Wearable assistance equipment intention recognition system, recognition method and assistance method

Abstract

The invention provides an intention recognition system, recognition method and assistance method for wearable assistance equipment, and the system comprises a flexible distributed sensing array which is configured to be distributed at a key part of an exoskeleton wearing system and is used for collecting an original sensing signal reflecting the muscle activity of a user; the processor is connected with the flexible distributed sensing array, the processor is configured to extract a multi-dimensional feature vector from the original sensing signal, and the multi-dimensional feature vector comprises at least two features selected from the following groups: an average absolute value, a root mean square, a variance, a slope, a peak value and short-time Fourier transform energy; and determining the action intention of the user based on the multi-dimensional feature vector. According to the method, the muscle deformation signal earlier than the joint macroscopic movement can be captured, so that the action intention of the user is predicted before the actual action of the user occurs, the advanced response of assistance is realized, and the naturalness and fluency of man-machine interaction are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of wearable devices and human-computer interaction technology, specifically relating to an intent recognition system, recognition method, and assistance method for wearable assistive devices. Background Technology

[0002] Exoskeleton robots, as advanced wearable devices, can enhance human function or provide mobility assistance to people with disabilities. One of their core technologies lies in how to accurately and quickly perceive the wearer's movement intentions. In existing technologies, exoskeleton robot systems typically use sensors such as motor encoders, inertial measurement units, or foot pressure pads to acquire the user's movement information. However, these sensors mainly capture "outcome variables" such as joint angles and limb postures after the action has already occurred. Therefore, there is an inherent lag in control response, making it difficult to achieve rapid assistance highly synchronized with the user's intentions, thus affecting the smoothness of human-computer interaction.

[0003] To address the issue of response lag, some approaches attempt to use bioelectrical signals such as surface electromyography (EMG) or electroencephalography (EEG) to predict intent. However, these signals are easily interfered with by environmental factors such as sweat and electromagnetic fields in practical applications, are extremely sensitive to electrode placement, exhibit significant individual variability, and require frequent calibration, resulting in poor robustness and practicality.

[0004] Other solutions also attempt to integrate information from multiple sensors, such as combining inertial measurement units and electromyography (EMG) sensors. However, their signal processing methods are often relatively simple. For example, they directly input the raw or simply processed signals into a classification model for one-step discrimination. This approach struggles to effectively distinguish between the user's "active muscle contractions" and the "passive interaction forces" between the device and the body, and it cannot fundamentally solve the problem of lag in intention prediction. Therefore, existing technologies still have significant shortcomings in terms of the sensing capabilities, predictive capabilities, and signal interpretation depth of wearable assistive devices, limiting the accuracy and safety of motion assistance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a wearable assistive device intention recognition system, recognition method, and assistive method.

[0006] According to the present invention, a wearable assistive device intent recognition system includes: A flexible distributed sensor array is configured to be distributed in key parts of the wearable assistive device to collect raw sensor signals that reflect the user’s muscle activity. A processor, connected to the flexible distributed sensing array, is configured to: Extract a multidimensional feature vector from the original sensing signal, the multidimensional feature vector including at least two features selected from the group consisting of: mean absolute value, root mean square, variance, slope, peak value and short-time Fourier transform energy; and determine the user's action intention based on the multidimensional feature vector.

[0007] In a preferred embodiment, the processor is configured to determine the user's intention in the following manner: The time series of the multidimensional feature vector is input into the prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles. The activation intensity sequence of one or more muscles is input into the judgment model to infer the intention of the action.

[0008] In a preferred embodiment, the processor is further configured to perform dynamic baseline compensation on the original sensing signal to eliminate low-frequency drift before extracting the multidimensional feature vector.

[0009] In a preferred embodiment, the flexible distributed sensor array covers one or more of the upper limb muscle groups, lower limb muscle groups, and trunk muscle groups.

[0010] An exoskeleton intent recognition method according to the present invention includes the following steps: By using a flexible distributed sensor array located in key parts of the wearable assistive device, raw sensor signals reflecting the user's muscle activity are collected. Extract a multidimensional feature vector from the original sensing signal, the multidimensional feature vector including at least two features selected from the group consisting of: mean absolute value, root mean square, variance, slope, peak value, and short-time Fourier transform energy; and determine the user's action intention based on the multidimensional feature vector.

[0011] In a preferred embodiment, the step of determining the user's intention specifically includes: The time series of the multidimensional feature vector is input into the prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles; and the activation intensity sequence of the one or more muscles is input into the decision model to infer the intention of the action.

[0012] In a preferred embodiment, dynamic baseline compensation is performed on the original sensing signal before extracting the multidimensional feature vector.

[0013] According to the present invention, a method for assisting a wearable assistive device includes the following steps: Acquire raw sensor signals; Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; Input the prediction model to estimate muscle activation intensity; The input judgment model infers the action intent; Based on the determined action intent, generate control commands to drive the actuator; The actuator executes control commands.

