Carrying unbalance loading assisting power distribution method based on artificial intelligence network and left-right assisting power ratio scheduling framework

Abstract

The invention provides an object-carrying unbalance loading assisting power distribution method based on an artificial intelligence network and a left-right assisting power ratio scheduling framework. According to the method, left and right lower limb joint movement, pelvic trunk posture, left and right foot sole stress and foot sole pressure center information are synchronously collected, gait phase recognition and phase alignment are carried out, the left and right support stress difference is calculated, and lateral sensitive characteristics are formed by the lateral migration amount of the foot sole pressure center and the lateral swing amount of the pelvic trunk; the features are input into an improved Kolmogov-Arnod network which meets phase consistency constraints and contains left and right coupling and contrast mapping structures, and the initial unbalance loading index and unbalance loading credibility are regressed; branch judgment is carried out according to a threshold value, when the credibility is met, micro-amplitude tentative assistance is applied, single-point correction is carried out on an unbalance loading index based on a gait response amount, a left-right assistance ratio is generated, and the left-right assistance ratio is mapped into a left-right asymmetric assistance instruction through amplitude limiting and change rate constraint to drive an execution mechanism; according to the method, left and right assisting power self-adaptive distribution under the unbalance loading working condition is achieved.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton-assisted control and gait perception, and in particular to a method for distributing load-carrying assistance based on an artificial intelligence network and a left-right assistance ratio scheduling framework. Background Technology

[0002] Lower limb exoskeletons are used to assist joints such as the hip, knee, and ankle during walking, reducing muscle burden and improving walking comfort. In everyday scenarios, users often engage in side-carrying behaviors such as carrying a bag with one hand or carrying items on their side, causing a shift in the body's center of gravity and supporting forces to the left and right sides, accompanied by compensatory gait behaviors such as lateral movement of the plantar pressure center and lateral swaying of the pelvis and trunk. To adapt to these conditions, the system typically needs to combine information such as inertial posture and plantar force to identify gait state and left-right load differences, and adjust the left and right assist output accordingly.

[0003] Existing exoskeleton control solutions mostly employ gait phase state machines and symmetrical assist strategies. After identifying the support phase and swing phase, they output left and right assist according to a preset curve or reference torque. Some solutions also use indicators such as plantar normal force, left and right support time, plantar pressure center trajectory, or torso posture to construct rules and correct left and right assist according to thresholds or proportional relationships. Some solutions introduce machine learning models to estimate load or gait parameters based on multi-sensor features, and then map the estimation results to left and right assist coefficients. The output is made smooth by limiting amplitude and rate of change, which is suitable for walking on flat ground and for a certain range of individual differences.

[0004] The above-mentioned scheme still has limitations when there is significant lateral compensation, such as when carrying objects on one side: First, the left and right features are often not aligned and expressed within a unified phase window, and the left and right comparison is easily affected by changes in walking speed and event detection jitter, resulting in misjudgment; Second, the estimation results usually lack quantitative output of reliability, and sensor noise or short-term anomalies may trigger unnecessary asymmetric assistance; Third, the left and right assistance ratio mostly depends on fixed rules or direct mapping of a single estimation, and lacks a mechanism to verify and correct the direction and magnitude of the off-center load under actual assistance, making it difficult for assistance scheduling to stably fit the off-center load state.

[0005] Therefore, those skilled in the art urgently need a method for assisting in the distribution of loads on an off-center basis. Summary of the Invention

[0006] One objective of this invention is to propose a method for distributing load-carrying assistance based on an artificial intelligence network and a left-right assistance ratio scheduling framework. The core technical problem to be solved by this application is: how to reliably determine the load-carrying imbalance and safely schedule the left-right assistance ratio online under the conditions of gait phase change and sensing uncertainty during the user's walking with a single hand or to one side, so that the exoskeleton can output asymmetrical assistance to match the lateral gait compensation caused by the load-carrying imbalance.

[0007] The off-center load-carrying assistance allocation method based on an artificial intelligence network and a left-right assist ratio scheduling framework according to an embodiment of the present invention includes:

[0008] S1. During the user's walking while carrying objects with one hand or to one side, acquire information on the joint movement of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet.

[0009] S2. Based on the joint motion information of the left and right lower limbs, the posture information of the pelvis and trunk, and the force and pressure center information of the left and right feet, gait phase recognition and phase alignment are performed. Based on the phase alignment results, the difference in support force between the left and right sides, the lateral migration of the pressure center of the foot, and the lateral sway of the pelvis and trunk are calculated to form laterally sensitive features.

[0010] S3. Input the lateral sensitive features into the improved Kolmogorov-Arnold network off-center load index regressor for inference. The improved Kolmogorov-Arnold network off-center load index regressor satisfies the phase-consistent input constraint and sets a left-right coupling mapping in the network structure to compare and map the contributions of the left and right supports within the same phase window. This directly regresses the lateral offset relationship caused by the off-center load to the initial off-center load index and generates an off-center load confidence level to characterize the reliability of the initial off-center load index.

[0011] S4. Based on the initial off-center load index and off-center load confidence, the assist ratio is determined. When the off-center load confidence is higher than the off-center load confidence threshold and the initial off-center load index is higher than the off-center load index threshold, a trial left and right assist ratio with a slight offset is generated and applied to form a trial gait response. Based on the trial gait response, the gait response quantity including the change in left and right force difference, the change in the lateral migration of the plantar pressure center and the change in the lateral swing of the pelvis and trunk is calculated. Based on the gait response quantity, the initial off-center load index is corrected once to obtain the corrected off-center load index. Then, the left and right assist ratio is generated from the corrected off-center load index. When the off-center load confidence is lower than the off-center load confidence threshold or the initial off-center load index is lower than the off-center load index threshold, the symmetrical left and right assist ratio is determined as the left and right assist ratio.

[0012] S5. Under the constraints of left and right assist ratio limit and left and right assist ratio change rate, map the left and right assist ratio to left and right asymmetric assist command.

[0013] S6. Send the left and right asymmetrical assist command to the exoskeleton actuator to drive the left and right side assist output.

[0014] Optionally, S1 is as follows:

[0015] A multi-sensor acquisition unit triggered by a unified clock is used to synchronously sample data from the left and right sides, so that the joint motion information of the left and right lower limbs, the posture information of the pelvis and trunk, and the force and pressure center information of the left and right feet have the same time reference.

[0016] Hip joint motion information, knee joint motion information, and ankle joint motion information are collected from the left and right lower limbs respectively, and the hip joint motion information, knee joint motion information, and ankle joint motion information are summarized to form the joint motion information of the left and right lower limbs.

[0017] Attitude angles and angular velocities were collected for the pelvis and trunk respectively, and the pelvic and trunk attitudes were mapped to the same body coordinate system to form pelvic and trunk attitude information.

[0018] The sequence of normal force and pressure center position of the left and right soles is collected respectively, and the force and pressure center information of the left and right soles is output according to the left and right sole channels.

[0019] Optionally, S2 is as follows:

[0020] The ground contact and takeoff events of the left and right feet are identified based on the ground contact threshold and takeoff threshold of the left and right feet, and the consistency of the ground contact and takeoff events is confirmed by combining the joint motion information of the left and right lower limbs, resulting in a gait phase sequence that includes the left support phase and the right support phase.

[0021] Based on the gait phase sequence, the pelvic and trunk posture information and the left and right foot plantar pressure center information are divided into left support phase window and right support phase window, and the data in each phase window are aligned to the same phase starting point to form a phase alignment result.

[0022] The left and right support force difference is calculated according to the phase window order based on the phase alignment result. The left and right support force difference is obtained by subtracting the right foot force from the left foot force. The lateral migration of the foot pressure center and the lateral swing of the pelvis and trunk are calculated based on the horizontal axis of the body coordinate system within the same phase window.

