Exoskeleton mountaineering intention recognition method and system and storage medium

Abstract

This invention provides a method, system, and storage medium for recognizing intentions during mountaineering using an exoskeleton, belonging to the field of exoskeleton robot control technology. The method includes signal preprocessing and compensation, hierarchical feature extraction, multimodal fusion recognition, and control output and safety protection. By performing 50Hz power frequency filtering and 20-450Hz bandpass filtering on electromyographic (EMG) signals, and combining IMU acceleration data, signal distortion caused by muscle deformation is compensated. Dynamic four-dimensional channel optimization is used to extract EMG features, and an improved attention mechanism model is used to extract multi-scale features. Intention recognition is achieved by fusing EMG, terrain, and other multimodal information based on a Bayesian fusion framework. Simultaneously, a triple safety protection strategy is implemented, including hard limit on joint angles, fall detection and protection, and signal loss fault tolerance. This improves the accuracy and control speed of motion intention recognition, while also possessing good safety performance, effectively enhancing the auxiliary effect and reliability of the exoskeleton in mountaineering.

Description

Technical Field

[0001] This invention belongs to the field of exoskeleton robot control technology, specifically relating to an exoskeleton mountaineering intention recognition method, system, and storage medium. Background Technology

[0002] In scenarios such as outdoor rescue, mountain operations, and extreme sports, the motion intent recognition performance of lower limb exoskeleton robots directly determines their enhancement effect on human movement. Mountain climbing, due to its high terrain complexity (slope -30°~45°, step height 0~30cm), strong dynamic movement (step frequency fluctuation ±50%), and high muscle load (quadriceps activation increases 2.3 times compared to flat ground), faces technical challenges in intent recognition that far exceed those of walking on flat ground.

[0003] During mountain climbing, lower limb muscles (especially the rectus femoris and biceps femoris) undergo significant deformation with joint flexion (up to 120°), leading to a nonlinear attenuation of surface electromyography (sEMG) signals. Experimental data shows that when the knee flexion angle increases from 30° to 90°, the sEMG signal-to-noise ratio (SNR) drops from 28dB to 11dB, resulting in a >35% decrease in the accuracy of traditional EMG recognition models. Existing technologies have fundamental limitations:

[0004] The electromyography prediction model proposed by the contact sensing invention compensates for signal fluctuations through adaptive sliding mode control, but the static muscle activation mapping relationship it relies on fails in dynamic deformation scenarios, and the RMSE of hip joint trajectory prediction increases from 7.94 to 13.2 (an increase of 68.8%). The time-domain-frequency domain feature fusion method used in the patent of Tianjin University of Science and Technology (202411442953.3) does not consider the influence of muscle fiber orientation changes on the signal, and the recognition accuracy drops to 78.3% in the squat mountaineering movement.

[0005] Although the capacitive sensing technology developed through non-contact sensing avoids skin contact, its measurement accuracy is affected by the dielectric constant of clothing (εr fluctuation ±20%) and the electrode spacing. In the scenario of wearing hiking pants, the slope recognition error reaches ±5° (3 times higher than the state of bare skin). Although flexible liquid metal electrodes (such as CN119033567A) have stretch adaptability, under high-frequency muscle vibration (hiking step frequency 1.2~2.5Hz), the signal baseline drift is >0.5mV, resulting in a motion switching recognition delay >150ms.

[0006] Current exoskeletons commonly use plantar sensors (pressure sensors, force-sensitive resistors, etc.) to assist in gait phase determination, but this fundamentally contradicts the safety standards for medical-grade devices.

[0007] The tensile and compressive pressure sensing device with biocompatibility issues has a skin contact area of ​​up to 12 cm², which violates the skin contact limit (≤5 cm²) for wearable devices in ISO 13482:2014 and is likely to cause allergic reactions (occurrence rate of about 8.3%). The plantar pressure signal synchronously collected by the invention of Chongqing University (202310581478.7) has a metal electrode corrosion rate of >10% / week in the sweat environment, which leads to signal drift.

[0008] The foot sensor module (including circuitry and packaging) adds an average of 320g to the weight of the exoskeleton, increasing energy consumption during mountain climbing by 15.7% (based on 10km load test data); and in gravel terrain, the sensor is subjected to impact loads (peak value up to 2.5 times body weight) resulting in a failure rate of >20% (outdoor experimental data from Wang Qining et al., 2019), which is much higher than in indoor scenarios (<5%).

[0009] In mountaineering, terrain parameters (slope, obstacle height) and human movement intentions (step length, force intensity) are strongly coupled (correlation coefficient r=0.82). Existing recognition architectures struggle to balance granularity and real-time performance. Random forest models can only distinguish five basic movement types and cannot recognize complex intentions such as climbing a 25° slope and crossing a 15cm obstacle. Insufficient granularity leads to an exoskeleton assistance timing deviation of >200ms. While modeling the correlation of electromyographic signals using graph structures improves the recognition accuracy to 92.1%, the model has 1.2MB of parameters, resulting in an inference latency of 280ms on embedded terminals (such as STM32H743), far exceeding the 150ms safety threshold for mountaineering scenarios. Dual-stream convolutional networks (202210091942.X) suffer from redundant multimodal feature splicing strategies, resulting in a power consumption of 4.2W and only supporting continuous operation for <2 hours, leading to excessive computational overhead.

[0010] Existing methods mostly rely on fixed terrain databases (such as preset slope levels) and do not achieve real-time perception of terrain parameters. In the unseen 35° scree slope scene, the motion recognition confusion rate reaches 31% (compared to only 8% in flat terrain scenes), which increases the risk of exoskeleton joint jamming.

[0011] The core principles of existing technologies are as follows: Electromyography (EMG) prediction model + adaptive sliding mode control. However, the limitation is that deformation compensation relies solely on muscle activation and does not consider terrain. A quantitative indicator is a 15% decrease in trajectory prediction accuracy when the knee is flexed at 90°. Non-contact capacitive sensing for motion mode recognition suffers from limitations due to interference from clothing dielectrics, resulting in low accuracy on steep slopes. A quantitative indicator is an 82% accuracy rate for recognizing motions on a 30° slope. Multi-feature fusion + genetic radial neural network fails to cover complex terrain movements, leading to poor generalization. A quantitative indicator is a confusion rate >25% for complex movement recognition. EMG structure + GNN model suffers from high computational cost and excessive latency. A quantitative indicator is an inference latency of 280ms and a power consumption of 3.5W. Tension / compression sensors for leg movements violate safety regulations and significantly increase weight. A quantitative indicator is a 320g weight gain for a 12cm² skin contact area.

[0012] Current exoskeleton mountaineering intention recognition technology faces three irreconcilable contradictions: Deformation-signal correlation gap: Existing models lack physical modeling of muscle deformation (displacement Δx = 0~5cm) and sEMG signal attenuation, and only compensate through data-driven methods, resulting in accuracy fluctuations of >20% in dynamic scenarios; Safety-performance conflict: The contradiction between foot contact sensing and the ISO13482 standard forces the system to compromise between safety (non-contact) and perception integrity (gait phase); Granularity-efficiency balance problem: The high latency of high-precision models (such as GNN) and the low recognition granularity of lightweight models (such as random forest) cannot match the dynamic response requirements of mountaineering scenarios (≤150ms latency + 12 types of composite action recognition).

[0013] Therefore, it is urgent to build a new perception framework to overcome the three major technical barriers of robust perception of muscle deformation, non-contact terrain parameter acquisition, and hierarchical efficient fusion, so as to achieve accurate, real-time, and safe recognition of mountaineering intentions by exoskeletons. Summary of the Invention

[0014] To address the issues of muscle deformation interference in signal perception, safety conflicts with foot contact sensors, and efficiency imbalances caused by terrain-motion coupling in exoskeleton robots, this invention provides a method, system, and storage medium for recognizing mountaineering intentions in exoskeleton robots.

[0015] To achieve the above objectives, the present invention provides an exoskeleton mountaineering intention recognition method, comprising:

[0016] During mountaineering, the climber wears an exoskeleton mountaineering robot, the method comprising:

[0017] Acquire raw surface electromyography signals of the target muscle groups of climbers, lower limb joint angles, acceleration, distance from the lower limb to the ground, trunk posture data, and air pressure and temperature data of the exercise environment.

