A habit adaptive assistance and multi-dimensional evaluation method for distribution network live-line work-oriented upper limb exoskeleton

Abstract

The application discloses a kind of upper limb exoskeleton habit adaptive assistance and multidimensional evaluation method for distribution network live working, belongs to exoskeleton robot control technical field, comprising: constructing the offline expert benchmark assistance mapping model based on behavior cloning network;Design online personal habit adaptive mechanism, according to the deviation degree of habit center of posture and smooth adjustment benchmark assistance generation adaptive assistance;Construct closed-loop fatigue prediction model based on surface electromyography and gated recurrent unit, introduce monotone smooth constraint fatigue gain coefficient dynamic compensation generation total control torque;Construct multidimensional evaluation system including single muscle activation, multi-muscle coordination and respiratory metabolism.The application realizes the dynamic assistance and fatigue dynamic compensation of individual habit, improves the smoothness of human-computer interaction and operation safety.

Description

Technical Field

[0001] This invention belongs to the field of exoskeleton robot control technology, and in particular relates to an upper limb exoskeleton habit-adaptive assistance and multi-dimensional evaluation method for live-line work in power distribution networks. Background Technology

[0002] As a multidisciplinary integrated intelligent system, exoskeleton robots have been widely used in medical rehabilitation, industrial production, and special operations. In the typical high-risk, high-intensity scenario of live-line work in power distribution networks, operators typically perform complex procedures such as voltage testing, installing insulating shields, wearing insulating gloves, and operating insulating rods within the insulated bucket of a truck, requiring frequent upper limb lifting and delicate operations in a confined space. Introducing upper-limb exoskeletons to provide assistance to operators can effectively reduce muscle load, delay fatigue accumulation, and improve operational stability and safety. Existing industrial upper-limb exoskeleton control strategies mostly employ constant torque output or simple models based on gravity compensation, which can achieve physical load reduction to some extent and have shown positive application effects in some structured scenarios.

[0003] However, existing control strategies still have significant shortcomings when directly applied to live-line work in power distribution networks. First, different operators have individualized habitual working postures and force application patterns. Traditional constant torque or simple gravity compensation methods lack adaptability to individual differences in habits. When the operator deviates from the preset posture, the exoskeleton can easily generate unexpected resistance, disrupting the smoothness of human-machine movement. Second, during prolonged lifting tasks, muscle fatigue exhibits dynamic cumulative characteristics. Existing strategies output a fixed auxiliary torque, which cannot dynamically compensate for the operator's time-varying fatigue, leading to a significant increase in subjective effort as the work time lengthens. Finally, existing assessment methods are mostly limited to simple torque output or single muscle activation analysis, failing to encompass the inherent multi-muscle synergistic neural control structure of the human body and lacking in-depth consideration of whether the assistance strategy disrupts the original neuromuscular coordination. If the exoskeleton alters the operator's muscle coordination pattern while providing assistance, it may cause physiological stiffness, increasing the risk of vibration at the end of the insulating rod, operational errors, and even electric shock, making it difficult to guarantee motion transparency and operational safety in high-voltage live-line working environments. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an upper limb exoskeleton habit-adaptive assistance and multi-dimensional evaluation method for live-line work in power distribution networks, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides an upper limb exoskeleton habit-adaptive assistance and multi-dimensional assessment method for live-line work in power distribution networks, comprising: S1: Construct an offline expert policy baseline assist mapping model based on a behavior cloning network, learn the mapping relationship between joint motion states and nonlinear baseline assist from expert teaching trajectories, and obtain the baseline assist; S2: Design an online personal habit adaptive mechanism to estimate the operator's steady-state habit center in real time during the operation, and dynamically adjust the benchmark assist according to the deviation between the current posture and the habit center through a smooth transition function to generate adaptive assist; S3: Construct a closed-loop physiological fatigue prediction model based on surface electromyography signals and gated circulatory unit network to predict muscle fatigue trends in real time and obtain predicted fatigue index values. S4: Perform dynamic fatigue compensation, introduce monotonic smoothness constraints to calculate fatigue gain coefficient, determine fatigue compensation torque based on the predicted fatigue index value, and superimpose the adaptive assistance with the fatigue compensation torque to generate the total control torque of the exoskeleton; wherein, the total control torque is used to provide the operator with comfortable and effortless assistance that adapts to the current task and personal habits in the live-line work scenario; S5: Construct a multi-dimensional neurophysiological comprehensive assessment system to quantitatively evaluate the control efficacy generated by S1 to S4 through single muscle activation assessment, multi-muscle synergy assessment and respiratory metabolism assessment, reflecting the actual assistive effect of the exoskeleton from the local muscle level to the whole body metabolic level.

