Exoskeleton safety redundancy strategy and switching method based on anomaly detection

Abstract

The present application relates to an exoskeleton safety redundancy strategy and switching method based on anomaly detection, which collects the tension state between the user's active control intention and the response execution of the exoskeleton system, defines potential anomalies when there is persistent asynchronous tension between the user's expected control direction and the system's inertial behavior as a redundancy strategy trigger precondition; After the anomaly is triggered, the system outputs control suggestions, and judges whether the control right should be taken over by the redundancy strategy, continue to maintain the main control, or enter the forced frozen state by combining the user's reverse force, posture offset trend behavior; Dynamically adjust the intervention level of the redundancy control strategy to realize continuous control evolution from low invasion to high takeover; If it is detected in the control evolution process that there is an irreconcilable confrontation trend between the user's continuous reverse intention input and the system control output, and the system intervention imbalance continues to expand, the system enters the soft frozen or conservative posture mode; In the frozen state, it is judged whether the human-machine trust relationship is rebuilt.

Description

Technical Field

[0001] This invention relates to the field of exoskeleton safety redundancy technology, specifically to an exoskeleton safety redundancy strategy and switching method based on anomaly detection. Background Technology

[0002] Existing technologies, such as the one disclosed in Chinese patent document CN114129392A, are adaptive redundant-drive exoskeleton rehabilitation robots with adjustable fingertip force. Their core innovation lies in collecting fingertip force through a finger linkage mechanism, servo motor drive, and pressure sensors on the palmar platform, combined with a control module to adaptively switch between active and passive modes, thereby ensuring grip stability and training effectiveness. However, this approach still has many shortcomings and drawbacks. First, the control logic of this solution mainly revolves around the magnitude and duration of fingertip force. Its anomaly judgment condition is singular, relying solely on the setting of fingertip force thresholds and safety thresholds, lacking multi-dimensional human-machine collaborative state recognition capabilities. In complex human-machine interaction scenarios, such as when the user experiences movement delays, abnormal electromyographic signals, or sudden changes in joint movement patterns, simply relying on force threshold adjustment is prone to misjudgment, potentially leading to delayed or even failed redundant control intervention, failing to proactively respond to potentially high-risk states.

[0003] Secondly, while the invention proposes switching between active and passive modes, the switching mechanism is based on binary threshold logic. It enters active mode when the threshold is reached and passive mode when the threshold is not reached. This rigid mode transition fails to reflect a gradual evolution from low intrusion to high control, potentially leading to sudden deprivation of user intent or excessive delegation of power, thus weakening the smoothness of human-computer interaction and user confidence. Thirdly, although the solution emphasizes compensatory control of fingertip position through a Disturbance Observer (DOB), it remains a local optimization targeting electrical and mechanical disturbances in the servo motor. It lacks a global analysis of the overall user behavior trend and cannot identify continuous indicators during action execution, such as trajectory repetition rate, directional consistency, or speed change rate. As a result, the system can only correct errors after they occur at the end of the action, failing to provide predictive protection when abnormal signs appear. This reactive compensation approach is inadequate for high-risk actions.

[0004] Fourth, this patent does not address the dynamic maintenance of the human-machine trust relationship; its logic remains solely based on mechanical judgments using sensor data, i.e., stopping or switching training modes when force exceeds a threshold. However, in long-term rehabilitation training or complex task execution, the user's trust in the exoskeleton is a crucial factor affecting user experience and rehabilitation outcomes. Frequent false triggers or excessive interventions by the system can easily lead to user resistance. Fifth, existing technologies are primarily designed for localized applications in hand function rehabilitation, focusing on finger grasping training and lacking adaptability for broader movement scenarios (such as gait assistance and lower limb exoskeleton control). Finally, regarding the flexibility of the control strategy, this solution's training mode selection leans towards static presets, requiring users to train under fixed action libraries and threshold conditions, failing to achieve dynamic adjustment based on individual differences or real-time conditions. This one-size-fits-all approach may lead to the risk of undertraining or overtraining for some patients. Summary of the Invention

[0005] The purpose of this invention is to provide an exoskeleton safety redundancy strategy and switching method based on anomaly detection, thereby solving some of the drawbacks and shortcomings pointed out in the background art.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: an exoskeleton safety redundancy strategy and switching method based on anomaly detection, including: collecting the tension state between the user's active control intention and the response execution of the exoskeleton system; when a continuous asynchronous tension is detected between the user's expected control direction and the system's inertial behavior, it is defined as a potential anomaly and used as a precondition for triggering the redundancy strategy.