[0018] The altitude of the exercise environment is calculated based on air pressure; the slope of the climber's location is calculated using altitude and trunk posture data; limb movement characteristics are extracted using lower limb joint angles and acceleration; terrain features of the exercise environment are extracted using the distance from the lower limbs to the ground, trunk posture data, air pressure, and slope; the tensile displacement of each muscle in the muscle group is estimated using acceleration, and the original surface electromyography (EMG) signal is compensated based on the tensile displacement and ambient temperature; the time-domain and frequency-domain features of the compensated surface EMG signal are extracted to obtain a multidimensional EMG feature vector of the surface EMG signal.

[0019] The multidimensional electromyographic feature vectors are aggregated according to muscle type to obtain aggregated feature groups for each muscle; the aggregated multidimensional electromyographic vector of the aggregated feature groups is calculated; the terrain pattern probability distribution vector is calculated using the limb change motion features and the terrain features of the movement environment; the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector are identified using a preset model to predict the future movement state marker of the climber.

[0020] The exoskeleton mountaineering robot identifies the climber's future movement intentions based on the terrain pattern probability distribution vector and future action state flags, and then generates control commands.

[0021] Preferably, the raw surface electromyography (EMG) signals of the target muscle groups of the climber are acquired using a flexible textile electrode array; the target muscle groups include the rectus femoris, vastus lateralis, and erector spinae; the flexible textile electrode array uses a silver-ammonia blended material and a serpentine yarn design, with an elongation greater than a set threshold 1 and a contact impedance less than a set threshold 2; a 9-channel electrode array is deployed in a 3x3 matrix in the rectus femoris and vastus lateralis, and a 6-channel electrode array is deployed in a 2x3 matrix in the erector spinae; IMU sensors deployed at the hip, knee, and waist of the exoskeleton are used to acquire the climber's lower limb joint angles, accelerations, and trunk posture data; a laser ToF module deployed at the lower end of the calf is used to acquire the distance from the climber's calf to the ground; and a barometer deployed at the waist is used to acquire the air pressure at the climber's location.

[0022] Preferably, before estimating the tensile displacement of each muscle in the muscle group using acceleration and compensating the original surface electromyography (EMG) signal based on the tensile displacement and ambient temperature, the method further includes sequentially filtering the original EMG signal using spatial domain, temporal domain, frequency domain, and physiological correlation, specifically including:

[0023] Spatial domain filtering: Calculate the angle between the signal acquisition direction of the original surface electromyography (EMG) signal of each channel and the direction of the main muscle fiber of the muscle group; determine the effective signal of the original surface EMG signal based on the angle;

[0024] Time-domain filtering: Extract the resting period signal of the effective signal channel, calculate the signal-to-noise ratio of the resting period signal, and remove electromyographic signal channels with a signal-to-noise ratio less than a set threshold of 3.

[0025] Frequency domain filtering: Calculate the preset frequency band energy of the remaining channel electromyography signals after spatial and temporal domain filtering, calculate the ratio of the preset frequency band energy to the total frequency band energy of the current channel, and use the ratio to remove the remaining channel electromyography signals;

[0026] Physiological correlation screening: Forced activation of channels 7-9 of the rectus femoris and channels 17-18 of the vastus lateralis;

[0027] The surface electromyographic signals of the channels, after being filtered by spatial, temporal, frequency, and physiological correlations, are merged and activated to obtain an activated channel index list that satisfies the spatial, temporal, frequency, and physiological correlations; the electromyographic signals of each channel in the activated channel index list are compensated according to the stretching displacement.

[0028] Preferably, the step of extracting the time-domain and frequency-domain features of the compensated surface electromyography (EMG) signal to obtain a multidimensional EMG feature vector specifically includes:

[0029] The temporal characteristics of each sampling point in the compensated activated channel electromyography signal are calculated, specifically including: calculating the mean absolute value, zero-crossing rate, slope sign change, waveform length, variance and root mean square of each sampling point, to obtain a 6-dimensional temporal feature vector for each activated channel;

[0030] The frequency domain features of each sampling point in the compensated activated channel electromyography signal are calculated, specifically including: calculating the median frequency, average frequency, power spectral entropy, wavelet packet energy, bandwidth power ratio and spectral peak of each sampling point, to obtain a 6-dimensional frequency domain feature vector for each activated channel;

[0031] By merging the 6-dimensional time-domain feature vector and the 6-dimensional frequency-domain feature vector of each active channel, a 12-dimensional feature vector for each active channel is obtained.

[0032] Preferably, the method further includes preprocessing the electromyographic signals of each channel in the activated channel index list, specifically including: using an adaptive notch filter and a fourth-order Butterworth bandpass filter to sequentially filter out the 50Hz power frequency of the electromyographic signals of each channel in the activated channel index list, and retaining the 20-450Hz frequency band signals in each channel to obtain the filtered channel electromyographic signals; and using the filtered channel electromyographic signals for compensation.

[0033] Preferably, a Support Vector Machine (SVM) classifier is used to calculate the terrain pattern probability distribution vector, which includes probability distribution vectors for walking on flat ground, ascending a gentle slope, climbing a steep slope, climbing stairs, and braking downhill. A lightweight Attention network model is used to identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyography vector to predict the climber's future action state markers. The future action state markers include standing up, sitting down, turning to avoid obstacles, and resting on the slope.

[0034] Preferably, the method further includes generating a lookupable torque curve database based on the terrain pattern probability distribution vector and future action state flags, and adjusting the direction and speed of the exoskeleton mountaineering robot in real time using the torque curve database.

[0035] Preferably, the control commands include joint angle hard limit protection control, fall detection and protection control, and signal loss fault tolerance protection control.

[0036] The present invention also provides an exoskeleton mountaineering intention recognition system, comprising:

[0037] The data acquisition module is used to acquire raw surface electromyographic signals of the target muscle groups of climbers, lower limb joint angles, acceleration, distance from the lower limbs to the ground, trunk posture data, and air pressure, slope value and temperature data of the exercise environment.

[0038] The feature extraction module is used to calculate the altitude of the exercise environment based on air pressure; calculate the slope value of the climber's location using altitude and trunk posture data; extract limb movement features using lower limb joint angles and acceleration; extract terrain features of the exercise environment using the distance from the lower limbs to the ground, trunk posture data, air pressure, and slope value; estimate the tensile displacement of each muscle in the muscle group using acceleration; compensate the original surface electromyography (EMG) signal based on the tensile displacement and ambient temperature; and extract the time-domain and frequency-domain features of the compensated surface EMG signal to obtain a multidimensional EMG feature vector of the surface EMG signal.

[0039] The robot control module is used to aggregate the multidimensional electromyographic feature vectors according to muscle type to obtain aggregated feature groups for each muscle; calculate the aggregated multidimensional electromyographic vector of the aggregated feature groups; calculate the terrain pattern probability distribution vector using the limb change motion features and the terrain features of the movement environment; identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector using a preset model to predict the future movement state flag of the climber; the exoskeleton climbing robot identifies the climber's future movement intention based on the terrain pattern probability distribution vector and the future movement state flag, and then generates control commands.

[0040] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any of the steps in the exoskeleton mountaineering intention recognition method.

[0041] The exoskeleton mountaineering intention recognition method provided by this invention has the following beneficial effects:

[0042] This system collects electromyographic (EMG) signals, kinematic characteristics, and terrain features from climbers. Breaking through the limitations of traditional single-sensor systems, it constructs a three-dimensional acquisition system encompassing physiological, kinematic, and environmental aspects, ensuring the data comprehensively depicts the complex characteristics of the mountaineering scenario. It calculates the slope of the climber's location by collecting data on the distance from the lower leg to the ground, the rate of change of air pressure at the climber's location, and the ambient temperature. Utilizing non-contact technology, it resolves the safety issues associated with traditional foot-contact sensors. Through channel filtering and compensation of EMG signals, it addresses the core challenge of EMG signal attenuation and decreased signal-to-noise ratio caused by significant muscle deformation during mountaineering, significantly improving EMG signal quality. It extracts multidimensional EMG features in the time and frequency domains of activated channels, ensuring single-channel characteristics are accurately represented. It can characterize muscle activation state in multiple dimensions; then aggregate features by muscle type to avoid data isolation between channels, effectively suppress abnormal data interference, and ensure that the aggregated features can represent the true activation level of the muscle group; at the same time, it splices limb change motion features and motion environment terrain features to form an independent terrain pattern vector, which not only preserves the correlation between terrain and motion, but also avoids dimensional redundancy caused by direct mixing with electromyographic features. Through structural optimization, it reduces feature redundancy, improves subsequent recognition efficiency, and solves the efficiency imbalance problem caused by terrain-motion coupling; the prediction process abandons the limitations of traditional single feature recognition, adopts dual-input preset model recognition of terrain patterns and electromyographic features, and outputs future action state flags, significantly improving the accuracy and foresight of intent recognition. Attached Figure Description