[0014] In a preferred embodiment, a flexible distributed sensor array arranged in the quadriceps, gluteus maximus and sole of the foot collects minute deformations of the skin surface and changes in sole pressure distribution caused by muscle contraction, and converts them into continuous electrical signals to form the original sensing signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the predictive model determines that the activation intensity of the quadriceps and gluteus maximus shows a continuous increasing trend; When the judgment model detects that the activation intensity of the quadriceps and gluteus maximus exceeds the preset threshold at the same time, and the center of pressure on the sole of the foot tends to move backward, it will judge the movement intention as getting up.

[0015] In a preferred embodiment, a flexible distributed sensor array arranged in the gluteus maximus and quadriceps femoris muscles collects minute deformations on the skin surface caused by muscle contraction and converts them into continuous electrical signals to form the original sensing signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the prediction model determines the activation intensity sequence of the gluteus maximus and quadriceps femoris, it shows a periodic fluctuation consistent with the stride frequency. The judgment model identifies the user's real-time step frequency by analyzing this periodic signal; at the same time, it estimates the user's force intensity in real time by monitoring the overall level of the peak amplitude and average absolute value of the root mean square. Based on the real-time decoded step frequency and force intensity, a periodic assist torque command is generated.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This application employs a flexible distributed sensor array to capture muscle deformation signals that precede macroscopic joint movements, thereby predicting the user's intention to move before the actual action occurs, achieving an early response to assistive devices, and greatly improving the naturalness and fluency of human-computer interaction.

[0017] 2. This application can effectively distinguish between situations where the signal variance is large when the user actively exerts force and situations where the signal variance is small when the external force is passively applied by extracting multi-dimensional feature vectors including variance from the sensing signal, thus solving the problem that traditional pressure sensors cannot distinguish the source of the signal and improving the accuracy of intent recognition.

[0018] 3. This application provides high-density, multi-dimensional sensing data through a large-area distributed sensor array, which can capture subtle changes and spatial distribution characteristics of local muscle groups, enabling the system to have a deeper and more comprehensive understanding of the user's physical condition.

[0019] 4. This application effectively improves the algorithm's noise resistance and adaptability to complex dynamic scenes through multi-feature fusion analysis and a two-step inference model. It is not sensitive to slight deviations in sensor attachment position and has better stability and practicality compared to the easily interfered bioelectric signal scheme. Attached Figure Description

[0020] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the deployment method of the flexible sensor array, which is the main feature of this invention. The sensor array can be worn independently on the user's body or clothing surface (left figure), or it can be embedded in the wearable system to form a local sensor array (right figure). Figure 2 This is a system block diagram illustrating the intent recognition and control method of this invention; Among them, the basic link is to combine the data / data sets collected by the traditional sensing schemes in the system and each group of flexible sensing arrays with the algorithm strategy / model to form a motion control strategy / scheme. Strategy closed loop: Based on the changes in motion strategy / scheme, the front-end sensor data changes accordingly, and the algorithm strategy and model can be further optimized. Through this "data-model-motion" loop, the effect of the algorithm strategy is continuously optimized, the algorithm model is continuously strengthened and trained, and the motion control scheme gradually reaches the optimal state. Figure 3 This is a flowchart illustrating the intent recognition method of the present invention. Figure 4 This is a schematic diagram illustrating the hardware architecture of the intent recognition system, which is the main feature of this invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0022] It is important to emphasize that the wearable assistive devices in this application refer to wearable devices capable of assisting humans or other living beings in specific scenarios. These include exoskeletons, assistive clothing, smart walking aids, and flexible assistive modules. In the following embodiments, exoskeletons will be used as an example for illustration.

[0023] Example 1 like Figures 1 to 4 As shown, a wearable assistive device intention recognition system according to the present invention includes: a flexible distributed sensor array configured to be distributed in key parts of the wearable assistive device for collecting raw sensor signals reflecting the user's muscle activity.

[0024] A processor, connected to a flexible distributed sensor array, is configured to: extract a multidimensional feature vector from raw sensor signals, the multidimensional feature vector including at least two features selected from the group consisting of: mean absolute value, root mean square, variance, slope, peak value, and short-time Fourier transform energy; and determine the user's action intention based on the multidimensional feature vector. The processor is configured to determine the user's action intention by: inputting a time series of the multidimensional feature vector into a prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles; and inputting the activation intensity sequence of one or more muscles into a decision model to infer the action intention.

[0025] Before extracting the multidimensional feature vector, the processor is also configured to perform dynamic baseline compensation on the original sensing signal to eliminate low-frequency drift. The flexible distributed sensing array covers one or more of the upper limb muscle groups, lower limb muscle groups, and trunk muscle groups. It should be noted that the upper limb muscle groups, lower limb muscle groups, and trunk muscle groups all contain multiple small muscle groups, and the flexible distributed sensing array of this application only needs to cover one or more small muscle groups.

[0026] According to the present invention, a method for intention recognition of a wearable assistive device, employing the aforementioned wearable assistive device intention recognition system, includes the following steps: By using a flexible distributed sensor array located in key parts of the wearable assistive device, raw sensor signals reflecting the user's muscle activity are collected. Extract a multidimensional feature vector from the original sensor signal. The multidimensional feature vector includes at least two features selected from the following group: mean absolute value, root mean square, variance, slope, peak value, and short-time Fourier transform energy. Based on the multidimensional feature vector, determine the user's action intention.

