; The weight parameters for the loop operation of the forget gate.

[0091] Input gate It operates similarly to a forget gate, with the specific expression being:

[0092]

[0093] In the formula: These are the bias parameters used in the input gate operation; These are the input weight parameters in the input gate operation; These are the weight parameters for the loop operation of the input gate.

[0094] Meanwhile, the operation method for the LSTM neuron state is as follows:

[0095]

[0096] In the formula: These are bias parameters in the operational state; These are the input weight parameters in the computation state; The weight parameters are used for iterative calculation of the operation state.

[0097] Finally, the output gate is operated as follows:

[0098]

[0099]

[0100] In the formula: These are the bias parameters used in the output gate operation; These are the input weight parameters in the output gate operation; This refers to the weight parameters for loop operations in the output gate operation.

[0101] As can be seen from the above computational mechanism, the LSTM neural network possesses the ability of "adaptive memory selection," making it better at processing sequential data than traditional feedforward or shallow recurrent networks, and exhibiting stronger nonlinear representation capabilities. Considering the characteristics of this task, the LSTM's gating structure (input gate, forget gate, output gate) can autonomously adjust information retention and forgetting strategies during training, effectively alleviating the gradient vanishing / exploding problem in long sequence data training. It is particularly suitable for temporal patterns like gait, which exhibit "short-cycle superimposed gradually changing trends," and performs robustly under gradually changing conditions such as gradual speed changes and increasing slope. Furthermore, LSTM has good tolerance for phase misalignment and local data loss: when short-term shocks or sensor occlusion cause data anomalies, the model can maintain stable representations based on contextual information without relying on precise event alignment. In practice, a moderately sized LSTM (such as the dual-hidden-layer network used in this invention) can cover the temporal receptive field required for the gait's dominant frequency and its lower-order harmonics, retaining sufficient expressive power to learn the coupling relationships between multi-channel data while controlling the number of parameters.

[0102] In terms of output layer design, LSTM can be connected to a classification head and a regression head respectively: the former uses the Softmax function to distinguish four types of motion patterns, and the latter regresses the knee joint angle in the form of Seq2One or Seq2Seq; thus, two tasks can be unified under the same model skeleton and feature representations can be shared, improving model training efficiency and generalization ability.

[0103] For the task of co-identifying motion patterns and knee angles, the motion pattern recognition network and the knee angle prediction network are trained independently with no parameter sharing: the former focuses on distinguishing four motion patterns, while the latter accurately identifies the knee angle given a specific pattern. Based on this design, the architecture of the first-level LSTM neural network classification model of this invention is as follows: Figure 6As shown, a dual LSTM layer network architecture is adopted, with the hierarchical connection relationship as follows: classification input layer, first LSTM layer, second LSTM layer, first ReLU activation function layer, first fully connected layer, and Softmax classification output layer. The classification input layer receives standardized four-channel temporal input data and converts it into a tensor format that the model can process. The first LSTM layer has 64 hidden units and a sequence output mode, which can capture the periodic structure of gait and the transition information of adjacent phases over a longer time scale, fully exploring the temporal dependencies in the data. The second LSTM layer has 32 hidden units and an end-state output mode, which compresses the sequence features output by the first LSTM layer into a compact global representation, effectively reducing the computational complexity and memory access overhead of the model. The first ReLU activation function layer introduces a nonlinear mapping relationship, enhances the feature expression capability of the model, and further alleviates the gradient vanishing problem. The output dimension of the first fully connected layer is set to 4, which corresponds one-to-one with the four preset motion mode categories (walking on flat ground, running on flat ground, walking uphill, and climbing stairs). The Softmax classification output layer calculates the posterior probabilities of the four motion modes (all values ​​are in the range of [0,1] and the sum is 1), and finally outputs the category with the highest probability as the motion mode recognition result, that is, generates the corresponding motion mode category label.

[0104] The model's training was completed offline. The training dataset consisted of real-world data from multiple groups of subjects across four movement modes, collected synchronously via a motion capture system and IMU sensors. It included 467 segments of walking on flat ground, 619 segments of running on flat ground, 413 segments of walking uphill, and 377 segments of climbing stairs, representing complete gait cycles. The data was randomly divided into training, validation, and test sets at a ratio of 75%, 10%, and 15%, respectively, and the data was randomly rearranged to eliminate order correlation. The Adam optimizer was used for training, with an initial learning rate of 0.003 and a maximum of 1000 iterations. The learning rate decreased by a factor of 0.8 every 100 iterations. Model parameters were dynamically adjusted by monitoring the validation set accuracy. To mitigate the risk of gradient explosion during long training sequences, a global gradient threshold of 0.1 was set, and gradients exceeding the threshold were pruned. The validation set data was evaluated every 10 iterations to monitor overfitting and convergence rate. The training algorithm terminated when the optimal fitness remained unchanged for 100 consecutive iterations, ensuring sufficient model training and avoiding redundant computation.