[0043] To more clearly illustrate the embodiments of the present invention and its design, the accompanying drawings required for these embodiments will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0044] Figure 1 This is a flowchart of an exoskeleton mountaineering intention recognition method according to an embodiment of the present invention;

[0045] Figure 2 This is a flowchart of an exoskeleton mountaineering intention recognition method according to an embodiment of the present invention;

[0046] Figure 3 This is a technical framework diagram of the exoskeleton mountaineering intention recognition system according to an embodiment of the present invention;

[0047] Figure 4 This is a general hardware architecture diagram of the multimodal data acquisition layer according to an embodiment of the present invention;

[0048] Figure 5 This is a data acquisition diagram of a multimodal sensor array according to an embodiment of the present invention;

[0049] Figure 6 This is a flowchart illustrating the feature processing of an embodiment of the present invention;

[0050] Figure 7 This is a hierarchical intent recognition engine architecture and information flow diagram according to an embodiment of the present invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand and implement the present invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the present invention and should not be construed as limiting the scope of protection of the invention.

[0052] This invention provides a method for recognizing the intention to climb mountains using an exoskeleton, specifically as follows: Figure 1 As shown, it includes:

[0053] The present invention constructs a four-layer closed-loop technical system consisting of a multimodal data acquisition layer, a feature extraction and preprocessing layer, a hierarchical intent recognition engine, and a control output and application layer. The system architecture is as follows: Figure 2 As shown, through the collaborative innovation of multimodal flexible sensing, dynamic anti-interference signal processing, hierarchical intent recognition, and embedded real-time control, this system systematically solves three core problems in motion intent recognition of mountaineering exoskeletons: muscle deformation interference, foot sensor dependence, and terrain-motion coupling efficiency imbalance. This system achieves collaboration through directional data flow: the acquisition layer obtains raw information from the human body and environment; the preprocessing layer cleans and analyzes the features of multi-source data; the recognition engine decodes intent in complex scenarios through hierarchical modeling; and finally, the control layer transforms decisions into execution commands and ensures system security.

[0054] S1, the multimodal data acquisition layer, constructs a comprehensive sensing network that coordinates human contact and non-contact perception. Its overall hardware architecture is as follows: Figure 3 As shown, it specifically includes:

[0055] The system collects raw surface electromyography (EMG) signals from each channel of each muscle in the target muscle group of the climber; it also collects data on the lower limb joint angles, accelerations, and trunk posture of the climber; it collects the distance from the lower leg to the ground; it collects the rate of change of air pressure and ambient temperature at the climber's location, and calculates the relative altitude change value from the rate of change of air pressure; and it uses the altitude change value and trunk posture data to calculate the slope value at the climber's location.

[0056] S101. Deployment of the electromyography (EMG) sensing system: 9-channel flexible textile electrodes (channels 1-9) arranged in a 3×3 matrix are attached to the rectus femoris muscle region (6×8cm), 9-channel electrodes (channels 10-18) arranged in a 3×3 matrix are attached to the vastus lateralis muscle region (5×6cm), and 6-channel electrodes (channels 19-24) arranged in a 2×3 matrix are attached to the erector spinae muscle region (4×5cm). The electrodes are made of silver-ammonia blended material (elongation ≥200%), with a single electrode diameter of 12mm and a spacing of 15mm between adjacent electrodes. They are fixed with medical silicone fixation straps (pretension 15N) to ensure skin contact impedance ≤10kΩ@100Hz.

[0057] To address signal distortion caused by electrode misalignment during motion, a multi-channel redundancy matrix and dynamic optimization mechanism are employed, as detailed in the appendix. Figure 4 , Figure 4 This diagram illustrates the deployment of an electromyography (EMG) sensing matrix for the lower limb exoskeleton. Based on principles of muscle electrophysiology and biomechanics, the matrix layout adapts to the direction of muscle fibers (pinnate arrangement of the rectus femoris at an angle of 15°-30°). The 15mm fine spacing increases electrode coverage to 98% at 90° knee flexion, greater than the 71% of traditional inventions. While traditional rigid conductors achieve >40% coverage, the serpentine conductor design in this embodiment ensures that the electrode resistance change is <3% under 12% strain.

[0058] S102. Deployment of the motion and environment sensing system: Data from the laser ToF, IMU, and barometer do not directly participate in the DWF fusion of electromyographic signals. Instead, they are used for terrain perception and kinematic modeling, indirectly affecting electromyographic signal processing (e.g., providing joint angles and accelerations for compensation in DWF). Simultaneously, this sensor data will also be fused with other features in the subsequent hierarchical intent recognition engine. Three sets of triaxial IMU sensors are fixed to the hip joint support (sagittal axis), knee joint support (coronal axis, primary and secondary backups), and lumbar support (horizontal reference). The laser ToF sensor is installed on the outer side of the lower leg support, 10cm from the ankle joint, with an installation tilt angle of 5°±1° forward. The barometer is fixed in the center of the lumbar support and equipped with a wind shield to reduce airflow interference.

[0059] S103, Signal Acquisition Figure 5The preprocessing and fusion workflow for multimodal signals is demonstrated, mainly including processing mechanisms for three types of signals: electromyography (EMG) signals, motion data, and terrain data, ultimately outputting a spatiotemporally aligned raw data stream. EMG signal processing employs a dual optimization mechanism: first, real-time impedance monitoring, which continuously detects electrode contact impedance, automatically isolating abnormal channels and triggering alarms when the impedance exceeds 20kΩ, ensuring the reliability of signal acquisition; second, dynamic gain control, which adaptively adjusts the amplification factor (1k-5k times) based on the EMG signal amplitude, ensuring effective amplification of weak signals and no distortion of strong signals. Motion data processing focuses on noise suppression and accuracy improvement. Based on IMU six-axis data (three-axis acceleration + three-axis angular velocity), a Madgwick filter is used for data fusion, outputting stable Euler angles (pitch, roll, and yaw). A 2-5Hz notch filter module is designed to effectively suppress vibration noise within the stride frequency range, addressing mechanical vibration interference in mountaineering scenarios.

[0060] Terrain data processing achieves environmental adaptability optimization; the ToF laser ranging module employs a multi-echo processing algorithm to suppress false detection interference from gravel terrain by filtering effective echo signals; barometer data incorporates a temperature drift correction mechanism based on a formula. Compensation for temperature changes ( T The influence of ambient temperature on air pressure measurement is investigated to improve the accuracy of altitude difference measurement.

[0061] Using the altitude difference data (ΔH) from the barometer and the torso pitch angle (θ) output by the IMU, the slope value is calculated through a geometric relationship and data fusion algorithm. Since there is a strong correlation between torso posture and terrain slope during mountaineering (the torso pitch angle can approximately reflect the direction and angle of terrain tilt), the fusion of the two can eliminate motion artifact interference and achieve accurate slope calculation.

[0062] After the above processing, the system outputs a spatiotemporally aligned raw data stream, including 24-channel electromyography signals, 9-axis IMU data (3-axis acceleration + 3-axis angular velocity + 3-axis magnetic field), 64-partition ToF point cloud data, and corrected barometric pressure data, providing high-quality input for subsequent hierarchical intent recognition.

[0063] S2, the feature extraction and preprocessing layer, receives the raw sensor data stream from the multimodal data acquisition layer. Its core task is to convert the multi-source heterogeneous raw data from 24-channel sEMG, IMU, laser ToF, and barometer into standardized, high-value feature vectors that can be used by the upper-layer recognition engine. The entire processing flow follows a divide-and-conquer strategy, using three parallel processing pipelines—a biosignal branch, a kinematics branch, and a terrain feature branch—to ultimately fuse and generate a unified 42-dimensional feature vector.

