Physiological signal-based rehabilitation exoskeleton adaptive control method and system

Abstract

The application provides a rehabilitation exoskeleton adaptive control method and system based on physiological signals, comprising: acquiring multi-modal physiological signal data of a wearable promoting actuator and preprocessing to obtain a synchronized multi-channel sensing signal sequence; then extracting feature data sets, performing similarity calculation with a pre-built motion intention template library, determining a motion intention category and intensity quantitative value to select a control mode, and completing physiological state classification and obtaining a state confidence; subsequently, judging the safety of a preset motion trajectory through a boundary detection algorithm, calculating a safety margin evaluation value, and adjusting the trajectory below a threshold; and finally calculating target driving parameters to generate real-time motion control instructions. The application realizes efficient browsing and analysis of massive monitoring data, provides a comprehensive intelligent control solution for wearable auxiliary equipment, and significantly improves the safety, effectiveness and individualization level of rehabilitation training.

Description

Technical Field

[0001] This invention belongs to the field of medical rehabilitation equipment technology, and in particular relates to an adaptive control method and system for rehabilitation exoskeleton based on physiological signals. Background Technology

[0002] Rehabilitation exoskeletons are important assistive devices in the field of neurorehabilitation, helping patients with limb dysfunction to conduct rehabilitation training and promoting the recovery of neurological function. However, traditional rehabilitation exoskeletons mostly use fixed-mode control, lacking the ability to perceive and adapt to the patient's real-time physiological state, and making it difficult to personalize them according to the patient's actual needs.

[0003] Most rehabilitation exoskeletons currently used in clinical practice rely on preset movement trajectories for mechanical assistance. They cannot recognize the patient's active movement intentions, nor can they conduct safety assessments and adjustments based on the patient's physiological state. This "mechanical" approach to rehabilitation training not only has limited effectiveness but may also lead to safety issues such as muscle spasms and joint injuries due to the lack of safety mechanisms.

[0004] In recent years, with the development of biosignal acquisition and processing technologies, attempts have begun to incorporate physiological data such as electromyography (EMG) signals into the control of rehabilitation exoskeletons. However, most of these approaches utilize only a single signal source and employ simple threshold judgment methods, making it difficult to cope with complex and variable clinical environments and individual patient differences. Especially for patients with neurological injuries, their physiological signals are often weak, unstable, and contain a significant amount of noise, making accurate and reliable control difficult with current technologies.

[0005] Furthermore, existing rehabilitation exoskeleton systems generally lack effective safety mechanisms. In particular, when patients have abnormal muscle tone or limited joint mobility, fixed movement patterns may cause discomfort or even injury, seriously affecting the continuity and safety of rehabilitation training.

[0006] Therefore, there is an urgent need for a rehabilitation exoskeleton control method that can integrate multiple physiological signals, achieve accurate intention recognition, ensure safety, and be adaptively adjustable, in order to improve the effectiveness and safety of rehabilitation training. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a rehabilitation exoskeleton adaptive control method and system based on physiological signals. This method can effectively integrate multimodal physiological signals to achieve accurate recognition of the patient's movement intentions, real-time assessment of physiological status, and dynamic planning of safe trajectories, thereby providing personalized, safe and effective rehabilitation training.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention provides a method for adaptive control of a rehabilitation exoskeleton based on physiological signals, comprising:

[0010] Acquire multimodal physiological signal data collected by a wearable actuation mechanism, and perform timestamp alignment, resampling, and standardization preprocessing on the multimodal physiological signal data to obtain a synchronized multi-channel sensor signal sequence;

[0011] Based on the synchronized multi-channel sensor signal sequence, a feature dataset is extracted, and the feature dataset is compared with a pre-built motion intention template library to calculate the similarity, determine the current motion intention category and the quantization value of the intention intensity, determine the control mode type based on the current motion intention category and the quantization value of the intention intensity, and perform physiological state classification to obtain the current physiological state category and state confidence.

[0012] Based on the current physiological state category, the state confidence level, and the preset motion trajectory, the safety of the preset motion trajectory is determined by a boundary detection algorithm, and a safety margin assessment value is calculated. When the safety margin assessment value is lower than the safety threshold, the preset motion trajectory is adjusted to obtain the adjusted motion trajectory.

[0013] Based on the safety margin assessment value, the adjusted motion trajectory, the control mode type, and the intent intensity quantification value, the target driving parameters of the actuator are calculated, and real-time motion control commands are generated.

[0014] The beneficial effects of this invention are as follows:

[0015] 1. This invention integrates multiple physiological information such as joint kinematic signals, joint dynamic signals, muscle electrophysiological signals, and peripheral circulatory physiological signals, thereby improving the accuracy and comprehensiveness of the perception of the patient's condition;

[0016] 2. By extracting features in both the time and frequency domains and combining them with an elastic alignment matching algorithm, high-precision recognition of patients' subtle active movement intentions was achieved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the adaptive control method for rehabilitation exoskeleton based on physiological signals provided in an embodiment of the present invention.

[0019] Figure 2This is a schematic diagram of the structure of the adaptive control system for rehabilitation exoskeleton based on physiological signals provided in an embodiment of the present invention. Detailed Implementation

[0020] Example 1

[0021] like Figure 1 As shown, the present invention provides an adaptive control method for a rehabilitation exoskeleton based on physiological signals, comprising the following steps:

[0022] Step S1: Acquire multimodal physiological signal data collected by the wearable actuation mechanism, and perform timestamp alignment, resampling and standardization preprocessing on the multimodal physiological signal data to obtain a synchronized multi-channel sensor signal sequence;

[0023] Step S2: Based on the synchronized multi-channel sensor signal sequence, extract the feature dataset, calculate the similarity between the feature dataset and the pre-built motion intention template library, determine the current motion intention category and the quantization value of the intention intensity, determine the control mode type according to the current motion intention category and the quantization value of the intention intensity, and perform physiological state classification to obtain the current physiological state category and state confidence.

[0024] Step S3: Based on the current physiological state category, the state confidence level, and the preset motion trajectory, the safety of the preset motion trajectory is determined by a boundary detection algorithm, and a safety margin assessment value is calculated. When the safety margin assessment value is lower than the safety threshold, the preset motion trajectory is adjusted to obtain the adjusted motion trajectory.

[0025] Step S4: Calculate the target driving parameters of the actuator based on the safety margin assessment value, the adjusted motion trajectory, the control mode type, and the intent intensity quantification value, and generate real-time motion control commands.

[0026] Specifically, firstly, multimodal physiological signal data from patients is collected using wearable actuators. These signals contain rich information on physiological state and motor intention, but due to being collected from different sensors, there are issues such as inconsistent sampling rates and time asynchrony. Therefore, these signals are first preprocessed by timestamp alignment, resampling, and standardization to unify signals from different sources to the same time reference and numerical range, resulting in a synchronized multi-channel sensor signal sequence, laying the foundation for subsequent analysis.

[0027] Secondly, a feature dataset is extracted based on the preprocessed multi-channel sensor signal sequence. This step extracts signal features from multiple dimensions, including the time and frequency domains, comprehensively capturing various signal attributes. Then, the extracted feature dataset is compared with a pre-established motion intention template library for similarity calculation. Matching analysis is used to determine the patient's current motion intention category (e.g., flexion, extension) and the quantified value of the intention intensity. Based on the identified motion intention and intensity, the method determines the corresponding control mode type, such as passive assistance, active coordination, or an adaptive hybrid mode. Simultaneously, the method also classifies physiological states based on signal features, assessing whether the patient is in a state of fatigue, normalcy, or excitement, and calculates the confidence level of the classification result.

[0028] Next, based on the identified physiological state category, state confidence level, and preset movement trajectory, the safety of the preset trajectory is evaluated using a boundary detection algorithm. Specifically, a safety margin assessment value is calculated, representing the distance between the current movement and the potential risk boundary. When the safety margin assessment value is lower than the preset safety threshold, it means that the preset trajectory may lead to unsafe situations (such as joint hyperextension, muscle spasms, etc.). In this case, the preset movement trajectory will be automatically adjusted to generate a new safe trajectory to ensure the safety of rehabilitation training.

[0029] Finally, taking into account the safety margin assessment value, the adjusted motion trajectory, the control mode type, and the quantified value of intent intensity, the target drive parameters of the actuators (such as motors, pneumatic systems, etc.) are calculated to generate real-time motion control commands. These commands directly control the movement of the exoskeleton device, providing precise force assistance and motion guidance.

[0030] This step is the final stage of the entire control process and a crucial step in translating all the preceding analysis results into actual control actions. First, the safety margin assessment value is considered. This value reflects the distance between the current action and the safety boundary; a higher value indicates a larger safety margin, allowing for more flexible control parameters. When the safety margin is low, safety factors are prioritized, and the actuation force and response speed are appropriately reduced to ensure patient safety.

[0031] Subsequently, the adjusted motion trajectory was analyzed; this trajectory represents the optimal path after safety assessment and optimization. The trajectory was decomposed into a set of key points over time, with each key point containing parameters such as the target joint angle, angular velocity, and expected torque. Based on these key points, a spline interpolation algorithm was used to generate a smooth and continuous motion curve, ensuring smooth and natural joint movement and avoiding discomfort caused by abrupt changes.

[0032] The control mode type is a crucial factor influencing the calculation of driving parameters. In passive assist mode, the system provides dominant guiding force, and the patient simply follows the movement. In active coordination mode, the system provides proportionally enhanced assistance based on the detected active force from the patient. In adaptive hybrid mode, the system dynamically balances active and passive components, adjusting the level of assistance in real time according to the patient's level of participation. Different modes require different configurations for trajectory tracking accuracy, stiffness parameters, and impedance control strategies.

[0033] The intensity of intent directly affects the level of assistance provided by the system. When the patient demonstrates a strong intention to move (quantity value close to 1), the level of assistance is reduced, allowing the patient more autonomy in control; conversely, when the intention is weak (quantity value close to 0), the level of assistance is increased to ensure the completion of the movement. This adjustment of assistance level based on intent intensity effectively avoids the problem of "over-assistance," promotes active patient participation, and improves rehabilitation outcomes.

[0034] Based on a comprehensive analysis of the above factors, the target drive parameters for each joint motor are further calculated. These parameters include target position, velocity curve, maximum permissible torque, and impedance control parameters (such as virtual stiffness and damping coefficient). The system also makes fine adjustments based on the patient's physiological state, for example, reducing the required speed and force in a fatigued state, maintaining standard parameters in a normal state, and appropriately increasing the challenge in an excited state.

[0035] The calculated target driving parameters are converted into real-time motion control commands that the device controller can directly execute. These commands use a standardized data structure, including information such as timestamps, target positions, speeds, accelerations, and torque limits, and are sent to the controllers of each joint motor via a high-speed communication bus. After receiving the commands, the controllers execute them precisely through an internal closed-loop control algorithm, while simultaneously providing feedback on the actual execution status, forming a complete closed-loop control system.

[0036] The entire control command generation process is continuously performed at a frequency of 50-200 times per second, ensuring that the system can respond to changes in patient status and environment within milliseconds. Simultaneously, a short-term memory buffer is maintained to store recent control commands and their execution effects, used to dynamically optimize subsequent command generation strategies, enabling continuous improvement in self-learning and adaptive capabilities.

[0037] The method of this invention, through a closed-loop feedback mechanism, enables continuous monitoring and adaptive adjustment of the patient's condition, significantly improving the personalization and safety of rehabilitation training. Compared with traditional methods, this method can identify weak motor intention signals, adapt to dynamic changes in the patient's abilities, predict and avoid potential safety risks, and provide more precise, safe, and effective rehabilitation training for patients with neurological injuries.

[0038] Example 2

[0039] In this embodiment, the acquisition of multimodal physiological signal data collected by the wearable actuation mechanism, followed by timestamp alignment, resampling, and standardization preprocessing of the multimodal physiological signal data to obtain a synchronized multi-channel sensor signal sequence, includes:

[0040] The wearable actuator collects multimodal physiological signal data, which includes joint kinematic signals, joint dynamic signals, muscle electrophysiological signals, and peripheral circulatory physiological signals. The joint kinematic signals include joint range of motion and angular velocity, the joint dynamic signals include joint torque and motion resistance, the muscle electrophysiological signals include multichannel surface electromyography signals, and the peripheral circulatory physiological signals include optical volumetric pulse signals.

[0041] Using the starting point of the motor motion cycle as the time reference anchor point, the timestamps of each channel signal in the multimodal physiological signal data are aligned, and signals with different sampling rates are interpolated to a unified time grid to obtain time-synchronized signal data.

[0042] The signals of each channel in the time-synchronized signal data are subjected to linear detrending processing and z-score normalization processing to eliminate DC components and amplitude differences, thereby obtaining the synchronized multi-channel sensor signal sequence.

[0043] Specifically, the acquisition of multimodal physiological signals is the foundation of adaptive control. Wearable actuators incorporate multiple miniature sensors capable of simultaneously acquiring four major categories of physiological signals, providing the system with comprehensive patient status information. Specifically, joint kinematic signals are primarily acquired through high-precision angle encoders and miniature gyroscopes, accurately capturing joint range of motion (in degrees, typically -180° to +180°) and angular velocity (in degrees / second, typically -500° / second to +500° / second). These signals directly reflect the limb's position and motion state, serving as the fundamental input to the control system.

[0044] Joint dynamic signals are acquired through torque sensors and pressure sensor arrays, including joint torque (in Newton-meters, typically ranging from -50 Nm to +50 Nm with a resolution of 0.05 Nm) and motion resistance (in Newtons). These signals reflect the interaction forces between the patient's limb and the exoskeleton and are important indicators for identifying the patient's level of active involvement and resistance status.

[0045] Muscle electrophysiological signals refer to multi-channel surface electromyography (EMG) signals, acquired through an electrode array attached to the surface of the patient's muscles, recording the potential changes generated during muscle activation. These signals typically have amplitudes in the microvolt range and high sampling rates (usually 1000Hz to 2000Hz), enabling them to anticipate the patient's movement intentions and serving as a core data source for the control system to recognize these intentions. This system employs an 8- to 16-channel EMG sensor array, covering the major motor muscle groups to capture more comprehensive muscle activation patterns.

[0046] Peripheral circulatory physiological signals mainly refer to optical volumetric pulse signals (PPG signals), which are collected by photoelectric sensors mounted at the point of contact between the device and the patient's skin. These sensors emit light of a specific wavelength and detect changes in light intensity reflected back from the tissue. These changes are directly related to blood flow and volume changes and can be used to assess the patient's heart rate, blood oxygen saturation, and peripheral circulatory status, providing important physiological load feedback for the system.