[0105] The first-level LSTM neural network classification model achieved an overall accuracy of 96.40% in independent validation by the subjects, with balanced performance across categories: For the flat walking mode, the precision was 98.11%, recall was 94.55%, and F1 score was 96.30%; for the flat running mode, the precision, recall, and F1 score were all 96.88%; for the uphill walking mode, the precision was 97.37%, recall was 96.10%, and F1 score was 96.73%; and for the stair climbing mode, the precision was 94.12%, recall was 97.56%, and F1 score was 95.81%. The category confusion matrix shows that the flat walking and flat running modes were the most separable, with almost no mutual misclassification. The uphill walking and stair climbing modes showed limited mutual misclassification in continuously ascending scenarios, mainly due to their similarities in support ratio, knee flexion / extension timing, and trunk forward leaning patterns, but there was no systematic bias, and it did not affect the overall recognition performance.

[0106] When the model is running, it receives preprocessed four-channel time-series input data in real time, quickly completes inference calculations on the embedded platform, and outputs the current motion mode category label. The entire recognition process is responsive and can accurately match the wearer's motion state switching, providing timely and reliable mode judgment basis for subsequent adaptive control.

[0107] Step 3: Knee joint angle prediction

[0108] The four-channel time-series input data preprocessed in step 1 is synchronously input into the pre-trained second-level LSTM neural network regression model. This model, with time-series feature learning as its core, can output the knee joint angle prediction sequence within a set future time window from the current moment based on current and historical motion signals in a forward-looking manner. This provides the driver with accurate and advance trajectory tracking targets, effectively compensating for system sensing and execution delays.

[0109] In this invention, the design of the second-level LSTM neural network regression model is isomorphic to the first-level model for motion pattern recognition, with adjustments made only to the number of neurons and the output layer structure based on the characteristics of the regression task. Figure 7As shown, its specific architecture includes a regression input layer, a third LSTM layer, a fourth LSTM layer, a second ReLU activation function layer, a second fully connected layer, and a regression output layer. The regression input layer receives standardized four-channel temporal input data and converts it into a format that the model can process. The third LSTM layer has 32 hidden units and a sequence output mode, preserving the time-series hidden states within the window and fully capturing the transitional features of gait rhythm and joint movement. The fourth LSTM layer has 16 hidden units and an end-state output mode, compressing the sequence features of the third LSTM layer into a compact global representation, ensuring feature expressiveness while controlling the number of parameters. The second ReLU activation function layer introduces a non-linear mapping relationship, enhancing the model's ability to fit complex motion patterns and alleviating the gradient vanishing problem. The second fully connected layer has an output dimension of 1, used to output a single-dimensional knee angle prediction value. The regression output layer ultimately outputs a knee angle prediction sequence with the same length as the input sequence, corresponding to the knee angle change trajectory over the next 100 milliseconds (i.e., 10 sampling points, adapted to a 100Hz sampling frequency) from the current moment.

[0110] The model's forward prediction capability is achieved through label construction and data alignment strategies during the training phase: When constructing the training dataset, the true knee angle sequence obtained through the optical motion capture system is shifted forward by 100 milliseconds relative to the four-channel temporal input data on the time axis to form training labels with forward time offset, ensuring that the model learns the mapping relationship of "current input - future angle"; to ensure that the input data sequence and the label sequence are strictly aligned in length and index during training, 10 frames with all zeros are padded at the beginning of the four-channel temporal input data, so that each time step of the input sequence matches the angle value of the corresponding forward time in the label sequence one by one, which not only ensures the effectiveness of model training, but also ensures the temporal consistency between the output sequence and the actual motion during online inference.

[0111] The loss function trained on this model is primarily based on mean squared error (MSE), used to minimize the deviation between the predicted and true angles. Its expression is as follows:

[0112]

[0113] In the formula: The true value of the knee angle is given after a look-ahead time of 0.1 seconds. The predicted knee joint angle output by the model is denoted by T, where T is the sequence length. Furthermore, based on actual control requirements, penalty terms for angle amplitude and angular velocity constraints can be added without altering the neural network architecture to ensure that the prediction results conform to the physiological limits of human knee joint movement and the rationality of the mechanism's motion, thus avoiding safety risks associated with excessive flexion or extension.

[0114] The training hyperparameters of the second-stage LSTM neural network regression model were determined through experimental optimization: the batch size was set to 1, and the Sequence Length=longest strategy was used to pad the samples at the end of the batch, while an invalid step size was ignored through a masking mechanism; the optimizer was Adam, the initial learning rate was set to 0.01, the maximum number of iterations was 300, and the learning rate was decreased by a factor of 0.5 every 100 iterations to balance the model's convergence speed and generalization ability; the validation set was evaluated every 10 parameter updates to monitor the training convergence process and the risk of overfitting. When the optimal fitness had not been updated for several consecutive times, training was terminated to avoid redundant computation.

[0115] The training dataset for the model is from the same source as the first-level motion pattern recognition model, also constructed based on measured data from multiple groups of subjects in four motion modes. The data is divided into training, validation, and test sets according to the subjects' independent strategies, with proportions of 75%, 10%, and 15%, respectively. All channels are normalized using offline fixed magnitude differences, and standardized parameters are estimated from the training set and reused during validation and testing to ensure data distribution consistency. In the model performance evaluation phase, the Pearson correlation coefficient is used. As a core indicator, the similarity between the predicted curve and the true curve is quantitatively evaluated, and its expression is as follows:

[0116]

[0117] In the formula: The true value is the sample mean from the perspective of the angle. is the mean of the predicted angle samples; n is the number of samples. Simultaneously, point-to-point error curves and key event deviations are plotted on gait time normalized coordinates (0-100%) to verify the stability of look-ahead prediction in the gait phase transition neighborhood.