[0064] The process involves extracting multidimensional kinematic features of the climber's lower limb joint angles and accelerations; extracting posture features from the climber's torso posture data; extracting the distance features from the climber's lower leg to the ground; extracting altitude and slope features based on the rate of change of air pressure at the climber's location; constructing multidimensional terrain features from posture, distance, altitude, and slope features; filtering the raw surface electromyography (EMG) signals from all channels; estimating the stretching displacement of each muscle group using acceleration; compensating the EMG signals of each activated channel based on the stretching displacement and ambient temperature; and extracting the time-domain and frequency-domain features of the compensated channel activation signals to obtain the multidimensional EMG feature vector for each activated channel. The processing flow is as follows: Figure 6 As shown, the following will be based on Figure 6 The explanation will follow the specific steps:

[0065] S201, a dynamic four-dimensional channel optimization system, addresses the issue of poor electrode-skin contact caused by severe muscle deformation, sweat secretion, and motion artifacts. It selects the channels with the best signal quality from 24 physical channels in real time for subsequent processing, serving as the first crucial barrier to ensure the robustness of sEMG signals. Inputting 24 channels of raw sEMG signals and the current knee joint angle (from the kinematic branch), a progressive selection strategy encompassing spatial, temporal, frequency, and physiological correlations is employed to output an index list of activated channels.

[0066] The four dimensions are processed as follows:

[0067] 1. Spatial domain screening: Calculate the angle between the signal of each channel and the direction of the main muscle fiber, and remove channels with an angle >15° to ensure that the electrode pickup direction is consistent with the direction of physiological signal generation.

[0068] 2. Temporal quality assessment: Calculate the signal-to-noise ratio (SNR) of the remaining channels. ,in For a 200ms signal window, the effective signal is... For 50ms resting period noise, only channels with SNR>25dB are retained.

[0069] 3. Frequency domain verification: Calculate the proportion of energy in the 80-120Hz frequency band to the total energy (20-450Hz), and remove channels with a proportion of less than 30% to filter out high-frequency noise or low-frequency motion artifacts caused by non-muscle contraction.

[0070] 4. Physiological correlation judgment: Under high knee flexion (knee joint angle ≥60°), the distal rectus femoris (ch7-9) and the posterior vastus lateralis (ch16-18) channels are forcibly activated because these channels can still maintain good contact during deformation. This is a redundancy backup mechanism based on prior knowledge.

[0071] S202. Remove the ubiquitous 50Hz mains power and its harmonic interference, which is the most common and essential source of interference in sEMG signal processing. The input is the activated channel sEMG signal after channel optimization, using an adaptive notch filter with a transfer function... as follows:

[0072]

[0073] in, The transfer function of a filter represents the system's frequency response in the Z-domain. Let represent the complex variable in the Z-transform, and let represent the delay operator in the discrete-time system. It is the digital angular frequency corresponding to 50Hz interference. For the corresponding channel sampling rate, r It is the polar radius (in this embodiment) r =0.995, corresponding to a quality factor Q =30, Q Used to control the bandwidth of the filter. The adaptive notch filter is implemented by cascading second-order IIR sections, which can effectively filter out the 50Hz component with minimal impact on adjacent frequency band signals, and outputs an sEMG signal after removing 50Hz power frequency interference.

[0074] The effective information of S203 and sEMG is mainly concentrated in the mid-frequency band (20-450Hz). This step aims to filter out low-frequency baseline drift, DC components and high-frequency thermal noise, and retain the core signal frequency band.

[0075] The input sEMG signal, after notch filtering, is processed using a fourth-order Butterworth bandpass filter. Its differential equation is... Represented as:

[0076]

[0077] Among them, For output signal, For input signal, 、 These are the filter coefficients. For discrete-time indexing, This represents the delay steps (indicating the relationship between the current and historical values). A forward-backward filtering method is used to achieve zero phase distortion, avoiding the delay and timing distortion introduced by the filtering process. A clean sEMG signal is output within the 20-450Hz frequency band.

[0078] S204. To address the core issue of sEMG amplitude decay caused by severe muscle deformation during mountaineering, a physical model is established to perform electromyography-acceleration dynamic deformation compensation, restoring the true level of muscle activation.

[0079] Input bandpass filtered sEMG signal ( ) Three-axis acceleration data from the thigh IMU ( By doubly integrating the thigh acceleration, the muscle stretch displacement is approximately estimated using the formula. Approximate calculation, where For thigh acceleration, This is the cumulative trapezoidal integral function, used for numerical integration. D The value represents the muscle stretch, and t represents time. The deformation coefficient is adjusted based on the real-time temperature. ,in The current temperature. These are laboratory calibration values ​​(rectus femoris: 0.023 / mm, vastus lateralis: 0.017 / mm, erector spinae: 0.012 / mm). According to the formula... The filtered electromyographic signal is compensated, where This is the electromyography signal after bandpass filtering. To compensate for the post-electromyographic signal, These are the deformation parameters after temperature compensation. D The amplitude of the sEMG signal, after deformation compensation correction, is used to measure muscle stretching, and its amplitude more accurately reflects the level of neuromuscular activity.

[0080] S205: Time-domain and frequency-domain feature extraction, extracting numerical features that can characterize muscle activity patterns and information from sEMG signals, compressing a time-domain signal into a low-dimensional, information-rich feature vector.

[0081] Input the deformed sEMG signal (with a 200ms time window). For each active channel, calculate the temporal features independently within a 200ms data window, and output a 6-dimensional temporal feature vector for each active channel. The following six calculation methods are used:

[0082] 1. Mean Absolute Value (MAV): ,in For the first in the window One sampling point, N The number of sampling points in the window. It reflects the average energy level of the signal.

[0083] 2. Zero-crossing rate (ZC): This roughly estimates the frequency change of the signal. Here, {⋅} is an indicator function, taking a value of 1 when the condition within the parentheses (adjacent sampling points have opposite signs and the amplitude difference exceeds a threshold) is met, and a value of 0 otherwise. threshold The set amplitude threshold (usually 0.05-0.1mV) is used to eliminate false positives caused by minute noise.

[0084] 3. Slope sign change (SSC): Reflects the sharpness and complexity of the waveform. ,in Indicates the first Electromyography signal values ​​at each sampling point Used to count the number of times the slope sign changes within a statistical window.

[0085] 4. Waveform Length (WL): , which represents the total complexity of the waveform in the time domain.

[0086] 5. Variance (VAR): ,in This is the average value of the signal within the window, used to measure the degree of fluctuation of the signal deviating from the mean and reflect the stability of the signal.

[0087] 6. Root Mean Square (RMS): Classic energy characteristics.

[0088] Frequency domain features (the power spectral density (PSD) needs to be obtained by performing an FFT transform on the data window first). The input is the deformation-compensated electromyographic signal (processed in a 200ms window). The following six frequency domain features are calculated independently for each activated channel, and a 12-dimensional feature vector (6-dimensional time domain + 6-dimensional frequency domain) is output for each activated channel. The calculation of the six frequency domain features is as follows:

[0089] 1. Median Frequency (MDF): The frequency point at which the total energy of the PSD is divided in half, i.e., the energy below this frequency accounts for 50% and the energy above this frequency accounts for 50%.

[0090] 2. Mean Frequency (MNF): The weighted average frequency of the PSD. ,in Indicates the power spectrum of the first j Frequency values ​​at each frequency point; Indicates the first j Power spectral density values ​​corresponding to each frequency point (unit: mV² / Hz). M This represents the total number of frequency points in the power spectrum.

[0091] 3. Power Spectral Entropy (PSE): Measures the complexity of the spectrum in the 80-120Hz main frequency band. ,in , Indicates the first frequency band within the 80-120Hz range Power percentage at each frequency point Indicates the frequency index within the 80-120Hz frequency band; A larger value indicates a more dispersed spectral distribution.

[0092] 4. Wavelet packet energy (80-160Hz): The coefficient energy of nodes [3,0] and [3,1] is extracted through three-layer db4 wavelet packet decomposition.

[0093] 5. Power-to-Band Ratio (PBR): Reflects the proportion of low-frequency components. ,in For frequency f The power spectral density at that location.

[0094] 6. Peak Spectrum (PSF): Find the maximum power value in the 50-150Hz range. .

[0095] S206: Feature dimensionality reduction and aggregation solves the problem of dynamic changes in the number of features caused by channel optimization, and aggregates channel-level features into more representative and stable muscle group-level features, reducing the burden on subsequent classifiers.