[0023] The left and right support force difference of the left and right support phase window, the left and right support force difference of the right and right support phase window, the lateral migration of the plantar pressure center of the left and right support phase windows, and the lateral sway of the pelvis and trunk of the left and right support phase windows are spliced ​​together in a fixed dimensional order to form a six-dimensional lateral sensitivity feature, which is used as the input of the improved Kolmogorov-Arnold network off-center loading index regressor.

[0024] Optionally, the six-dimensional lateral sensitive features are extracted from the phase alignment results in one step using a phase-first extraction formula, which is as follows:

[0025] ;

[0026] in, For the first Six-dimensional lateral sensitivity features per gait cycle. The difference in force between the left and right supports of the left support phase window. The difference in force between the left and right supports of the right support phase window. This represents the lateral migration of the plantar pressure center within the left support phase window. This represents the lateral migration of the plantar pressure center within the right support phase window. This represents the lateral sway of the pelvis and trunk within the left support phase window. The lateral sway of the pelvis and trunk within the right support phase window. This represents the standard number of sampling points for the phase window. The discrete index within the phase window and its value range is to , Left support phase window Inner The normal force on the left foot sole at each resampling point Left support phase window Inner Normal force on the right foot sole at each resampling point For right support phase window Inner The normal force on the left foot sole at each resampling point For right support phase window Inner Normal force on the right foot sole at each resampling point Left support phase window Inner The center of pressure on the left foot at each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis For right support phase window Inner The right plantar pressure center of each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis Left support phase window Inner The combined roll angle of each resampling point, For right support phase window Inner The combined roll angle of each resampled point.

[0027] Optionally, S3 specifically refers to:

[0028] Lateral sensitive features are organized into six-dimensional lateral sensitive features in a fixed dimensional order, and phase consistency input constraint verification is performed on the six-dimensional lateral sensitive features. The phase consistency input constraint verification is used to confirm that the six-dimensional lateral sensitive features correspond to the left support phase window and the right support phase window respectively and are expressed under the same phase starting point.

[0029] The six-dimensional lateral sensitive features are input into the left-right coupling mapping layer, and coupling mapping is performed sequentially according to the physical quantity categories of left-right support force difference, lateral migration of plantar pressure center, and lateral swing of pelvis and trunk.

[0030] Within each physical quantity category, one-dimensional learnable piecewise function units are configured for the left support phase window dimension and the right support phase window dimension, respectively. The two dimensions are nonlinearly transformed using the same set of one-dimensional learnable piecewise function unit parameters to achieve left-right coupling mapping where the left and right dimensions share one-dimensional learnable piecewise function unit parameters.

[0031] The nonlinear transformation results within each physical quantity category are weighted and summed to obtain a two-dimensional coupling representation of that physical quantity category. The two-dimensional coupling representations of the three physical quantity categories are then summarized in a fixed order to form a six-dimensional coupling representation.

[0032] The six-dimensional coupling representation is input into the contrast mapping layer, and the contrast quantity between the left support phase window coupling representation and the right support phase window coupling representation is calculated under the same phase window definition according to the physical quantity category, forming a three-dimensional contrast representation. The three-dimensional contrast representation corresponds to the force path contrast, the plantar pressure center path contrast, and the pelvic trunk lateral movement path contrast in sequence.

[0033] The three-dimensional contrastive representation is input into the regression output layer. The regression output layer uses a function neuron structure to map the three-dimensional contrastive representation. The function neuron structure configures a one-dimensional learnable piecewise function unit for each dimension of the input, performs a nonlinear transformation, and then performs a weighted summation to output a one-dimensional initial bias index.

[0034] The three-dimensional contrastive representation is input into the confidence output layer. The confidence output layer uses the same function neuron structure as the regression output layer to map the three-dimensional contrastive representation and outputs a one-dimensional bias confidence.

[0035] Optionally, S4 specifically refers to:

[0036] Receive the initial off-center load index and off-center load confidence level, and compare them with the off-center load index threshold and the off-center load confidence level threshold respectively, and output the assist ratio branch determination result;

[0037] When the result of the assist ratio branch determination is a trial branch, the off-load direction is determined according to the initial off-load index, and a preset trial offset amplitude is introduced based on the symmetrical left and right assist ratio to generate the trial left and right assist ratio. The trial left and right assist ratio increases the assist ratio on the side corresponding to the off-load direction and decreases the assist ratio on the other side to keep the assist ratio benchmark consistent.

[0038] The trial left and right assistance ratio is applied to the preset trial phase window to form a trial gait response. Within the trial phase window, the left and right lower limb joint motion information, pelvic trunk posture information, left and right foot plantar force and plantar pressure center information are obtained in the same way as in step one.

[0039] The same gait phase recognition and phase alignment as in step two are performed on the trial gait response. The difference in left and right support force, the lateral migration of the plantar pressure center and the lateral swing of the pelvis and trunk are calculated on the phase alignment results to obtain the trial lateral sensitivity characteristics.

[0040] The lateral sensitivity features and the anterior lateral sensitivity features are used to calculate the changes in left and right force difference, the lateral migration of the plantar pressure center, and the lateral swing of the pelvis and trunk according to the same phase window, and the changes are fused in a fixed order to form the gait response quantity.

[0041] Based on the gait response and the preset trial bias amplitude, a single-point correction is performed on the initial off-center load index. The single-point correction is limited to a single correction action and outputs the corrected off-center load index. Then, the left and right assist ratios are generated according to the corrected off-center load index through the preset off-center load index to assist ratio mapping relationship.

[0042] When the result of the assist ratio branch determination is a symmetrical branch, the symmetrical left and right assist ratio is directly determined as the left and right assist ratio.

[0043] Optionally, the single-point correction corrects the initial off-load index based on the gait response and a preset trial bias amplitude, and outputs the corrected off-load index. The single-point correction is calculated using a verification formula, specifically:

[0044] ;

[0045] in, For the first The corrected off-load index for each gait cycle, For the first The initial off-load index for each gait cycle, To preset the correction gain, , , Preset fusion weights are used to weight and fuse changes in left-right force differences, changes in the lateral migration of the plantar pressure center, and changes in lateral swaying of the pelvis and trunk. , , It is a dimensionless constant. This represents the normalized change in the difference in force between the left and right sides. This represents the normalized lateral migration of the plantar pressure center. This represents the normalized lateral swaying variation of the pelvis and trunk. To test the bias amplitude, which is a dimensionless value, To test the left-side assist ratio, To test the right-side assist ratio, To test the absolute value of the difference in the left and right assist ratios, This is the probe direction symbol, used to align the correction direction with the direction in which the probe left and right assist ratio is applied. To test the phase window The average time of the sum of the normal forces on the soles of the left and right feet. To preset the lateral length parameter of the foot, This serves as the preset roll angle normalization benchmark.

[0046] Optional, S5 specifically includes:

[0047] Input the left and right assist ratio obtained in step four into the amplitude limiting module, and truncate the left and right assist ratio according to the amplitude limiting constraint to obtain the amplitude-limited left and right assist ratio.

[0048] The left and right assist ratio of the amplitude limit is differentially compared with the left and right assist ratio of the previous control cycle. The difference result is then updated with a speed limit based on the constraint of the change rate of the left and right assist ratio to obtain the amplitude limit and speed limit left and right assist ratio.

[0049] Apply the limiting and speed-limiting left and right assist ratios to the preset symmetrical assist baseline command, and scale the left and right drive channels proportionally to obtain the target assist outputs on the left and right sides.