[0027] The steps for determining a user's action intent specifically include: inputting the time series of multidimensional feature vectors into a prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles; and inputting the activation intensity sequence of one or more muscles into a decision model to infer the action intent. Dynamic baseline compensation is also performed on the original sensor signal before extracting the multidimensional feature vectors.

[0028] According to a wearable assistive device assistance method provided by the present invention, the above-mentioned wearable assistive device intention recognition system includes the following steps: Collect raw sensor signals.

[0029] Perform dynamic baseline compensation and standardization.

[0030] Calculate multidimensional feature vectors.

[0031] Input the prediction model to estimate muscle activation intensity.

[0032] The input judgment model infers the action intent; Based on the determined action intent, control commands are generated to drive the actuator.

[0033] The actuator executes control commands.

[0034] The predictive model estimates that the activation intensity of the quadriceps and gluteus maximus muscles shows a continuously increasing trend. When the judgment model detects that the activation intensity of the quadriceps and gluteus maximus muscles simultaneously exceeds a preset threshold, and the center of pressure on the foot tends to shift backward, it will interpret the movement intention as getting up.

[0035] The predictive model outputs an activation intensity sequence of the gluteus maximus and quadriceps femoris muscles, which exhibits periodic fluctuations consistent with stride frequency. The decision model analyzes this periodic signal to identify the user's real-time stride frequency; simultaneously, by monitoring the overall level of the root mean square peak amplitude and mean absolute value, it estimates the user's exertion intensity in real time. Based on the real-time decoded stride frequency and exertion intensity, it generates an assist torque command that also exhibits periodicity.

[0036] Example 2: Based on Example 1, such as Figures 1 to 4As shown, this embodiment provides a wearable assistive device intention recognition system and method for daily activity safety assistance scenarios for the elderly and people with reduced mobility. This embodiment aims to solve the technical problem of users easily losing balance or falling during daily activities such as getting up from a seat or going up or down a slope. By accurately recognizing the user's movement intention in advance, the system can provide timely and stable assistance, thereby preventing potential dangers.

[0037] In one embodiment of this application, a schematic diagram of the hardware architecture of an exoskeleton intent recognition system is shown. The system is integrated into a lightweight lower limb exoskeleton robot, which includes: multiple flexible distributed sensor arrays, a flexible encapsulation structure for fixing the sensor arrays, several regional acquisition nodes, a central controller, and actuators disposed at the joints.

[0038] Specifically, the flexible distributed sensor array is the core component of this system for achieving advance intent perception. To comprehensively capture muscle deformation signals closely related to lower limb activities such as standing and walking, this embodiment strategically deploys multiple flexible distributed sensor arrays at key user locations, including the quadriceps femoris on the front of the thigh, the gluteus maximus in the buttocks, the gastrocnemius on the back of the calf, and the sole of the foot, particularly the forefoot and heel areas, for sensing changes in the center of pressure. It is understood that each flexible distributed sensor array consists of multiple independent sensing pixels. These pixels can be implemented based on principles such as piezoresistive, capacitive, piezoelectric, or optical fiber, and can sensitively detect surface deformation or pressure changes caused by muscle contraction or bulging within their coverage area.

[0039] To ensure the sensor array adheres stably and reliably to the user's body surface, this embodiment employs a strap-type flexible encapsulation structure. This structure can be designed as an embedded component, directly integrating the flexible distributed sensor array into the exoskeleton's straps or the lining that contacts the body. This flexible encapsulation structure not only provides fixation, but its own elasticity also ensures a moderate and uniform preload between the sensor array and the skin, which is crucial for acquiring high-quality sensor signals.

[0040] Each flexible distributed sensor array is connected to one or more area acquisition nodes via signal cables. The area acquisition node is typically a small microcontroller unit configured to perform high-speed sampling, pre-amplification, filtering, and analog-to-digital conversion on the raw analog signals output by the connected sensor arrays, and to send the digitized signals with precise timestamps to the central controller via bus technology.

[0041] The central controller, acting as the "brain" of the entire system, is typically positioned at the user's waist to facilitate connection of various components and maintain overall balance. The central controller can be a high-performance embedded computer or a system-on-a-chip, containing a processor and memory. The memory stores pre-trained intent recognition models and control algorithms, which the processor then executes.

[0042] An actuator is a component that provides physical assistance. In the lower limb exoskeleton of this embodiment, the actuator is typically a motor or hydraulic / pneumatic drive unit mounted at the hip or knee joint, which receives control commands from a central controller and outputs precise assist torque.

[0043] Taking the typical action of a user "getting up from the seat" as an example, the working process of intent recognition and power assist control in this embodiment is explained in detail: Flexible distributed sensor arrays deployed in the quadriceps, gluteus maximus, and sole of the foot collect minute deformations of the skin surface and changes in plantar pressure distribution caused by muscle contraction, and convert them into continuous electrical signals to form raw sensor signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the predictive model determines that the activation intensity of the quadriceps and gluteus maximus shows a continuous increasing trend; When the judgment model detects that the activation intensity of the quadriceps and gluteus maximus exceeds the preset threshold at the same time, and the center of pressure on the sole of the foot tends to move backward, it will judge the movement intention as getting up.