[0006] Preferably, step S1 specifically includes: collecting the motion state vector and corresponding assist torque during the expert teaching process as supervision labels, constructing the input sequence using a sliding window, flattening the input and standardizing it, and using a multilayer perceptron network to realize regression prediction from motion state to assist torque.

[0007] Preferably, step S2 specifically includes: determining attitude stability by calculating the standard deviation of joint angles within a sliding window; extracting a state stability vector when the stability condition is met; updating the habit center using an exponential moving average; calculating the normalized offset distance of the current attitude relative to the habit center; constructing a piecewise quadratic continuous decay smooth transition function based on the offset distance using soft and hard thresholds; and applying the smooth transition function to the reference assist to obtain adaptive assist.

[0008] Preferably, step S3 specifically includes: acquiring multi-channel surface electromyography (EMG) signals, calculating the root mean square (RMS) after filtering and rectification, and performing weighted summation of the multi-channel RMS signals; normalizing the weighted values ​​using the maximum voluntary contraction test reference value, and calculating the fatigue index; aligning the exoskeleton kinematic data with the surface EMG signals using timestamps, constructing an input sequence window, using a gated recurrent unit network to model the temporal dependence of fatigue, and outputting continuous fatigue index prediction values.

[0009] Preferably, step S4 specifically includes: setting an upper limit for the fatigue gain coefficient constrained by the human biomechanical attenuation law; dynamically modulating the adaptive assist according to the fatigue index prediction value to generate the total output torque of the exoskeleton; the calculation formula for the dynamic modulation is: total output torque = adaptive assist × (1 + fatigue gain coefficient × fatigue index prediction value), wherein the fatigue gain coefficient changes monotonically and smoothly between zero and the upper limit.

[0010] Preferably, the single muscle activation assessment includes: calculating the root mean square amplitude of the surface electromyography signal after filtering, full-wave rectification and normalization, and assessing the local load reduction effect of the exoskeleton on the target force-generating muscle group.

[0011] Preferably, the multi-muscle coordination assessment includes: performing non-negative matrix decomposition on the electromyographic envelope signal matrix to obtain a coordination weight matrix and a coordination activation coefficient matrix; and using a cosine similarity algorithm to calculate the similarity of coordination vectors before and after wearing the exoskeleton, so as to determine whether the exoskeleton increases the burden on the central nervous system and changes the muscle coordination force exertion mode.

[0012] Preferably, the respiratory metabolic assessment includes: collecting oxygen consumption rate and carbon dioxide excretion rate, and calculating the human energy consumption rate using a modified Weir formula.

[0013] Preferably, the multi-muscle synergy assessment further includes: determining the optimal number of synergies by calculating the variance explained rate, adjusting the number of synergies from the minimum value to the number of muscles, and selecting the minimum number of synergies with a global variance explained rate greater than or equal to a preset threshold.

[0014] In a second aspect, the present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs an offline expert strategy benchmark assistance mapping model based on a behavior cloning network in step S1. It learns the mapping relationship between joint motion state and nonlinear benchmark assistance from the expert teaching trajectory, so that the benchmark assistance output by the exoskeleton can accurately reproduce the typical force exertion pattern of the expert in unstructured work scenarios, and provides a reference benchmark that conforms to the mechanical nature of the operation task for subsequent adaptive adjustment.

[0016] This invention designs an online personal habit adaptive mechanism through step S2, which estimates the operator's steady-state habit center in real time and dynamically adjusts the baseline assistance based on the deviation of the current posture from the habit center using a smooth transition function to generate adaptive assistance. This mechanism ensures that the operator receives sufficient assistance when in a habitual posture, and the assistance automatically and smoothly weakens when the operator actively deviates from the habitual posture to make transitional movements, thereby eliminating unexpected resistance and improving the compliance and fluency of human-computer physical interaction.