[0050] The left and right target assist outputs are converted into left and right asymmetrical assist commands that match the drive channel of the exoskeleton actuator, and the left and right asymmetrical assist commands are kept continuous within the phase window corresponding to step four.

[0051] Optionally, step S6 specifically includes:

[0052] The left and right asymmetric assistance commands obtained in step five are sent to the exoskeleton actuators according to the left and right drive channels, respectively, and the left and right asymmetric assistance commands are associated with the current gait phase sequence to determine the execution timing.

[0053] The exoskeleton actuator drives the left and right side assist outputs according to the left and right asymmetrical assist commands, and maintains the timing consistency of the left and right side assist outputs by using the continuous control method within the phase window during the assist output process;

[0054] During the assisted output process, the motion information of the left and right lower limb joints, the posture information of the pelvis and trunk, the force on the left and right soles and the information of the sole pressure center are acquired simultaneously, and the information is aligned with the corresponding phase window to maintain phase consistency input constraints.

[0055] The synchronously acquired information on the joint movement of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet will be used as the input for step one of the subsequent cycles.

[0056] The beneficial effects of this invention are:

[0057] (1) This method proposes an improved Kolmogorov-Arnold network eccentricity index regression method. By introducing phase consistency constraints on the input side and setting left-right coupling mapping and contrast mapping in the network structure, the same physical quantity is represented in a consistent nonlinear manner by sharing piecewise function parameters in the left and right support phase windows. Then, under the same phase definition, the force path, the center path of plantar pressure, and the lateral movement path of the pelvis and trunk are compared and contrasted. Thus, the lateral offset relationship caused by eccentricity is regressed into an eccentricity index in the form of "same phase and comparable". Compared with common learning / rule schemes based on unilateral index or direct splicing of features, this method explicitly embeds the formation process of left-right differences into the network calculation link, reduces the impact of the incomparability of left and right sides caused by gait changes and phase drift on the eccentricity judgment, and provides a usable reliable basis for subsequent auxiliary scheduling by simultaneously outputting the eccentricity confidence.

[0058] (2) This method proposes a left-right assist ratio scheduling and single-point correction mechanism based on off-center load confidence. When the off-center load confidence meets the threshold and the off-center load index exceeds the threshold, the estimation result is not directly mapped to asymmetric assist. Instead, a trial left-right assist ratio with a slight offset is first generated and applied to a preset trial phase window. Then, the trial lateral sensitive features are extracted using the phase recognition and phase alignment method consistent with conventional recognition. The changes in left-right force difference, lateral migration of plantar pressure center, and lateral sway of pelvis and trunk are calculated and fused into gait response quantities, which are used to perform a single-point correction on the initial off-center load index. This mechanism allows the off-center load direction and magnitude to be verified and corrected in the actual gait response when "assistance has been applied". This is different from the existing methods that rely on fixed rules or output left-right difference assist in a single estimation. From the perspective of control closed loop, it reduces the possibility of misjudgment triggering unnecessary asymmetric assist and maintains symmetric assist when the confidence is insufficient or the off-center load is not significant. It adapts to sensing uncertainty and short-term anomalies at the strategy level.

[0059] (3) This method proposes a left-right asymmetric assist command generation method for exoskeleton execution. The determined left-right assist ratio is updated under amplitude and rate of change constraints and mapped to the left-right channel scaling of the symmetric assist reference command to ensure that the left-right difference assist changes smoothly within a safe range. At the same time, the continuous output within the phase window is used and associated with the gait phase sequence to determine the execution timing, so that the asymmetric assist and the feature expression after phase alignment are consistent. Through the overall link of "multi-sensor synchronous acquisition - phase alignment feature - off-center load index and confidence inference - branch scheduling and trial correction - amplitude and speed limiting mapping - phase association execution", this method focuses on the off-center load identification and assist allocation problem in the off-center carrying scenario, realizes the online adaptive scheduling of the left-right assist ratio, and improves the control stability and safety of use under phase change and signal uncertainty conditions. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 The flowchart shows a method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework proposed in this invention.

[0062] Figure 2 This is a flowchart of the multi-sensor acquisition unit synchronous sampling process for a load-carrying off-center assist allocation method based on artificial intelligence network and left-right assist ratio scheduling framework proposed in this invention.

[0063] Figure 3 This is a flowchart of gait phase recognition, phase alignment, and six-dimensional lateral sensitive feature extraction for a load-carrying off-center assist allocation method based on artificial intelligence network and left-right assist ratio scheduling framework proposed in this invention.

[0064] Figure 4 This is a flowchart of the inference process of an improved Kolmogorov-Arnold network off-center load index regressor for a load-carrying off-center load allocation method based on an artificial intelligence network and a left-right assist ratio scheduling framework proposed in this invention.

[0065] Figure 5 This is a flowchart of the load-carrying off-center assist branch determination and trial gait response correction based on an artificial intelligence network and left-right assist ratio scheduling framework proposed in this invention.

[0066] Figure 6 This is a flowchart illustrating the left-right assist ratio constraint and left-right asymmetric assist command generation process of a load-carrying off-center assist allocation method based on an artificial intelligence network and a left-right assist ratio scheduling framework proposed in this invention. Detailed Implementation

[0067] In Example 1, reference Figures 1 to 6 A method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework, comprising:

[0068] S1. During the user's walking while carrying objects with one hand or to one side, acquire information on the joint movement of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet.

[0069] S2. Based on the joint motion information of the left and right lower limbs, the posture information of the pelvis and trunk, and the force and pressure center information of the left and right feet, gait phase recognition and phase alignment are performed. Based on the phase alignment results, the difference in support force between the left and right sides, the lateral migration of the pressure center of the foot, and the lateral sway of the pelvis and trunk are calculated to form laterally sensitive features.

[0070] S3. Input the lateral sensitive features into the improved Kolmogorov-Arnold network off-center load index regressor for inference. The improved Kolmogorov-Arnold network off-center load index regressor satisfies the phase-consistent input constraint and sets a left-right coupling mapping in the network structure to compare and map the contributions of the left and right supports within the same phase window. This directly regresses the lateral offset relationship caused by the off-center load to the initial off-center load index and generates an off-center load confidence level to characterize the reliability of the initial off-center load index.

[0071] S4. Based on the initial off-center load index and off-center load confidence, the assist ratio is determined. When the off-center load confidence is higher than the off-center load confidence threshold and the initial off-center load index is higher than the off-center load index threshold, a trial left and right assist ratio with a slight offset is generated and applied to form a trial gait response. Based on the trial gait response, the gait response quantity including the change in left and right force difference, the change in the lateral migration of the plantar pressure center and the change in the lateral swing of the pelvis and trunk is calculated. Based on the gait response quantity, the initial off-center load index is corrected once to obtain the corrected off-center load index. Then, the left and right assist ratio is generated from the corrected off-center load index. When the off-center load confidence is lower than the off-center load confidence threshold or the initial off-center load index is lower than the off-center load index threshold, the symmetrical left and right assist ratio is determined as the left and right assist ratio.

[0072] S5. Under the constraints of left and right assist ratio limit and left and right assist ratio change rate, map the left and right assist ratio to left and right asymmetric assist command.

[0073] S6. Send the left and right asymmetrical assist command to the exoskeleton actuator to drive the left and right side assist output.