[0096] The 12-dimensional features of all activated channels are input, but instead of using all channel features directly, they are aggregated by muscle group region: the same feature (e.g., MAV value of all activated channels belonging to the same muscle (rectus femoris, vastus lateralis, erector spinae) is collected; for each feature group, the median is calculated. The median is less sensitive to outliers and more robust than the mean. The medians of all features are concatenated sequentially to form the electromyographic feature vector. Output a fixed 18-dimensional electromyographic feature vector. (3 muscle groups × 6 features).

[0097] The final output of this layer is a 42-dimensional fused feature vector composed of 18-dimensional electromyographic features, 6-dimensional articular kinematic features, and 18-dimensional topographic features. This vector is updated every 200ms and directly input into the hierarchical intent recognition engine.

[0098] The S3 hierarchical intent recognition engine is the core decision-making unit of this invention. It receives a 42-dimensional fused feature vector from the feature extraction and preprocessing layers. Through a hierarchical and frequency-based processing strategy, it decomposes the complex motion intent recognition task into three levels: terrain pattern judgment, action state recognition, and underlying signal assurance. This significantly reduces computational complexity while maintaining accuracy, meeting the real-time requirements of embedded platforms. The engine's innovation lies in its adoption of a "pre-environment constraint on action" cognitive logic. It first identifies relatively stable, slowly changing terrain patterns, then analyzes faster-changing human action intentions under these environmental constraints, and finally provides high-fidelity data for upper-level decision-making through continuous underlying physiological signal compensation. Its core architecture and information flow are as follows: Figure 7 As shown.

[0099] The multidimensional electromyographic feature vectors of each activated channel are aggregated according to muscle type to obtain aggregated feature groups for each muscle. The features of each dimension in all aggregated feature groups are sorted, and the median of each dimension feature in each aggregated feature group is taken. The medians of all features in each aggregated feature group are concatenated to obtain the aggregated multidimensional electromyographic vector of each aggregated feature group. Multidimensional terrain features and multidimensional kinematic features are concatenated, and the terrain pattern probability distribution vector of the concatenated features is calculated. A preset model is used to identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector, outputting the future movement state flag of the climber. Based on the terrain pattern probability distribution and the future movement state flag, the movement intention control commands of the exoskeleton climbing are identified.

[0100] S301, Terrain Master Pattern Recognition Layer: Terrain patterns are the most stable and highest-level prior information in mountaineering, directly determining the basic assistance modes that the exoskeleton needs to provide (such as high-torque climbing, low-torque walking, and negative-torque braking). This layer's role is to initially perceive and classify the environment, providing crucial contextual constraints for subsequent action recognition and preventing absurd errors such as misidentifying a steep slope as a squat. Its processing cycle is set at 200ms to ensure rapid response to terrain changes.

[0101] The input consists of 24 environmental and kinematic features from a 42-dimensional feature vector, specifically including: 18-dimensional terrain features and 6-dimensional kinematic features. The 18-dimensional terrain features include 12-dimensional distance features extracted using laser Time-of-Flight (ToF), 3-dimensional altitude / slope features calculated by a barometer, and 3-dimensional attitude (pitch, roll, and heading) features provided by an IMU. The 6-dimensional kinematic features are the 3-dimensional angles and 3-dimensional angular velocities of the hip and knee joints calculated by an IMU.

[0102] The terrain main pattern recognition layer does not directly use electromyography (EMG) features because muscle activity is a result of intent rather than an environmental cause, focusing instead on environmental and fundamental kinematic information. A Support Vector Machine (SVM) is used as the classifier, with its core decision function being... ,in, It is the input 24-dimensional feature vector. It is a Lagrange multiplier. These are sample labels. It is a bias term. This is the kernel function, where N is the number of feature dimensions. A Radial Basis Function (RBF) kernel is chosen: The RBF kernel can effectively handle the mapping relationship between characteristics and nonlinearity in this problem. The optimal hyperparameters (penalty factor C=1.0, kernel coefficient γ) are determined by grid search. They represent the first and second elements in the training dataset, respectively. The SVM outputs a decision value that distinguishes five terrain patterns using a one-vs-one strategy, based on the given samples and the current input sample.

[0103] Output a terrain pattern probability distribution vector containing 5 elements. These correspond to the estimated probabilities of walking on flat ground, ascending a gentle slope (5°-25°), climbing a steep slope (>25°), climbing stairs (15-25cm), and braking downhill, respectively. The category with the highest probability is the currently identified dominant terrain pattern. Indicates the probability of terrain patterns. These represent the estimated probabilities of walking on flat ground, ascending a gentle slope (5°-25°), climbing a steep slope (>25°), climbing stairs (15-25cm), and braking downhill, respectively.

[0104] The terrain pattern recognition layer achieves a terrain pattern recognition accuracy of 93.6% in extreme scenarios such as 35° granite slopes, providing reliable environmental prior information for the entire system and significantly reducing the search space for complex intents.

[0105] S302, ActionStateRecognitionLayer: Under the constraint of known terrain patterns, this layer is responsible for recognizing specific human actions (such as standing up, sitting down, turning, and resting) performed in that environment. These actions change more frequently than the terrain itself, but require a higher level of precision; therefore, the processing cycle is set to 500ms to allow for more complex calculations and higher recognition granularity. Its role is to refine coarse terrain assistance into precise action assistance.

[0106] The input to the action state recognition layer is 18-dimensional electromyographic features, which are aggregated from the feature extraction layer. This directly reflects the activation state and pattern of the neuromuscular system and is the most direct manifestation of action intention. The input is the classification result from the terrain master pattern recognition layer. This prior knowledge is embedded into the recognition process. For example, when recognizing a steep slope, the system pays more attention to the explosive activation patterns of the rectus femoris and vastus lateralis muscles. The 18-dimensional electromyographic features are concatenated with the One-Hot encoding of the terrain pattern to form an extended feature vector. A lightweight Attention model optimized for embedded platforms is used to process the input features, linearly transforming them into query, key, and value vectors through a learnable weight matrix.

[0107] (3)

[0108] in, x Electromyographic characteristics, These represent the query, key, and value vectors, respectively. Let represent the learnable weight matrix. Calculate the attention weights: ,in This is the dimension of the key vector, used for scaling to prevent gradient vanishing. Let be the attention weight, representing the th i The importance of each feature, the weighted summation output needs to satisfy... .

[0109] By compressing the feature dimensions (e.g., setting the dimensions of Q, K, and V to 8), the model parameter size is kept below 32KB, significantly reducing computational and storage overhead. The network outputs basic action probabilities. To further improve reliability, rule-based threshold decision-making is introduced.

[0110] Standing / sitting: Final confirmation is made by combining the pitch angle change rate (>30° / s) of the lumbar IMU.

[0111] Turning obstacle avoidance: The judgment is made by combining the yaw angular velocity threshold (>30° / s) of the lumbar IMU and the difference in activation of bilateral electromyography (such as left / right erector spinae muscles) (>40%).

[0112] Slope rest: Under the premise of identifying sloping terrain, detect electromyographic activity levels that remain below the resting threshold for more than 1.5 seconds.

[0113] Output binary flags (0 / 1) for three action states (stand up / sit down, turn to avoid obstacles, and rest on a slope). At any given time, only one flag is usually active. Under the constraint of prior terrain knowledge, this layer achieves a misclassification rate of less than 8% for the three auxiliary action states, realizing accurate fine-grained intent recognition.

[0114] S303, the Muscle Deformation Compensation Layer, is not an independent recognition layer, but rather a continuous underlying signal protection system. It operates continuously, independent of 200ms or 500ms cycles. Its core function is to counteract the attenuation effect of muscle deformation on sEMG signals in real time, ensuring that the 18-dimensional electromyographic features input to the action recognition layer are high-fidelity and undistorted, thus guaranteeing the accuracy of the final intent recognition from the data source.

[0115] Input raw triaxial acceleration data from the leg IMU ( The original sEMG signal after bandpass filtering (( This process is completed in the feature extraction layer, see S204 for details. This engine layer mainly reflects its logical location and continuous operation characteristics. Its core is the channel-specific deformation compensation model: ,in Channel numbers (1-24). The input is the raw triaxial acceleration data from the leg IMU. Differential calibration based on muscle group (rectus femoris 0.023 / mm, vastus lateralis 0.017 / mm, erector spinae 0.012 / mm) was used to correct for signal attenuation caused by muscle deformation. The channels are numbered. The model runs continuously, independently correcting the signal for each channel.