[0017] This invention constructs a closed-loop physiological fatigue prediction model based on surface electromyography signals and a gated recurrent unit network in step S3 to predict muscle fatigue trends in real time. Combined with the fatigue gain coefficient with monotonic smoothness constraints introduced in step S4, the adaptive assistance and fatigue gain coefficient are integrated to generate the total control torque of the exoskeleton. This dynamic fatigue compensation mechanism can automatically adjust the output torque as muscle fatigue accumulates over time, compensating for the decrease in human force due to fatigue within safe limits, while avoiding excessive assistance and effectively delaying the muscle fatigue process during long-term work.

[0018] This invention constructs a multi-dimensional neurophysiological comprehensive assessment system through step S5, which quantitatively evaluates control efficacy from three dimensions: single muscle activation, multi-muscle synergy, and respiratory metabolism. This system can not only verify the local load reduction effect of the exoskeleton on the target muscle group, but also ensure that the assistance strategy does not damage the original neuromuscular control structure of the human body through muscle synergy analysis, and comprehensively quantify the assistance efficacy by combining whole-body metabolic energy consumption indicators, providing an objective evaluation basis for the operational safety in high-voltage live-line working environments. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a habit-adaptive exoskeleton control framework based on multi-scenario on-site expert demonstrations according to an embodiment of the present invention; wherein, (a) is multi-scenario on-site data acquisition, (b) is offline BC network pre-training, and (c) is an online real-time inference architecture based on intent triggering; Figure 2 This is a schematic diagram of the multi-dimensional neurophysiological comprehensive assessment method according to an embodiment of the present invention; wherein, (a) is the sensor layout for 7-channel surface electromyography (sEMG) and portable respiratory metabolic data acquisition, and (b) is the comprehensive data processing flow and assessment indicators, covering single muscle activation, multi-muscle synergy and whole body energy consumption; Figure 3 This is a schematic diagram of the angle-torque distribution characteristics learned by the behavioral cloning network in a complete live-line working process according to an embodiment of the present invention. Figure 4This is a schematic diagram illustrating the dynamic response of the online habit-adaptive power assistance strategy in continuous unstructured motion according to an embodiment of the present invention, wherein (a) represents the actual output torque. Compared with the reference torque (a) Relationship; (b) is the pose offset distance and smooth transition function (c) represents the change; (c) represents the steady-state time. Accumulation and habit center The update mechanism; Figure 5 The following is a dynamic evaluation of the exoskeleton fatigue compensation strategy in this embodiment of the invention, wherein (a) is the tracking and reconstruction performance of the GRU network for smooth fatigue state labels in offline testing; and (b) is the actual torque response during online human-computer interaction.

[0020] Figure 6 This is a schematic diagram comparing the root mean square (RMS) amplitude of surface electromyography (sEMG) under different experimental conditions in this embodiment of the invention; wherein, (a) is the average RMS amplitude of the seven-channel muscle; and (b) is the sliding RMS curve of the time change during the 4-minute insulating rod operation task. Figure 7 This is a schematic diagram illustrating the variance explanation rate under different numbers of cooperation in the operation task of the insulating rod according to an embodiment of the present invention; Figure 8 This invention relates to the synergistic action of muscles during the operation of an insulating rod under conditions C0, C1, and C2, as described in an embodiment of the invention. (a) represents the first synergy under condition C0; (b) represents the first synergy under condition C1; (c) represents the first synergy under condition C2; (d) represents the second synergy under condition C0; (e) represents the second synergy under condition C1; and (f) represents the second synergy under condition C2. Figure 9 This is a schematic diagram of the cosine similarity between the muscle coordination vectors of the exoskeleton assisted condition and the baseline non-assisted condition in an embodiment of the present invention. Figure 10 This is a schematic diagram comparing the whole-body respiratory metabolic load under three experimental conditions (C0, C1, and C2) in an embodiment of the present invention; where (a) is the energy consumption rate and (b) is the oxygen consumption rate. Figure 11 This is a schematic diagram of the cosine similarity of muscle synergy under three short-time tasks with and without exoskeleton assistance, according to an embodiment of the present invention; wherein, (a) is voltage testing; (b) is operation with insulated gloves; and (c) is installation of insulating shield. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] Example 1 like Figure 1 As shown, this embodiment provides an upper limb exoskeleton habit-adaptive assistance and multi-dimensional assessment method for live-line work in power distribution networks. It interacts with the upper exoskeleton real-time inference terminal (Python) via socket communication through intent judgment and command transmission from the lower-level control terminal (C#), and sends back control commands for real-time torque distribution. This is implemented in conjunction with surface electromyography (sEMG) acquisition equipment and a portable respiratory metabolism testing instrument. Specifically, it includes: S1: Construct an offline expert policy baseline assist mapping model based on a behavior cloning network, learn the mapping relationship between joint motion states and nonlinear baseline assist from expert teaching trajectories, and obtain the baseline assist; Further, step S1 specifically includes: collecting the motion state vector and corresponding assist torque during the expert teaching process as supervision labels, constructing the input sequence using a sliding window, flattening the input and standardizing it, and using a multilayer perceptron network to realize regression prediction from motion state to assist torque.