[0044] The process begins with step S301: acquiring raw sensor signals. When the user wearing the exoskeleton prepares to get up from the seat, even before any macroscopically visible movement occurs, the quadriceps and gluteus maximus muscles, responsible for knee and hip extension, begin isometric or eccentric contractions to accumulate power. Simultaneously, to stabilize the body's center of gravity, subtle changes occur in the pressure distribution on the soles of the feet; for example, the center of pressure shifts slightly backward. At this time, the individual pixels in the flexible distributed sensor array positioned in the corresponding muscle groups and on the soles of the feet capture these minute deformations of the skin surface and changes in foot pressure distribution caused by muscle contraction in real time, converting them into continuous electrical signals to form the raw sensor signal sequence s_i(t), where i represents the i-th sensor pixel and t represents time.

[0045] Next, the process proceeds to step S302: dynamic baseline compensation and standardization. Since slight body movements during wear or minor changes in strap tightness over time can cause low-frequency baseline drift in the original sensor signal sequence s_i(t), interfering with subsequent feature extraction, the central controller first performs dynamic baseline compensation on the received original sensor signal. Specifically, this is done using the formula s_i'(t) = s_i(t) - b_i(t), where s_i'(t) is the compensated signal, and the dynamic baseline b_i(t) is calculated in real-time by applying an average or median filter over a sliding time window to the original signal s_i(t). This step effectively filters out slowly changing, unintended signal interference. After baseline compensation is completed, the system will also perform Gaussian filtering and other noise reduction processing on the signal s_i'(t), and standardize it (e.g., normalize it to the [0,1] or [-1,1] interval) to obtain a clean and normalized standard signal sequence ˆs(n), which will prepare for subsequent feature extraction.

[0046] Then, the system executes the crucial step S303: calculating the multidimensional feature vector f_i. To deeply extract physical information related to motion intent from the signal, the central controller does not directly use the preprocessed signal, but instead calculates a multidimensional feature vector f_i for each pixel's standard signal sequence ˆs(n) within a preset time window w (e.g., 100 milliseconds). This vector contains multiple features with explicit physical meaning; in this embodiment, these features include at least: The average absolute value, which reflects the average intensity of the signal, is very sensitive to the sustained and gradual exertion of muscle force (such as the preparatory phase before standing up).

[0047] The root mean square of the signal energy is directly related to the amplitude of muscle contraction, i.e. the degree of muscle bulging, and this value increases significantly during the power exertion phase of standing up.

[0048] The variance reflects the degree of dispersion or volatility of the signal. This feature can be used to effectively distinguish between active muscle contraction (usually accompanied by rapid recruitment and tremor of muscle fibers, resulting in a large signal variance) and passive external pressure (usually smooth, with a small signal variance).

[0049] The slope obtained by calculating the rate of change of the signal within a time window is used to determine the speed of the movement. For movements requiring explosive force, such as standing up, the signal slope will be large at the initial stage of force exertion. The peak value, reflecting the maximum value of the signal within the window, is used to detect the location of the largest local protrusion of the muscle, which helps to locate the core force exertion area. Furthermore, the short-time Fourier transform energy of the signal is calculated by performing a short-time Fourier transform on the signal within a specific frequency band (e.g., the 5-50 Hz band related to muscle physiological activity). This helps to filter out high-frequency noise or friction signals unrelated to muscle activity. By calculating these features, the system transforms the one-dimensional time-series signal into a more informative and higher-dimensional feature vector.

[0050] The process then proceeds to a two-step intent inference phase. The first step, S304, involves inputting a prediction model h() to estimate the muscle activation intensity A_m(t). The central controller takes a sequence of feature vectors f_i(tT:t) of one or more relevant pixels over a continuous time period (e.g., the past 200 milliseconds) as input to a pre-trained prediction model h(). This model h() can be a time-series model, such as a Long Short-Term Memory network or a temporal convolutional network, which excels at learning dynamic patterns from time-series data. The output of model h() is the normalized activation intensity A_m(t) of the corresponding muscle (e.g., the quadriceps). This activation intensity value is understood to be a quantitative estimate of the current level of muscle exertion. During the standing up process, the system observes a continuous increasing trend in the activation intensity A_m(t) of the quadriceps and gluteus maximus.

[0051] In one feasible implementation: the prediction model h() is a time series regression model based on long short-term memory networks.

[0052] Training steps: 1. Training data collection In the experimental setting, subjects wore the exoskeleton system described in this application, and simultaneously collected: raw signals from a flexible distributed sensor array (for feature extraction) and muscle activation intensity as a reference label.

[0053] The reference labels were obtained by rectifying, filtering, and normalizing the surface electromyography (EMG) signals to obtain the muscle activation intensity.