[0118] The results of testing the second-level LSTM neural network regression model under four motion modes are as follows:

[0119] (1) Flat ground walking motion mode: its loss function descent curve and prediction results are as follows Figure 8 As shown. Figure 8 (a) shows that during the flat walking training process, the loss function value drops rapidly within 50 steps, rebounds slightly around 60 steps, then drops rapidly again, and converges completely around 80 steps. Figure 8In (b), the predicted results are aligned to the true value 0.1s ahead, with an 8s window. The knee joint angle trajectory during flat walking exhibits the clearest periodicity, showing a stable "flexion-extension-swing-support" phase change pattern. The phase relationships of each peak remain stable, and the peak and trough amplitudes are basically consistent. The relative error is mainly concentrated in the gait window during the support-swing transition, when the knee joint movement amplitude is relatively small. Furthermore, due to the stability of the input signal measured by the IMU under flat walking mode, based on the Pearson correlation coefficient, the overall time history accuracy of the predicted knee joint angle curve reaches 97.15%.

[0120] (2) Flat ground running exercise mode: its loss function descent curve and prediction results are as follows Figure 9 As shown. Figure 9 In (a), the loss function value for running on flat ground shows a jump at about 50 iterations and continues to decrease, and enters a stable convergence interval after about 100 iterations; Figure 9 (b) shows the predicted sequence aligned to the true value with a look-ahead time of 0.1s, displayed within a 7s time window. The results show that, except for the first gait cycle, the peak and trough positions and phase order of the remaining cycles are consistent with the true values. The phase deviation in the first cycle mainly stems from zero padding at the beginning of the input sequence: under conditions of insufficient real starting data and higher running stride frequency and steeper contact pulses, the model makes a relatively conservative estimate of the upcoming phase transition; from the second cycle onwards, as the input information becomes complete, the error converges rapidly and remains stable, demonstrating good synchronization. The overall prediction accuracy, evaluated using the Pearson correlation coefficient, is 96.44%.

[0121] (3) Climbing walking exercise mode: its loss function descent curve and prediction results are as follows Figure 10 As shown. Figure 10 In (a), the loss function value decreases relatively steadily during training, reaching a low point around 50 steps, and then achieves stable convergence. Figure 10 (b) shows the predicted sequence aligned to the ground truth with a look-ahead time of 0.1s, displayed within a 20s time window. The challenge in uphill walking scenarios lies in the gradual change in rhythm and the drift in knee joint posture. From the start to the stable phase, the knee joint's duration, peak amplitude, and stride frequency all change slowly with slight adjustments in slope and speed. A comparison of prediction and actual measurements reveals that some cycles in the middle of the predicted knee joint angle trajectory exhibit underestimation or overestimation of amplitude. This is mainly concentrated in the transition zone, where the actual knee joint angle peak and duration are still adjusting. The measured Euler angle trajectories of the thigh and lower leg from the input IMU sensor are also relatively less stable than those in flat ground walking or running. Nevertheless, the overall phase sequence remains consistent with the main peak and trough positions. Based on the Pearson correlation coefficient, the overall accuracy of the prediction results reaches 86.06%.

[0122] (4) Staircase climbing motion mode: its loss function descent curve and prediction results are as follows Figure 11 As shown. Figure 11 In (a), the loss function value during training exhibits multiple jumps within the first 100 steps, but generally shows a stable downward trend, reaching a low point around 150 steps before entering a slow and stable convergence phase. Given that stair climbing has a repetitive structure of "discrete steps - short pendulum phase - re-contact", the measured Euler angles on the X-axis of the IMU show rapid increases and decreases when lifting the foot and stepping over, while the Euler angles on the Y-axis show perceptible amplitudes when changing steps and adjusting the body laterally. The training loss plateaus around 150 steps, indicating that the model can adapt to the sequence characteristics of "repetitive rather than completely cyclical". Figure 11 (b) shows the comparison between predicted and measured knee angles (prediction aligned to the true value 0.1s ahead, window 20s): most gait cycles showed good alignment in peak amplitude and phase, with some peaks showing deviations and occasional overestimation, mainly attributed to the fitting of the "sudden rise-rapid fall" segment being affected by the instantaneous instability of the measured signal. Overall, no systematic phase lag was observed. Based on the Pearson correlation coefficient, the overall accuracy reached 93.71%.

[0123] Overall, the knee joint angle prediction sequence output by the model can accurately capture the "flexion-extension" pattern of the knee joint in different motion modes. The 100ms look-ahead setting provides sufficient response time for the actuator, effectively compensating for the inherent sensing, calculation and execution delays of the system. Moreover, the prediction results have no cross-cycle error accumulation, providing a high-quality trajectory reference for subsequent adaptive control execution.