[0024] Specifically, in the offline phase of the Behavioral Cloning (BC) network, motion data from experienced operators performing standard tasks while wearing upper limb exoskeletons in actual work environments are collected as the training dataset. At each moment, define the kinematic state vector for each frame. and supervision labels as follows: (1) (2) in, and These are the angles of the left and right shoulder joints, respectively. and The angular velocities of the left and right shoulder joints. Characterizing bilateral posture differences, and The assist torque for the left and right shoulder joints used by operators wearing exoskeletons in field operations.

[0025] To capture short-term dynamic relationships, a length of [length missing] is used. Sliding window constructs input sequence and the torque at the end of the window As a supervisory label.

[0026] To match the input structure of a multilayer perceptron (MLP) and improve training stability, the input is flattened into... A 3D vector is generated, and the input and output are standardized using z-scores respectively: (3) (4) in and These are the mean and standard deviation of the five-dimensional input, respectively. and These are the mean and standard deviation of the two-dimensional output, respectively. It is a very small constant used to avoid singularities.

[0027] A behavioral cloning strategy network is constructed using an MLP with two hidden layers to achieve regression prediction from motion state to assist torque: (5) in, The normalized torque output predicted by the network. This represents all the learnable parameters of the model. The network has an input dimension of 5N=100, contains two hidden layers (256 and 128 units respectively) and an output layer (2 units), and the hidden layers use ReLU activation.

[0028] S2: Design an online personal habit adaptive mechanism to estimate the operator's steady-state habit center in real time during the operation, and dynamically adjust the benchmark assist according to the deviation between the current posture and the habit center through a smooth transition function to generate adaptive assist; Further, step S2 specifically includes: determining attitude stability by calculating the standard deviation of joint angles within a sliding window; extracting the state stability vector when the stability condition is met; updating the habit center using an exponential moving average; calculating the normalized offset distance of the current attitude relative to the habit center; constructing a piecewise quadratic continuous decay smooth transition function based on the offset distance using soft and hard thresholds; and applying the smooth transition function to the reference assist to obtain adaptive assist.

[0029] Specifically, during the online operation phase, the system determines posture stability by calculating the standard deviation of the left and right joint angles within a sliding window. When the standard deviation is less than a set threshold and the accumulated stabilization time exceeds a specific threshold, it is determined that the operator has established a steady-state operating habit. At this point, the state stability vector is extracted. : (6) (7) in, and These are the average angles of the right and left shoulders within the window, respectively. The difference in pose between the two sides within an N-frame window. It represents the average difference in pose between the two sides within an N-frame window.

[0030] Define habit center as When the steady-state condition is met, an exponential moving average (EMA) is used for online updates: (8) in, The update factor for EMA, .

[0031] Define the normalized offset distance of the current pose relative to the habit center as: : (9) in, The scaling factor for the components is obtained statistically from the training data and is used to eliminate the differences in the dimensions of different components. It is a very small constant used to avoid singularities.