[0047] Because signals collected by different sensors have different sampling rates and time bases, synchronization is performed using the start point of the motor motion cycle as a time reference anchor point to ensure the time consistency of multimodal data. Here, the "start point of the motor motion cycle" refers to the moment when the exoskeleton device starts executing a complete motion cycle. This moment is clearly timestamped in the system, and all control commands are strictly synchronized with it, making it an ideal time reference point.

[0048] First, the sampling moment closest to the reference anchor point in each signal stream is identified and marked as the synchronization start point. Then, a timestamp alignment operation is performed on all signal streams, adjusting the timestamps of each signal to make their offsets relative to the synchronization start point consistent. For signals with different sampling rates (e.g., electromyography signals are typically 1000Hz, while angle signals may be 100Hz), the system uses a cubic spline interpolation algorithm to resample all signals to a unified time grid (usually selecting the highest sampling rate as the unified standard), ensuring that data points correspond one-to-one in time.

[0049] After time synchronization is completed, the scale differences between different signals need to be addressed. First, linear detrending is performed, subtracting the linear fitting trend line of the signal to eliminate baseline drift and DC components caused by sensor drift or slow changes in patient posture, making the signal more stable. Second, z-score normalization is applied, converting each channel signal into a standard form with a mean of 0 and a standard deviation of 1, eliminating amplitude differences between different signal types, allowing subsequent feature extraction and pattern recognition algorithms to process various signals fairly. The normalized signal sequence retains the waveform characteristics and time-varying properties of the original signal, while exhibiting better comparability and stability, laying the foundation for subsequent feature extraction and pattern recognition.

[0050] Example 3

[0051] In this embodiment, the step of extracting the feature dataset based on the synchronized multi-channel sensor signal sequence includes:

[0052] The synchronized multi-channel sensor signal sequence is segmented using a sliding window technique, with the window length set to 0.5 to 3 seconds and the overlap rate to 50% to 75%, to obtain time window signal segments.

[0053] Based on the time window signal segment, the mean, variance, peak value, signal envelope value and inter-channel Pearson correlation coefficient of each channel are calculated to obtain the time domain feature vector;

[0054] Based on the time window signal segment, perform spectrum analysis on the signal of each channel, calculate the energy ratio, peak frequency position and spectral centroid within the predefined functional frequency band, and obtain the frequency domain feature vector;

[0055] The angles of each degree of freedom of the joint in the joint kinematic signal, the torque value in the joint dynamic signal, and the electromyographic envelope value in the muscle electrophysiological signal are combined to construct a six-dimensional to ten-dimensional state space vector;

[0056] The time-domain feature vector, the frequency-domain feature vector, and the state-space vector are concatenated column-wise to obtain the feature dataset.

[0057] Specifically, extracting effective features from the pre-processed multi-channel sensor signal sequence is a crucial step in motion intent recognition. Due to the non-stationary nature of physiological signals, a sliding window segmentation technique is first employed to divide the continuous signal into a series of short time segments for analysis. Sliding window segmentation is a fundamental technique in time-series signal processing; by sliding a fixed-length analysis window across the signal, a long sequence is decomposed into multiple short sequences for processing. In this system, the window length is set considering the temporal characteristics of human motion intent expression and physiological signal changes. A window that is too short will lead to unstable extracted features, making it difficult to capture complete motion patterns; a window that is too long will reduce the system response speed and increase control latency. Experiments have verified that a window length of 0.5 to 3 seconds achieves a good balance between temporal resolution and feature stability. Lower limb movements typically use a longer window (2-3 seconds), while fine upper limb movements use a shorter window (0.5-1 second).

[0058] The overlap ratio of a sliding window refers to the proportion of signal shared between two adjacent windows. A high overlap ratio provides smoother feature changes, reduces boundary effects, and improves the system's sensitivity to changes in motion intent, but it also increases computational load. Choosing an overlap ratio of 50% to 75% strikes a balance between response sensitivity and computational efficiency. For example, with a 2-second window length and a 75% overlap ratio, feature results are output every 0.5 seconds, ensuring the control system can respond promptly to changes in patient intent.

[0059] For each time window signal segment, features are extracted from both the time and frequency domains. Time-domain features are calculated directly from the original signal waveform, including the mean (reflecting the signal baseline level), variance (reflecting the signal fluctuation intensity), peak value (reflecting the maximum activation level of the signal), signal envelope value (reflecting the overall activation profile of the signal), and Pearson correlation coefficients between channels (reflecting the synergistic relationship between different muscles or joints). The signal envelope value, which refers to the outer contour curve of the signal amplitude variation, is obtained through Hilbert transform and low-pass filtering. It can smoothly characterize the overall activation level of the signal, making it particularly suitable for characterizing the activity intensity of electromyographic signals.

[0060] Frequency domain features, extracted by performing a spectral transform on the signal, can reveal the periodic components and frequency characteristics contained within the signal. First, a Fast Fourier Transform or Wavelet Transform is applied to each channel signal to obtain its spectral representation. Then, the energy percentage, peak frequency position, and spectral centroid within a predefined functional frequency band are calculated. The predefined functional frequency band refers to a frequency range divided according to the characteristics of physiological signals. For example, for electromyography (EMG) signals, it is typically divided into a low-frequency band (0-20Hz, mainly reflecting the firing rate of action potentials in motor units), a mid-frequency band (20-100Hz, containing most of the EMG signal energy), and a high-frequency band (100-400Hz, reflecting the conduction velocity of muscle fibers). The peak frequency position refers to the frequency point with the highest energy in the spectrum, and the spectral centroid is the "center of gravity" of the spectral energy distribution. These two indicators together reflect the frequency characteristics of the signal and are of great value in identifying different types of motor intentions.

[0061] In addition to the statistical features mentioned above, a state space vector is constructed, which is a multidimensional vector that comprehensively represents the dynamic state of the system. The state space vector integrates the angles of each degree of freedom of the joints from the joint kinematic signals (such as the current angles of the hip, knee, and ankle joints), the torque values ​​from the joint dynamic signals (reflecting the force state of the joint), and the electromyographic envelope values ​​from the muscle electrophysiological signals (reflecting the muscle activation level). Depending on the number of controlled joint degrees of freedom and the number of muscles monitored, the state space vector typically has six to ten dimensions, comprehensively capturing the current movement state and muscle activation patterns of the patient's limbs, providing a rich information foundation for intent recognition.

[0062] Finally, the temporal feature vectors (dimension approximately 5 times the number of channels), frequency domain feature vectors (dimension approximately 3 times the number of channels), and state space vectors (dimension 6-10) are concatenated column-wise to form a high-dimensional feature dataset. This multi-source feature fusion strategy fully leverages the complementary advantages of different types of features, comprehensively capturing the patient's motor intention information and improving the accuracy and robustness of recognition. The final dimension of the feature dataset is typically tens to hundreds of dimensions, varying depending on the specific sensor configuration and feature selection strategy.

[0063] Example 4

[0064] In this embodiment, the step of calculating the similarity between the feature dataset and the pre-built motion intent template library to determine the current motion intent category and intent intensity quantification value includes:

[0065] Standard motion intention templates are extracted from the pre-built motion intention template library. These standard motion intention templates are formed by preprocessing and feature extraction of standard active motion signals collected clinically.

[0066] The current time window feature vector sequence in the feature dataset is compared one by one with each of the standard motion intention templates in the pre-built motion intention template library. The Euclidean distance, cosine similarity or dynamic time warping distance between the two are calculated to obtain the similarity score for each template.

[0067] Select the template corresponding to the maximum value from all the similarity scores to obtain the best matching template identifier and the highest similarity score;

[0068] Based on the motion intent type corresponding to the best matching template identifier, the current motion intent category is obtained;

[0069] The highest similarity score is normalized and mapped to the interval between 0 and 1 to obtain the quantized value of the intent intensity.

[0070] Specifically, motion intention recognition is the core component of adaptive control. This embodiment employs a template matching-based method to identify the patient's motion intention. First, standard motion intention templates are extracted from a pre-built motion intention template library. These standard templates are formed by clinically collecting active motion signals from healthy subjects and patients at different rehabilitation stages, followed by the same preprocessing and feature extraction procedures as described above. The template library typically includes various common rehabilitation motion types, such as standing, sitting, climbing stairs, descending stairs, and walking on flat ground in lower limb rehabilitation; and grasping, extending, rotating, and lifting in upper limb rehabilitation. Each motion type is further divided into multiple variant templates based on different execution speeds, intensities, and ranges to accommodate individual differences among patients.

[0071] A standard motion intent template is a time series of multidimensional feature vectors that represents the typical change patterns of physiological signal features during a specific motion. Each template contains complete feature changes before, during, and after the motion begins, comprehensively capturing the dynamic expression of motion intent. The template library is constructed using hierarchical clustering and representative sample selection techniques to ensure the distinguishability between templates and the representativeness within each template class. Typically, a complete template library contains 10-20 basic motion types, with 5-10 variant templates for each type, totaling 50-200 standard templates.

[0072] To identify the current movement intention, the sequence of feature vectors extracted from the current signal window is compared one by one with all standard templates in the template library, and the similarity between them is calculated. Considering the time-varying characteristics of physiological signals and individual differences, multiple similarity calculation methods are supported: Euclidean distance is suitable for direct distance measurement in feature space and is simple and efficient to calculate; cosine similarity focuses on the similarity of vector directions and is not sensitive to amplitude scaling; while dynamic time warping distance (DTW) can handle nonlinear scaling on the time axis and adapt to the speed differences of different patients performing the same action.

[0073] Dynamic time warping distance (DTDM) is a special sequence similarity metric that finds the optimal non-linear alignment between two time series, even if the two series have different lengths or temporal distortions (some parts are faster than others). Its core idea is to construct a cumulative distance matrix and use a dynamic programming algorithm to find an alignment path that minimizes the total distance. This method is particularly suitable for rehabilitation scenarios because patients' performance speed is often unstable and their movements are stretched or compressed in time compared to a standard template.

[0074] A similarity score is calculated for each standard template. Then, the template with the highest score (or the lowest score for distance metrics, which will be appropriately converted later) is selected as the best match. The identifier of the best matching template directly points to a specific motion intention type, such as "sit down" or "stand up," which the system uses to determine the current motion intention category.

[0075] In addition to determining the category of intent, it is also necessary to quantify the intensity of the patient's motor intent, which is crucial for adjusting the level of assistance. The highest similarity score is normalized and mapped to a standard range of 0 to 1. This process first determines the effective range of scores (minimum acceptable threshold and ideal maximum value) based on historical matching statistics, and then converts the raw scores into quantified values ​​of intent intensity through linear or non-linear mapping (such as sigmoid function mapping). Values ​​close to 1 indicate that the patient is exhibiting a very clear motor intent and should be provided with minimal assistance; values ​​close to 0 indicate that the intent is unclear or very weak and requires more assistance.

[0076] This template-matching-based intent recognition method is intuitive and highly interpretable, while also adapting to the diversity and instability of movement patterns in rehabilitation patients through techniques such as dynamic time warping. It also supports incremental learning, allowing new typical samples to be continuously added during use to optimize the template library and improve adaptability to specific patients.

[0077] Example 5

[0078] In this embodiment, determining the control mode type based on the current motion intention category and the quantized value of the intention intensity includes:

[0079] The highest similarity score is compared with a preset effective intent threshold. When the highest similarity score is higher than the preset effective intent threshold, it is determined that a valid active motion intent has been detected; otherwise, it is determined that no valid active motion intent has been detected.

[0080] When it is determined that no valid active motion intention is detected, the control mode type is set to passive assist mode;

[0081] When a valid active motion intention is detected and the intensity quantization value of the intention is lower than the first intensity threshold, the control mode type is set to active coordination mode.

[0082] When a valid active motion intention is detected and the quantized value of the intention intensity is higher than the second intensity threshold, the control mode type is set to adaptive hybrid mode.

[0083] Specifically, the selection of the control mode directly affects the method and extent of assistance provided by the exoskeleton device, and is a key aspect of personalized rehabilitation training. This embodiment describes an automatic control mode selection mechanism based on movement intention recognition results, which can provide the most suitable assistance method according to the patient's real-time status. The entire process first determines whether a valid active movement intention has been detected by comparing the highest similarity score with a preset effective intention threshold.

[0084] A preset effective intention threshold is a boundary value used to determine whether a patient has expressed a clear motor intention. It is typically determined through extensive clinical trials and expert evaluation. This threshold setting must avoid misinterpreting noise and unconscious activity as effective intention (too low a threshold) while also avoiding ignoring weak but genuine motor intentions (too high a threshold). In practice, the preset effective intention threshold is usually set between 60% and 75% of the similarity score, with the specific value adjusted individually based on the patient's rehabilitation stage and functional status. For patients in the early stages of rehabilitation, the threshold is appropriately lowered to capture weaker expressions of intention; while for patients in the later stages of rehabilitation, the threshold is raised to encourage more explicit active control.

[0085] When the highest similarity score is lower than the preset effective intent threshold, it is determined that no effective active movement intent has been detected. This may be because the patient currently has no intention to move, or the intent expression is too weak to be reliably identified. In this case, the control mode type is automatically set to passive assist mode. Passive assist mode is a device-led control mode where the exoskeleton device guides the patient to complete movements according to a preset movement trajectory and speed. The patient only needs to relax and follow the device's movements. This mode is suitable for training when the patient's active movement ability is severely limited or in a state of fatigue. It can ensure basic exercise volume and joint range of motion, prevent muscle atrophy and joint stiffness, and promote neural remodeling through proprioceptive input. In passive assist mode, the provided assist torque typically covers 80% to 100% of the torque required for the movement, ensuring that the movement can be completed smoothly.

[0086] Once a valid active movement intention is detected, an appropriate active control mode needs to be selected based on the quantified value of the intention intensity. The quantified value of intention intensity reflects the clarity and strength of the patient's movement intention and is a crucial basis for determining the level of assistance. When the quantified value of intention intensity is below the first intensity threshold (usually set between 0.3 and 0.5), it indicates that although the patient has expressed a clear movement intention, the intensity is weak, and more assistance may be required to complete the movement. In this case, the control mode type is set to active coordination mode. Active coordination mode is a patient-led, device-assisted control mode that provides proportional force assistance based on the detected patient movement intention, realizing the rehabilitation concept of "assistance" rather than "replacement." In this mode, 30% to 70% of the torque required for the movement is typically provided, with the remainder requiring active generation by the patient. The magnitude of the assisting torque is dynamically adjusted according to the quantified value of intention intensity; the stronger the intention, the less assistance is provided, encouraging the patient to maximize their initiative.