[0054] 2. Feature Sequence Construction Dynamic baseline compensation and standardization are performed on the flexible sensor array signal. Multidimensional feature vectors are extracted within the sliding time window. The feature vectors of multiple consecutive time windows are concatenated to form the time series input f_i(tT:t).

[0055] 3. Supervise learning and training Using time series features as input and reference muscle activation intensity at corresponding time points, the mean squared error is used as the loss function for training.

[0056] 4. Model output normalization The muscle activation intensity output by the model is constrained and normalized to the [0,1] interval, and this output value is used as an estimate of the activation intensity of the corresponding muscle at the current time.

[0057] Next is step S305: Input the judgment model F(·) to infer the action intention Action(t). The central controller combines the activation intensity sequences from multiple key muscle groups (quadriceps, gluteus maximus, gastrocnemius, etc.) and the pressure center location information extracted from the plantar sensor array into a high-dimensional multi-dimensional action feature vector M(t) = [A_quadriceps(t), A_gluteus maximus(t), ..., P_plantar center(t)]. Then, this vector M(t), which integrates the states of multiple body parts, is input into the judgment model F(·). The model F(·) can be a classifier, such as a support vector machine, decision tree, or a small feedforward neural network. Its role is to infer the user's current macroscopic action intention Action(t) based on the overall muscle activation pattern input. When the model F(·) detects that the activation intensity of the quadriceps and gluteus maximus simultaneously exceeds a preset threshold, and the plantar pressure center has a tendency to move backward, it will determine the action intention Action(t) as "get up". According to experimental data, this recognition process can be completed 40 to 60 milliseconds earlier than the macroscopic movement of the user's torso actually lifting off the seat.

[0058] In one feasible implementation: the decision model is a small feedforward neural network classification model.

[0059] Training steps: 1. Action Sample Collection and Labeling In the experimental environment, the subjects wore the exoskeleton system described in this application, collected raw signals from the flexible distributed sensor array, and performed a variety of standardized actions. Each data segment was labeled with its corresponding macroscopic action intention (standing up, walking, squatting, running, standing still) by manual or semi-automatic means.

[0060] 2. Feature Sample Construction The trained model h() is used to convert the sensor feature sequence into an activation intensity sequence of multiple muscles, and the plantar pressure center location information is extracted at each time point to construct the action feature vector M(t).

[0061] 3. Supervised Categorized Training The action feature vector M(t) is used as input, and the action intention category Action(t) is used as the supervision label. The cross-entropy loss function is used for classification training.

[0062] 4. Model Deployment The trained model F(·) is deployed to the central controller, and during real-time operation, the action intent determination result is output according to the current multi-muscle group activation mode.

[0063] Finally, the system executes step S306: calculating and outputting the assist torque τ. Once the central controller recognizes the "intention to stand up," the assist control module is immediately triggered. This module calculates the desired assist torque τ based on the recognized intention (standing up), the real-time activation intensity A_m(t) of each muscle, and the preset individualized parameters in memory. The calculation follows a function τ=g(P,biomech_params,safety_limits), where P is the confidence level of the intention, biomech_params includes biomechanical parameters such as the user's weight and leg length, and safety_limits defines safety boundaries such as the maximum assist torque and the fastest torque increase rate to ensure a smooth and safe assist process. The calculated assist torque command is sent to the actuators in the hip and knee joints, driving them to smoothly output supporting force, helping the user overcome initial inertia and gravity, and smoothly and safely complete the standing action, thereby effectively avoiding dizziness and imbalance caused by insufficient or excessive force exertion.

[0064] The entire system forms a closed loop. The actions of the actuators and changes in the user's physical state are again captured by the flexible distributed sensor array, forming feedback. Accordingly, the system can use this closed-loop data from sensor execution to fine-tune and optimize the prediction model h() and the decision model F(·) online, making them more adaptable to the habits and characteristics of specific users.

[0065] It should be noted that the sensor data transmission methods of the technical solution in this application include wired transmission and / or wireless transmission.

[0066] Example 3: Based on Embodiment 1, this embodiment demonstrates the application of the system and method provided in this application in complex industrial scenarios, particularly in high-altitude operations. In such scenarios, workers typically need to carry heavy loads and perform large-amplitude, high-risk movements such as climbing, bending over, and raising their arms. This embodiment aims to provide workers with dynamic load-bearing assistance and real-time balance protection through precise intent recognition, thereby reducing muscle strain and preventing falls from heights due to imbalance.

[0067] In this embodiment, the intent recognition system of this application is integrated into a full-body industrial exoskeleton robot covering the upper and lower limbs and torso. Its flexible distributed sensor array has a wider deployment range, covering multiple key muscle groups such as the deltoid muscles of the shoulder, the erector spinae muscles of the lower back, the hips, and the quadriceps and hamstrings of the thighs through an embedded flexible encapsulation structure (e.g., integrated into the exoskeleton's lining fabric). Actuators are distributed across multiple joints such as the shoulder, elbow, waist, hip, and knee to provide full-body assistance. The central controller is located in the back or waist area.

[0068] Taking the specific action of "climbing scaffolding" as an example, the workflow of this embodiment is described as follows.