[0074] In this embodiment, step S1 specifically includes:

[0075] The exoskeleton system is equipped with multiple sensor acquisition units, including sensors for the left and right lower limb joints, pelvis, trunk, and left and right foot soles. These units are driven by a unified clock trigger signal, which is a periodic trigger pulse generated by the controller's internal clock. Each sensor acquisition unit generates the same timestamp in each trigger cycle. timestamp Used to characterize the At the next sampling time, the controller will timestamp Data frames containing left and right lower limb joint motion information, pelvic and trunk posture information, and left and right foot plantar force and plantar pressure center information are written, so that all kinds of information have the same time reference. The controller uses the same trigger pulse to simultaneously start the data reading of the left and right channels and completes the data packet output within the same control cycle, so that subsequent gait phase recognition and phase alignment can establish a phase consistency relationship between the left and right sides and multiple sensor sources.

[0076] Regarding the acquisition of joint motion information for the left and right lower limbs, joint sensors for the left and right lower limbs are respectively deployed at the left hip joint, left knee joint, left ankle joint, right hip joint, right knee joint, and right ankle joint. Each joint sensor outputs the joint angle. With joint angular velocity Among them, joint angle The joint angular velocity corresponds to the angle value of the joint around the preset axis of rotation. Joint angle For the rate of change over time, the controller uses the joint angle difference between two adjacent timestamps and divides it by the timestamp interval to obtain the joint angular velocity. The joint angles and angular velocities of the left hip, knee, and ankle joints were summarized in a fixed order. The joint angles and angular velocities of the right hip, knee, and ankle joints are summarized in a fixed order. The fixed order is: hip joint angle, hip joint angular velocity, knee joint angle, knee joint angular velocity, ankle joint angle, ankle joint angular velocity, and joint angle. Adopting a definition method consistent with the mechanical zero point of the exoskeleton, so that... and It can be directly used for motion consistency verification in subsequent gait phase recognition;

[0077] Regarding the acquisition of pelvic and trunk posture information, the pelvic sensor and trunk sensor respectively output posture angles. With attitude angular velocity attitude angle It includes three components: roll angle, pitch angle, and yaw angle, and attitude angular velocity. Includes angular velocity components around three axes, with the controller preset in a body coordinate system. Body coordinate system The longitudinal axis aligns with the direction of human movement, the lateral axis points to the left side of the body, and the vertical axis points vertically upward. The controller obtains the sensor coordinate system to the body coordinate system through wear calibration. The fixed rotation relationship, the wearing calibration includes guiding the user to maintain a standing upright posture and collecting the posture output of the pelvic sensor and trunk sensor within a preset time window, and then interpolating the posture output with the body coordinate system. The pelvic sensor output is mapped to pelvic posture based on the upright alignment to determine a fixed rotational relationship. Angular velocity of pelvic posture Mapping the torso sensor output to torso posture angular velocity of torso posture and will , , , Pelvic and trunk posture information is compiled in a fixed order. By mapping pelvic pose and torso pose to the same body coordinate system This allows the subsequent lateral sway of the pelvis and trunk to be within the body coordinate system. Input data is obtained uniformly on the horizontal axis;

[0078] Regarding the acquisition of information on the force and center of pressure on the left and right soles, the left and right sole sensors output the normal force on the soles respectively. , It also outputs the sequence of plantar pressure center locations. , The plantar pressure center location sequence is calculated from the pressure values ​​and geometric positions of the plantar pressure measuring points. Specifically, the geometric positions of each measuring point are weighted and summed, and then normalized using the total plantar pressure to obtain the plantar pressure center location that can be directly used for transfer calculations. The controller will... , Mapping from the foot sensor coordinate system to the body coordinate system ,make , Includes body coordinate system The controller calculates the longitudinal and lateral coordinate components, using the lateral coordinate component as input data for the subsequent lateral migration of the plantar pressure center. Finally, the controller outputs the left and right plantar force and plantar pressure center information via the left and right plantar channels, respectively. and and with , , Together, they form the input sequence for subsequent gait phase recognition, phase alignment, and lateral sensitivity feature construction.

[0079] In this embodiment, step S2 specifically includes:

[0080] The controller uses the time-series data output from step one as input to perform gait phase recognition and phase alignment, and constructs lateral sensitive features. The time-series data includes the normal force on the left plantar surface. Force on the right foot in the normal direction Sequence of left plantar pressure center location Sequence of the location of the right plantar pressure center Pelvic and trunk posture information and information on the movement of the left and right lower limb joints. , ,in For the first The controller uses a body coordinate system and a timestamp for each sampling time. and the body coordinate system The horizontal axis is used as the calculation axis for the lateral migration of the plantar pressure center and the lateral sway of the pelvis and trunk, so that the plantar pressure center and the pelvic trunk posture are expressed under the same horizontal axis.

[0081] In terms of gait phase recognition, the controller presets a grounding threshold. With ground threshold Among them, the landing threshold Used to determine when the foot enters a support state under normal force on the sole, ground clearance threshold. Used to determine the state of foot normal force withdrawal from support, and to set Greater than To create threshold hysteresis, the controller scans sequentially over time. and ,when From less than Become greater than or equal to And in continuous Within each sampling point, maintain a value greater than or equal to At that time, determine the moment of the left foot contact with the ground. ,when From greater than Change to less than or equal to And in continuous Within each sampling point, maintain less than or equal to At that time, determine the moment of the left foot leaving the ground. Right foot contact event Moment of right foot off the ground The threshold determination process is the same as that for the left foot, where... The controller uses a preset number of continuous sampling points to suppress single-point noise triggering. It extracts the angular velocity of the ipsilateral ankle joint from the joint motion information of the left and right lower limbs. and Within the confirmation window before and after each ground contact event and ground lift event, the ankle joint angular velocity amplitude is checked to see if it is less than the preset ankle joint confirmation threshold. When the threshold determination result and the ankle joint confirmation result are simultaneously satisfied, the ground contact event or ground departure event is confirmed to be valid, and the controller uses the confirmed result... , , , Constructing gait phase sequences And define the left support phase as from arrive The time interval is defined as the right support phase from arrive Time interval;

[0082] Regarding phase alignment, the controller bases its actions on the gait phase sequence. Pelvic and trunk posture information Sequence of left and right plantar pressure center positions , The left support phase window is obtained by segmentation. With right support phase window ,in Using the gait cycle number as the starting point, the controller aligns the start point of each phase window to the ground contact event time corresponding to that phase window, and unifies the number of sampling points within the phase window to the preset standard number of sampling points for the phase window. When the original number of sampling points in the phase window is equal to... When there is inconsistency, the controller will adjust the phase window accordingly. , , , , Linear interpolation resampling is performed in chronological order to ensure that each left support phase window... With each right support phase window All form a length of The discrete sequence is used to form a phase-aligned result;

[0083] In terms of laterally sensitive feature calculation and organization, the controller operates in the body coordinate system. Extraction , The lateral coordinate components are used to obtain the lateral coordinate sequence of the left foot. With right foot lateral coordinate sequence Discrete sequences are formed within the left and right support phase windows after phase alignment. , ,in As a discrete index within the phase window, the controller uses pelvic and torso pose information. Extracting pelvic roll angle sequence With the torso roll angle sequence And form according to the same resampling index within each phase window. , , , The controller for each discrete index Calculate the combined roll angle and The combined roll angle is obtained by linearly weighting the pelvic roll angle and the trunk roll angle, with the weighting coefficients being preset constants. and and satisfy and The sum of these values ​​is equal to one, thus ensuring that the combined roll angle retains its angular dimensions and can be directly used for the start and end differential within the phase window. The controller further resamples the foot normal force within each phase window to obtain... , , , Based on this, a phase extraction formula is used to obtain six-dimensional lateral sensitive features in one step. and will As input to the improved Kolmogorov-Arnold network biased exponential regressor to satisfy the phase-consistent input constraint:

[0084] ;

[0085] in, For the first Six-dimensional lateral sensitivity features per gait cycle. The difference in force between the left and right supports of the left support phase window. The difference in force between the left and right supports of the right support phase window. This represents the lateral migration of the plantar pressure center within the left support phase window. This represents the lateral migration of the plantar pressure center within the right support phase window. This represents the lateral sway of the pelvis and trunk within the left support phase window. The lateral sway of the pelvis and trunk within the right support phase window. This represents the standard number of sampling points for the phase window. The discrete index within the phase window and its value range is to , Left support phase window Inner The normal force on the left foot sole at each resampling point Left support phase window Inner Normal force on the right foot sole at each resampling point For right support phase window Inner The normal force on the left foot sole at each resampling point For right support phase window Inner Normal force on the right foot sole at each resampling point Left support phase window Inner The center of pressure on the left foot at each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis For right support phase window Inner The right plantar pressure center of each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis Left support phase window Inner The combined roll angle of each resampling point, For right support phase window Inner The combined roll angle of each resampled point.