[0087] When the quantified value of intent intensity exceeds the second intensity threshold (typically set between 0.7 and 0.8), it indicates that the patient has expressed a very clear and strong motor intent and is capable of taking on greater active control responsibility. At this point, the system sets the control mode type to adaptive hybrid mode. Adaptive hybrid mode is an advanced control mode that combines the advantages of patient active control and intelligent device assistance. In this mode, the system continuously monitors the patient's performance and dynamically adjusts the control strategy. When the patient performs well, assistance is gradually reduced, and appropriate resistance may even be increased to provide a challenge; when the patient shows fatigue or difficulty, the assistance intensity is promptly increased to avoid frustration and over-fatigue. Adaptive hybrid mode typically provides only 0% to 30% of the torque required for the movement as basic assistance, but it is rapidly adjusted based on real-time performance assessments to maximize the recovery and remodeling of the patient's neuromuscular system.

[0088] It is worth noting that there is a transition range (0.3-0.5 to 0.7-0.8) between the first and second intensity thresholds. The intent intensity quantization value within this range triggers an intermediate state control mode, which determines which mode to use based on historical training data and the current state, or applies a weighted mixture of the two modes. This continuous transition design avoids abrupt mode switching and provides a smoother user experience.

[0089] The automatic selection of control modes is not only based on the results of a single intent recognition, but also incorporates short-term memory mechanisms to analyze the trend of intent pattern changes over the past 30 seconds to 2 minutes, avoiding training discontinuity caused by frequent switching. It also considers the cumulative training load throughout the day and the patient's fatigue level, appropriately reducing the control difficulty in the later stages of training to ensure safe and effective training.

[0090] Example 6

[0091] In this embodiment, the similarity calculation adopts a flexible alignment matching method based on time series similarity measurement, specifically including:

[0092] The temporal feature vector sequence in the feature dataset is non-linearly aligned with each standard template sequence in the pre-built motion intent template library. A cumulative distance matrix is ​​constructed using a dynamic programming algorithm, and the optimal alignment path is calculated to obtain the minimum distance metric value that allows time warping.

[0093] Based on the minimum distance metric, an adaptive threshold determination mechanism is used to determine whether there is a valid pattern match. The adaptive threshold is dynamically adjusted according to the statistical characteristics of the distance distribution within the pre-built motion intent template library to obtain the matching confidence and pattern category identifier.

[0094] Based on the matching confidence and the pattern category identifier, when the matching confidence exceeds the confirmation threshold, the signal pattern of the corresponding matching event is included as a new sample in the pre-built motion intent template library. The representative template is updated through cluster analysis to achieve incremental learning and evolution optimization of the pre-built motion intent template library.

[0095] Specifically, the elastic alignment matching method used in similarity calculation is an advanced pattern matching technique designed specifically for processing time series data. It can effectively address challenges such as inconsistent movement speeds and different completion times of movements among patients during rehabilitation training. The core of elastic alignment matching is to allow non-linear deformation along the time axis to find the optimal correspondence between two sequences, thereby more accurately assessing their similarity.

[0096] In practice, the feature vector sequence of the current time window is first non-linearly aligned with each standard template sequence in the pre-built motion intent template library. "Non-linear alignment" here means allowing the two sequences to stretch and deform locally in the time dimension to find the best match. For example, to complete the "stand up" action, a healthy person might only need 2 seconds, while a patient might need 5 seconds or more, and the speed at different stages might be uneven (slow at the beginning, fast in the middle). Traditional fixed-time-window comparisons struggle to handle this situation, while flexible alignment can automatically adapt to such temporal deformations.

[0097] The cumulative distance matrix is ​​constructed using a dynamic programming algorithm, which is the core step in achieving flexible time alignment. The cumulative distance matrix is ​​a two-dimensional matrix where each element D(i, j) represents the minimum cumulative distance required to optimally align the first i points of the current sequence with the first j points of the template sequence. This matrix is ​​filled in a bottom-up manner, with the value of each cell depending on the minimum of three possible preceding cells (left, bottom, or bottom-left diagonal) plus the distance of the current point pair. This process can be formally described as a dynamic programming recursion, and the final element D(m, n) in the upper right corner of the matrix represents the minimum distance metric allowed for time warping between the two complete sequences.

[0098] In addition to the final distance value, the optimal alignment path is obtained by backtracking the cumulative distance matrix. This path starts from the top right corner of the matrix, and at each step, it selects the smallest value among three adjacent cells on the left, bottom, or bottom left diagonal, until it reaches the bottom left corner. The shape of the path visually reflects the temporal correspondence between the two sequences: the diagonal portion represents velocity matching, the horizontal segment represents the current sequence stalling (the template continues to advance), and the vertical segment represents the template stalling (the current sequence continues to advance). Analyzing this path can help understand the specific differences between the patient and the standard pattern, providing a basis for rehabilitation assessment and training.

[0099] To avoid unreasonable matching due to excessive time distortion, a series of constraints are imposed, such as global path slope constraints (limiting the overall velocity ratio within a reasonable range), local continuity constraints (limiting the variation in the correspondence between adjacent points), and Sakoe-Chiba bandwidth constraints (limiting the maximum distance the alignment path deviates from the diagonal). These constraints ensure that the matching results conform to the physical laws of human motion, improving the accuracy and reliability of recognition.

[0100] After obtaining the minimum distance metric, it is necessary to determine whether this distance represents a valid pattern match. An adaptive thresholding mechanism is used here, rather than a simple fixed threshold. The adaptive threshold is a dynamically adjusted decision boundary that automatically adjusts based on the statistical characteristics of the distance distribution within the pre-built motion intent template library. Specifically, the distance matrix between various types of templates within the template library is calculated periodically, the average distance within each class (intra-class distance) and the average distance between different classes (inter-class distance) are analyzed, and the decision threshold is set based on these statistics.

[0101] Intra-class distance reflects the natural variability of the same motor intention, while inter-class distance reflects the discriminative power between different motor intentions. Ideally, intra-class distance should be significantly smaller than inter-class distance, indicating clear classification boundaries. A baseline threshold is set based on the distribution of intra-class distances (e.g., mean plus twice the standard deviation), and then fine-tuned by incorporating the patient's historical performance and current rehabilitation stage. This adaptive thresholding mechanism can adapt to individual differences and rehabilitation progress among different patients, improving the flexibility and adaptability of identification.

[0102] An adaptive threshold is used to determine the match confidence score and pattern category identifier. The match confidence score is a value between 0 and 1, representing the reliability of the current match result. It considers not only the distance value of the best match but also the distance difference (discrimination) with the second-best match. The pattern category identifier points to a specific motion intent category in a pre-built template library, such as "standing" or "sitting". These two outputs together constitute the core result of motion intent recognition.

[0103] One key innovation is the implementation of incremental learning and evolutionary optimization of the pre-built motor intention template library. When the confidence level of a match exceeds the confirmation threshold (usually set relatively high, such as 0.85 or 0.9), it is considered a high-quality match, and the corresponding signal pattern is likely to represent the patient's typical presentation. At this point, this signal pattern is added as a new sample to the pre-built motor intention template library, enriching the sample diversity of the template library.

[0104] The addition of new samples triggers the template library update process. Cluster analysis methods (such as K-means clustering or hierarchical clustering) are used to group all samples within the same category, identifying a representative subset of samples as the updated templates. This process can be seen as "compressing" and "refining" the template library, avoiding the computational burden caused by the unlimited growth of the number of templates, while retaining key pattern variation information.

[0105] Through this incremental learning mechanism, the template library can gradually adapt to the movement characteristics of specific patients, capture individual-specific patterns, and improve recognition accuracy. Simultaneously, it periodically analyzes the distribution characteristics of newly added samples, adjusts intra- and inter-class thresholds, and optimizes the overall classification boundary, achieving evolutionary optimization of the template library. This adaptive learning capability allows the system to continuously improve over long-term use, providing patients with increasingly precise rehabilitation assistance.

[0106] It is worth noting that, to ensure the quality of incremental learning, new samples undergo rigorous quality control. In addition to confidence requirements, signal quality (signal-to-noise ratio), motion integrity, and compatibility with existing templates are also checked. Only samples that simultaneously meet all quality criteria are included in the template library, ensuring that the incremental learning process is not degraded by noisy samples.

[0107] Example 7

[0108] In this embodiment, the extraction of the frequency domain feature vector includes a fast spectrum analysis method based on the sparsity assumption, specifically including:

[0109] The synchronized multi-channel sensor signal sequence is segmented and windowed, with a window length of 256 to 2048 sampling points and an overlap rate of 50%, to obtain a multi-channel time window signal matrix.

[0110] Based on the multi-channel time window signal matrix, the sparse spectrum recovery algorithm is used to randomly hash and bucket the signals of each channel, calculate the energy amplitude of each frequency component, compare the energy amplitude with an energy threshold, and when the energy amplitude is greater than the energy threshold, the corresponding frequency component is determined to be a significant energy frequency component. Hash collision is used to locate all the significant energy frequency components to obtain a sparse spectrum coefficient set, wherein the energy threshold is set to 10% to 50% of the mean of the energy amplitude of all frequency components.

[0111] The sparse spectral coefficient set is divided into predefined functional frequency bands, including a low frequency band of 0-5Hz, a mid frequency band of 5-50Hz, and a high frequency band of 50-250Hz. The energy proportion, peak frequency position, and cross-channel coherence index of each frequency band are calculated to obtain the frequency domain feature vector.

[0112] Specifically, the entire process of frequency domain feature vector extraction is a crucial step in motion intent recognition systems. This method leverages the sparsity of bioelectrical signals in the frequency domain, employing efficient algorithms to rapidly extract key frequency components, significantly reducing computational complexity and enabling real-time processing while preserving the signal's key features.

[0113] First, the synchronized multi-channel sensor signal sequence is segmented and windowed. Windowing is a fundamental step in time-frequency analysis, dividing a long sequence into a series of short time segments to capture the time-varying characteristics of the signal. In this embodiment, a window length range of 256 to 2048 sampling points is selected. This range is set based on two considerations: ensuring frequency resolution (the longer the window, the higher the frequency resolution) and ensuring temporal resolution (the shorter the window, the more accurate the time positioning). The specific window length is dynamically adjusted according to the signal sampling rate and the target frequency range. For example, when the sampling rate is 1000Hz, a 512-point window can provide a frequency resolution of approximately 2Hz, suitable for analyzing most bioelectrical signals. The window overlap rate is set to 50%, meaning that adjacent windows share half of the data points. This overlap setting can reduce information loss caused by window edge effects and provide smoother time-frequency analysis results. Through windowing, the original continuous signal is converted into a signal matrix with multiple time windows, where each row represents a channel and each column represents a time point.

[0114] Next, a sparse spectrum recovery algorithm is used to process the signals from each channel. The sparse spectrum recovery algorithm is a signal processing method based on compressed sensing theory. Its core assumption is that most bioelectrical signals are sparse in the frequency domain, meaning that the signal energy is mainly concentrated on a few frequency components. Based on this assumption, instead of performing detailed calculations on all frequency points, a random hashing bucketing technique is used to quickly locate the frequency points with significant energy, greatly reducing computational complexity.

[0115] Random hashing bucketing is a key technique in sparse spectrum recovery algorithms. It divides the frequency domain into multiple "buckets" by designing specific hash functions. Applying these hash functions to the original signal is equivalent to mapping different frequency components to different buckets and accumulating the energy of the corresponding component in each bucket. Due to the sparsity of the frequency domain, most buckets have very little energy (containing only noise), while a few buckets have significant energy (containing effective signal components). The energy amplitude of each bucket is calculated and compared with a preset energy threshold. The energy threshold is typically set to 10% to 50% of the average energy amplitude of all frequency components. This range is based on extensive experimental data analysis and effectively distinguishes between signal and noise. When the energy amplitude of a bucket exceeds the threshold, it is considered that the bucket contains significant energy frequency components and requires further analysis.

[0116] To accurately locate significant energy frequency components, the hash collision principle is utilized. A hash collision refers to the phenomenon where different frequency components are mapped to the same bucket. By designing multiple different hash functions, the same frequency component will be mapped to different buckets, while the probability of different frequency components being mapped to multiple identical buckets simultaneously is extremely low. The energy distribution patterns of the buckets under different hash functions are analyzed, and the position and amplitude of significant energy frequency components are accurately recovered by solving linear equations or iterative optimization algorithms, ultimately obtaining a sparse spectral coefficient set. Compared to the traditional Fast Fourier Transform (FFT), which requires calculating all frequency points, this method reduces the computational complexity from O(N log N) to O(K log N), where K is the number of significant frequency components, which is usually much smaller than the total number of frequency points N, thus significantly improving computational efficiency.

[0117] After obtaining the sparse spectral coefficient set, it is divided into predefined functional frequency bands. The frequency band division is based on the differences in physiological information carried by different frequency ranges. This embodiment defines three main functional frequency bands: the low-frequency band (0-5Hz) mainly contains the basic rhythm of limb movement and intention preparation signals; the mid-frequency band (5-50Hz) contains signals related to active muscle contraction and coordination control; and the high-frequency band (50-250Hz) contains high-frequency components related to rapid muscle fiber activation and fine control. This division method enables the system to specifically analyze the signal characteristics generated by different physiological processes.

[0118] For each functional frequency band, three key indicators are calculated: band energy percentage, peak frequency location, and cross-channel coherence index. Band energy percentage is the ratio of the energy in that band to the total energy, reflecting the relative importance of different frequency components; peak frequency location is the frequency point with the highest energy within the band, reflecting the dominant rhythm; and cross-channel coherence index quantifies the degree of synchronization between signals from different channels in a specific frequency band, reflecting multi-muscle coordinated control characteristics. These indicators together constitute a frequency domain feature vector, providing rich frequency domain information for subsequent motion intent recognition.

[0119] It is worth noting that the parameter settings for frequency domain feature extraction are fine-tuned based on the usage scenario and individual differences. For example, for scenarios focusing on fine movements, the weight of high-frequency bands is increased; for signals with low signal-to-noise ratios, the energy threshold is appropriately increased to filter out more noise. The effectiveness of feature extraction is also evaluated regularly, and the frequency band division and index calculation methods are adjusted through feature importance analysis to ensure that the extracted features can distinguish different motion intentions to the greatest extent.