[0069] First, in steps S301 and S302, when a worker prepares to lift one leg onto a higher level of scaffolding, the quadriceps, gluteus maximus, and core muscles of the supporting leg contract to stabilize the body, while the hip flexors and quadriceps of the lifting leg exert force to lift the lower limb. These rapid muscle contractions and deformations are captured in real-time by a flexible distributed sensor array at the corresponding location and preprocessed by the central controller, including dynamic baseline compensation and normalization.

[0070] In step S303, the central controller performs multi-dimensional feature extraction on the preprocessed signal. In rapid, explosive movements like climbing, the "slope" feature is particularly important. At the moment the worker exerts force and lifts their leg, the slope value of the corresponding muscle sensor signal increases significantly, far exceeding the level during stillness or slow movement. Simultaneously, the "root mean square" value exhibits a periodic, pulse-like increase with the rhythm of climbing, while the "average absolute value" reflects the average muscle tension throughout the climbing process. Furthermore, the "variance" feature plays a crucial role in distinguishing between active exertion and passive imbalance in this embodiment. During active climbing, muscles contract in a coordinated manner, resulting in a relatively regular signal pattern; however, if the body unexpectedly sways or tilts, stabilizing muscle groups such as the lower back will perform emergency, irregular, tense contractions to maintain balance, at which point their corresponding variance value will abnormally increase.

[0071] In steps S304 and S305, the system performs a two-step intent inference. The prediction model h() (e.g., a temporal convolutional network) estimates the activation intensity sequence A_m(t) of multiple muscles, including the deltoid, erector spinae, gluteus maximus, and quadriceps, in real time based on the input feature vector sequence. Subsequently, the decision model F(·) (e.g., a hidden Markov model) receives these multidimensional muscle activation intensity sequences. When model F(·) identifies a synergistic pattern of increased activation intensity in the supporting leg's gluteus maximus and quadriceps, coupled with a rapid increase in activation intensity in the lifting leg's hip flexors (reflected by high slope features), it infers that the user's action intent is "climbing force."

[0072] In step S306, once the intention to "climb and exert force" is detected, the assist control module immediately calculates and issues control commands. On one hand, it drives the hip and knee joint actuators of the lower limb exoskeleton to synchronously provide support and lifting torque, effectively distributing the load on the worker's legs. On the other hand, the system continuously monitors the activation intensity and variance of the erector spinae muscles in the lower back. If the system detects an abnormal increase in the variance of the erector spinae muscles during the non-active force exertion phase, this usually indicates that the torso is becoming unbalanced. At this time, the assist control module enters a "balance protection" mode, which may immediately adjust the assist distribution of the leg actuators to help restore the center of gravity, or lock certain joints to provide rigid support, while simultaneously issuing an alarm through vibration or sound to remind the worker to pay attention to their posture.

[0073] Through this intelligent closed-loop control of "intention prediction - precise assistance - balanced protection", the system in this embodiment can not only effectively reduce the physical exertion and muscle strain of workers, but also significantly improve the safety of operations in dangerous environments such as high altitudes.

[0074] Example 4: Based on Embodiment 1, this embodiment applies the technical solution of this application to the professional strength training scenario of athletes, aiming to provide an analysis and feedback tool for quantifying muscle exertion patterns and performing real-time movement correction. Unlike the previous two embodiments, the system in this embodiment may not include an actuator with assistive function, but mainly serves as a high-precision data acquisition, analysis, and feedback device. Its core lies in using the perception and intent recognition methods of this application to deeply analyze the athlete's movement techniques.

[0075] In this embodiment, for specific training exercises, such as the squat movement in weightlifting or powerlifting, the system employs a flexible, layered encapsulation structure. This structure, similar to a compression garment or protective gear, allows a flexible, distributed sensor array to precisely and tightly cover the athlete's target muscle groups. For squat training, the sensor array will focus on covering the erector spinae, gluteus maximus, quadriceps, and core abdominal muscles of the lower back. The central controller can be an external computer or tablet that wirelessly connects to the athlete's worn device for data processing and results display.

[0076] Taking the movement correction in "squat training" as an example, the working process of this embodiment is described as follows.

[0077] A proper squat requires athletes to keep their core muscles engaged and back straight throughout the squatting and standing motion, primarily relying on the gluteus maximus and quadriceps for power. A common mistake is "compensation by bending forward," where, during the standing phase, due to insufficient gluteal and leg strength or incorrect force application, the hips extend prematurely, and the upper body is "pulled" up using the strength of the lower back. This incorrect movement not only reduces training effectiveness but also greatly increases the risk of lumbar spine injury.

[0078] Before training begins, coaches can store a "standard squat model" in the system. This model defines the ideal temporal relationship and amplitude range of activation intensity of key muscle groups (erector spinae, gluteus maximus, quadriceps femoris) under standard movement.

[0079] During training, athletes perform squatting movements. Steps S301 to S304 are similar to those in the previous embodiment: a flexible distributed sensor array covering each muscle group monitors muscle deformation in real time, and the central controller accurately calculates the real-time activation intensity sequence A_m(t) of each muscle group through preprocessing, feature extraction (such as mean absolute value, root mean square, variance, etc.) and prediction model h().