[0086] In this embodiment, step S3 specifically includes:

[0087] The controller obtains the first Six-dimensional lateral sensitivity features per gait cycle Afterwards, Inference is performed using an improved Kolmogorov-Arnold network biased index regressor, with six-dimensional lateral sensitivity features. Organized according to a fixed dimensional order to ,in The difference in force between the left and right supports of the left support phase window. The difference in force between the left and right supports of the right support phase window. This represents the lateral migration of the plantar pressure center within the left support phase window. This represents the lateral migration of the plantar pressure center within the right support phase window. This represents the lateral sway of the pelvis and trunk within the left support phase window. The lateral sway of the pelvis and trunk within the right support phase window;

[0088] The controller performs phase-consistency input constraint verification on the six-dimensional laterally sensitive features. The phase-consistency input constraint verification includes dimension order verification and phase start point verification. Dimension order verification is used to confirm... to The arrangement is consistent with the preset fixed-dimensional order. The phase start point verification is used to confirm that the left support phase window and the right support phase window come from the same step cycle sequence number. The controller reads the generated data. The time of left foot contact with the ground recorded. With the moment of right foot contact ,in For the first The moment of left foot ground contact during a gait cycle For the first The controller determines the timing of the right foot's ground contact event during each gait cycle based on... and The consistency of the period sequence number confirms that the left support phase window and the right support phase window are expressed under the same phase starting point reference.

[0089] In the implementation of the left-right coupling mapping layer, the controller will The input is a left-right coupling mapping layer, which performs coupling mapping processing sequentially according to three physical quantity categories: left-right support force difference, lateral migration of the plantar pressure center, and lateral sway of the pelvis and trunk. For each physical quantity category, the left-right coupling mapping layer configures a one-dimensional learnable piecewise function unit. The left and right support phase window dimensions of this physical quantity category undergo nonlinear transformation using the same set of one-dimensional learnable piecewise function unit parameters, thereby achieving shared one-dimensional learnable piecewise function unit parameters for both left and right dimensions. The one-dimensional learnable piecewise function unit is implemented using a piecewise linear method, and the controller presets the number of segments for this unit. ,in To determine the number of adjacent intervals into which the input axis is divided, the controller... Set each interval Learnable node positions to With the corresponding Each learnable node output value to ,in to Arranged in numerical order, when the input falls into two adjacent node positions. and When the defined interval is reached, the controller outputs a value at the node. and Linear interpolation is performed between the points, and the transformation result is output. Through this implementation, the parameters of the one-dimensional learnable piecewise function unit are obtained from... Determined through training, and directly invoked during inference;

[0090] Within each physical quantity category, the controller transforms the left support phase window dimension and the right support phase window dimension using a shared-parameter one-dimensional learnable piecewise function unit to obtain two transformation results. These two transformation results are then weighted and summed to obtain a two-dimensional coupling representation. This two-dimensional coupling representation is achieved through two sets of independent learnable weighting coefficients. The first set of learnable weighting coefficients outputs the first coupling dimension of the physical quantity category, and the second set outputs the second coupling dimension. This results in a two-dimensional coupling representation for each physical quantity category. The controller denotes the two-dimensional coupling representation of the physical quantity category with the force difference between the left and right supports as follows: , and will Defined as the coupling mapping result of the left support phase window contribution corresponding to this physical quantity category, Defined as the coupling mapping result of the right support phase window contribution corresponding to this physical quantity category, the controller obtains the physical quantity category of the lateral migration of the plantar pressure center in the same way. , And the physical quantity categories of pelvic trunk lateral sway. , The controller will to The data are summarized in a fixed order to form a six-dimensional coupling representation. ,in This is the output of the left-right coupling mapping layer;

[0091] In the implementation of the contrast mapping layer, the controller couples the six dimensions into a representation. Input the contrast mapping layer, and calculate the contrast quantities of the left support phase window coupling representation and the right support phase window coupling representation under the same phase window definition according to the physical quantity category, to form a three-dimensional contrast representation. Specifically, the controller uses the following categories to classify the physical quantities of the force difference between the left and right supports: reduce Obtain the force path comparison quantity The physical quantity category of lateral migration of the plantar pressure center is adopted. reduce Obtain the comparison of the center path of plantar pressure. The physical quantity categories for lateral sway of the pelvis and trunk are adopted. reduce Obtain the comparison of lateral movement paths of the pelvis and trunk. The controller will , , A three-dimensional contrast representation is formed by assembling the components in a fixed order. ,in The three dimensions correspond to the force path comparison, the foot pressure center path comparison, and the pelvic and trunk lateral movement path comparison in sequence.

[0092] In the implementation of the output layer, the output layer adopts a dual-output structure, including a regression output layer and a confidence output layer. The controller will perform three-dimensional comparative representation. The input is a regression output layer, and the regression output layer uses a functional neuron structure. Mapping is performed, and the functional neuron structure is a mapping... Each dimension is configured with a one-dimensional learnable piecewise function unit and a nonlinear transformation is performed. Then, the three-dimensional transformation results are weighted and summed to output a one-dimensional initial off-center load index. ,in For the first The initial off-center load index for each gait cycle is compared to the same three-dimensional characterization by the controller. The input confidence level output layer uses the same functional neuron structure and computational process as the regression output layer, and outputs a one-dimensional bias confidence level. ,in For the first The offloading confidence level for each gait cycle, the parameters of the one-dimensional learnable piecewise function unit and the weighted summation weights are determined by training, and during inference the controller directly calls the parameters to complete the mapping from the six-dimensional lateral sensitive features to the initial offloading index and offloading confidence level.

[0093] In this embodiment, step S4 specifically includes:

[0094] After completing the inference of the improved Kolmogorov-Arnold network off-center load exponential regressor, the controller receives the first... One-dimensional initial bias index for each gait cycle With one-dimensional off-center load confidence ,in The controller presets the off-center load index threshold as the current gait cycle number. With off-center load confidence threshold , respectively and , and The comparison is performed, and the output of the assist ratio branch determination result is shown. The controller also presets the symmetrical left and right assist ratios. The symmetrical left and right assist ratio This is a symmetrical reference value for the left and right side assist ratios, used to directly output consistent assist ratios on both sides when there is no reliable evidence of off-center loading.