[0032] Then, a smooth transition function based on deviation is introduced. Using soft threshold and hard threshold Construct a piecewise quadratic continuous decay boundary: (10) in, As a habit-based factor, the dynamic arbitration weight, when the deviation... When it is smaller (within the area of ​​personal habits), When it approaches 1, it outputs the full benchmark assist; when As the threshold is gradually increased and exceeds the soft threshold, the function ensures the weight of the exoskeleton intervention. Continuous decay maintains extremely high human-machine compliance during transitional movements that deviate from the habitual posture.

[0033] Applying this transition function to the denormalized reference help yields the online adaptive help: (11) Among them, when When smaller, When the output is close to 1, the output is close to the baseline assist. When it increases, Less than 1, thus reducing the assist when deviating from the habitual posture.

[0034] S3: Construct a closed-loop physiological fatigue prediction model based on surface electromyography signals and gated circulatory unit network to predict muscle fatigue trends in real time and obtain predicted fatigue index values. Further, step S3 specifically includes: acquiring multi-channel surface electromyography (EMG) signals, calculating the root mean square (RMS) after filtering and rectification, and performing weighted summation of the multi-channel RMS signals; normalizing the weighted values ​​using the maximum voluntary contraction test reference value, and calculating the fatigue level index; aligning the exoskeleton kinematic data with the surface EMG signals using timestamps, constructing an input sequence window, using a gated recurrent unit network to model the temporal dependence of fatigue, and outputting continuous fatigue index prediction values.

[0035] Specifically, multi-channel surface electromyography (sEMG) signals were acquired during the operator's work process. After filtering and full-wave rectification, the root mean square (RMS) of the signals was calculated using a sliding window, and the multi-channel RMS was then weighted and summed. (12) (13) in, Let N be the EMG signal processed at the i-th sampling point at time t, and N be the window length. Number of sampling points within, Let p be the RMS of the p-th channel at time t. For the corresponding weights.

[0036] Subsequently, the RMS weighted value was normalized using the MVC test reference value to determine the fatigue level index. As shown below: (14) (15) in, It is a very small constant used to avoid singularities. This is the reference value for the maximum voluntary contraction of the m-th muscle. This represents the operator's average normalized activation level during the initial stage of operation on the insulating rod (i.e., before fatigue occurs). The set limit fatigue threshold, This indicates that the muscles are in an initial, relatively resting or fatigue-free baseline exertion state. This indicates increased muscle activation relative to the resting phase, corresponding to a higher level of fatigue.

[0037] The exoskeleton kinematic data and electromyographic signals are time-stamped to construct an input sequence window of length L. The time-series dependency of fatigue is modeled using a gated recurrent unit (GRU) network: (16) in, These are the predicted fatigue index values ​​output by the GRU model. This represents the complete set of learnable parameters for the GRU fatigue prediction network. Additionally, define... These are the optimal model parameters obtained through training.

[0038] S4: Perform dynamic fatigue compensation, introduce monotonic smoothness constraints to calculate fatigue gain coefficient, determine fatigue compensation torque based on the predicted fatigue index value, and superimpose the adaptive assistance with the fatigue compensation torque to generate the total control torque of the exoskeleton; wherein, the total control torque is used to provide the operator with comfortable and effortless assistance that adapts to the current task and personal habits in the live-line work scenario; Further, step S4 specifically includes: setting an upper limit for the fatigue gain coefficient constrained by the human biomechanical attenuation law; dynamically modulating the adaptive assist according to the predicted fatigue index value to generate the total output torque of the exoskeleton; the calculation formula for the dynamic modulation is: total output torque = adaptive assist × (1 + fatigue gain coefficient × predicted fatigue index value), wherein the fatigue gain coefficient changes monotonically and smoothly between zero and the upper limit.

[0039] Specifically, a fatigue gain coefficient K is set (constrained by the biomechanical attenuation law of the human body, the maximum fatigue compensation gain is usually set to a safe upper limit of K=0.3). Based on the real-time fatigue prediction index... Total output torque of the exoskeleton Perform dynamic modulation: (17) Under this mechanism, when the system predicts fatigue accumulation, it automatically adds a corresponding proportion of compensating torque (up to 30% increase) to compensate for the decline in human force exertion, while preventing interactive instability caused by excessive assistance. The generated total torque command is sent to the exoskeleton's underlying control loop in real time.