[0080] The core of this embodiment lies in the variant application of step S305. Here, the system does not infer a macroscopic movement intention, but rather compares the real-time multi-muscle group activation sequence pattern with a preset "standard squat model." For example, the standard model might specify that during the standing phase, the activation intensity of the gluteus maximus and quadriceps should reach its peak first, while the activation intensity of the erector spinae should remain at a lower level to maintain stability.

[0081] If the system detects during the comparison that the average absolute value or root mean square value of the erector spinae muscles rises prematurely and abnormally rapidly in the early stages of the standing phase, even exceeding the activation level of the gluteus maximus, the system will determine that the athlete has made the erroneous movement of "bending over to compensate".

[0082] Once an incorrect movement is detected, the system immediately triggers a feedback mechanism (replacing the assist control in step S306). For example, it may provide a vibration alert via a wristband connected to the system, or a voice prompt via Bluetooth headphones: "Core engaged, leg power, back straight!" This instant feedback helps athletes adjust and correct their movements immediately during the exercise.

[0083] After training, the system can generate a detailed quantitative analysis report. The report can include charts that visually display the time-series curves of activation intensity for each muscle group during each squat, the contribution ratio of the main muscle groups, and the force coordination score. This objective and precise data, based on muscle levels, is difficult to obtain through traditional video analysis or coach's visual observation, providing athletes and coaches with a scientific basis to optimize training programs and effectively prevent sports injuries.

[0084] Example 5: Based on Embodiment 1, this embodiment describes the application of the system and method provided in this application in periodic and rhythmic sports such as running and cycling. The aim is to provide enhanced assistance that is perfectly synchronized with the user's exercise rhythm and intensity, thereby improving athletic performance and delaying fatigue. Its core principle lies in utilizing the feature extraction method of this application to accurately decode the user's movement frequency and force intensity from rhythmic muscle activity signals, thereby achieving adaptive synchronized assistance.

[0085] A flexible distributed sensor array deployed in the gluteus maximus and quadriceps femoris muscles collects minute deformations on the skin surface caused by muscle contraction and converts them into continuous electrical signals to form the original sensing signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the prediction model determines the activation intensity sequence of the gluteus maximus and quadriceps femoris, it shows a periodic fluctuation consistent with the stride frequency. The judgment model identifies the user's real-time step frequency by analyzing this periodic signal; at the same time, it estimates the user's force intensity in real time by monitoring the overall level of the peak amplitude and average absolute value of the root mean square. Based on the real-time decoded step frequency and force intensity, a periodic assist torque command is generated.

[0086] As an alternative implementation, the system of this application is integrated onto a lightweight lower limb motor exoskeleton. A flexible, distributed sensor array is primarily positioned on the gluteus maximus and quadriceps muscles, which contribute the most to running or cycling, via a strap-like flexible encapsulation structure. Actuators are located at the hip and knee joints to provide auxiliary torque.

[0087] Taking "running" as an example, the workflow of this embodiment can be described in detail as follows.

[0088] When a user starts running, their gluteus maximus and quadriceps muscles exhibit periodic, explosive contractions and relaxations. In steps S301 and S302, the sensor array captures these rhythmic muscle deformation signals, which are then acquired and preprocessed by the central controller.

[0089] In the feature extraction stage of step S303, the multidimensional feature vector calculated by the central controller exhibits distinct periodic characteristics. Specifically, the root mean square (RMS) value increases sharply in a pulse-like manner with each push-off, and decreases again during the relaxation phase of the swing leg; the frequency of this pulse directly corresponds to the user's stride frequency, while its peak amplitude is positively correlated with the user's running intensity (i.e., the magnitude of the force exerted with each push-off). Simultaneously, the average absolute value calculated over a slightly longer time window reflects the average muscle tension and overall exercise intensity throughout the run; for example, the overall level of this value increases significantly from jogging to sprinting. Furthermore, the slope during start-up or sudden acceleration serves as a powerful indicator of "acceleration intention," as the upward slope of the RMS signal is very steep at these times.

[0090] In the intent inference phase of steps S304 and S305, the system understands the user's movement state by analyzing the dynamic changes of the aforementioned features. The prediction model h() outputs the activation intensity sequence of the gluteus maximus and quadriceps femoris muscles, which exhibits periodic fluctuations consistent with the stride frequency. The judgment model F(·) then analyzes this periodic signal (e.g., by extracting the dominant frequency through Fourier transform or autocorrelation analysis) to accurately identify the user's real-time stride frequency. Simultaneously, by monitoring the overall level of the peak amplitude and mean absolute value of the root mean square, model F(·) can estimate the user's force intensity in real time. Combining this information, the system can not only identify the macroscopic intent of "running" but also decode the two key parameters of running, "rhythm" and "intensity," in real time.