[0124] Step 4: Adaptive Control Execution

[0125] By combining the motion pattern recognition results from step 2 with the knee joint angle prediction sequence from step 3, collaborative control commands are generated according to preset logic. Through the synergy of stiffness adaptive adjustment and precise trajectory tracking, adaptive assistance synchronized with the wearer's movement intentions is achieved, ensuring the exoskeleton's human-machine compatibility and assistance effectiveness in different movement scenarios. The specific execution process is as follows:

[0126] First, a stiffness adjustment command is generated based on the motion pattern recognition results. The controller invokes a preset motion pattern-actuator equivalent stiffness level mapping relationship, which is determined through human biomechanical characteristic analysis and experimental verification: In the flat ground walking mode, the human knee joint requires low mechanical resistance in the swing phase to ensure the naturalness of the movement, so it is mapped to a low stiffness level (equivalent stiffness of approximately 8201.2 N / m, corresponding to the actuator's sequence 2 state, i.e., activating the a / b rope combination); in the flat ground running mode, the movement speed is fast and the ground impact load is large, requiring high stiffness to maintain joint stability, so it is mapped to a high stiffness level (equivalent stiffness of approximately 24469.1 N / m, corresponding to the actuator's "sequence 4" state, i.e., activating the a / b / c / d rope combination); in the uphill walking and stair climbing modes, the human body needs to overcome additional gravitational potential energy to do work, and high stiffness can improve power transmission efficiency and reduce transmission delay, so it is also mapped to a high stiffness level. Through this mapping logic, the controller generates targeted stiffness adjustment commands to match the optimal stiffness characteristics for different motion modes. This stiffness adjustment is achieved through a variable stiffness mechanism in the actuator, that is, by stepping in or out of the combination of flexible fiber ropes, the number of flexible fiber ropes participating in power transmission is changed, thereby achieving rapid switching of equivalent stiffness. The response time of the entire adjustment process is on the same order of magnitude as the natural recruitment time of human muscles, adapting to the needs of dynamic motion mode switching.

[0127] Simultaneously, the knee joint angle prediction sequence output from step 3 is used as the desired trajectory. Inverse kinematics calculations are then used to generate a trajectory tracking command. This command includes the target angular displacement, angular velocity, and torque parameters of the motor, clarifying the motor's motion law to drive the knee joint to move along the predicted angle trajectory. The trajectory tracking command and stiffness adjustment command are synchronously transmitted to the exoskeleton's execution control system, driving the variable stiffness mechanism of the actuator to work in conjunction with the motor: the variable stiffness mechanism operates according to the stiffness adjustment command, achieving graded switching of the actuator's equivalent stiffness; the motor adjusts its speed and direction according to the trajectory tracking command, transmitting power to the knee joint actuator through the transmission components, outputting an auxiliary torque synchronized with the predicted angle trajectory, providing precise assistance to the knee joint during the support or extension phases.

[0128] like Figure 12As shown, during the entire control execution process, the system ensures control accuracy and stability through a dual closed-loop feedback mechanism: On the one hand, the motor encoder collects the actual motion state of the motor in real time and feeds back the angular displacement and angular velocity data to the execution control system. The control system, based on the PID closed-loop control algorithm, compares the actual motion state with the expected trajectory and dynamically adjusts the motor output torque to ensure that the deviation between the actual motion trajectory of the knee joint and the predicted angle trajectory is controlled within the allowable range, thereby improving trajectory tracking accuracy. On the other hand, the dual IMU sensors continuously collect the motion signals of the thigh and calf and feed them back to the two-level LSTM neural network in real time to realize the dynamic update of the motion pattern recognition results and the knee joint angle prediction sequence, thereby triggering the real-time adjustment of control commands. This ensures that the stiffness characteristics and motion trajectory of the exoskeleton are always adapted to the wearer's real-time motion state, avoiding human-machine confrontation caused by motion mode switching or posture changes.

[0129] Example 2

[0130] This embodiment provides a knee exoskeleton system that is adapted to the adaptive control method based on LSTM multi-motion pattern recognition and joint angle prediction in the aforementioned embodiment 1. The system mainly consists of a controller, two inertial measurement units (six-axis IMU sensors), a driver, a mechanical body, and a power module. The components are modularly integrated, taking into account lightweight, portability, and reliability.

[0131] Specifically, in this invention, the controller serves as the core computing and control unit, employing a hierarchical architecture of host computer and slave computer to balance the performance requirements of complex algorithm execution and real-time control execution. The host computer is a Raspberry Pi embedded single-board computer, which is small in size, low in power consumption, and has sufficient computing power to meet the offline deployment and real-time inference requirements of the LSTM neural network model. The host computer pre-deploys a trained and optimized first-level LSTM neural network classification model and a second-level LSTM neural network regression model. It receives four-channel time-series input data transmitted from dual IMU sensors in real time via a USB-to-TTL module. After continuously executing model inference, it outputs the current motion mode category label and a knee joint angle prediction sequence for the next 100 milliseconds.

[0132] Furthermore, the lower-level machine uses an STM32 control board, which has rich peripheral interfaces (including CAN bus, serial port, and GPIO) and efficient real-time processing capabilities. It establishes communication with the upper-level machine via serial port. The serial communication protocol is set to a baud rate of 115200, using a five-byte data frame format with a frame header of 0x55 and a frame tail of 0x7D. The second byte, CmdID (0x01 for motion mode instructions and 0x02 for angle control instructions), distinguishes the instruction type, ensuring the stability and accuracy of data transmission. The angle signal employs a high-precision mapping and conversion mechanism during transmission: the Raspberry Pi uses the formula:

[0133]

[0134] The angle value is converted into unsigned 16-bit data for transmission. The STM32 receives this data and uses the following formula:

[0135]

[0136] When converted to actual angles, the angular resolution reaches approximately 0.0165°, meeting the requirements for high-precision trajectory tracking. Furthermore, the host computer sends control commands at 10ms intervals, and the slave computer ensures timely communication response by configuring serial port interrupts, preventing message omissions or slave computer overload.