[0116] The corrected sEMG signal is output and fed back to the feature extraction layer for subsequent feature calculation. This layer is the core technical guarantee for dealing with mountaineering scenarios, enabling the system to maintain an overall recognition accuracy of 92.1% even under severe deformation conditions with knee flexion >60°.

[0117] The engine's final decision is a combination of terrain patterns and various action states, with the output potentially being {Terrain: Steep Slope Climbing, Action: Turning and Obstacle Avoidance}. The exoskeleton controller then generates a complex assist strategy that integrates high torque output and unilateral differential adjustment. Through hierarchical processing and decoupled scheduling (terrain layer 200ms / 70% computing power, action layer 500ms / 20% computing power, compensation layer continuous / 10% computing power), the engine achieves an average latency of 118ms and low power consumption of 1.8W on the TIAM2634 chip. This perfectly meets the ≤150ms safety response threshold requirement for mountaineering scenarios, completely resolving the contradiction between excessive computational overhead of high-precision models and insufficient recognition granularity of lightweight models.

[0118] S4, the control output and application layer, is the execution terminal and security guarantee closed loop of this invention. It receives decision commands (terrain patterns and action states) from the hierarchical intent recognition engine and converts them into precise torque outputs from the exoskeleton joints. Simultaneously, it ensures the reliability and safety of the human-computer interaction system through multiple safety protection mechanisms. This layer's role is to realize the final conversion of motion intent into physical assistance, serving as a crucial bridge connecting "perception-decision" and action. Its core objective is to achieve safe, compliant, and efficient human-computer interaction, while ensuring that the system prioritizes the wearer's safety in any abnormal situation.

[0119] Dynamic Joint Torque Mapping transforms the abstract intentions (such as climbing a steep slope) identified by the upper layer into specific, executable joint motor torque commands. Different terrains and combinations of movements require drastically different assist curves. The role of this module is to establish this intelligent mapping relationship from intention to torque, making the assist conform to the laws of human biomechanics and achieving an "adaptive" assist effect.

[0120] The input intent recognition result comes from the final output of the recognition engine, including the dominant terrain pattern (one of five categories) and activated action state flags (one or more of three categories). The IMU provides real-time hip and knee joint angles and angular velocities. This module employs a model-based dynamic torque mapping strategy, the core of which is a pre-generated, lookup-table database of torque curves. The processing flow is as follows:

[0121] S401. Based on the {terrain, motion} combination, select a baseline torque-angle curve from the database. For example: {terrain: steep slope climbing, motion: none} -> Call the steep slope ascent torque curve: This curve requires a peak torque of 120 Nm during knee extension (approximately 60° flexion) to assist the wearer in overcoming gravity during climbing. {terrain: step climbing, motion: none} -> Call the step climbing double-peak torque curve: This curve provides the first torque peak (80 Nm) at initial knee flexion (approximately 30°) to guide leg lift, and a second, larger torque peak (100 Nm) during primary extension (approximately 60°) to complete the step climbing motion. {terrain: downhill braking, motion: none} -> Call the negative torque damping curve: In this case, the motor does not provide positive power, but instead generates a damping force (proportional coefficient) opposite to the direction of movement. This simulates the eccentric contraction of muscles during downhill walking to absorb impact, stabilize gait, and reduce knee joint load.

[0122] S402. The selected baseline curve will be fine-tuned based on the real-time joint angular velocity. The greater the angular velocity, the faster the user intends to move, and the system will proportionally increase the torque output slightly to match the user's movement rhythm and ensure the smoothness of the assist.

[0123] The output sends the desired torque command to the tandem elastic actuator (SEA) of the hip and knee joints. This system achieves precise and dynamic matching between intention and assistance, with the assistance conforming to the body's natural movement patterns. It avoids the abruptness of traditional on / off control or constant assistance, significantly improving the user experience. Real-world testing shows that when climbing a 35° slope, this mapping strategy can reduce the user's quadriceps electromyographic activity level by approximately 30%, demonstrating significant energy-saving and efficiency-enhancing effects.

[0124] S5. Safety Protection Mechanism: As a medical assistive device, safety is paramount. This mechanism continuously monitors the system status and, in the event of any abnormal situation that could endanger the wearer or the device itself, immediately overrides the normal torque output and switches to safety mode, serving as the system's "ultimate insurance." It utilizes raw data from all sensors and the system's internal status, employing a triple independent safety protection strategy:

[0125] S501, Hard Joint Limit: Prevents exoskeleton joint movement from exceeding mechanical or physiological limits, thus avoiding strains or hardware damage. Before sending drive commands, the underlying controller (such as a Cortex-M series MCU) independently verifies whether the desired angle is within the safe range (hip joint: [-20°, 80°]; knee joint: [0°, 120°]). If the limit is exceeded, the command is immediately clamped to the boundary value, and the clamped safe angle / torque command is output.

[0126] S502, Fall Detection and Protection: Detects instability, slips, and other emergencies, and intervenes immediately to prevent the wearer from falling and getting injured. It monitors the combined acceleration of the lumbar IMU in real time. When a peak acceleration >3g (3 times the acceleration due to gravity) is detected and lasts for more than 50ms, a fall event is determined to have occurred.

[0127] It immediately triggers an emergency state, switching all joint controllers to high-damping mode (equivalent to "brake"), greatly increasing joint movement resistance, quickly locking the exoskeleton, providing support for the wearer and buying them reaction time.

[0128] S503, Signal Loss Fault Tolerance: Even when the primary signal source (such as sEMG) fails, the system can still degrade to prevent sudden loss of support and subsequent user imbalance. It continuously monitors sEMG signal quality. If the system detects the loss of all valid sEMG signals for more than 10 consecutive frames (i.e., 200ms), it is determined that the sensor has detached or is severely malfunctioning.

[0129] The output automatically and smoothly switches from "sEMG-dominated mode" to "IMU-dominated rhythmic walking mode." This mode, based on a pre-stored gait phase model, generates basic walking assistance solely from the rhythmic movements of the IMU. While it cannot recognize complex intentions, it ensures the user's basic walking ability and provides a safety net for returning home. A multi-layered, in-depth defense safety system is constructed, encompassing mechanical limiting, active intervention, and system fault tolerance, 100% preventing secondary injuries caused by system failures and fully meeting the risk control requirements of medical safety standards such as ISO 13482.

[0130] Embedded Real-time Scheduling & Deployment ensures that all the complex algorithms and control logic described above can run stably and in real-time on resource-constrained embedded platforms, meeting the high reliability requirements of exoskeleton systems. The computational load and real-time requirements of each task are input, and a hierarchical task scheduler is deployed on the TIAM2634 chip based on the FreeRTOS real-time operating system. CPU resources and priorities are statically allocated according to the urgency and importance of the tasks.

[0131] High-priority task (priority 9): Safety protection task (especially fall detection), which runs at the highest frequency (1kHz) and can interrupt any other task to ensure absolute timeliness of safety response.

[0132] Medium priority task (priority 7): Joint torque mapping and control task (100Hz), receives the intention result and calculates the output torque to ensure the real-time performance of the assist.

[0133] Low-priority tasks (priority 5): Intent recognition tasks (terrain recognition 5Hz, action recognition 2Hz). These two tasks are triggered at fixed intervals and are computed asynchronously in the background.

[0134] It provides precise scheduling and management of CPU resources. Ultimately, with a resource limit of 128KB RAM, the number of parameters of the entire system (including the recognition model and control logic) is compressed to 48KB, the average recognition and control latency is controlled within 118ms, the overall power consumption is <1.8W, and it can work continuously for more than 4.5 hours on a single charge, perfectly achieving a balance between performance and power consumption and meeting productization requirements.

[0135] To address the problem of muscle deformation interference, this invention constructs a sensing architecture that combines hardware anti-interference with algorithmic compensation. At the hardware anti-interference level, a basic anti-deformation capability is achieved by deploying a 24-channel flexible textile electrode array in the rectus femoris (9 channels), vastus lateralis (9 channels), and erector spinae (6 channels). This electrode array employs a serpentine wiring design, with an elongation rate >200% and a contact impedance <10kΩ@100Hz, enabling it to adapt to the dynamic deformation process of the muscle.