[0091] In the assist control phase of step S306, the assist control module no longer responds to one-off movements but enters a "rhythm synchronization" mode. Based on the real-time decoded stride frequency and force intensity, it generates periodic assist torque commands. Specifically, each time a push-off force is detected (i.e., a rapid increase in the root mean square value), the central controller instructs the actuators of the hip and knee joints to output an assist torque pulse that matches the user's own force intensity. The timing, amplitude, and duration of the assist are highly synchronized with the user's own muscle exertion pattern.

[0092] Understandably, this adaptive synchronous assistance works like adding an extra set of "external muscles" to the user, effectively sharing the energy consumption of the primary muscle groups without interfering with or disrupting the user's own movement patterns. This not only helps users run faster and farther but also significantly delays the onset of muscle fatigue and improves endurance performance. When the user decelerates or stops, the system detects a decrease in the amplitude and frequency of the root mean square pulse and automatically and smoothly reduces or stops the assistance, achieving seamless and natural collaboration between the user and the machine.

[0093] The above descriptions are merely some preferred embodiments of this application and are not intended to limit the scope of this application. Those skilled in the art can make various modifications and variations within the spirit and principles of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0094] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0095] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0096] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A wearable assistive device intent recognition system, characterized in that, include: A flexible distributed sensor array is configured to be distributed in key parts of the wearable assistive device to collect raw sensor signals that reflect the user’s muscle activity. A processor, connected to the flexible distributed sensing array, is configured to: Extract a multidimensional feature vector from the original sensing signal, the multidimensional feature vector including at least two features selected from the group consisting of: mean absolute value, root mean square, variance, slope, peak value and short-time Fourier transform energy; and determine the user's action intention based on the multidimensional feature vector.

2. The wearable assistive device intent recognition system according to claim 1, characterized in that, The processor is configured to determine the user's intent in the following ways: The time series of the multidimensional feature vector is input into the prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles. The activation intensity sequence of one or more muscles is input into the judgment model to infer the intention of the action.

3. The wearable assistive device intent recognition system according to claim 1, characterized in that, Before extracting the multidimensional feature vector, the processor is also configured to perform dynamic baseline compensation on the original sensing signal to eliminate low-frequency drift.

4. The wearable assistive device intent recognition system according to claim 1, characterized in that, The flexible distributed sensor array covers one or more of the following areas: upper limb muscle groups, lower limb muscle groups, and trunk muscle groups.

5. A method for intent recognition in a wearable assistive device, characterized in that, The wearable assistive device intent recognition system according to any one of claims 1 to 4 includes the following steps: By using a flexible distributed sensor array located in key parts of the wearable assistive device, raw sensor signals reflecting the user's muscle activity are collected. Extract a multidimensional feature vector from the original sensing signal, the multidimensional feature vector including at least two features selected from the group consisting of: mean absolute value, root mean square, variance, slope, peak value, and short-time Fourier transform energy; and determine the user's action intention based on the multidimensional feature vector.

6. The wearable assistive device intention recognition method according to claim 5, characterized in that, The specific steps to determine a user's intent include: The time series of the multidimensional feature vector is input into the prediction model to estimate the activation intensity of one or more muscles, forming a corresponding activation intensity sequence of one or more muscles; and the activation intensity sequence of the one or more muscles is input into the decision model to infer the intention of the action.

7. The wearable assistive device intent recognition method according to claim 5 or 6, characterized in that, Before extracting the multidimensional feature vector, dynamic baseline compensation is also performed on the original sensing signal.

8. A method for assisting with a wearable assistive device, characterized in that, The wearable assistive device intention recognition system according to any one of claims 1 to 4, the assistive method includes the following steps: Acquire raw sensor signals; Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; Input the prediction model to estimate muscle activation intensity; The input judgment model infers the action intent; Based on the determined action intent, control commands are generated to drive the actuator; The actuator executes control commands.

9. The wearable assistive device assistance method according to claim 8, characterized in that, Flexible distributed sensor arrays deployed in the quadriceps, gluteus maximus, and sole of the foot collect minute deformations of the skin surface and changes in plantar pressure distribution caused by muscle contraction, and convert them into continuous electrical signals to form raw sensor signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the predictive model determines that the activation intensity of the quadriceps and gluteus maximus shows a continuous increasing trend; When the judgment model detects that the activation intensity of the quadriceps and gluteus maximus exceeds the preset threshold at the same time, and the center of pressure on the sole of the foot tends to move backward, it will judge the movement intention as getting up.

10. The wearable assistive device assistance method according to claim 8, characterized in that, A flexible distributed sensor array deployed in the gluteus maximus and quadriceps femoris muscles collects minute deformations on the skin surface caused by muscle contraction and converts them into continuous electrical signals to form the original sensing signals. Perform dynamic baseline compensation and standardization; Calculate multidimensional feature vectors; If the prediction model determines the activation intensity sequence of the gluteus maximus and quadriceps femoris, it shows a periodic fluctuation consistent with the stride frequency. The judgment model identifies the user's real-time step frequency by analyzing this periodic signal; at the same time, it estimates the user's force intensity in real time by monitoring the overall level of the peak amplitude and average absolute value of the root mean square. Based on the real-time decoded step frequency and force intensity, a periodic assist torque command is generated.

Images