[0137] In this invention, both inertial measurement units use six-axis IMU sensors (model IMU-600), which are fixedly installed directly above the thigh clamp and calf clamp of the mechanical body, respectively, and are calibrated in accordance with the method of Example 1 before use.

[0138] like Figures 13 to 15 As shown, the actuator, as the execution mechanism of the system of this invention, adopts a muscle-like variable stiffness design, which can dynamically adjust the equivalent stiffness and output auxiliary torque based on the controller command. Its structural design simulates the recruitment mechanism of the human quadriceps muscle and mainly consists of a motor 1, a rope winding wheel 2, multiple parallel flexible fiber ropes 3, a slider 4, a guide rail 5, a return spring 6, and a stiffness adjustment mechanism 7. All components are integrated on the housing (approximately 200mm × 80mm × 60mm). The housing is fixedly connected to the thigh clamp of the mechanical body by bolts to ensure structural stability.

[0139] Specifically, motor 1 is a Sigma SG-6010C robot joint motor. This motor has high power density, a rated torque of 6.5 N·m, a rated speed of 170±10% rpm, a maximum speed of 490±10% rpm, and an operating voltage of 24V, providing sufficient driving force to meet the assistance requirements of different movement modes. The output shaft of motor 1 is directly connected to the rope winding wheel 2. The surface of the rope winding wheel 2 has four evenly distributed grooves to separate and fix multiple flexible fiber ropes 3, preventing the rope bundles from tangling or rubbing against each other.

[0140] The flexible fiber rope 3 is made of aramid fiber rope (also known as Kevlar fiber rope). This material is approximately five times stronger than steel wire and has a density of only 1.44 g / cm³. It combines advantages such as light weight, impact resistance, fatigue resistance, and corrosion resistance, and can maintain stable mechanical properties during high-frequency stretching and resetting processes. In this embodiment, four aramid fiber ropes (clockwise labeled a, b, c, and d) of different diameters are connected in parallel, with diameters of 1.0 mm, 1.5 mm, 2.0 mm, and 2.5 mm, respectively. Their corresponding elastic moduli, as determined by experiments, are 3198.3 N / m, 5002.6 N / m, 6981.3 N / m, and 9286.6 N / m, respectively. Seven levels of stiffness adjustment (a, a / b, a / b / c, a / b / c / d, b / c / d, c / d, d) can be achieved through different combinations, as shown in Table 1 below. Figure 15 As shown in Table 2 below, the stiffness range covers from 3198 N / m to 24469 N / m (corresponding to the rope group from only d rope to all four ropes involved). One end of each flexible fiber rope 3 is wound and fixed in the groove of the rope winding wheel 2, and the other end is connected to the rope group fixing ring on the slider 4 to form a parallel transmission structure.

[0141] Table 1 Rope Group Intervention Combination Table

[0142]

[0143] Table 2 Driver rope assembly and its corresponding output stiffness

[0144]

[0145] The slider 4 is slidably mounted on two sets of guide rails 5 laid inside the housing. The guide rails ensure that the slider can only move in a straight line, with its direction of movement parallel to the thigh axis, to simulate the linear contraction drive law of the human quadriceps muscle. Four return springs 6 are embedded inside the slider. When the motor 1 rotates clockwise to pull the flexible fiber rope 3, the slider 4 moves towards the motor, and the return springs 6 are compressed and store elastic potential energy. When the motor reverses or stops pulling, the return springs release the elastic potential energy, pushing the slider 4 to move in the opposite direction to achieve reset, ensuring the bidirectional movement flexibility of the actuator.

[0146] The stiffness adjustment mechanism 7 is used to realize the graded intervention and withdrawal of the flexible fiber ropes 3 groups, thereby adjusting the equivalent stiffness of the actuator. It mainly consists of a servo motor 71, a clutch shaft 72, and a cam 73. Among them, the servo motor 71 is a Nangu-RC-10kg model, which has a rapid response and can complete a 40° rotation within about 24 milliseconds to meet the requirement of rapid stiffness switching. The servo motor 71 is fixedly connected to the cam 73 through the clutch shaft 72. When the servo motor 71 receives the stiffness adjustment command from the controller, it drives the clutch shaft 72 to rotate the cam 73 to a preset angle. The outer contour of the cam 73 selectively contacts and tensions different combinations of flexible fiber ropes 4, so that the intervened flexible fiber ropes are connected to the transmission chain to participate in power transmission, while the uninterrupted flexible fiber ropes remain in a slack state, realizing the rapid switching of the equivalent stiffness of the actuator and adapting to the requirements of dynamic motion mode switching.