[0136] At the algorithmic compensation level, signal robustness is improved through the synergistic effect of dynamic four-dimensional channel optimization and deformation compensation model. Dynamic four-dimensional channel optimization selects effective signal channels from four dimensions: spatially, channels with an angle <15° to the main muscle fiber direction are prioritized; temporally, channels with an activation signal-to-noise ratio (SNR) >25dB and a correlation with the acceleration signal <0.4 are selected; in the frequency domain, channels with an energy share >30% in the 80-120Hz frequency band are retained; and physiologically, the synergy between electromyographic activation and joint movement is verified (e.g., the activation intensity of the rectus femoris muscle during knee extension needs to be >0.2 × maximum value). Simultaneously, a channel-specific electromyographic signal-acceleration coupled deformation compensation model is established. , For one of them Channel numbers (1-24). Differential calibration based on muscle group (rectus femoris 0.023 / mm, vastus lateralis 0.017 / mm, erector spinae 0.012 / mm) was used to correct for signal attenuation caused by muscle deformation.

[0137] The architecture demonstrates significant technical effectiveness. Even under strong deformation scenarios with the knee bent at 90°, the signal-to-noise ratio of surface electromyography (sEMG) can still be maintained at ≥25dB ​​(≤11dB in similar scenarios with traditional inventions), and the motion intention recognition accuracy reaches 92.1%, which is 17.6% higher than the 78.3% of traditional inventions, effectively solving the signal attenuation problem caused by muscle deformation.

[0138] To completely eliminate reliance on foot contact sensors, this invention constructs a non-contact terrain sensing architecture, achieving accurate perception of foot status and terrain parameters through multi-source sensor data fusion. For non-contact foot status detection, a laser time-of-flight (ToF) module is integrated at the lower leg, directly detecting the foot's height off the ground based on the time-of-flight principle, with a detection accuracy of ±2cm. Simultaneously, it integrates joint angle data (error <3°) calculated by three-axis IMUs of the hip, knee, and waist, and through collaborative analysis of multi-dimensional kinematic information, it determines the foot's ground contact status in real time, ensuring the accuracy of motion state recognition.

[0139] In terms of real-time terrain parameter perception, an invention combining a barometer and an IMU is used: using the elevation difference data collected by the barometer (resolution 0.1 hPa) and the pitch angle information output by the IMU, the slope value in the range of 0°-45° is calculated in real time through a data fusion algorithm, with a resolution of 0.1°. This achieves dynamic perception of terrain parameters in complex mountain climbing scenarios without relying on a preset terrain database.

[0140] The architecture achieves significant technical benefits, with a ground contact detection accuracy of 96%. By eliminating the foot contact sensor, the exoskeleton is 320g lighter overall, and the total skin contact area is controlled within 5cm². It is 100% compliant with ISO13482 medical safety standards and completely replaces the traditional foot contact sensor.

[0141] To address the strong coupling between terrain and motion, this invention designs a three-level collaborative recognition architecture, which achieves accurate parsing of motion intent in complex scenes through layered processing.

[0142] The terrain main pattern recognition layer has a processing cycle of 200ms. It takes 12 laser point cloud areas, 6-dimensional IMU kinematic data and 3-dimensional air pressure slope features as input, and uses SVM support vector machine (RBF kernel) for classification. It outputs the probability distribution of five terrain patterns: walking on flat ground, walking on gentle slopes (5°-25°), climbing steep slopes (>25°), climbing stairs (15-25cm), and braking downhill. This provides an environmental constraint basis for the upper-level action recognition.

[0143] Raw data was acquired using multimodal hardware including laser ToF (lower limb distance), lumbar IMU (torso posture), and barometer (barometric pressure). After preprocessing such as noise suppression, temperature drift correction, and slope calculation, 12-dimensional distance features, 3-dimensional barometric pressure-slope features, and 3-dimensional posture correlation features were extracted and finally aggregated into 18-dimensional terrain features.

[0144] The action state recognition layer updates every 500ms, inputting 18-dimensional electromyographic features (6 dimensions each for rectus femoris, vastus lateralis, and erector spinae) and terrain master pattern recognition results. It performs feature association analysis through a lightweight Attention network with 32KB of parameters, and outputs three types of auxiliary action state markers: standing up / sitting down, turning around to avoid obstacles, and resting on a slope, to achieve refined recognition of action intentions under terrain constraints.

[0145] The muscle deformation compensation layer operates continuously, outputting corrected electromyographic features in real time, providing high-fidelity physiological signal support for the terrain master pattern recognition layer and the action state recognition layer, and ensuring the consistency of cross-level feature fusion.

[0146] The hierarchical mechanism has significant technical effects, achieving a terrain recognition accuracy of 93.6% in a 35° granite slope scene, with a false recognition rate of less than 8% for composite intent (terrain-action combination). Through the fusion processing of 42-dimensional multi-source features (18-dimensional electromyography + 6-dimensional kinematics + 18-dimensional terrain), it effectively solves the problem of the imbalance in recognition efficiency caused by the coupling of terrain and action.

[0147] To meet the real-time and low-power requirements of the exoskeleton embedded terminal, this invention deploys a hierarchical task scheduler on the TIAM2634 chip (128KB RAM), achieving dynamic allocation of computing power through priority division: the terrain main pattern recognition task is triggered every 200ms, occupying 70% of CPU resources; the action state update task is triggered every 500ms, occupying 20% ​​of CPU resources; the electromyography compensation task runs continuously as a background process, occupying the remaining computing power. This strategy compresses the model parameter size to 48KB, controls power consumption to <1.8W (57% lower than the GNN model), and ensures that the 118ms recognition latency meets the 150ms safe response threshold.

[0148] Compared with traditional inventions, this invention achieves significant breakthroughs in core performance indicators. In terms of muscle deformation anti-interference, the sEMG signal-to-noise ratio at 90° knee flexion is improved from ≤11dB in traditional inventions to ≥25dB, an increase of 127%. Foot sensing completely eliminates reliance on contact sensors, achieving full laser replacement. Terrain recognition accuracy reaches 93.6% in a 35° slope scenario, an improvement of 14.1% compared to 82% in a traditional 30° slope scenario. In terms of real-time performance, the recognition latency is reduced from 280ms in the GNN model to 118ms, an improvement of 58%. Power consumption is reduced from 4.2W in a dual-stream convolutional network to 1.8W, a reduction of 57%.

[0149] The core innovations of this invention are as follows: First, it constructs a physical-data fusion sensing framework, overcoming the disconnect between muscle deformation and signal attenuation through flexible electrode dynamic channel selection (hardware interference immunity) and electromyography-accelerometer coupling model (algorithm compensation); Second, it designs a contactless terrain sensing chain, integrating laser ranging (foot status), IMU (joint angle), and barometer (slope) to achieve full ISO13482 compliance; Third, it proposes a hierarchical computing power allocation mechanism, solving the balance problem between recognition granularity and efficiency through decoupled scheduling of terrain classification (200ms / 70% computing power) and action recognition (500ms / 20% computing power). This invention systematically overcomes the three major technical bottlenecks in mountaineering exoskeleton motion intention recognition, and its measured performance comprehensively surpasses the boundaries of existing technologies.

[0150] Based on the same inventive concept, this invention also provides an exoskeleton mountaineering intention recognition system, comprising:

[0151] The data acquisition module is used to acquire raw surface electromyographic signals of the target muscle groups of climbers, lower limb joint angles, acceleration, distance from the lower limbs to the ground, trunk posture data, and air pressure, slope value and temperature data of the exercise environment.

[0152] The feature extraction module is used to calculate the altitude of the exercise environment based on air pressure; calculate the slope value of the climber's location using altitude and trunk posture data; extract limb movement features using lower limb joint angles and acceleration; extract terrain features of the exercise environment using the distance from the lower limbs to the ground, trunk posture data, air pressure, and slope value; estimate the tensile displacement of each muscle in the muscle group using acceleration; compensate the original surface electromyography (EMG) signal based on the tensile displacement and ambient temperature; and extract the time-domain and frequency-domain features of the compensated surface EMG signal to obtain a multidimensional EMG feature vector of the surface EMG signal.

[0153] The robot control module is used to aggregate the multidimensional electromyographic feature vectors according to muscle type to obtain aggregated feature groups for each muscle; calculate the aggregated multidimensional electromyographic vector of the aggregated feature groups; calculate the terrain pattern probability distribution vector using the limb change motion features and the terrain features of the movement environment; identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector using a preset model to predict the future movement state flag of the climber; the exoskeleton climbing robot identifies the climber's future movement intention based on the terrain pattern probability distribution vector and the future movement state flag, and then generates control commands.