[0120] Example 8

[0121] In this embodiment, the step of classifying physiological states to obtain the current physiological state category and state confidence includes:

[0122] The frequency domain feature vector is extracted from the feature dataset to construct a Gaussian mixture probability model. The pre-stored parameters of the Gaussian mixture probability model are obtained. The parameters of the Gaussian mixture probability model include the mean vector, covariance matrix and mixture weights of each basic state. The basic states include fatigue state, normal state and excitement state.

[0123] Substitute the frequency domain feature vector into the Gaussian mixture probability model, calculate the Gaussian probability density value of the frequency domain feature vector in each basic state, and calculate the posterior probability of each basic state using Bayes' theorem to obtain the state posterior probability distribution.

[0124] Based on the state posterior probability distribution, the state with the highest posterior probability is selected as the current physiological state category by the maximum posterior criterion, and the posterior probability value corresponding to the state with the highest posterior probability is used as the state confidence.

[0125] Based on the current physiological state category, a preset state-parameter mapping table is queried to obtain the motion speed adjustment coefficient, torque limit parameter, and recommended motion amplitude value for the current physiological state category, and personalized control parameter recommendation value is generated.

[0126] Specifically, the physiological state classification method based on frequency domain feature vectors analyzes the user's physiological state through a probabilistic model and dynamically adjusts system control parameters based on the state assessment results, achieving intelligent adaptation in human-computer interaction. This state perception mechanism enables exoskeleton devices to sense the user's fatigue level and mental state, providing more humane and safer auxiliary control.

[0127] First, frequency domain feature vectors are extracted from the feature dataset. The frequency domain feature vectors mentioned in the previous embodiment include the energy proportions, peak frequencies, and cross-channel coherence indices of the low, mid, and high frequency bands. These indices contain rich information about physiological states. For example, fatigue is typically characterized by a decrease in high-frequency components, an increase in low-frequency components, and reduced inter-channel coherence; while excitement is characterized by an increase in mid-to-high frequency energy and enhanced inter-channel synergy. Using these frequency domain feature variation patterns, an automatic classification of physiological states is performed through a Gaussian mixture probability model.

[0128] Gaussian mixture models are powerful probability density estimation tools that can represent complex multimodal data distributions through a weighted combination of multiple Gaussian distributions. In this system, each basic physiological state (fatigue, normal, and excited) is represented by a multidimensional Gaussian distribution, and the entire model is a mixture of these three basic state distributions. The core parameters of the model include: the mean vector for each state (representing the average value of the feature vectors in that state), the covariance matrix (representing the degree of variation and correlation of each dimension of the features), and the mixture weights (representing the prior probability of each state).

[0129] These model parameters are obtained through offline training, which uses a large amount of labeled state feature data and iteratively optimizes the parameter values ​​using the Expectation-Maximization (EM) algorithm. The trained Gaussian mixture model parameters are pre-stored in the system for real-time state classification. For different users, the system has an initial adaptation period during which user data is collected and model parameters are fine-tuned to achieve personalized state recognition.

[0130] During real-time classification, the extracted frequency domain feature vector is substituted into a Gaussian mixture model to calculate the Gaussian probability density value of the feature vector under each base state. The Gaussian probability density value reflects the degree of matching between the currently observed feature and the typical features of each state; a higher value indicates a greater likelihood of belonging to that state. However, the probability density value itself is not a probability and needs to be converted into a posterior probability using Bayes' theorem.

[0131] Bayes' theorem is a fundamental formula in probability theory used to update hypothetical probabilities. It combines prior knowledge and new evidence to calculate posterior probabilities. In this system, the prior probability is given by mixed weights, and the new evidence is the currently observed frequency domain feature vector. Bayes' theorem combines the two to calculate the conditional probability (posterior probability) of the user being in each basic state given the current feature vector. These posterior probability values ​​constitute the state posterior probability distribution, comprehensively describing the uncertainty of the current physiological state.

[0132] Based on the posterior probability distribution of the states, the maximum a posteriori (MAP) criterion is used to select the most probable state. The MAP criterion is a decision rule that selects the hypothesis with the highest posterior probability as the optimal decision; in this system, it selects the state with the highest posterior probability as the current physiological state category. Simultaneously, this highest posterior probability value is used as the state confidence score, reflecting the reliability of the classification result. A higher confidence score indicates a more certain classification; a lower confidence score indicates that the current state may be in a transitional region between multiple basic states, leading to more cautious adjustments to the control parameters.

[0133] After the physiological state is classified, the system queries a pre-defined state-parameter mapping table based on the current physiological state category to obtain the corresponding control parameter adjustment values. The state-parameter mapping table is a predefined lookup table in the system that maps different physiological states to corresponding control parameter adjustment strategies. The main control parameters to be adjusted include: motion speed adjustment coefficient, torque limit parameter, and recommended motion amplitude value.

[0134] The movement speed adjustment coefficient directly affects the execution speed of exoskeleton-assisted movements and will be adjusted according to the user's state: the speed is reduced when the user is fatigued (coefficient less than 1) to give the user more reaction time; the speed is appropriately increased when the user is excited (coefficient greater than 1) to meet the user's need for a faster pace; and the standard speed is maintained when the user is in a normal state (coefficient approximately equal to 1).

[0135] The torque limiting parameter controls the maximum assist force or resistance that the exoskeleton can apply, and is a core parameter for safety control. In a fatigued state, the resistance setting will be reduced and the upper limit of assist force will be increased to reduce the burden on the user; in an excited state, the resistance may be appropriately increased to provide a greater challenge, but the overall torque range will be strictly controlled to prevent safety risks caused by excessive force.

[0136] The recommended range of motion guides the system in adjusting the range of motion. In a fatigued state, it is recommended to reduce the range of motion to avoid overextension of the joints; in a normal or excited state, the range of motion may be appropriately increased according to the training goals, but always within safe boundaries.

[0137] These parameter adjustments are not simple discrete state transitions, but rather a continuous adjustment process that considers the confidence level of the state. For example, when the system determines with 80% confidence that the user is in a state of mild fatigue and 20% probability of being in a normal state, the final parameter adjustment will be a weighted average of the fatigue state parameter and the normal state parameter, with weights of 0.8 and 0.2, respectively. This smooth transition mechanism avoids abrupt changes in parameters and provides a more natural human-computer interaction experience.

[0138] Finally, personalized control parameter recommendations are generated based on state assessment and parameter mapping. These recommendations are then passed to the control strategy generation module, influencing the real-time control behavior of the exoskeleton device. The entire process achieves a closed-loop feedback from physiological signal analysis to state perception, and then to adaptive adjustment of control parameters, enabling the exoskeleton device to intelligently adapt to the user's real-time state changes and provide a safe, comfortable, and efficient assistive experience.

[0139] It's important to note that the state-parameter mapping table is not fixed; it records user feedback and usage history, and gradually optimizes the mapping relationship through reinforcement learning or rule updates. For example, if a specific user is repeatedly observed to perform well even in the system's perceived "mild fatigue" state, the user's parameter mapping will be gradually adjusted to make parameter adjustments in the "mild fatigue" state more gentle. This personalized learning mechanism further enhances the system's adaptability, enabling it to adapt to the different state performance characteristics and preferences of various users.

[0140] Example 9

[0141] In this embodiment, the step of determining the safety of the preset movement trajectory based on the current physiological state category, the state confidence level, and the preset movement trajectory using a boundary detection algorithm, calculating a safety margin assessment value, and adjusting the preset movement trajectory when the safety margin assessment value is lower than a safety threshold to obtain an adjusted movement trajectory includes:

[0142] Step S31: Construct a multi-dimensional joint space. The dimensions of the multi-dimensional joint space include the angle, angular velocity and torque of each degree of freedom of the joint. In the multi-dimensional joint space, the boundaries of the dynamic safe region and the taboo region are defined according to the current physiological state category and the state confidence. The multi-dimensional joint space is partitioned into an octree space, and a hierarchical index structure is established to obtain a spatial index database.

[0143] Step S32: Obtain the coordinates of the current joint state point and the direction vector of the preset motion trajectory. Use a point positioning algorithm to query the spatial index database to obtain the spatial region to which the coordinates of the current joint state point belong, and obtain the region type identifier. The region type identifier includes a safe region identifier, a boundary region identifier, and a forbidden region identifier.

[0144] Step S33: Based on the region type identifier and the direction vector of the preset motion trajectory, emit a detection ray along the direction vector of the preset motion trajectory. Calculate the shortest distance from the current position to the forbidden region boundary by intersecting the detection ray with the forbidden region boundary. Use the shortest distance as the safety margin assessment value.

[0145] Step S34: When the safety margin assessment value is lower than the safety threshold, the preset motion trajectory is corrected by adjusting the trajectory parameters through the gradient descent method so that the corrected trajectory maintains a safe distance from the taboo boundary, thus obtaining the adjusted motion trajectory.

[0146] Specifically, a trajectory safety detection and adjustment method based on multi-dimensional joint space analyzes the user's physiological state and joint dynamic characteristics to assess the safety of preset motion trajectories in real time and makes intelligent adjustments when necessary, ensuring safety and comfort during human-computer interaction. This proactive safety mechanism is an indispensable key technology for wearable robot systems, effectively preventing usage risks caused by improper device control.

[0147] First, a multi-dimensional joint space is constructed, which forms the foundational environment for safety assessment. Unlike a simple three-dimensional physical space, the multi-dimensional joint space is a high-dimensional abstract space. Its dimensions are determined by the degrees of freedom of the joints, with each degree of freedom corresponding to a dimension in the space, typically including three physical quantities: angle, angular velocity, and torque. For example, for a single-joint system with 3 degrees of freedom, its multi-dimensional joint space is a 9-dimensional space (3 angle dimensions, 3 angular velocity dimensions, and 3 torque dimensions); for multi-joint systems, the dimensions increase further. In this high-dimensional space, any point represents a complete state of the joint system, while a trajectory represents the dynamic process of state change.

[0148] After constructing the multidimensional joint space, the boundaries of safe and contraindicated regions are dynamically defined based on the current physiological state category and state confidence level. Safe regions refer to the set of states where joint movement will not cause any discomfort or potential harm; contraindicated regions refer to the set of states that may cause discomfort, pain, or safety hazards to the user; boundary regions are the transitional areas between the two and require careful control. The boundaries of these regions are not fixed but dynamically adjusted according to the physiological state. For example, in a state of fatigue, the safe region will shrink accordingly, and more state points will be classified as contraindicated regions; state confidence level affects the "rigidity" of the boundaries—high confidence level produces clear boundaries, while low confidence level makes the boundaries more blurred, forming wider transitional areas.

[0149] Since multidimensional joint spaces typically have high dimensionality, performing security assessments directly across the entire space is inefficient. Therefore, an octree space partitioning technique is used to hierarchically index the space. Octree space partitioning is a data structure that recursively divides a space into eight subspaces. Although named "octree," the concept can be generalized to spaces of any dimension; in high-dimensional spaces, each node splits into two subspaces. d The system has _ _ child nodes, where _ _ is the spatial dimension. This partitioning method divides the space into blocks of different sizes. Larger blocks are used for areas with clear boundaries between safe and forbidden regions, while smaller blocks are used for areas with complex boundaries, creating an adaptive resolution. Each block is marked as a safe region, a boundary region, or a forbidden region, and its associated safety parameters are stored. This hierarchical index structure significantly improves spatial query efficiency, enabling the system to respond to state changes in real time.

[0150] The specific implementation process of octree space partitioning is as follows: starting from the root node, representing the entire space, evaluate whether the node completely belongs to a single region type (safe, boundary, or taboo); if so, mark it directly and stop further splitting; if not, split the node into multiple child nodes, recursively perform the above evaluation on each child node, until a preset maximum depth is reached or all nodes have been clearly classified. This adaptive partitioning strategy ensures that computational resources are concentrated in key regions near the boundary, while coarse-grained representation is used for clearly defined safe or taboo regions, balancing accuracy and efficiency.

[0151] After the spatial index database is established, the coordinates of the current joint state point and the direction vector of the preset motion trajectory are obtained. The current joint state point is a point in the multi-dimensional joint space, representing the complete state of the joint at the current moment; the direction vector indicates the direction and intensity of the expected motion, generated by the control system based on the user's intention and auxiliary strategies. A point localization algorithm queries the spatial index database to find the spatial region to which the current state point belongs, obtaining the region type identifier. The point localization algorithm utilizes the hierarchical structure of an octree, starting from the root node and progressively determining the child nodes to which the point belongs until a leaf node is reached, obtaining the region type identifier of that leaf node. The computational complexity of this step is logarithmic, allowing for rapid completion even in high-dimensional spaces.

[0152] Based on the region type identifier and the direction vector of the preset motion trajectory, a trajectory safety assessment is performed. If the current point is already in a forbidden zone, a safety intervention will be triggered immediately; if the current point is in a safe zone, the system needs to assess whether moving along the preset trajectory will enter the forbidden zone. The assessment method is to emit a probe ray along the direction vector and calculate the intersection of the ray with the boundary of the forbidden zone. The probe ray is a virtual line segment extending from the current point along the direction of motion, and its length is usually set to 1.5 to 3 times the expected motion distance to provide sufficient predictability.

[0153] The intersection calculation between the detection ray and the forbidden region boundary employs a ray tracing algorithm. This algorithm is efficiently implemented in an octree space, traversing the ray path node by node to detect whether the ray enters a node marked as a forbidden region. When an intersection is detected, the distance from the current position to the intersection point is calculated as a safety margin assessment value. The safety margin assessment value is a key safety indicator that quantifies the distance between the current movement trend and potential risks; a smaller value indicates a closer proximity to the potential risk. The safety margin assessment value is compared with a preset safety threshold, which is typically dynamically adjusted based on the movement speed; the faster the speed, the larger the required safety margin.

[0154] When the safety margin assessment value is lower than the safety threshold, the preset trajectory is considered to pose a safety risk and requires trajectory adjustment. Trajectory adjustment employs gradient descent, an optimization algorithm that iterates along the inverse direction of the objective function's gradient to find a local minimum. In this system, the objective function is designed as the negative value of the distance between the trajectory and the taboo region boundary, plus a penalty term for deviation from the original preset trajectory. The goal of this design is to maintain the intent of the original trajectory as much as possible while ensuring that the corrected trajectory maintains a sufficiently safe distance from the taboo boundary.