[0147] like Figure 13 As shown, the mechanical body of this invention includes a thigh clamp, a calf clamp, and a transmission unit hinged to both the calf clamp and the drive output end. The thigh and calf clamps are custom-modeled based on the contours of the human lower limb using 3D scanning technology, employing a design strategy of being thicker in the middle (approximately 8mm at the center) and thinner at the edges (approximately 3mm at the edges). This ensures structural rigidity while improving fit through flexible edge deformation. The clamps are made of high-strength nylon and manufactured using 3D printing. Furthermore, both the thigh and calf clamps are equipped with BOA knob straps for fixation, which, together with elastic straps, allow for quick tightening and adjustment. This effectively improves the fit and anti-slip capability of the clamps to the limb, reduces the relative displacement of the human-machine interface during movement, and minimizes interference with the movement trajectory. Preferably, the top of the calf clamp is designed with a sensor expansion platform, which can be used to accommodate additional monitoring sensors or functional modules, enhancing the system's scalability.

[0148] The power module of this invention adopts a multi-power independent power supply scheme to provide stable power to each component of the system: a 24V power supply powers the motor 1 of the driver, meeting the high power output requirements of the motor 1; an 8.4V power supply powers the servo motor 71 of the stiffness control mechanism, ensuring the servo motor 71's fast response and stable operation; a 5V power supply powers the Raspberry Pi host computer, balancing low power consumption and computing power output; the Raspberry Pi provides auxiliary power to the STM32 control board via a serial port, avoiding the impact of a single power supply failure or current fluctuation on the control link. The power module integrates overcurrent and overvoltage protection functions to prevent abnormal voltage from damaging electronic components. It also features a lightweight design, placing it in a backpack along with the control unit, bringing the overall center of gravity of the device closer to the wearer's center of gravity and improving wearing comfort.

[0149] Experimental verification

[0150] To verify the efficacy of this invention, test subjects wore the knee exoskeleton of this invention to complete four exercise modes: flat ground walking, flat ground running, uphill walking, and stair climbing. The actuators used aramid fiber ropes with diameters of 1mm, 1.5mm, 2mm, and 2.5mm, respectively. The experimental parameters were set as follows: the flat ground walking mode used the stiffness level of activation sequence 2, with a treadmill speed of 3m / s; the flat ground running, uphill walking, and stair climbing modes all used the stiffness level of activation sequence 4, with the treadmill speed at 6m / s for flat ground running and the incline of 10° and speed at 3m / s for uphill walking; the stair climbing experiment used a 4-step platform with a single step height of 16cm, and the test subjects had to repeatedly go up and down to complete the experiment. Throughout the experiment, images were captured to observe the movement trajectory, and valid data for at least 10 complete movement cycles were ensured for each movement mode. At the same time, the knee joint angle was recorded using a motion capture system, and the human-machine kinematic characteristics were analyzed to verify the synchronization between the prototype and limb movement. Electromyography (EMG) data of the rectus femoris (RF), vastus lateralis (VL), and vastus medialis (VM) muscles were collected using EMG signal sensors in each movement mode to evaluate the compensatory effect of the prototype on the biomimetic target muscles.

[0151] I. Results of knee joint angle synchronization test

[0152] The knee joint motion angles were captured using a motion capture system under four different motion modes, with the control group (without exoskeleton) and the experimental group (with exoskeleton). The specific results are as follows:

[0153] 1. Flat ground walking mode (e.g.) Figure 16 As shown in the figure): the knee joint motion trajectories of the two groups are basically overlapping. In the early stage of the stance phase, the knee joint angle curve of the experimental group is lower than that of the control group and has a larger slope, indicating that the exoskeleton can achieve torque and velocity compensation during knee extension. In the middle and late stages of the stance phase, the peak position and phase of the main peak of the two groups are basically the same, with only slight differences in amplitude. The error band (green and blue shading) shows that the dispersion within the sample is small, indicating that the exoskeleton has little interference with the natural gait and good human-machine kinematic compatibility.

[0154] 2. Flat ground running mode (e.g.) Figure 17 As shown in the figure): The knee joint angle data of the two groups are generally in phase over time. Although the control group has a higher dispersion of error band due to changes in the running posture of the subjects, data screening or collection errors, the trends of the two are well aligned. The curve of the experimental group has a slightly lower angle and a slightly higher peak amplitude in the early support phase, the joint angle changes faster, and the error band is more convergent overall. This indicates that the joint angle trajectory during running after wearing the prototype is consistent with the natural trajectory trend, and the movement is more regular.

[0155] 3. Stair climbing mode (e.g.) Figure 18As shown): The overall shape of the knee joint angle trajectory of the two groups is basically the same. In the later stage of the support, the knee joint enters a deep flexion state to complete the step-over movement, and in the end of the swing, it quickly extends to complete the step. In the wearing state, there is a slight deviation in the amplitude and phase of the support phase, but the flexion and extension processes coincide with the trajectory of the unwearing state. The error bands of the two groups show that the gait angle difference caused by the step transition is small, indicating that the prototype has little interference with the natural joint movement and high synchronization of the angle trajectory timing.

[0156] 4. Hill Climbing Walking Mode (e.g.) Figure 19 As shown in the figure): the knee joint angles of the two groups are generally in phase over time, and the peak and trough values ​​of the trajectory curves are basically the same; the error band of the control group is wider, reflecting that the knee joint angle under natural conditions is affected by the slope and lower limb posture changes, and has a higher degree of randomness; after wearing the prototype, although it has a certain constraint effect on the random fluctuation of the knee joint posture in natural gait, it does not interfere with the natural gait trajectory, and makes the angle trajectory show higher temporal regularity and lower dispersion.