[0040] S5: Construct a multi-dimensional neurophysiological comprehensive assessment system to quantitatively evaluate the control efficacy generated by S1 to S4 through single muscle activation assessment, multi-muscle synergy assessment and respiratory metabolism assessment, reflecting the actual assistive effect of the exoskeleton from the local muscle level to the whole body metabolic level.

[0041] Furthermore, the single muscle activation assessment includes: calculating the root mean square amplitude of the surface electromyography signal after filtering, full-wave rectification and normalization, and assessing the local load reduction effect of the exoskeleton on the target force-generating muscle group.

[0042] Furthermore, the multi-muscle coordination assessment includes: performing non-negative matrix decomposition on the electromyographic envelope signal matrix to obtain a coordination weight matrix and a coordination activation coefficient matrix; determining the optimal number of coordinations by calculating the variance explanation rate, adjusting the number of coordinations from the minimum value to the number of muscles, and selecting the minimum number of coordinations with a global variance explanation rate greater than or equal to a preset threshold; and using a cosine similarity algorithm to calculate the similarity of coordination vectors before and after wearing the exoskeleton to determine whether the exoskeleton increases the burden on the central nervous system and changes the muscle coordination force exertion mode.

[0043] Furthermore, the respiratory metabolic assessment includes: collecting oxygen consumption rate and carbon dioxide excretion rate, and calculating the human energy consumption rate using a modified Weir formula.

[0044] Specifically, to quantify the effectiveness of the above control strategies, this embodiment constructs a comprehensive evaluation system covering single muscle activation, multi-muscle synergy, and systemic metabolism.

[0045] Single muscle activation level: The effect of exoskeleton on local load reduction of core muscle groups (anterior deltoid, middle deltoid, posterior deltoid, biceps brachii, triceps brachii, trapezius, and pectoralis major) was evaluated by calculating the filtered and normalized RMS amplitude of electromyography.

[0046] Central neuromuscular coordination level: Electromyographic envelope signal matrix Perform nonnegative matrix factorization (NMF): (18) Where m represents the number of muscles, n is the number of data points, and k represents the current experimental conditions. For the collaborative weight matrix, Each element represents the relative contribution of each muscle in a certain synergy. For the co-activation coefficient, For the number of collaborations. Each line represents how co-activation changes over time.

[0047] To eliminate amplitude differences between different subjects and different experimental trials, we The summation and normalization were performed, and for ease of subsequent processing... The matrix was subjected to summation normalization and peak normalization respectively: (19) (20) in, , and .

[0048] The optimal number of collaborators was determined by calculating the variance explained (VAF), adjusting the number of collaborators from 1 to 7, and selecting global [the appropriate option]. The minimum number of collaborations ≥ 95%, VAF is defined as: (twenty one) in, This represents the Frobenius norm.

[0049] Furthermore, the cosine similarity algorithm was used to calculate the similarity of the collaborative vectors before and after wearing the exoskeleton, in order to verify whether the exoskeleton disrupts the original motion control laws of the human body.

[0050] Systemic respiratory and metabolic level: To intuitively quantify the overall metabolic reduction level of the exoskeleton replacing human mechanical work, a modified Weir formula was used, combined with the collected oxygen consumption rate. and carbon dioxide emission rate Calculate the rate of energy expenditure in the human body: (twenty two) The beneficial effects of this embodiment: This embodiment proposes a behavior cloning (BC) assistance strategy based on on-site expert teaching, which overcomes the limitations of traditional fixed impedance control and achieves accurate mapping from joint motion state to nonlinear reference assistance. It can closely match the real mechanical requirements of unstructured scenarios in live-line work of distribution networks.

[0051] This embodiment incorporates an online adaptive mechanism based on individual user habits. By capturing the center of habitual posture in real time, it smoothly adjusts the assistance based on the individual differences of different operators. This achieves dynamic torque distribution—providing full assistance when it aligns with habits and automatically reducing or avoiding assistance when it deviates from habits—significantly improving the safety and smoothness of human-computer physical interaction.