[0155] During gradient descent, the key points closest to the taboo boundary in the trajectory are first identified. Then, multiple candidate adjustment schemes are generated around these points. The safety margin and trajectory deviation of each scheme are evaluated, and the optimal adjustment direction is selected. The adjustment process is iterative, with trajectory parameters fine-tuned and the safety margin recalculated each time, until the safety threshold requirement is met or the maximum number of iterations is reached. The final adjusted trajectory satisfies the safety requirements while preserving the original trajectory's intended motion to the greatest extent possible.

[0156] In practical applications, the trajectory adjustment process is continuous and smooth, constantly monitoring the safety margin during movement and dynamically adjusting the trajectory as needed. This real-time safety monitoring and trajectory optimization mechanism effectively prevents potential safety risks while maintaining the smoothness and naturalness of human-computer interaction. For example, when a sudden increase in the user's muscle tension is detected, the restricted area is immediately expanded, and the movement trajectory is adjusted accordingly to avoid applying excessive torque; when joint movement is detected to be approaching a certain limit, the speed is reduced in advance and the trajectory direction is corrected to prevent overextension of the joint.

[0157] It is worth noting that the rules for defining safe and prohibited areas will be continuously optimized based on usage. By recording user feedback and changes in physiological signals, personalized safety boundaries can be learned, ensuring that the safety mechanism is both effective and does not over-interfere, providing users with a safe and comfortable assistance experience.

[0158] Example 10

[0159] In this embodiment, real-time anomaly monitoring based on the feature dataset and historical database is also included. When an anomaly is detected, an early warning signal and recommended intervention measures are generated, specifically including:

[0160] The historical data in the historical database is hierarchically segmented according to multiple time scales such as day, hour, and minute to obtain a multi-level time interval division structure;

[0161] For each time interval in the multi-level time interval division structure, the signal statistics, abnormal event counts and representative feature samples in the corresponding time interval are calculated to obtain the interval feature summary, and a kd-tree spatial index is constructed based on the interval feature summary.

[0162] Extract the current time window feature vector from the feature dataset, calculate the deviation from the historical normal pattern, and when the deviation exceeds the anomaly detection threshold, determine that an anomaly has been detected and use the current time window feature vector as the anomaly feature vector.

[0163] Based on the abnormal feature vector, a k-nearest neighbor query is initiated in the kd-tree spatial index, with k set to a value of 5 to 20, and the set of historical time intervals with the closest feature distance is returned.

[0164] For each coarse-grained time interval in the set of historical time intervals, the system recursively enters the next level of index for location until a specific historical anomaly record is located. The system then returns the specific historical anomaly record and its corresponding subsequent development information, and generates the warning signal and the recommended intervention measures.

[0165] Specifically, by analyzing the differences between current physiological signal characteristics and historical patterns, potential anomalies can be detected in a timely manner, and possible development trends can be predicted based on similar historical cases, providing early warnings and intervention suggestions. This predictive safety assurance mechanism is an advanced function of intelligent auxiliary systems, capable of intervening before abnormal situations fully manifest, significantly improving system security and user experience.

[0166] First, the data in the historical database is hierarchically segmented according to multiple time scales, forming a multi-level time interval division structure. This hierarchical time index structure is the data foundation of the anomaly detection system. It divides the continuous historical data stream into time segments of different granularities, facilitating rapid location and retrieval. A top-down hierarchical strategy is adopted, first dividing by day, then by hour within each day, and then by minute within each hour, forming a tree-like time index structure. This multi-level division method can capture patterns and anomalies at different time scales.

[0167] For each time interval in the multi-level time interval division structure, three key indicators are calculated: signal statistics, anomaly event counts, and representative feature samples, collectively referred to as the interval feature summary. Signal statistics include statistical features such as mean, variance, kurtosis, and skewness, describing the overall distribution characteristics of the signal within the time interval. Anomaly event counts record the frequency of various anomaly events detected within the interval, reflecting the degree of anomaly. Representative feature samples are representative feature vectors selected from the interval, typically including both normal and anomaly samples, used for subsequent similarity comparisons. These interval feature summaries significantly compress the original data volume while retaining key information, enabling efficient processing and retrieval of large amounts of historical data.

[0168] A kd-tree spatial index is constructed based on interval feature summaries. A kd-tree (k-dimensional tree) is a data structure used to organize points in a k-dimensional space, particularly suitable for range queries and nearest neighbor searches in multidimensional spaces. In this system, each interval feature summary is considered a point in the multidimensional space, and the point's coordinates are determined by the values ​​of various indicators in the summary. The kd-tree recursively divides the space into two halves, alternating between different dimensions as the partitioning criteria, forming a binary tree. Each non-leaf node of the tree is a partitioning hyperplane, dividing the space into two parts; leaf nodes contain the actual interval feature summary data. This index structure enables nearest neighbor searches in high-dimensional feature spaces to be performed efficiently with logarithmic complexity.

[0169] The construction process of a kd-tree is as follows: First, the dimension with the largest variance is selected as the partition dimension of the root node. The median of all points along this dimension is calculated as the partition value, dividing the point set into two subsets. Then, the above process is recursively executed for each subset, selecting the dimension with the largest variance in the current subset as the partition dimension each time, until a termination condition is met (such as the subset size being less than a preset threshold or reaching the maximum depth). This construction strategy ensures the balance of the tree and optimizes query performance.

[0170] During real-time anomaly monitoring, a time window feature vector is extracted from the current feature dataset, and its deviation from historical normal patterns is calculated. The deviation calculation employs algorithms such as Mahalanobis distance or local outlier factor, which consider the correlation between feature dimensions and local density distribution, and more accurately reflect the degree of data anomaly than simple Euclidean distance. When the deviation exceeds a preset anomaly detection threshold, an anomaly is detected, and the current feature vector is marked as an anomalous feature vector. The anomaly detection threshold is typically determined based on statistical analysis of historical data and may be dynamically adjusted with increased usage time and the learning of individual characteristics.

[0171] After detecting an anomalies, it is necessary to analyze their severity and potential trends to generate appropriate early warnings and intervention recommendations. To this end, based on the anomaly feature vector, a k-nearest neighbor query is initiated in the kd-tree spatial index to search for the most similar cases in historical data. The k-nearest neighbor query is an algorithm that finds the k nearest points to a target point in a multi-dimensional space. The value of k is set to 5 to 20; this range is selected based on experimental verification, providing sufficient reference cases while avoiding interference from noisy points.

[0172] The k-nearest neighbor query leverages the properties of kd-trees, recursively searching from the root node. It utilizes a hyperplane to quickly eliminate a large number of points that are unlikely to be neighbors, significantly improving query efficiency. The query returns a set of historical time intervals with the closest feature distances; these intervals represent historical cases most similar to the current anomaly pattern.

[0173] For the returned set of historical time intervals, further fine-grained positioning is performed. Since the aforementioned query may return coarse-grained time intervals (such as days or hours), it is necessary to recursively enter the next level of index, gradually narrowing the positioning range until the specific historical anomaly record is located. For example, if the initial query returns data for a certain day, it will further search in the hourly index within that day, and then delve into the minute-level index, ultimately pinpointing the historical anomaly event most similar to the current anomaly.

[0174] Once a specific historical anomaly record is located, its subsequent development information is retrieved. This subsequent development information refers to the evolution of the historical anomaly after its occurrence, including the duration of the anomaly, changes in severity, whether it led to more serious consequences, intervention measures taken, and their effects. This information is crucial for predicting the potential development path of the current anomaly and for formulating intervention strategies.

[0175] Based on historical anomaly records and subsequent development information, early warning signals and recommended interventions are generated. Early warning signals include the type of anomaly, its severity, potential development trends, and suggested indicators for attention. Recommended interventions are tailored to the specific circumstances of the current user, based on effective methods for handling similar anomalies in the past. For example, if an abnormal pattern of electromyographic signals is detected, and historical data shows that such an anomaly often leads to motor incoordination, it might be recommended to temporarily reduce the assisted intensity while increasing stability control.

[0176] Early warnings and interventions are generated using a hybrid rule-based and case-based approach. First, experience is extracted from historical cases, then predefined safety rules and current environmental factors are combined to formulate the final intervention strategy. The severity of the intervention measures is proportional to the severity and certainty of the anomaly; minor or uncertain anomalies may only trigger increased monitoring frequency, while clearly defined severe anomalies will lead to immediate protective intervention.

[0177] The entire anomaly monitoring and early warning system possesses self-learning and optimization capabilities. After each anomaly detection and intervention, the results are recorded and the historical database is updated, continuously enriching the case library and optimizing detection parameters. As usage time increases, the system's ability to recognize anomaly patterns for specific users will continuously improve, and the accuracy and timeliness of early warnings will correspondingly increase. This continuous learning mechanism enables the system to adapt to the characteristics and needs of different users, providing personalized security guarantees.

[0178] It is worth noting that anomaly monitoring systems not only focus on obvious emergencies but also pay special attention to capturing subtle early signs of anomalies. By analyzing the coordinated change patterns of multimodal signals, potential anomaly trends can be detected before problems become apparent, achieving true predictive protection. For example, subtle changes in the coherence between certain channels may be detected. Although these changes may not yet affect overt behavior, historical data shows that they are often precursors to uncoordinated subsequent actions, allowing control strategies to be adjusted in advance.

[0179] Example 11

[0180] In this embodiment, the construction of the pre-built motion intent template library includes a template compression method based on representative selection, specifically including:

[0181] The motion intent events collected in history are acquired, and the signals of each motion intent event are segmented by sliding window, and time-domain and frequency-domain features are extracted to obtain the feature vectors of each motion intent event. The vector set composed of all the feature vectors is defined as the feature space.

[0182] Based on the feature space, the cosine similarity value between each feature vector in the feature space is calculated, and a similarity graph is constructed. The vertices of the similarity graph are each motion intention event, and the edges connect event pairs whose cosine similarity value exceeds the connection threshold.

[0183] Based on the similarity graph, for each vertex in the similarity graph, the number of vertices whose cosine similarity value with the vertex exceeds the connection threshold and has not yet been covered by other vertices is counted, and the number of vertices is used as the coverage contribution of the vertex.

[0184] The representative node set is initialized to an empty set, and the uncovered vertex set is initialized to all vertices of the similarity graph. The vertex with the largest coverage contribution is selected from the uncovered vertex set and added to the representative node set. Vertices whose cosine similarity value with the vertex with the largest coverage contribution exceeds the connection threshold are removed from the uncovered vertex set and marked as covered vertices. The coverage contribution of the remaining vertices in the uncovered vertex set is recalculated. The selection and removal operations are repeated until the uncovered vertex set is empty, resulting in the minimum representative node set.

[0185] Extract the signal pattern and feature vector of the motion intention event corresponding to each representative node in the minimum representative node set, and store the signal pattern and feature vector as a standard motion intention template in the pre-built motion intention template library.

[0186] Specifically, regarding the motion intent template compression method based on representative selection, this method constructs an efficient motion intent template library by intelligently selecting the most representative samples. This significantly reduces the number of templates while maintaining recognition performance, thereby improving system efficiency. This optimization is particularly important for resource-constrained wearable devices, as it can significantly reduce storage requirements and computational burden while ensuring complete coverage of the motion intent space by the template library.

[0187] First, we acquire historical datasets of motor intention events. This historical data comes from various sources, including early user records, clinical trial data, and publicly available datasets. A motor intention event is a complete signal record generated when a user performs a specific action intention; it typically includes multimodal data such as electromyography (EMG), electroencephalography (EEG), and acceleration signals, as well as corresponding action labels and timestamps. This raw data forms the foundation for constructing the template library.

[0188] For each motion intent event, preprocessing and feature extraction are performed first. The preprocessing stage includes signal denoising, baseline drift correction, and bandpass filtering to improve signal quality. Subsequently, a sliding window segmentation technique is used to divide the continuous signal. The sliding window is a commonly used technique in signal processing; it divides a long time series into multiple short segments by moving a window of fixed length along the time axis. In this system, the window length is typically set to 100-300 milliseconds, a range that captures complete motion intent features while maintaining good temporal resolution. The window overlap rate is set to 50-75% to ensure that key features at the window boundaries are not missed.

[0189] For each signal segment within a window, both time-domain and frequency-domain features are extracted simultaneously to obtain a comprehensive feature representation. Time-domain features primarily include statistics such as mean, variance, peak value, zero-crossing rate, and waveform length; these features directly describe the signal's amplitude and morphological characteristics. Frequency-domain features are obtained through Fast Fourier Transform or Wavelet Transform, including power spectral density, dominant frequency components, and bandwidth ratios; these features reflect the signal's frequency composition and rhythmic characteristics. These two types of features complement each other, together constituting a comprehensive description of the motion intent signal.

[0190] For each motion intent event, the feature vectors of all windows are integrated to form a feature representation of that event. The integration method can be simple concatenation or statistical aggregation, depending on the specific application requirements. The feature vectors of all motion intent events together constitute a high-dimensional feature space, where each point represents a motion intent event, and the distance between points reflects the similarity between events.

[0191] Based on the constructed feature space, cosine similarity values ​​between feature vectors are calculated to construct a similarity graph. Cosine similarity is an indicator that measures the directional similarity between two vectors. It is calculated by taking the cosine of the angle between the two vectors, ranging from -1 to 1. A larger value indicates greater similarity in the vector directions. Cosine similarity is particularly suitable for similarity measurement in high-dimensional feature spaces, is insensitive to vector length, and can effectively capture directional differences in feature distribution. The cosine similarity between all vector pairs in the feature space is calculated. When the similarity value exceeds a predetermined connection threshold, the two motion intention events are considered to be significantly similar. The connection threshold is typically set between 0.8 and 0.95, and can be adjusted according to the accuracy requirements of the application scenario. A higher threshold will produce a stricter similarity judgment, ensuring that only very close samples are considered similar.

[0192] Similarity graphs are an intuitive method for representing the structure of a feature space, where vertices represent motion intent events and edges connect event pairs with similarity exceeding a threshold. This graph structure clearly demonstrates the clustering relationships and connectivity in the motion intent space, providing a topological foundation for subsequent representative sample selection. The graph density reflects the redundancy of the dataset; densely connected regions represent a large number of similar samples and are the primary target for compression.