[0157] II. Results of Knee Joint Torque Compensation Experiment

[0158] Torque tests were conducted in a flat-ground walking mode. Data was collected using a ground force feedback measurement platform, and the knee joint torque comparison results between wearing and not wearing the exoskeleton were calculated (e.g., Figure 20 (As shown). The experiment showed that after wearing the prototype, the average peak knee joint torque per unit body weight of the subjects decreased by about 0.21 Nm / kg, indicating that the prototype can compensate for and reduce the peak of the quadriceps extension demand during the support phase. The torques of the two groups were basically the same during the swing phase, indicating that the underactuated mode has little interference with the non-weight-bearing phase, and the data dispersion of the two groups was low, further verifying the effectiveness of the prototype in assisting the knee.

[0159] III. Results of Electromyographic Signal Compensation Experiment

[0160] Electromyography (EMG) data were collected under four different movement modes using an EMG signal sensor. The compensatory effect of the prototype on the quadriceps femoris muscle was analyzed, and the specific results are as follows:

[0161] 1. Flat ground walking mode (e.g.) Figure 21 As shown): In the unworn and worn prototype states, the peak timing of the electromyographic signals of the three muscles is basically the same; after wearing, the overall electromyographic amplitude of the three muscles generally decreases, with RF showing the most significant decrease (about 20%), followed by VL and VM, indicating that the prototype can effectively compensate for the quadriceps load; and the error band is more convergent in the worn state, indicating that the behavioral pattern is more fixed during the gait cycle, the muscle activation is more stable, and the physiological activation timing of the muscles is not changed.

[0162] 2. Flat ground running mode (e.g.) Figure 2As shown in the figure): In the unworn group, the peak and phase of RF and VL were basically the same, while the peak of VM was lower; in the worn group, VL and VM showed a dominant trend, with similar activation timing and peak values, while the RF amplitude decreased significantly, only about 50% of the peak values ​​of VL and VM, and much lower than the peak values ​​of the unworn group. This indicates that the prototype can significantly reduce the RF recruitment level in running mode, with little impact on VL and VM.

[0163] 3. Stair climbing mode (e.g.) Figure 23 As shown in the figure): In the unworn group, the electromyographic signals of all muscles burst out in a concentrated manner during the ascent, and the peak value and peak phase were basically at the same level; after wearing the exoskeleton, the muscle activation sequence remained basically unchanged, the overall electromyographic amplitude decreased to about 80% of that in the unworn state, and the electromyographic distribution was redistributed - the RF peak increase became dominant, the VL and VM peak values ​​were relatively reduced, and the error band dispersion also converged, indicating that the prototype can provide partial compensation for knee extension movement without affecting the physiological activation sequence of the quadriceps femoris.

[0164] 4. Hill Climbing Walking Mode (e.g.) Figure 24 As shown): The overall peak distribution and phase sequence of the two groups of electromyographic signals are basically consistent; the overall peak values ​​of RF and VL in the wearing group decreased by about 6%, while the amplitude of VM remained basically unchanged. Moreover, the electromyographic activation levels of RF and VL before and after wearing the device increased and decreased at the first peak, which indicates that the prototype can share part of the load of the quadriceps muscle during the knee extension phase, provide assistance with the physiological recruitment sequence, and reduce the peak activation level of the muscle group.

[0165] IV. Experimental Conclusions

[0166] In summary, the multi-scenario wearable experiment covered four basic movement modes. The experimental results showed that the range of motion (ROM) of the knee exoskeleton of the present invention can fully cover the knee flexion angle requirements of each movement mode. The knee angle trajectory in the wearing and unwearing states has good temporal synchronization, excellent human-machine kinematic compatibility, and only slightly constrains the random fluctuations of knee joint posture, without changing the subject's natural joint movement trajectory.

[0167] The above description is merely a specific embodiment of the present invention, intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments without creative effort will be readily apparent to those skilled in the art; the basic principles defined herein can be applied to other embodiments without departing from the spirit and scope of the present invention.

[0168] It should be understood that this invention is not limited to the specific content described above, and various equivalent modifications or substitutions can be made without departing from its scope of protection. The scope of protection of this invention is defined only by the appended claims.

Claims

1. A knee exoskeleton adaptive control method based on LSTM multi-motion mode recognition and joint angle prediction, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: Motion signals of the ipsilateral thigh and calf segments of the knee exoskeleton wearer are acquired by two inertial measurement units, and the motion signals are screened, filtered and normalized to form multi-channel time-series input data; Step 2, Motion Pattern Recognition: The multi-channel temporal input data obtained in Step 1 is input into the pre-trained first-level LSTM neural network classification model. The first-level LSTM neural network classification model identifies the current motion pattern category of the wearer in real time based on the temporal characteristics of the input data. Step 3, Knee joint angle prediction: Input the multi-channel time series input data obtained in Step 1 into the pre-trained second-level LSTM neural network regression model. The second-level LSTM neural network regression model outputs the knee joint angle prediction sequence within a set future time window from the current moment in a look-ahead manner. Step 4, Adaptive Control Execution: Based on the current motion mode identified in Step 2, and based on the preset mapping relationship between the motion mode and the equivalent stiffness level of the actuator, an adaptive stiffness adjustment command is generated to adjust the equivalent stiffness of the actuator in the knee exoskeleton; simultaneously, the knee joint angle prediction sequence output in Step 3 is used as the desired trajectory to generate a trajectory tracking command; combining the stiffness adjustment command and the trajectory tracking command, the actuator is controlled to output a corresponding auxiliary torque to achieve adaptive assistance synchronized with the wearer's movement intention.