[0154] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the exoskeleton mountaineering intention recognition method provided above.

[0155] Specific limitations regarding the computational system for the exoskeleton mountaineering intention recognition method can be found in the limitations of the exoskeleton mountaineering intention recognition method described above, and will not be repeated here. Each module in the aforementioned exoskeleton mountaineering intention recognition system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0156] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for recognizing the intention to climb mountains using an exoskeleton, characterized in that, During mountaineering, the climber wears an exoskeleton mountaineering robot, the method comprising: Acquiring raw surface electromyography (EMG) signals of the target muscle groups of a climber, lower limb joint angles, acceleration, distance from the lower limb to the ground, trunk posture data, and air pressure and temperature data of the exercise environment, specifically includes: acquiring raw EMG signals of the target muscle groups of the climber through a flexible textile electrode array; the target muscle groups include the rectus femoris, vastus lateralis, and erector spinae; the flexible textile electrode array adopts a silver-ammonia blended material and a serpentine wiring design, with an elongation greater than a set threshold 1 and a contact impedance less than a set threshold 2; a 9-channel flexible electrode array deployed in a 3×3 matrix is ​​attached to the rectus femoris. Textile electrode array, i.e., channels 1-9; a 9-channel flexible textile electrode array deployed in a 3×3 matrix attached to the vastus lateralis muscle, i.e., channels 10-18; a 6-channel flexible textile electrode array deployed in a 2×3 matrix attached to the erector spinae muscle, i.e., channels 19-24; IMU sensors deployed at the hip, knee, and waist of the exoskeleton to collect the climber's lower limb joint angles, acceleration, and trunk posture data, respectively; a laser ToF module deployed at the lower end of the calf to collect the distance from the climber's calf to the ground; and a barometer deployed at the waist to collect the air pressure at the climber's location. The altitude of the exercise environment is calculated based on air pressure; the slope of the climber's location is calculated using altitude and trunk posture data; limb movement characteristics are extracted using lower limb joint angles and acceleration; terrain features of the exercise environment are extracted using the distance from the lower limbs to the ground, trunk posture data, air pressure, and slope; the tensile displacement of each muscle in the muscle group is estimated using acceleration, and the original surface electromyography (EMG) signal is compensated based on the tensile displacement and ambient temperature; the time-domain and frequency-domain features of the compensated surface EMG signal are extracted to obtain a multidimensional EMG feature vector of the surface EMG signal. The multidimensional electromyographic feature vectors are aggregated according to muscle type to obtain aggregated feature groups for each muscle; the aggregated multidimensional electromyographic vector of the aggregated feature groups is calculated; the terrain pattern probability distribution vector is calculated using the limb change motion features and the terrain features of the movement environment; a lightweight attention network model is used to identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector to predict the future movement state marker of the climber. The exoskeleton mountaineering robot identifies the climber's future movement intentions based on the terrain pattern probability distribution vector and future action state flags, and then generates control commands.

2. The exoskeleton mountaineering intention recognition method according to claim 1, characterized in that, Before estimating the stretching displacement of each muscle in the muscle group using acceleration and compensating the original surface electromyography (EMG) signal based on the stretching displacement and ambient temperature, the method further includes sequentially filtering the original EMG signal using spatial domain, temporal domain, frequency domain, and physiological correlation. Specifically, this includes: Spatial domain filtering: Calculate the angle between the signal acquisition direction of the original surface electromyography (EMG) signal of each channel and the direction of the main muscle fiber of the muscle group; determine the effective signal of the original surface EMG signal based on the angle; Time-domain filtering: Extract the resting period signal of the effective signal channel, calculate the signal-to-noise ratio of the resting period signal, and remove electromyographic signal channels with a signal-to-noise ratio less than a set threshold of 3. Frequency domain filtering: Calculate the preset frequency band energy of the remaining channel electromyography signals after spatial and temporal domain filtering, calculate the ratio of the preset frequency band energy to the total frequency band energy of the current channel, and use the ratio to remove the remaining channel electromyography signals; Physiological correlation screening: Forced activation of channels 7-9 of the rectus femoris and channels 17-18 of the vastus lateralis; The surface electromyographic signals of the channels, after being filtered by spatial, temporal, frequency, and physiological correlations, are merged and activated to obtain an activated channel index list that satisfies the spatial, temporal, frequency, and physiological correlations; the electromyographic signals of each channel in the activated channel index list are compensated according to the stretching displacement.

3. The exoskeleton mountaineering intention recognition method according to claim 1, characterized in that, The extraction of time-domain and frequency-domain features of the compensated surface electromyography (EMG) signal yields a multidimensional EMG feature vector, specifically including: The temporal characteristics of each sampling point in the compensated activated channel electromyography signal are calculated, specifically including: calculating the mean absolute value, zero-crossing rate, slope sign change, waveform length, variance and root mean square of each sampling point, to obtain a 6-dimensional temporal feature vector for each activated channel; The frequency domain features of each sampling point in the compensated activated channel electromyography signal are calculated, specifically including: calculating the median frequency, average frequency, power spectral entropy, wavelet packet energy, bandwidth power ratio and spectral peak of each sampling point, to obtain a 6-dimensional frequency domain feature vector for each activated channel; By merging the 6-dimensional time-domain feature vector and the 6-dimensional frequency-domain feature vector of each active channel, a 12-dimensional feature vector for each active channel is obtained.

4. The exoskeleton mountaineering intention recognition method according to claim 2, characterized in that, It also includes preprocessing the electromyographic signals of each channel in the active channel index list, specifically including: using an adaptive notch filter and a fourth-order Butterworth bandpass filter to sequentially filter out the 50Hz power frequency of the electromyographic signals of each channel in the active channel index list, and retaining the 20-450Hz frequency band signals in each channel to obtain the filtered channel electromyographic signals; and using the filtered channel electromyographic signals for compensation.

5. The exoskeleton mountaineering intention recognition method according to claim 1, characterized in that, The terrain pattern probability distribution vector is calculated using a support vector machine (SVM) classifier. The terrain pattern probability distribution vector includes the probability distribution vectors for walking on flat ground, walking up a gentle slope, climbing a steep slope, climbing stairs, and braking downhill. A lightweight attention network model is used to identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyography vector to predict the future action state markers of the climber. The future action state markers include standing up, sitting down, turning around to avoid obstacles, and resting on the slope.

6. The exoskeleton mountaineering intention recognition method according to claim 5, characterized in that, It also includes generating a lookupable torque curve database based on the terrain pattern probability distribution vector and future action state flags, and adjusting the direction and speed of the exoskeleton mountaineering robot in real time using the torque curve database.

7. The exoskeleton mountaineering intention recognition method according to claim 1, characterized in that, The control commands include joint angle hard limit protection control, fall detection and protection control, and signal loss fault tolerance protection control.

8. A system for implementing the exoskeleton mountaineering intention recognition method of claim 1, characterized in that, include: The data acquisition module is used to acquire raw surface electromyography signals of the target muscle groups of climbers, lower limb joint angles, acceleration, distance from the lower limbs to the ground, trunk posture data, and air pressure, slope value and temperature data of the exercise environment. The feature extraction module is used to calculate the altitude of the exercise environment based on air pressure; calculate the slope value of the climber's location using altitude and trunk posture data; and extract limb movement features using lower limb joint angles and acceleration. Using the distance from the lower limbs to the ground, trunk posture data, air pressure, and slope values, terrain features of the exercise environment are extracted; The tensile displacement of each muscle in the muscle group is estimated using acceleration, and the original surface electromyography (EMG) signal is compensated based on the tensile displacement and ambient temperature. The time-domain and frequency-domain features of the compensated surface EMG signal are extracted to obtain the multidimensional EMG feature vector of the surface EMG signal. The robot control module is used to aggregate the multidimensional electromyographic feature vectors according to muscle type to obtain aggregated feature groups for each muscle; calculate the aggregated multidimensional electromyographic vector of the aggregated feature groups; calculate the terrain pattern probability distribution vector using the limb change motion features and the terrain features of the movement environment; identify the terrain pattern probability distribution vector and the aggregated multidimensional electromyographic vector using a preset model to predict the future movement state flag of the climber; the exoskeleton climbing robot identifies the climber's future movement intention based on the terrain pattern probability distribution vector and the future movement state flag, and then generates control commands.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images