[0193] Based on the constructed similarity graph, a coverage contribution is calculated for each vertex. Coverage contribution is a measure of a vertex's ability to "represent" other vertices, defined as the number of vertices with a similarity exceeding a connection threshold that are not yet covered by other vertices. Intuitively, vertices with high coverage contributions are located at the center of their cluster, covering (or representing) a large number of similar samples; while vertices with low contributions are located on the periphery, or the samples they might cover have already been covered by other vertices. By calculating coverage contributions, the system can identify the most representative samples and prioritize their inclusion in the template library.

[0194] The calculation of coverage contribution needs to consider the covered state, which is a dynamically changing attribute. Initially, all vertices are marked as uncovered; as representative nodes are selected, more and more vertices are marked as covered, and the coverage contribution of the remaining vertices changes accordingly. This dynamic update mechanism ensures that each selection is the optimal choice in the current state, avoiding redundant coverage.

[0195] The selection of representative nodes employs a greedy algorithm, a classic approach to solving the set-covering problem. First, the representative node set is initialized to an empty set, and the set of uncovered vertices consists of all vertices. Then, the following steps are repeated: The vertex with the highest coverage contribution is selected from the set of uncovered vertices and added to the representative node set; this vertex and all vertices it can cover are removed from the uncovered set and marked as covered; the coverage contribution of the remaining uncovered vertices is recalculated. This process continues until all vertices are covered, i.e., the set of uncovered vertices is empty.

[0196] While greedy algorithms do not guarantee a globally optimal solution, they typically provide near-optimal results in practical applications and are far more computationally efficient than exhaustive search. Theoretically, for the set covering problem, the approximation ratio of a greedy algorithm is ln(n), where n is the total number of elements, meaning the solution found by the algorithm is at most ln(n) times the size of the optimal solution. In the practical application of this system, this approximation is usually perfectly acceptable.

[0197] Through the greedy selection process described above, the minimum representative node set is finally obtained. The number of nodes in this set is usually much smaller than the original dataset size, and the compression ratio can reach 80-95%, depending on the redundancy of the original data and the set connection threshold. A lower connection threshold will produce a higher compression ratio, but may reduce the accuracy of template coverage; a higher threshold will retain more templates, improving accuracy but with weaker compression effect. The system will balance this trade-off according to the actual application requirements.

[0198] Finally, the original signal patterns and feature vectors of motion intent events corresponding to each node in the minimum representative node set are extracted and stored as standard motion intent templates in a pre-built motion intent template library. The signal patterns retain complete information about the time-domain waveform, facilitating direct template matching; the feature vectors provide compressed feature representations, suitable for rapid similarity calculation. These two forms complement each other, meeting the needs of different scenarios.

[0199] The completed motion intent template library is not static but is continuously optimized and updated as the system is used. When new motion intent samples are collected, their similarity to existing templates is evaluated. If a new sample cannot be well covered by existing templates (similarity is below a certain threshold), it is added to the template library. At the same time, the usage frequency and effectiveness of existing templates are regularly evaluated, and templates that have not been used for a long time or have poor recognition performance are removed to keep the template library concise and efficient.

[0200] This template compression method based on representativeness selection enables the system to maintain a comprehensive and efficient motion intent template library with limited computing and storage resources, providing a reliable reference foundation for real-time motion intent recognition. Compared with traditional full-sample storage or simple random sampling, this method can intelligently retain the most informative samples, maintaining recognition performance while significantly reducing the number of templates, making it particularly suitable for resource-constrained wearable device applications.

[0201] Example 12

[0202] This embodiment also includes a multi-object resource scheduling method based on coverage optimization, specifically including:

[0203] Obtain the current state vectors of multiple monitored objects, wherein the current state vectors include risk level, physiological stability, time interval between the last monitoring, and parameters of the current training phase;

[0204] Calculate the state similarity between each monitored object, define the attention substitution relationship between objects based on the state similarity, and construct an object relationship graph;

[0205] Calculate the attention demand weight for each monitored individual, the attention demand weight taking into account risk level, time interval and status change rate;

[0206] Based on the object relationship graph and the weights of the concerns, an integer linear programming method is used to solve the minimum coverage set problem with weight constraints to obtain a priority concern object list. The objects in the priority concern object list can directly or through similarity cover the needs of all monitored objects.

[0207] Based on the list of priority objects and real-time status change events, the priority queue sorting is dynamically updated to generate nursing resource allocation suggestions.

[0208] Specifically, for multi-object monitoring scenarios, the system optimizes the allocation strategy of limited resources by intelligently analyzing the similarity of object states and the urgency of their needs, ensuring maximum overall monitoring efficiency. This resource scheduling mechanism is particularly suitable for scenarios with limited nursing staff and numerous monitored objects, significantly improving resource utilization efficiency and monitoring quality.

[0209] First, obtain the current state vector for all monitored subjects. Monitored subjects refer to users who need to use wearable robotic assistive function training devices, and may be patients, recovering patients, or health trainers. The current state vector is a multi-dimensional data structure describing the current comprehensive condition of the monitored subject, containing four key types of information: risk level, physiological stability, last attention interval, and current training stage parameters. The risk level indicates the likelihood of the subject experiencing an accident or requiring emergency intervention, typically categorized into low, medium, and high levels, or more detailed classifications. Physiological stability quantifies the stability of the subject's physiological indicators, including heart rate variability, blood pressure fluctuations, and electromyographic signal stability. The last attention interval records the time elapsed since the last direct attention from a caregiver, reflecting the potential risk of insufficient attention. Current training stage parameters include information such as training type, intensity, completion rate, and adaptability score, reflecting the subject's current training status and progress.

[0210] This status information comes from multiple sources, including real-time physiological signals collected by wearable devices, historical data recorded by the system, subjective assessments by medical staff, and self-reports from the subjects. This heterogeneous data will undergo standardization processing, converting indicators from different sources and with different dimensions into standardized scores to facilitate subsequent calculations and comparisons.

[0211] After obtaining the state vectors, the state similarity between each monitored object is calculated. State similarity is an indicator of how close the current states of two monitored objects are. High similarity means that the two objects face similar risks and needs, and may be suitable for similar attention strategies. Similarity calculation uses weighted Euclidean distance or Mahalanobis distance, with the weights of different dimensions set according to clinical importance and practical experience. For example, risk level and physiological stability usually receive higher weights because they are directly related to safety; while parameters from the training phase may have relatively lower weights. Similarity values ​​are usually normalized to the [0, 1] interval, with larger values ​​indicating greater similarity.

[0212] Based on calculated state similarity, we define the substitutability relationship between objects and construct an object relationship graph. The substitutability relationship is the core concept of this method; it defines the possibility that "object A can, to some extent, substitute for object B." When two objects are highly similar in state, paying attention to one object may simultaneously satisfy some of the attention needs for the other object. This "one-time attention, double benefit" effect is the basis of resource optimization.

[0213] The definition of substitutability is not a simple binary relation, but a fuzzy relation with intensity. When the similarity between two objects exceeds a preset threshold (usually 0.8 or 0.85), substitutability is considered to exist, and the similarity value is used as the substitution intensity. Substitutability intensity reflects the degree to which object A can substitute for object B. Identical objects have a substitution intensity of 1 (100% substitution), and the substitution intensity decreases accordingly as the similarity decreases.

[0214] An object relationship graph is a mathematical structure representing the substitutive relationships among all monitored individuals. Nodes in the graph represent monitored individuals, directed edges represent substitutive relationships, and the weights of the edges represent the strength of substitution. This graph structure visually illustrates similarity clustering and potential resource-sharing opportunities within the entire monitored group. Object relationship graphs can be dense or sparse, depending on the homogeneity of the monitored individual group. Highly homogeneous groups (such as patients of similar age or with similar conditions) typically form dense graphs, providing more opportunities for resource optimization; while highly heterogeneous groups form sparser graphs with limited space for resource sharing.

[0215] While constructing the object relationship graph, a attention requirement weight is calculated for each monitored object. The attention requirement weight is a comprehensive indicator measuring the urgency of an object's current need for attention and is a key parameter in subsequent optimization processes. The attention requirement weight comprehensively considers three main factors: risk level, time interval, and state change rate. Risk level directly affects attention requirement; high-risk objects naturally require more attention. The time interval reflects the fairness of attention; objects that have not received attention for a long time will have their weight increased to ensure no object is neglected indefinitely. The state change rate captures the dynamic characteristics of an object's state; objects with rapidly changing states typically require more frequent attention to allow for timely adjustments to intervention strategies.

[0216] The calculation of demand weights employs a non-linear weighted model. Risk levels typically influence weights exponentially; for example, a high-risk object might have twice the weight of a medium-risk object, and a medium-risk object might have 1.5 times the weight of a low-risk object. The impact of time intervals is also non-linear, conforming to the principle of "increasing marginal utility," with the weight growth rate accelerating as the period of no attention increases. The impact of the rate of change of state is quantified through short-term volatility indicators, with objects experiencing significant fluctuations receiving additional weight. These non-linear relationships ensure that the objects requiring the most attention receive sufficiently high weights, giving them an advantage in resource allocation.

[0217] Based on the constructed object relationship graph and the calculated attention requirement weights, an integer linear programming method is used to solve the minimum coverage set problem with weight constraints. The minimum coverage set problem is a classic problem in combinatorial optimization, aiming to find the smallest subset of vertices such that all vertices in the graph are either in this subset or adjacent to some vertex in the subset. In this system, this problem is extended to a version with weight constraints: finding the subset of vertices with the largest sum of attention requirement weights, such that all monitored objects are either in this subset (directly monitored) or have a sufficiently strong substitution relationship with some vertex in the subset (covered by similarity).

[0218] Integer linear programming is a powerful tool for solving such combinatorial optimization problems. It formalizes the problem as an optimization problem of a linear objective function under integer constraints. In this system, the decision variables are binary variables representing whether to select specific objects to focus on; the objective function is to maximize the sum of the focus demand weights of the selected objects; the constraints include coverage constraints (each object must be directly focused on or covered by similarity) and resource constraints (the number of objects to be focused on simultaneously does not exceed the number of available resources). Solving this integer programming problem yields the optimal set of objects to focus on.

[0219] Since integer linear programming problems are typically NP-hard, finding an exact solution for a large group of monitored objects can be time-consuming. In real-time systems, approximate or heuristic algorithms, such as greedy algorithms or local search, can be used to obtain near-optimal solutions within an acceptable timeframe. In practical applications, the system dynamically selects an appropriate solution strategy based on the number of monitored objects and available computational resources.

[0220] The solution yields a priority list of objects that can directly or through similarity cover the needs of all monitored objects. Objects in this list are ordered by priority, determined by both the weight of the need and the coverage contribution. Coverage contribution refers to an object's ability to cover other objects through similarity relationships. Objects with higher coverage contributions are prioritized because focusing on them yields greater overall benefits.

[0221] The priority list is the result of static planning, but the actual monitoring environment is dynamically changing. To adapt to this dynamism, the attention queue is dynamically updated based on the priority list and real-time status change events. Real-time status change events include sudden anomalies (such as abnormal fluctuations in physiological indicators), training phase transitions (such as moving from warm-up to main training), and user requests (such as pressing the help button). These events trigger immediate reordering of the attention queue to ensure that the most urgent needs are responded to promptly.

[0222] Dynamic updates employ a sliding window approach, allowing for local adjustments to respond to urgent changes while maintaining overall planning stability. A short-term history of attention is maintained to avoid frequent switching between the same object; this "sticky" mechanism improves attention consistency and efficiency. Simultaneously, the fairness of attention allocation is monitored to ensure no object remains in a low-priority state for an extended period and is ignored.

[0223] Ultimately, recommendations for allocating nursing resources are generated to guide the work arrangements of healthcare professionals. These recommendations include priorities, key areas of focus, estimated duration of attention, and key areas for subsequent observation. The recommendations are presented in a concise and clear format, typically including graphical representations (such as priority heatmaps) and textual explanations, to facilitate quick understanding and implementation by healthcare professionals.

[0224] This multi-object resource scheduling method based on coverage optimization significantly improves monitoring efficiency, especially in resource-constrained scenarios. By identifying groups of objects with similar states, it optimizes the order of attention, satisfying multiple similar needs with a single attention, reducing redundant work. Simultaneously, dynamic priority adjustment ensures timely response to urgent needs, maintaining overall monitoring quality. Notably, the scheduling strategy continuously learns and optimizes based on actual usage. By recording the effects of attention behaviors and feedback from monitored objects, it can adjust the similarity calculation model, the definition of substitution strength, and the method for calculating demand weights, making the scheduling strategy increasingly aligned with actual needs and providing increasingly accurate resource allocation recommendations.

[0225] Example 13

[0226] In this embodiment, the generation of the early warning signal includes a multi-source information fusion decision-making mechanism, specifically including:

[0227] By integrating the current movement intention category, the quantified value of the intention intensity, the safety margin assessment value, the current physiological state category, the state confidence level, and the specific historical anomaly records and their subsequent development information, the weight coefficients of each information source are determined through the analytic hierarchy process (AHP), and a weighted comprehensive anomaly score is calculated to obtain a multidimensional anomaly index vector.

[0228] Based on the multidimensional abnormal indicator vector, the degree of abnormality is divided into normal, mild warning, moderate warning and severe warning levels by a grading judgment rule, the warning level is determined, and the warning signal and the corresponding clinical significance interpretation are output.

[0229] Based on the warning level and the intervention-effect correlation records in the historical intervention effect database, the historical case with the smallest distance from the current situation is queried by similar case matching, and the effective intervention strategy in the historical case is extracted to generate the recommended intervention measures.

[0230] Specifically, regarding the early warning decision-making mechanism based on multi-source information fusion, this mechanism achieves accurate assessment and timely early warning of potential risks by comprehensively analyzing information from different dimensions. This multi-dimensional risk assessment method significantly improves the accuracy and relevance of early warnings, providing strong protection for the safe operation of wearable robot-assisted systems.

[0231] First, six key information sources are integrated: current movement intention category, quantified intention intensity value, safety margin assessment value, current physiological state category, state confidence level, and specific historical anomaly records and their subsequent development information. These information sources come from different functional modules of the system, each reflecting different aspects of the user's state and the system's operational security. The current motion intention category is the output of the motion intention recognition module, indicating the user's current action intention, such as "elbow flexion" or "wrist extension." The intention intensity quantification value further describes the intensity level of the intention, usually expressed as a percentage, reflecting the degree of determination of the user's action intention. The safety margin assessment value comes from the trajectory safety detection module, quantifying the distance between the current motion trajectory and the danger boundary; the smaller the value, the closer to the potential risk area. The current physiological state category is the output of the physiological state classification module, indicating whether the user is in a state such as "fatigue," "normal," or "excitement." The state confidence level indicates the degree of certainty about the physiological state classification result; high confidence level indicates that the classification is reliable, while low confidence level indicates that the classification result is uncertain. Specific historical anomaly records and their subsequent development information come from the anomaly monitoring module, which includes historical anomaly cases similar to the current state and their development trajectories, providing a reference for predicting the evolution of the current situation.