[0052] This embodiment constructs a closed-loop physiological fatigue prediction and dynamic compensation mechanism based on sEMG-GRU. By predicting the operator's muscle fatigue state and adjusting the assist gain in real time in conjunction with smooth monotonic constraints, it automatically fills the torque output loss caused by fatigue accumulation within safe limits (such as a maximum of 30%), effectively delaying the accumulation of muscle fatigue under long-term high-risk operations and significantly reducing the operator's overall metabolic energy consumption.

[0053] This embodiment establishes a multi-dimensional neuro-physiological comprehensive assessment system covering both local and global aspects, filling the gap in existing single assessment indicators for special operations exoskeletons. This system not only objectively verifies the physical load-reduction effectiveness of the exoskeleton, but also ensures through muscle synergy analysis that the assistance strategy does not disrupt the body's original neural control structure, thereby effectively avoiding physiological operational stiffness and guaranteeing motion transparency and operational safety under high-pressure working conditions.

[0054] Figure 1 This is a habit-adaptive exoskeleton control framework based on multi-scenario on-site expert demonstrations; where (a) is multi-scenario on-site data acquisition, (b) is offline BC network pre-training, and (c) is an online real-time inference architecture based on intent triggering.

[0055] Figure 2 This is a multi-dimensional neurophysiological comprehensive assessment method; (a) is the sensor deployment for 7-channel surface electromyography (sEMG) and portable respiratory metabolic data acquisition. (b) is the comprehensive data processing workflow and assessment indicators, covering single muscle activation, multi-muscle synergy, and total body energy consumption.

[0056] Figure 3 This is a schematic diagram of the angle-torque distribution characteristics learned by the behavioral cloning network during a complete live-line working process.

[0057] Figure 4 The dynamic response of the online adaptive power assist strategy in continuous unstructured motion is given; where (a) represents the actual output torque. Compared with the reference torque (a) Relationship; (b) is the pose offset distance and smooth transition function (c) represents the change; (c) represents the steady-state time. Accumulation and habit center The update mechanism.

[0058] Figure 5 To evaluate the dynamic effectiveness of the exoskeleton fatigue compensation strategy, (a) shows the tracking and reconstruction performance of the GRU network for smooth fatigue state labels in offline testing; (b) shows the actual torque response during online human-computer interaction.

[0059] The results of the multidimensional assessment are as follows: (1) In terms of muscle activation, combined with Figure 6 (a) and (b) show that during long-term operation with an insulated rod, the average RMS amplitude and the sliding RMS curve over time of the core weight-bearing muscle groups under conditions C1 and C2 are significantly lower than the baseline of C0, and condition C2 further delays the rise in activation level through fatigue compensation. (2) Regarding muscle synergy, combined with Figure 7 It can be seen that there is no significant difference in the variance explained (VAF) under different conditions, and the optimal number of synergies remains at 2, indicating that wearing an exoskeleton does not impose an additional computational burden on the nervous system. Figure 8 As shown in (af), the first and second cooperative features remain essentially consistent under conditions C0, C1, and C2. Figure 9As shown, the cosine similarity of the muscle synergy vectors between the exoskeleton-assisted and baseline conditions remains at a high level, proving that the control strategy provides physical assistance without disrupting the original neural control patterns of the human body. (3) In terms of respiratory metabolism, combined with Figure 10 (a) and (b) show that during long-duration work, the total energy consumption rate and oxygen consumption rate under conditions C1 and C2 are significantly lower than those under condition C0, with the metabolic energy consumption reduction under condition C2 being the most significant, which directly verifies the system's overall load reduction efficiency. (4) Regarding short-duration task compatibility, combined with Figure 11 As shown in (a), (b), and (c), in short-term auxiliary tasks such as voltage detection, operation of insulating gloves, and installation of insulating shields, the cosine similarity of muscle coordination between exoskeleton assistance and non-exoskeleton conditions is extremely high, indicating that the system still has excellent kinematic transparency and operational compliance when switching between different working conditions.