2. The knee exoskeleton adaptive control method based on LSTM multi-motion mode recognition and joint angle prediction according to claim 1, characterized in that, In step 1, the motion signal specifically refers to Euler angle data representing sagittal flexion-extension and coronal oscillation postures measured by two six-axis IMU sensors. The Euler angle data includes the Euler angles of the X-axis and Y-axis of the thigh segment IMU sensor and the Euler angles of the X-axis and Y-axis of the calf segment IMU sensor. The four Euler angle data together constitute four-channel timing input data.

3. The adaptive control method for knee exoskeleton based on LSTM multi-motion mode recognition and joint angle prediction according to claim 1, characterized in that, In step 2, the first-level LSTM neural network classification model is a dual LSTM layer network structure, which includes a classification input layer for receiving multi-channel time-series data, a first LSTM layer, a second LSTM layer, a first ReLU activation function layer, a first fully connected layer, and a Softmax classification output layer connected in sequence. The first LSTM layer has 64 hidden units and its output mode is sequence output; the second LSTM layer has 32 hidden units and its output mode is end state output; the output dimension of the first fully connected layer is equal to the number of preset motion pattern categories; the Softmax classification output layer is used to output motion pattern category labels.

4. The adaptive control method for knee exoskeleton based on LSTM multi-motion pattern recognition and joint angle prediction according to claim 3, characterized in that, The preset exercise mode categories include four types: flat walking, flat running, uphill walking, and stair climbing.

5. The knee exoskeleton adaptive control method based on LSTM multi-motion pattern recognition and joint angle prediction according to claim 1, characterized in that, In step 3, the second-level LSTM neural network regression model is a dual LSTM layer network structure, which includes, in sequence: a regression input layer for receiving multi-channel time series data, a third LSTM layer, a fourth LSTM layer, a second ReLU activation function layer, a second fully connected layer, and a regression output layer; The third LSTM layer has 32 hidden units and its output mode is sequence output; the fourth LSTM layer has 16 hidden units and its output mode is end state output; the second fully connected layer has an output dimension of 1, and the regression output layer is used to output the knee joint angle prediction sequence.

6. The adaptive control method for knee exoskeleton based on LSTM multi-motion mode recognition and joint angle prediction according to claim 1, characterized in that, In step 3, the second-level LSTM neural network regression model outputs the knee joint angle prediction sequence in a look-ahead manner. This is achieved through data preprocessing during the model's training phase, specifically: The true knee joint angle sequence is shifted forward by a fixed look-ahead time offset relative to the multi-channel time-series input data to form training labels. During model training, zero-value frames are padded to the front end of the multi-channel time-series input data to align the input data sequence during training with the label sequence having look-ahead time offset in length and index.

7. The adaptive control method for knee exoskeleton based on LSTM multi-motion pattern recognition and joint angle prediction according to claim 6, characterized in that, The fixed look-ahead time offset is 100 milliseconds, corresponding to 10 sampling points at a sampling frequency of 100 Hz.

8. A knee exoskeleton system controlled by the adaptive control method of any one of claims 1 to 7, characterized in that, include: The controller is used to perform motion pattern recognition, knee joint angle prediction, and adaptive control. Two inertial measurement units are fixedly installed on the thigh and lower leg segments of the knee exoskeleton, respectively, to collect motion signals and transmit them to the controller; The actuator, controlled by the controller, has a stiffness adjustment function for adjusting the equivalent stiffness to match the current motion mode according to the instructions of the controller and outputting an auxiliary torque.

9. The knee exoskeleton system of claim 8, wherein, The driver includes: The motor (1) is fixed to the housing; The rope reel (2) is connected to the output shaft of the motor (1); Multiple flexible fiber ropes (3) are arranged in parallel, and one end of each flexible fiber rope (3) is wound around the winding wheel (2). The slider (4) is connected to the other end of all the flexible fiber ropes (3); The guide rail (5) is laid on the housing and is used to guide the linear motion of the slider (4); A reset spring (6) acts on the slider to provide a reset force; The stiffness adjustment mechanism (7) includes a cam (73) driven by a servo motor (71) through a clutch shaft (72). The servo motor (71) drives the cam (73) to rotate to different preset angle positions according to the stiffness adjustment command. The outer contour of the cam (73) selectively engages different combinations of the flexible fiber ropes (3) so that the engaged flexible fiber ropes (3) are tensioned and connected to the transmission chain, while the unengaged flexible fiber ropes (3) remain slack, so as to achieve graded adjustment of the equivalent stiffness of the driver.

10. The knee exoskeleton system of claim 9, wherein, The controller adopts a hierarchical architecture, including: The embedded single-board computer, serving as the host computer, is used to deploy and run the first-level LSTM neural network classification model and the second-level LSTM neural network regression model to perform motion pattern recognition and knee joint angle prediction accordingly. As a lower-level STM32 control board, it receives instructions from the upper-level computer via serial communication and executes the servo control of the motor (1) in the driver and the drive of the servo motor (71) in the stiffness adjustment mechanism (7).

Images