[0095] When the assist ratio branch determination result is a trial branch, the controller determines the initial off-center load index. The controller will determine the off-center load direction. A positive value is defined as the load shifting to the left. A negative value is defined as the load shifting to the right, thus obtaining the side corresponding to the off-center load direction. The controller presets a trial offset amplitude. And with symmetrical left and right assist ratios Generate a baseline to test the left and right assist ratios and ,in To test the left-side assist ratio, To test the right-side assist ratio, the generation method involves adding an auxiliary force on the corresponding side in the off-center loading direction. Reduce on the other side and make Keep as Because the test of the left and right assist ratios used a symmetrical offset method with one increasing and the other decreasing, the controller confirmed... and They are not equal, thus ensuring that subsequent actions are based on... There is no zero-difference input when characterizing the probe direction;

[0096] The controller will test the left and right assist ratio. , Apply to the preset probe phase window To generate a trial gait response, the trial phase window Starting from the next support phase window after the assist ratio branch determination, and covering a complete support phase window, the probe disturbance is confined to a single support phase window. The controller operates within the probe phase window. The system acquires the same left and right lower limb joint motion information, pelvic and trunk posture information, and left and right foot plantar force and plantar pressure center information as in step one. It then performs gait phase recognition and phase alignment on this information as in step two to obtain trial lateral sensitivity features. The controller simultaneously reads the lateral sensitive features prior to the application of the probe. ,in To generate the six-dimensional lateral sensitivity features obtained before probing the left and right assist ratio, and and All satisfy the phase-consistent input constraint and are expressed at the same phase starting point;

[0097] The controller will probe lateral sensitive features. Lateral sensitivity features before probing Calculate gait response based on the same phase window. Gait response The controller first processes changes in the left-right force difference, the lateral shift of the plantar pressure center, and the lateral swaying of the pelvis and trunk, following a fixed sequence. and The corresponding dimension is differentially analyzed to obtain three varying original values. These three varying original values ​​are then normalized to form a scalar input that can be directly used for a single-point correction. Specifically, the controller employs a trial phase window. The time average of the sum of the normal forces on the inner and outer soles of the feet is used as the force normalization benchmark. The preset plantar lateral length parameter was used as the migration normalization benchmark. Using a preset roll angle normalization benchmark The controller divides the original value of the change in the difference in force between the left and right sides by... The normalized change in the left and right force differences was obtained. Divide the original value of the lateral migration change of the plantar pressure center by Normalized changes in the lateral migration of the plantar pressure center Divide the original value of the lateral sway change of the pelvis and trunk by Normalized changes in lateral pelvic and trunk sway Thus , , All are dimensionless numerical values ​​and can be fused within the same correction expression;

[0098] The controller is based on gait response. With the trial bias amplitude For the initial off-center load index Perform a single-point correction and output the correction off-center load index. A single-point correction is limited to a single correction action and exits the correction link after the current gait cycle ends. A single-point correction is calculated using a verification formula:

[0099] ;

[0100] in, For the first The corrected off-load index for each gait cycle, For the first The initial off-load index for each gait cycle, To preset the correction gain, , , Preset fusion weights are used to weight and fuse changes in left-right force differences, changes in the lateral migration of the plantar pressure center, and changes in lateral swaying of the pelvis and trunk. , , It is a dimensionless constant. This represents the normalized change in the difference in force between the left and right sides. This represents the normalized lateral migration of the plantar pressure center. This represents the normalized lateral swaying variation of the pelvis and trunk. To test the bias amplitude, which is a dimensionless value, To test the left-side assist ratio, To test the right-side assist ratio, To test the absolute value of the difference in the left and right assist ratios, This is the probe direction symbol, used to align the correction direction with the direction in which the probe left and right assist ratio is applied. To test the phase window The average time of the sum of the normal forces on the soles of the left and right feet. To preset the lateral length parameter of the foot, This serves as a preset roll angle normalization benchmark;

[0101] The controller obtains the corrected off-center load index Then, the left and right assist ratios are generated based on the preset mapping relationship between the off-center load index and the assist ratio. and The mapping relationship is implemented using a piecewise linear mapping method. The controller presets multiple off-center load index segment endpoints and corresponding assist ratio difference endpoints, and performs... Linear interpolation is performed on the segment to obtain the difference in the left and right assist ratios, and then the left and right assist ratios are symmetrically calculated. Based on this, the difference in the left and right assist ratios is allocated to the left assist ratio. Right-side assist ratio ,make Keep as ;

[0102] When the assist ratio branch determination result is a symmetrical branch, the controller does not execute the trial phase window. Exploring laterally sensitive features Gait response Compared to a single-point correction, this directly adjusts the symmetrical left and right assist ratio. The values ​​are respectively assigned to the left-side assist ratio. Right-side assist ratio and will , The output is used in step five for generating subsequent left-right asymmetric assist commands.

[0103] In this embodiment, step S5 specifically includes:

[0104] After obtaining the left and right assist ratios in step four, the controller will adjust the left assist ratio. Right-side assist ratio Input limiting module, where The proportional coefficient used to characterize the relative symmetrical assist reference command of the left drive channel. The proportional coefficient used to characterize the relative symmetrical assist reference command of the right drive channel is preset by the limiting module to limit the left and right assist ratios. The left and right assist ratio limiting constraints are determined by the lower limit of the left assist ratio. Left-side assist ratio limit Lower limit of right-side assist ratio Upper limit of right-side assist ratio Composition, controller Perform truncation, the truncation rule is when Less than When Set as ,when Greater than When Set as All other cases remain unchanged, and the controller... Following the same truncation rule as the left side, the controller outputs the truncated result as the left and right assist ratios for amplitude limiting. and ,in The left-side assist ratio is limited. The right-side assist ratio is limited;

[0105] The controller performs a speed limit update by differentiating the left and right assist ratios of the limiting function with those of the previous control cycle. The controller defines the control cycle index as follows: And read the left-side assist ratio output from the previous control cycle to the drive channel. Right-side assist ratio The controller calculates the differential component. and The controller constrains the left and right assist ratio change rate. , Speed ​​limit updates are performed, and the left and right assist ratio change rate constraint is adjusted based on the control cycle duration. Maximum rate of change of the assist ratio Confirmed, the controller sets the allowable variation per cycle to... The controller respectively , Implement symmetrical limiting, so that No more than and No more than The controller will limit the amplitude. , Superimposed on each , The left and right assist ratios are obtained by limiting amplitude and speed. and ,in To limit the amplitude and speed, the left-side assist ratio, The right-side assist ratio is for limiting amplitude and speed;

[0106] The controller applies the amplitude and speed limiting left and right assist ratios to the preset symmetrical assist reference command. ,in For timestamp The corresponding sampling time corresponds to the same auxiliary output reference amount on both sides, timestamp. For the first The timestamp of the next sampling moment, the symmetric assist reference command Expressed using a joint target assist torque sequence or a joint target assist force sequence, the controller performs proportional scaling on the left drive channel. Multiply Get the target on the left to help output The controller performs proportional scaling on the right drive channel, turning... Multiply Get the target's output on the right. When the symmetrical assist reference command When it is a multi-joint vector, the controller uses the same scaling factor for each joint component in the vector to ensure that the left and right drive channels scale the full joint assist amplitude at the same ratio within the same control cycle.

[0107] The controller converts the left and right target assist outputs into asymmetrical left and right assist commands that match the drive channels of the exoskeleton actuators. When the exoskeleton actuators are motor-driven, the controller will... , The controller converts the actuator torque constant and transmission ratio into left and right current commands, and then outputs them after superimposing the preset bias compensation of the drive channel. When the exoskeleton actuator is hydraulically or pneumatically driven, the controller will... , After converting the valve control gain into left and right valve opening commands or left and right pressure commands, the controller outputs them. The controller uses the phase window corresponding to step four to maintain the continuity of the left and right asymmetrical assist commands within the phase window, specifically updating once at the start of the phase window. , Within the phase window, it continuously outputs the left-right asymmetric assist command calculated by the ratio, until the phase window ends and the next ratio update and output begins.