[0232] These multi-source information sources each provide security-related clues from different perspectives, but directly interpreting and synthesizing this heterogeneous information is extremely challenging for human operators. It is necessary to integrate them into a comprehensive score to quickly assess the current overall risk level. Therefore, the analytic hierarchy process (AHP) is used to determine the weighting coefficients of each information source.

[0233] The Analytic Hierarchy Process (AHP) is a structured decision-making method that decomposes complex problems into hierarchical structures and determines the relative importance of each element through pairwise comparisons. In this system, the application of AHP involves three steps: First, establishing a hierarchical structure, decomposing the early warning decision problem into a target layer (comprehensive risk assessment), a criterion layer (six types of information sources), and possible sub-criterion layers; second, constructing a judgment matrix, comparing the importance of each pair of information sources through expert evaluation or learning based on historical data, quantifying it on a scale of 1-9; finally, calculating the weight coefficients of each information source, typically using the eigenvalue method to solve for the eigenvector corresponding to the largest eigenvalue of the judgment matrix, which, after normalization, becomes the weight coefficient.

[0234] The weighting coefficients are set by comprehensively considering the reliability, timeliness, and risk relevance of the information source. For example, the safety margin assessment value usually receives a higher weight because it directly quantifies the proximity to the safety boundary; while the weight of the physiological state category is dynamically adjusted according to its confidence level, increasing with high confidence and decreasing with low confidence, reflecting the impact of information reliability on decision-making. Furthermore, the weights are also personalized based on the usage environment and user characteristics. For example, for users with weaker muscle strength, the weight of physiological states may be increased to more cautiously assess fatigue risk.

[0235] After determining the weighting coefficients, a weighted comprehensive anomaly score is calculated. This is not a simple linear weighted sum, but rather employs a more complex nonlinear fusion model that considers the interaction effects between information sources. For example, when the intensity of the movement intention is high and the safety margin is low, the risk score will increase additionally, reflecting the synergistic risks that this combination may bring; similarly, when the physiological state is fatigued and historical records show that errors are more likely to occur under similar conditions, the risk score will also increase accordingly. This nonlinear fusion better reflects the complexity of actual risk assessment and can capture comprehensive risk patterns that are difficult to reflect with a single indicator.

[0236] The result of the fusion calculation is not a single anomaly score, but a multi-dimensional anomaly indicator vector. This vector contains anomaly indicators across multiple dimensions, such as safety anomaly indicators, physiological anomaly indicators, and so on. Figure 1 Indicators such as those for pathogenicity abnormalities each have their specific clinical significance and intervention guidance value. Multidimensional representation retains information about the type of abnormality, not just its severity, which is crucial for subsequent targeted interventions.

[0237] Based on a multidimensional anomaly indicator vector, the degree of anomaly is classified into different levels through a hierarchical judgment rule. This hierarchical judgment is not a simple threshold comparison, but a complex set of decision rules that considers the combination patterns and severity of indicators across various dimensions. The rules typically employ fuzzy logic or decision tree structures to handle the interactions and trade-offs between indicators. For example, high anomalies in some dimensions may be partially offset by normal values ​​in other dimensions; and certain specific combinations of indicators may directly trigger high-level warnings, even if individual indicator values ​​are not extreme.

[0238] Warning levels are typically divided into four levels: normal, mild warning, moderate warning, and severe warning. A normal state indicates all indicators are within safe ranges and the system can operate normally. A mild warning indicates a slight abnormality, requiring increased monitoring frequency and preparation for intervention, but current activities can continue. A moderate warning indicates a significant abnormality, suggesting suspension of current activities and necessary adjustments. A severe warning indicates a serious abnormality; the system may automatically suspend activities and implement protective measures, requiring immediate intervention from medical personnel. Each warning level has clearly defined triggering conditions and response protocols to ensure consistency and effectiveness between warnings and interventions.

[0239] Once the warning level is determined, a warning signal and its corresponding clinical significance are output. The warning signal includes multimodal cues such as visual, auditory, and tactile cues to ensure effective communication of risk information across various environments. The clinical significance explanation clearly outlines the cause of the warning, including key abnormal indicators, potential risk factors, and recommended areas of focus. This explanatory output helps users and healthcare professionals understand the reasons for the warning and make more informed response decisions.

[0240] The design of the warning output follows the principle of "low intrusion and high information content." Mild warnings use non-invasive prompts, such as a change in the status light color or slight vibration, to avoid unnecessary tension and interference; while severe warnings use more obvious prompts to ensure sufficient attention. At the same time, all warnings are accompanied by concise explanations, such as "Abnormal fluctuations in electromyographic signals detected, possibly related to fatigue," to help users and medical personnel quickly understand the nature and severity of the warning.

[0241] In addition to early warning, targeted intervention recommendations are also needed. To this end, based on the current early warning level and multidimensional anomaly indicator vectors, similar cases are searched in the historical intervention effectiveness database. This database is a knowledge resource accumulated by the system, recording intervention measures taken in various past anomaly situations and their effectiveness evaluations. Each record includes information such as anomaly characteristic description, intervention measures, intervention results, and effectiveness evaluation, forming a complete knowledge unit of "problem-solution-effect".

[0242] Similar case matching employs a distance-based nearest neighbor search method. The feature distance between the current anomaly indicator vector and historical cases is calculated, typically using Mahalanobis distance or weighted Euclidean distance, considering the relevance and importance of features. Distance calculation pays particular attention to the matching degree of key anomaly dimensions, ensuring that the found historical cases are similar to the current situation in key risk factors. The system returns multiple historical cases with the smallest feature distances, typically 5-10 of the most similar cases, as a reference basis for intervention decisions.

[0243] This process extracts previously effective intervention strategies from matched historical cases. It doesn't simply involve selecting the intervention method from a single most similar case, but rather analyzing the commonalities and individual effects of interventions across multiple similar cases to synthesize the most likely combination of effective intervention strategies. The success rate, magnitude of effect, and applicability of each intervention strategy in similar cases are evaluated, prioritizing strategies with high success rates and significant effects, while also considering the fit with the current environment and user characteristics.

[0244] The generated recommended interventions include multiple levels: automatically implemented adjustments, such as reducing assistance level or modifying control parameters; user-implemented adjustments, such as adjusting posture or providing rest suggestions; and professional interventions that may require healthcare personnel, such as adjusting training plans or professional assessments. Each intervention recommendation is accompanied by a concise explanation of the rationale and expected effects to help decision-makers understand the basis and purpose of the recommendation.

[0245] It is worth noting that the entire early warning and intervention recommendation system possesses learning and adaptive capabilities. It records the results of each early warning and intervention, including assessments of early warning accuracy and intervention effectiveness, continuously updating and optimizing the internal model. For example, if a certain type of early warning occurs frequently but rarely requires actual intervention, the early warning threshold for that type of situation may be adjusted to reduce false alarms; if a certain intervention measure is significantly effective among a specific user group, its recommendation priority in similar situations will be increased. This closed-loop learning mechanism enables the system to continuously improve the accuracy of early warnings and the effectiveness of interventions, providing increasingly personalized and precise support.

[0246] The multi-source information fusion-based early warning and decision-making mechanism significantly improves the comprehensiveness and accuracy of risk assessment by integrating status information from different dimensions. Compared with single-indicator monitoring, this multi-dimensional fusion analysis can capture complex risk patterns, reduce false alarms and missed alarms, and provide more reliable and interpretable early warnings. Simultaneously, the intervention recommendation mechanism based on historical cases fully utilizes accumulated experience and knowledge, helping users and healthcare professionals quickly identify potentially effective intervention options, improving the efficiency and effectiveness of responding to abnormal situations. This intelligent early warning and intervention support system is a crucial guarantee for the safety and effectiveness of wearable robot assistive function training devices, providing users with comprehensive protection and optimization support.

[0247] Example 14

[0248] This embodiment also includes a large-scale data visualization method based on sparse sampling, specifically including:

[0249] A set of historical time-series data points generated by long-term monitoring is obtained, and a representative score is calculated for each data point in the feature space. The representative score comprehensively considers the extreme value characteristics, rate of change characteristics, and abnormal characteristics of the data points.

[0250] Based on the representativeness score, highly representative data points are selected using an importance sampling algorithm to obtain a sparse keypoint subset. The number of data points in the sparse keypoint subset is 1% to 10% of the number of data points in the historical time series data point set.

[0251] Based on the zoom level parameters of the user interface, the sampling density of the sparse keypoint subset is adaptively adjusted to display low-density keypoints in the macro view and high-density keypoints in the micro view, generating multi-resolution display data.

[0252] Based on the user's point selection or box selection operation, a nearest neighbor query or range query is launched to quickly locate the data point closest to the selected position or all data points within the boxed area in the historical time series data point set, return detailed information and support interactive exploration.

[0253] Specifically, a large-scale data visualization method based on sparse sampling addresses the challenges of computational efficiency and information representation in visualizing massive time-series data by intelligently selecting the most representative data points for display. This visualization technology enables healthcare professionals to quickly browse and analyze long-term monitoring data, identify key patterns and trends, and provide strong support for clinical decision-making.

[0254] First, it's crucial to acquire a set of historical time-series data points generated from long-term monitoring. Long-term monitoring data is typically massive and complex in dimensions. A standard rehabilitation training process can last for weeks or months, generating hours of multi-channel sensor data daily, with sampling rates ranging from tens to thousands of hertz. For example, a system using 10 sensor channels at a sampling rate of 1000 Hz will generate approximately 288 million data points during 8 hours of training per day. This data volume far exceeds the number of pixels on display devices and human perception capabilities, making it impossible to effectively convey information through simple full-data presentation. Traditional data downsampling methods, such as uniform sampling or simple aggregation (e.g., mean, maximum, etc.), while reducing the amount of data, often lose key local features and abnormal patterns, leading to the omission of important information.

[0255] To address this issue, a representativeness score is calculated for each data point in the feature space. The representativeness score is a comprehensive indicator that measures the information content and importance of a data point, used to identify the most valuable key data points for presentation. The calculation of the representativeness score comprehensively considers three types of key characteristics: extreme value characteristics, rate of change characteristics, and outlier characteristics.

[0256] Extreme value characteristics refer to the locations of maximum or minimum values ​​of data points in time or feature dimensions. These extreme points often mark inflection points or peaks in signals and have significant physical or physiological meaning. The deviation of extreme values ​​within a local window is calculated, assigning higher extreme value scores to points that are significantly higher or lower than the local mean. The size of the local window is adaptively adjusted according to the data characteristics to ensure that extreme value features at different time scales are captured.

[0257] Rate of change characteristics focus on the speed of signal change near a data point, typically quantified by calculating local derivatives or difference values. Regions with high rates of change usually contain rich information about mode transitions. Calculating multi-scale rate of change indices considers both instantaneous changes (high-frequency changes within a short time window) and trend changes (low-frequency fluctuations within a long time window), ensuring that various change patterns can be appropriately represented.

[0258] Anomalies focus on the degree to which data points deviate from the overall distribution or expected pattern. Anomalies may indicate equipment malfunctions, unusual user reactions, or rare events, and are therefore of particular interest. A combination of statistical methods and machine learning techniques is used to assess the degree of anomalies, such as calculating local density, isolated forest scores, or autoencoder reconstruction errors, to identify potential anomalies from multiple perspectives.

[0259] These three types of characteristics each reflect the importance of different dimensions of the data points, and the final representative score is calculated through a weighted combination. The weight settings will be adjusted according to the application scenario and user needs. For example, anomaly detection tasks may emphasize anomaly characteristics, while trend analysis may place more emphasis on the rate of change characteristics. The representative score is usually normalized to the [0, 1] interval for easy comparison and threshold setting.

[0260] Based on the calculated representativeness score, a subset of sparse key points is obtained by selecting highly representative data points through an importance sampling algorithm. Importance sampling is a biased sampling technique that determines the probability of a sample being selected based on its importance weight (in this case, its representativeness score). Unlike simple threshold screening, importance sampling retains the randomness of probabilistic selection, ensuring that highly representative points are prioritized while allowing some moderately representative points to be selected, thus maintaining the overall representativeness of the data distribution.

[0261] The importance sampling implementation employs an improved reservoir sampling algorithm, suitable for streaming data processing without prior knowledge of the total data volume. A target sampling ratio is set (typically 1% to 10% of the original data), and the selection probability of each point is dynamically adjusted based on the representativeness score. During the sampling process, multiple subsets of varying density are maintained to support subsequent multi-resolution display.

[0262] Importance sampling yields a sparse subset of key points, which represents only a small fraction (1%-10%) of the original data, but contains the vast majority of informational key points. This significant reduction in data volume enables complex interactive visualizations while preserving the key features and overall structure of the data.

[0263] To cater to different analysis scenarios and user needs, the sampling density of sparse keypoint subsets is adaptively adjusted based on the zoom level parameters of the user interface. This multi-resolution display mechanism follows the visualization principle of "overview first, details later," providing appropriate information density at different zoom levels. When users view a macro view (such as global trends over several weeks or months), low-density keypoints are displayed, typically selecting a small number of points with the highest representative scores to outline the overall trend and main characteristics. When users zoom in to view a micro view (such as local details over several minutes or hours), high-density keypoints are displayed, including more moderately representative points, presenting more detailed fluctuations and patterns.

[0264] The generation of multi-resolution display data employs a hierarchical caching structure, pre-compiling and storing keypoint sets at different density levels. Typically, 4-6 density levels are maintained, ranging from extremely sparse (approximately 0.1% of the original data) to relatively dense (approximately 10% of the original data), covering various scaling requirements. This pre-computation strategy avoids the computational overhead of real-time resampling, ensuring smooth responsiveness in interactive operations.

[0265] In terms of presentation, various visual encoding methods are employed to enhance the expressiveness of key points. Representative scores can be encoded through visual attributes such as point size, color, or transparency, making important points stand out visually. For multidimensional data, multiple visualization views are provided, such as timeline charts, scatter plots, and heatmaps, allowing users to switch between different representation methods according to their analysis needs. Furthermore, visual annotations are supported, automatically marking important events and abnormal areas to help users quickly locate points of interest.