[0060] Example 2 This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0061] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adaptive upper limb exoskeleton assistance and multidimensional evaluation for live-line work in power distribution networks, characterized in that, Includes the following steps: S1: Construct an offline expert policy baseline assist mapping model based on a behavior cloning network, learn the mapping relationship between joint motion states and nonlinear baseline assist from expert teaching trajectories, and obtain the baseline assist; S2: Design an online personal habit adaptive mechanism to estimate the operator's steady-state habit center in real time during the operation, and dynamically adjust the benchmark assist according to the deviation between the current posture and the habit center through a smooth transition function to generate adaptive assist; S3: Construct a closed-loop physiological fatigue prediction model based on surface electromyography signals and gated circulatory unit network to predict muscle fatigue trends in real time and obtain predicted fatigue index values. S4: Perform dynamic fatigue compensation, introduce monotonic smoothness constraints to calculate fatigue gain coefficient, determine fatigue compensation torque based on the predicted fatigue index value, and superimpose the adaptive assistance with the fatigue compensation torque to generate the total control torque of the exoskeleton; wherein, the total control torque is used to provide the operator with comfortable and effortless assistance that adapts to the current task and personal habits in the live-line work scenario; S5: Construct a multi-dimensional neurophysiological comprehensive assessment system to quantitatively evaluate the control efficacy generated by S1 to S4 through single muscle activation assessment, multi-muscle synergy assessment and respiratory metabolism assessment, reflecting the actual assistive effect of the exoskeleton from the local muscle level to the whole body metabolic level.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: collecting motion state vectors and corresponding assist torques during expert teaching as supervision labels, constructing input sequences using a sliding window, flattening the inputs and standardizing them, and using a multilayer perceptron network to achieve regression prediction from motion state to assist torque.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: determining attitude stability by calculating the standard deviation of joint angles within a sliding window; extracting the state stability vector when the stability condition is met; updating the habit center using an exponential moving average; calculating the normalized offset distance of the current attitude relative to the habit center; constructing a piecewise quadratic continuous decay smooth transition function based on the offset distance using soft and hard thresholds; and applying the smooth transition function to the reference assist to obtain adaptive assist.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: acquiring multi-channel surface electromyography (EMG) signals, calculating the root mean square (RMS) after filtering and rectification, and performing weighted summation of the multi-channel RMS signals; normalizing the weighted values ​​using the maximum voluntary contraction test reference value, and calculating the fatigue level index; aligning the exoskeleton kinematic data with the surface EMG signals using timestamps, constructing an input sequence window, using a gated recurrent unit network to model the temporal dependence of fatigue, and outputting continuous fatigue index prediction values.

5. The method according to claim 1, characterized in that, Step S4 specifically includes: setting an upper limit for the fatigue gain coefficient constrained by the human biomechanical attenuation law; dynamically modulating the adaptive assist according to the predicted fatigue index value to generate the total output torque of the exoskeleton; the calculation formula for the dynamic modulation is: total output torque = adaptive assist × (1 + fatigue gain coefficient × predicted fatigue index value), wherein the fatigue gain coefficient changes monotonically and smoothly between zero and the upper limit.

6. The method according to claim 1, characterized in that, The single muscle activation assessment includes: calculating the root mean square amplitude of the surface electromyography signal after filtering, full-wave rectification and normalization, and assessing the local load reduction effect of the exoskeleton on the target force-generating muscle group.

7. The method according to claim 1, characterized in that, The multi-muscle coordination assessment includes: performing non-negative matrix decomposition on the electromyographic envelope signal matrix to obtain the coordination weight matrix and the coordination activation coefficient matrix; and using the cosine similarity algorithm to calculate the similarity of the coordination vectors before and after wearing the exoskeleton in order to determine whether the exoskeleton increases the burden on the central nervous system and changes the muscle coordination force exertion mode.

8. The method according to claim 1, characterized in that, The respiratory metabolic assessment includes collecting oxygen consumption rate and carbon dioxide excretion rate, and calculating the human energy consumption rate using a modified Weir formula.

9. The method according to claim 7, characterized in that, The multi-muscle synergy assessment also includes: determining the optimal number of synergies by calculating the variance explained rate, adjusting the number of synergies from the minimum value to the number of muscles, and selecting the minimum number of synergies with a global variance explained rate greater than or equal to a preset threshold.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-9.

Images