[0108] In this embodiment, step S6 specifically includes:

[0109] After the controller generates the left-right asymmetric assist command in step five, it records the left-right asymmetric assist command of the left drive channel as... The asymmetrical assist command of the right drive channel is recorded as ,in For the first The timestamp of the next control update time. To control index updates, the controller also maintains the current gait phase sequence. Phase window identifier , where gait phase sequence A phase window identifier for a chronologically recorded sequence of ground contact and ground lift events. To control update timing The number of the phase window to which it belongs;

[0110] The controller continuously updates the gait phase sequence using the forces on the left and right soles during the assist output period. The controller at the sensor sampling time Obtain the force on the left foot. Force on the right foot ,in For the first The timestamp of the next sensor sampling moment For sensor sampling indexing, the controller presets a grounding threshold. With ground threshold ,when From less than Become greater than or equal to Record the left foot's ground contact event at the time. From greater than Change to less than or equal to The controller records the left foot leaving the ground event. The same rules are used to record the right foot ground contact event and the right foot ground lift-off event. The controller writes these events into the gait phase sequence in chronological order. The phase window starts with the ground contact event and ends with the ground lift event, forming a phase window with adjacent start and end points. This allows the gait phase sequence to be used at any given time. Determine the phase window identifier ;

[0111] The controller will , Phase window identifier The data is then associated and encapsulated to form a timing-tagged drive data frame, and sent to the exoskeleton actuator according to the left and right drive channels respectively, so that the exoskeleton actuator can determine the timing based on the phase window identifier. Determine the execution timing and hold interval of the left-right asymmetric assist command;

[0112] After receiving asymmetrical assistance commands from the left and right sides, the exoskeleton actuator marks the phase window. Within the corresponding phase window, the left and right side assist outputs are synchronously driven. Specifically, the exoskeleton actuator latches the phase window after detecting the arrival of the phase window start time. and The system continuously outputs latch values ​​within the phase window until the phase window ends, thus maintaining the timing consistency of the left and right assist outputs using the continuous control method within the phase window. When the phase window ends and the next phase window begins, the exoskeleton actuator updates the latch values ​​according to the next phase window identifier, achieving discrete updates at the phase window boundary and continuous output within the phase window.

[0113] During the assisted output process, the controller simultaneously acquires information on the joint motion of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet. The controller records the joint motion information of the left lower limb as follows: The right lower limb joint movement information is recorded as The pelvic trunk posture information is recorded as The force applied to the sole of the left foot is recorded as The force applied to the sole of the right foot is recorded as The information on the center of pressure on the left foot is recorded as follows: The information on the center of pressure on the right foot is recorded as follows: The controller calculates the phase window identifier for each set of synchronously sampled data. ,in For timestamps In gait phase sequence The controller will assign the phase window number below. and Binding storage allows subsequent construction of laterally sensitive features to be directly segmented and aligned according to the phase window identifier, thereby maintaining phase consistency input constraints;

[0114] The controller synchronously acquires and binds the phase window identifier to the data, which is written into a circular buffer. The circular buffer organizes data segments using the phase window identifier as an index. When the controller starts step one in the next control cycle, it reads the data from the circular buffer in chronological order. , , , , , , As the input for step one in subsequent cycles, and using the timestamp. Phase window identifier Used for gait phase recognition and phase alignment processing in the next cycle.

[0115] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework, characterized in that, include: S1. During the user's walking while carrying objects with one hand or to one side, acquire information on the joint movement of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet. S2. Based on the joint motion information of the left and right lower limbs, the posture information of the pelvis and trunk, and the force and pressure center information of the left and right feet, gait phase recognition and phase alignment are performed. Based on the phase alignment results, the difference in support force between the left and right sides, the lateral migration of the pressure center of the foot, and the lateral sway of the pelvis and trunk are calculated to form laterally sensitive features. S3. Input the lateral sensitive features into the improved Kolmogorov-Arnold network off-center load index regressor for inference. The improved Kolmogorov-Arnold network off-center load index regressor satisfies the phase-consistent input constraint and sets a left-right coupling mapping in the network structure to compare and map the contributions of the left and right supports within the same phase window. This directly regresses the lateral offset relationship caused by the off-center load to the initial off-center load index and generates an off-center load confidence level to characterize the reliability of the initial off-center load index. S4. Based on the initial off-center load index and off-center load confidence, the assist ratio is determined. When the off-center load confidence is higher than the off-center load confidence threshold and the initial off-center load index is higher than the off-center load index threshold, a trial left and right assist ratio with a slight offset is generated and applied to form a trial gait response. Based on the trial gait response, the gait response quantity including the change in left and right force difference, the change in the lateral migration of the plantar pressure center and the change in the lateral swing of the pelvis and trunk is calculated. Based on the gait response quantity, the initial off-center load index is corrected once to obtain the corrected off-center load index. Then, the left and right assist ratio is generated from the corrected off-center load index. When the off-center load confidence is lower than the off-center load confidence threshold or the initial off-center load index is lower than the off-center load index threshold, the symmetrical left and right assist ratio is determined as the left and right assist ratio. S5. Under the constraints of left and right assist ratio limit and left and right assist ratio change rate, map the left and right assist ratio to left and right asymmetric assist command. S6. Send the left and right asymmetrical assist command to the exoskeleton actuator to drive the left and right side assist output.

2. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, S1 specifically refers to: A multi-sensor acquisition unit triggered by a unified clock is used to synchronously sample data from the left and right sides, so that the joint motion information of the left and right lower limbs, the posture information of the pelvis and trunk, and the force and pressure center information of the left and right feet have the same time reference. Hip joint motion information, knee joint motion information, and ankle joint motion information are collected from the left and right lower limbs respectively, and the hip joint motion information, knee joint motion information, and ankle joint motion information are summarized to form the joint motion information of the left and right lower limbs. Attitude angles and angular velocities were collected for the pelvis and trunk respectively, and the pelvic and trunk attitudes were mapped to the same body coordinate system to form pelvic and trunk attitude information. The sequence of normal force and pressure center position of the left and right soles is collected respectively, and the force and pressure center information of the left and right soles is output according to the left and right sole channels.

3. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, S2 specifically refers to: The ground contact and takeoff events of the left and right feet are identified based on the ground contact threshold and takeoff threshold of the left and right feet, and the consistency of the ground contact and takeoff events is confirmed by combining the joint motion information of the left and right lower limbs, resulting in a gait phase sequence that includes the left support phase and the right support phase. Based on the gait phase sequence, the pelvic and trunk posture information and the left and right foot plantar pressure center information are divided into left support phase window and right support phase window, and the data in each phase window are aligned to the same phase starting point to form a phase alignment result. The left and right support force difference is calculated according to the phase window order based on the phase alignment result. The left and right support force difference is obtained by subtracting the right foot force from the left foot force. The lateral migration of the foot pressure center and the lateral swing of the pelvis and trunk are calculated based on the horizontal axis of the body coordinate system within the same phase window. The left and right support force difference of the left and right support phase window, the left and right support force difference of the right and right support phase window, the lateral migration of the plantar pressure center of the left and right support phase windows, and the lateral sway of the pelvis and trunk of the left and right support phase windows are spliced ​​together in a fixed dimensional order to form a six-dimensional lateral sensitivity feature, which is used as the input of the improved Kolmogorov-Arnold network off-center loading index regressor.

4. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 3, characterized in that, The six-dimensional lateral sensitive features are extracted from the phase alignment results in one step using a phase-first extraction formula. The phase-first extraction formula is as follows: ； in, For the first Six-dimensional lateral sensitivity features per gait cycle. The difference in force between the left and right supports of the left support phase window. The difference in force between the left and right supports of the right support phase window. This represents the lateral migration of the plantar pressure center within the left support phase window. This represents the lateral migration of the plantar pressure center within the right support phase window. This represents the lateral sway of the pelvis and trunk within the left support phase window. The lateral sway of the pelvis and trunk within the right support phase window. This represents the standard number of sampling points for the phase window. The discrete index within the phase window and its value range is to , Left support phase window Inner The normal force on the left foot sole at each resampling point Left support phase window Inner The right foot normal force at each resampling point For right support phase window Inner The normal force on the left foot sole at each resampling point For right support phase window Inner Normal force on the right foot sole at each resampling point Left support phase window Inner The center of pressure on the left foot at each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis For right support phase window Inner The right plantar pressure center of each resampling point is located in the body coordinate system. The coordinate values ​​on the horizontal axis Left support phase window Inner The combined roll angle of each resampling point, For right support phase window Inner The combined roll angle of each resampled point.

5. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, S3 specifically refers to: Lateral sensitive features are organized into six-dimensional lateral sensitive features in a fixed dimensional order, and phase consistency input constraint verification is performed on the six-dimensional lateral sensitive features. The phase consistency input constraint verification is used to confirm that the six-dimensional lateral sensitive features correspond to the left support phase window and the right support phase window respectively and are expressed under the same phase starting point. The six-dimensional lateral sensitive features are input into the left-right coupling mapping layer, and coupling mapping is performed sequentially according to the physical quantity categories of left-right support force difference, lateral migration of plantar pressure center, and lateral swing of pelvis and trunk. Within each physical quantity category, one-dimensional learnable piecewise function units are configured for the left support phase window dimension and the right support phase window dimension, respectively. The two dimensions are nonlinearly transformed using the same set of one-dimensional learnable piecewise function unit parameters to achieve left-right coupling mapping where the left and right dimensions share one-dimensional learnable piecewise function unit parameters. The nonlinear transformation results within each physical quantity category are weighted and summed to obtain a two-dimensional coupling representation of that physical quantity category. The two-dimensional coupling representations of the three physical quantity categories are then summarized in a fixed order to form a six-dimensional coupling representation. The six-dimensional coupling representation is input into the contrast mapping layer, and the contrast quantity between the left support phase window coupling representation and the right support phase window coupling representation is calculated under the same phase window definition according to the physical quantity category, forming a three-dimensional contrast representation. The three-dimensional contrast representation corresponds to the force path contrast, the plantar pressure center path contrast, and the pelvic trunk lateral movement path contrast in sequence. The three-dimensional contrastive representation is input into the regression output layer. The regression output layer uses a function neuron structure to map the three-dimensional contrastive representation. The function neuron structure configures a one-dimensional learnable piecewise function unit for each dimension of the input, performs a nonlinear transformation, and then performs a weighted summation to output a one-dimensional initial bias index. The three-dimensional contrastive representation is input into the confidence output layer. The confidence output layer uses the same function neuron structure as the regression output layer to map the three-dimensional contrastive representation and outputs a one-dimensional bias confidence.

6. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, S4 specifically refers to: Receive the initial off-center load index and off-center load confidence level, and compare them with the off-center load index threshold and the off-center load confidence level threshold respectively, and output the assist ratio branch determination result; When the result of the assist ratio branch determination is a trial branch, the off-load direction is determined according to the initial off-load index, and a preset trial offset amplitude is introduced based on the symmetrical left and right assist ratio to generate the trial left and right assist ratio. The trial left and right assist ratio increases the assist ratio on the side corresponding to the off-load direction and decreases the assist ratio on the other side to keep the assist ratio benchmark consistent. The trial left and right assistance ratio is applied to the preset trial phase window to form a trial gait response. Within the trial phase window, the left and right lower limb joint motion information, pelvic trunk posture information, left and right foot plantar force and plantar pressure center information are obtained in the same way as in step one. The same gait phase recognition and phase alignment as in step two are performed on the trial gait response. The difference in left and right support force, the lateral migration of the plantar pressure center and the lateral swing of the pelvis and trunk are calculated on the phase alignment results to obtain the trial lateral sensitivity characteristics. The lateral sensitivity features and the anterior lateral sensitivity features are used to calculate the changes in left and right force difference, the lateral migration of the plantar pressure center, and the lateral swing of the pelvis and trunk according to the same phase window, and the changes are fused in a fixed order to form the gait response quantity. Based on the gait response and the preset trial bias amplitude, a single-point correction is performed on the initial off-center load index. The single-point correction is limited to a single correction action and outputs the corrected off-center load index. Then, the left and right assist ratios are generated according to the corrected off-center load index through the preset off-center load index to assist ratio mapping relationship. When the result of the assist ratio branch determination is a symmetrical branch, the symmetrical left and right assist ratio is directly determined as the left and right assist ratio.

7. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 6, characterized in that, The single-point correction is based on the gait response and a preset trial bias amplitude to correct the initial off-load index and output the corrected off-load index. The single-point correction is calculated using a verification formula, which is specifically as follows: ； in, For the first The corrected off-load index for each gait cycle, For the first The initial off-load index for each gait cycle. To preset the correction gain, , , Preset fusion weights are used to weight and fuse changes in left-right force differences, changes in the lateral migration of the plantar pressure center, and changes in lateral swaying of the pelvis and trunk. , , It is a dimensionless constant. This represents the normalized change in the difference in force between the left and right sides. This represents the normalized lateral migration of the plantar pressure center. This represents the normalized lateral swaying variation of the pelvis and trunk. To test the bias amplitude, which is a dimensionless value, To test the left-side assist ratio, To test the right-side assist ratio, To test the absolute value of the difference in the left and right assist ratios, This is the probe direction symbol, used to align the correction direction with the direction in which the probe left and right assist ratio is applied. To test the phase window The average time of the sum of the normal forces on the soles of the left and right feet. To preset the lateral length parameter of the foot, This serves as the preset roll angle normalization benchmark.

8. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, S5 specifically refers to: Input the left and right assist ratio obtained in step four into the amplitude limiting module, and truncate the left and right assist ratio according to the amplitude limiting constraint to obtain the amplitude-limited left and right assist ratio. The left and right assist ratio of the amplitude limit is differentially compared with the left and right assist ratio of the previous control cycle. The difference result is then updated with a speed limit based on the constraint of the change rate of the left and right assist ratio to obtain the amplitude limit and speed limit left and right assist ratio. Apply the limiting and speed-limiting left and right assist ratios to the preset symmetrical assist baseline command, and scale the left and right drive channels proportionally to obtain the target assist outputs on the left and right sides. The left and right target assist outputs are converted into left and right asymmetrical assist commands that match the drive channel of the exoskeleton actuator, and the left and right asymmetrical assist commands are kept continuous within the phase window corresponding to step four.

9. The method for allocating off-center load assistance based on an artificial intelligence network and a left-right assist ratio scheduling framework according to claim 1, characterized in that, Step S6 is as follows: The left and right asymmetric assistance commands obtained in step five are sent to the exoskeleton actuators according to the left and right drive channels, respectively, and the left and right asymmetric assistance commands are associated with the current gait phase sequence to determine the execution timing. The exoskeleton actuator drives the left and right side assist outputs according to the left and right asymmetrical assist commands, and maintains the timing consistency of the left and right side assist outputs by using the continuous control method within the phase window during the assist output process; During the assisted output process, the motion information of the left and right lower limb joints, the posture information of the pelvis and trunk, the force on the left and right soles and the information of the sole pressure center are acquired simultaneously, and the information is aligned with the corresponding phase window to maintain phase consistency input constraints. The synchronously acquired information on the joint movement of the left and right lower limbs, the posture of the pelvis and trunk, and the force and pressure center of the left and right feet are used as the input for step one of the subsequent cycles.

Images