[0266] In addition to static display, it also supports rich interactive exploration functions, enabling users to actively discover data patterns of interest. Based on the user's point selection or bounding selection, it fires nearest neighbor queries or range queries to quickly locate target data points in the original historical time series data point set. Point selection usually triggers nearest neighbor queries, returning the actual data point closest to the selected location and its detailed information; bounding selection triggers range queries, returning all data points within the bounded area for regional analysis and statistics.

[0267] To support efficient query operations, spatial index structures, such as kd-trees or R-trees, are built on the original data. These index structures organize the data into hierarchical spatial partitions, enabling fast location of data points in the target region with logarithmic complexity, maintaining good response performance even on massive datasets. Furthermore, a multi-level caching strategy is employed, preloading frequently used intervals into memory to further improve query speed.

[0268] The query results not only include the raw data values ​​but also rich contextual information and analysis results. For example, for point-based queries, detailed attributes of the selected points (timestamp, values ​​of each dimension, etc.), local statistical characteristics (mean, variance, etc.), relevant event records, and automatically generated explanatory text are returned. This information helps users fully understand the meaning and context of the selected data points. For bounding box queries, a statistical summary of the region is calculated, major patterns and outliers are identified, and a comprehensive description of the region's characteristics is provided.

[0269] During the interactive exploration process, various advanced analytical functions are supported, such as similar pattern search, trend prediction, and correlation analysis. Users can select a pattern of interest and automatically search for similar segments in the entire dataset; or select a time interval and request the system to predict subsequent trends; or select two variables to analyze their correlation and causal relationship. These analytical functions are calculated in real time using the system's data mining engine, and the results are presented intuitively in a visual format.

[0270] To enhance the system's adaptability and personalization, users can adjust numerous visualization and interaction parameters. For example, users can modify the weights of various features in the representative score calculation to emphasize data characteristics they are more interested in; they can adjust the sampling density to find a balance between performance and detail; and they can set custom focus areas and thresholds to receive targeted reminders and suggestions. The system remembers user preferences and automatically applies them in subsequent sessions, providing a consistent user experience.

[0271] This large-scale data visualization method based on sparse sampling addresses key challenges in medical monitoring data analysis, enabling professionals to efficiently browse and analyze massive amounts of monitoring data within limited display space and processing power. By intelligently filtering the most representative data points, it preserves key features and overall trends while significantly reducing computational and visual burdens. Multi-resolution display and interactive exploration capabilities further enhance the system's flexibility and adaptability, meeting diverse analytical needs.

[0272] Example 15

[0273] like Figure 2 As shown, the present invention also provides an adaptive control system for a rehabilitation exoskeleton based on physiological signals, comprising:

[0274] The data acquisition and preprocessing module 10 is used to acquire multimodal physiological signal data collected by the wearable actuation mechanism, and to perform timestamp alignment, resampling and standardization preprocessing on the multimodal physiological signal data to obtain a synchronized multi-channel sensor signal sequence.

[0275] The motion intention recognition and state classification module 20 is used to extract feature datasets based on the synchronized multi-channel sensor signal sequence, calculate the similarity between the feature datasets and the pre-built motion intention template library, determine the current motion intention category and the quantization value of the intention intensity, determine the control mode type based on the current motion intention category and the quantization value of the intention intensity, and perform physiological state classification to obtain the current physiological state category and state confidence.

[0276] The safety trajectory planning module 30 is used to determine the safety of the preset motion trajectory based on the current physiological state category, the state confidence level and the preset motion trajectory, and to calculate the safety margin assessment value. When the safety margin assessment value is lower than the safety threshold, the preset motion trajectory is adjusted to obtain the adjusted motion trajectory.

[0277] The control command generation module 40 is used to calculate the target driving parameters of the actuator based on the safety margin assessment value, the adjusted motion trajectory, the control mode type and the intent intensity quantification value, and generate real-time motion control commands.

[0278] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive control system for a rehabilitation exoskeleton based on physiological signals, characterized in that, include: The data acquisition and preprocessing module is used to acquire multimodal physiological signal data collected by the wearable actuation mechanism, and to perform timestamp alignment, resampling and standardization preprocessing on the multimodal physiological signal data to obtain a synchronized multi-channel sensor signal sequence. The motion intention recognition and state classification module is used to extract feature datasets based on the synchronized multi-channel sensor signal sequence, calculate the similarity between the feature datasets and the pre-built motion intention template library, determine the current motion intention category and the quantized value of the intention intensity, determine the control mode type based on the current motion intention category and the quantized value of the intention intensity, and perform physiological state classification to obtain the current physiological state category and state confidence. The safe trajectory planning module is used to determine the safety of the preset motion trajectory based on the current physiological state category, the state confidence level, and the preset motion trajectory using a boundary detection algorithm, calculate a safety margin assessment value, and adjust the preset motion trajectory when the safety margin assessment value is lower than a safety threshold to obtain an adjusted motion trajectory. This includes: constructing a multi-dimensional joint space, the dimensions of which include the angles, angular velocities, and torques of each joint degree of freedom; defining dynamic safe and forbidden regions in the multi-dimensional joint space based on the current physiological state category and the state confidence level; performing octree spatial partitioning on the multi-dimensional joint space; establishing a hierarchical index structure to obtain a spatial index database; and obtaining the coordinates of the current joint state point and the direction of the preset motion trajectory. The vector is used to query the spatial index database for the spatial region to which the coordinates of the current joint state point belong, using a point positioning algorithm, to obtain a region type identifier. The region type identifier includes a safe region identifier, a boundary region identifier, and a forbidden region identifier. Based on the region type identifier and the direction vector of the preset motion trajectory, a probe ray is emitted along the direction vector of the preset motion trajectory. The shortest distance from the current position to the forbidden boundary is calculated by intersecting the probe ray with the forbidden region boundary. This shortest distance is used as the safety margin assessment value. When the safety margin assessment value is lower than the safety threshold, the preset motion trajectory is corrected by adjusting the trajectory parameters using a gradient descent method to ensure that the corrected trajectory maintains a safe distance from the forbidden boundary, thus obtaining the adjusted motion trajectory. The control command generation module is used to calculate the target driving parameters of the actuator based on the safety margin assessment value, the adjusted motion trajectory, the control mode type, and the intent intensity quantification value, and generate real-time motion control commands.

2. The system according to claim 1, characterized in that, The data acquisition and preprocessing module includes: The wearable actuator collects multimodal physiological signal data, which includes joint kinematic signals, joint dynamic signals, muscle electrophysiological signals, and peripheral circulatory physiological signals. The joint kinematic signals include joint range of motion and angular velocity, the joint dynamic signals include joint torque and motion resistance, the muscle electrophysiological signals include multichannel surface electromyography signals, and the peripheral circulatory physiological signals include optical volumetric pulse signals. Using the starting point of the motor motion cycle of the wearable actuator as the time reference anchor point, the timestamps of each channel signal in the multimodal physiological signal data are aligned, and signals with different sampling rates are interpolated to a unified time grid to obtain time-synchronized signal data. The signals of each channel in the time-synchronized signal data are subjected to linear detrending processing and z-score normalization processing to eliminate DC components and amplitude differences, thereby obtaining the synchronized multi-channel sensor signal sequence.

3. The system according to claim 2, characterized in that, The motion intent recognition and state classification module, when extracting the feature dataset, includes: The synchronized multi-channel sensor signal sequence is segmented using a sliding window technique, with the window length set to 0.5 to 3 seconds and the overlap rate to 50% to 75%, to obtain time window signal segments. Based on the time window signal segment, the mean, variance, peak value, signal envelope value and inter-channel Pearson correlation coefficient of each channel are calculated to obtain the time domain feature vector; Based on the time window signal segment, perform spectrum analysis on the signal of each channel, calculate the energy ratio, peak frequency position and spectral centroid within the predefined functional frequency band, and obtain the frequency domain feature vector; The angles of each degree of freedom of the joint in the joint kinematic signal, the torque value in the joint dynamic signal, and the electromyographic envelope value in the muscle electrophysiological signal are combined to construct a six-dimensional to ten-dimensional state space vector; The time-domain feature vector, the frequency-domain feature vector, and the state-space vector are concatenated column-wise to obtain the feature dataset.

4. The system according to claim 3, characterized in that, The motion intent recognition and state classification module, when performing the similarity calculation, determining the current motion intent category, and the quantification value of the intent intensity, includes: Standard motion intention templates are extracted from the pre-built motion intention template library. These standard motion intention templates are formed by preprocessing and feature extraction of standard active motion signals collected clinically. The current time window feature vector sequence in the feature dataset is compared one by one with each of the standard motion intention templates in the pre-built motion intention template library. The Euclidean distance, cosine similarity or dynamic time warping distance between the two are calculated to obtain the similarity score for each template. Select the template corresponding to the maximum value from all the similarity scores to obtain the best matching template identifier and the highest similarity score; Based on the motion intent type corresponding to the best matching template identifier, the current motion intent category is obtained; The highest similarity score is normalized and mapped to the interval between 0 and 1 to obtain the quantized value of the intent intensity.

5. The system according to claim 4, characterized in that, When determining the control mode type, the motion intention recognition and state classification module includes: The highest similarity score is compared with a preset effective intent threshold. When the highest similarity score is higher than the preset effective intent threshold, it is determined that a valid active motion intent has been detected; otherwise, it is determined that no valid active motion intent has been detected. When it is determined that no valid active motion intention is detected, the control mode type is set to passive assist mode; When a valid active motion intention is detected and the intensity quantization value of the intention is lower than the first intensity threshold, the control mode type is set to active coordination mode. When it is determined that a valid active motion intention is detected and the quantized value of the intention intensity is higher than the second intensity threshold, the control mode type is set to adaptive hybrid mode; Wherein, the first intensity threshold is less than the second intensity threshold, and when the intent intensity quantization value is between the first intensity threshold and the second intensity threshold, a smooth transition control between the active collaborative mode and the adaptive hybrid mode is adopted.

6. The system according to claim 5, characterized in that, The motion intent recognition and state classification module employs a flexible alignment matching method based on time series similarity metrics when performing the similarity calculation, specifically including: The temporal feature vector sequence in the feature dataset is non-linearly aligned with each standard template sequence in the pre-built motion intent template library. A cumulative distance matrix is ​​constructed using a dynamic programming algorithm, and the optimal alignment path is calculated to obtain the minimum distance metric value that allows time warping. Based on the minimum distance metric, an adaptive threshold determination mechanism is used to determine whether there is a valid pattern match. The adaptive threshold is dynamically adjusted according to the statistical characteristics of the distance distribution within the pre-built motion intent template library to obtain the matching confidence and pattern category identifier. Based on the matching confidence and the pattern category identifier, when the matching confidence exceeds the confirmation threshold, the signal pattern of the corresponding matching event is included as a new sample in the pre-built motion intent template library. The representative template is updated through cluster analysis to achieve incremental learning and evolution optimization of the pre-built motion intent template library.

7. The system according to claim 6, characterized in that, The motion intent recognition and state classification module employs a fast spectrum analysis method based on the sparsity assumption when extracting the frequency domain feature vector, specifically including: The synchronized multi-channel sensor signal sequence is segmented and windowed, with a window length of 256 to 2048 sampling points and an overlap rate of 50%, to obtain a multi-channel time window signal matrix. Based on the multi-channel time window signal matrix, the sparse spectrum recovery algorithm is used to randomly hash and bucket the signals of each channel, calculate the energy amplitude of each frequency component, compare the energy amplitude with an energy threshold, and when the energy amplitude is greater than the energy threshold, the corresponding frequency component is determined to be a significant energy frequency component. Hash collision is used to locate all the significant energy frequency components to obtain a sparse spectrum coefficient set, wherein the energy threshold is set to 10% to 50% of the mean of the energy amplitude of all frequency components. The sparse spectral coefficient set is divided into predefined functional frequency bands, including a low frequency band of 0-5Hz, a mid frequency band of 5-50Hz, and a high frequency band of 50-250Hz. The energy proportion, peak frequency position, and cross-channel coherence index of each frequency band are calculated to obtain the frequency domain feature vector.

8. The system according to claim 7, characterized in that, The motion intent recognition and state classification module, when performing the physiological state classification, obtaining the current physiological state category and the state confidence level, includes: The frequency domain feature vector is extracted from the feature dataset to construct a Gaussian mixture probability model. The pre-stored parameters of the Gaussian mixture probability model are obtained. The parameters of the Gaussian mixture probability model include the mean vector, covariance matrix and mixture weights of each basic state. The basic states include fatigue state, normal state and excitement state. Substitute the frequency domain feature vector into the Gaussian mixture probability model, calculate the Gaussian probability density value of the frequency domain feature vector in each basic state, and calculate the posterior probability of each basic state using Bayes' theorem to obtain the state posterior probability distribution. Based on the state posterior probability distribution, the state with the highest posterior probability is selected as the current physiological state category by the maximum posterior criterion, and the posterior probability value corresponding to the state with the highest posterior probability is used as the state confidence. Based on the current physiological state category, a preset state-parameter mapping table is queried to obtain the motion speed adjustment coefficient, torque limit parameter, and recommended motion amplitude value for the current physiological state category, and personalized control parameter recommendation value is generated.

9. The system according to claim 1, characterized in that, It also includes an anomaly monitoring module, used for real-time anomaly monitoring based on the feature dataset and historical database. When an anomaly is detected, it generates an early warning signal and recommended intervention measures, specifically including: The historical data in the historical database is hierarchically segmented according to multiple time scales such as day, hour, and minute to obtain a multi-level time interval division structure; For each time interval in the multi-level time interval division structure, the signal statistics, abnormal event counts and representative feature samples in the corresponding time interval are calculated to obtain the interval feature summary, and a kd-tree spatial index is constructed based on the interval feature summary. Extract the current time window feature vector from the feature dataset, calculate the deviation from the historical normal pattern, and when the deviation exceeds the anomaly detection threshold, determine that an anomaly has been detected and use the current time window feature vector as the anomaly feature vector. Based on the abnormal feature vector, a k-nearest neighbor query is initiated in the kd-tree spatial index, with k set to a value of 5 to 20, and the set of historical time intervals with the closest feature distance is returned. For each coarse-grained time interval in the set of historical time intervals, the system recursively enters the next level of index for location until a specific historical anomaly record is located. The system then returns the specific historical anomaly record and its corresponding subsequent development information, and generates the warning signal and the recommended intervention measures.

Images