[0007] After an anomaly is triggered, the system allows the main controller and redundant controller to output control suggestions simultaneously. It also combines the user's reverse force application and attitude deviation trend behavior to determine whether control should be taken over by the redundant strategy, continue to maintain the main control, or enter a forced freeze state. Based on the clarity of the user's current behavior and the responsiveness of the system intervention, it constructs a coupling relationship between behavioral tendency and control intervention level, dynamically adjusts the intervention level of the redundant control strategy, and realizes continuous control evolution from low intrusion to high takeover.

[0008] If an irreconcilable antagonistic trend is detected between the user's continuous reverse intention input and the system's control output during the control evolution process, and the system intervention ratio imbalance continues to expand, then the system will actively withdraw control and enter a soft freeze or conservative posture mode, terminating further control command output. In the frozen state, the system will determine whether the human-machine trust relationship has been rebuilt by identifying whether the user's behavior has re-demonstrated active synchronization or posture surrender actions.

[0009] Furthermore, the determination of the tension state includes: identifying the response time difference between the user's active force signal and the exoskeleton system's execution response, wherein the time difference exceeding a preset dynamic response threshold is considered a potential abnormal state; when determining an abnormal state, the joint determination result of at least three indicators is referenced, including: joint angle deviation, changes in electromyographic activation mode, and consistency of action response.

[0010] Furthermore, during the parallel output phase of the main controller and the redundant controller, the control output weights of the two are dynamically allocated based on the execution trend of the user's actions. When determining the ownership of control, the user's action compliance performance is taken into account. If the user continuously performs actions consistent with the output direction of the main controller within a unit time window, the main controller is given priority.

[0011] Furthermore, the intervention level of the redundancy control strategy is adjusted in stages according to the stability index of user action trends, including the rate of change of action speed, trajectory repetition rate, and consistency of behavior direction; when the redundancy control strategy intervenes, the output intervention intensity is adjusted according to the overlap rate between the user's actual execution trajectory and the target trajectory.

[0012] Furthermore, the rate of change of movement speed is the standard deviation of the user's gait speed within a preset time window, and the redundancy control strategy intervention level is increased by one level when the standard deviation continuously exceeds the set fluctuation threshold; the trajectory repetition rate is calculated by comparing the trajectory overlap area and path difference of two consecutive movement cycles, and if the repetition rate shows a decreasing trend in multiple consecutive cycles, the redundancy strategy level is increased.

[0013] Furthermore, the consistency of the behavioral direction is calculated by detecting the change in the angle between the expected movement direction and the actual movement direction within each action cycle, and the average value of this angle change is used as a consistency reference index; the initial level of the redundancy control strategy is determined by the user's historical action trend stability score, which is the weighted average of multiple action stability indices over past control cycles.

[0014] Furthermore, when calculating the overlap rate between the execution trajectory and the target trajectory, the actual trajectory is projected onto the corresponding time period of the target trajectory for matching and comparison based on a preset time synchronization window; the output intervention intensity adjustment uses the overlap rate as a reference signal and adopts a proportional adjustment method to perform linear interpolation of the output value, so that the change in the controller output intensity smoothly transitions over multiple consecutive cycles; when the trajectory overlap rate is lower than a set threshold, and the user's behavior direction is continuously opposite to the system control output direction for more than two consecutive cycles, the redundancy control strategy automatically enters the highest intervention level.

[0015] Furthermore, the time synchronization window has an asymmetric structure, extending the user trajectory forward by multiple frames and backward by a few frames, so that the projection matching focuses on the early trend of user behavior; the first derivative change rate is monitored as the trajectory overlap rate continues to change, and when the downward trend of the overlap rate suddenly reverses or steepens sharply, the redundancy intervention level is triggered in advance.

[0016] Furthermore, before the actual trajectory is projected onto the target trajectory, the time scale of the target trajectory is adaptively adjusted based on the user's current step frequency to align the rhythms in high-frequency motion states; when the overlap rate is lower than the historical average and the descent rate exceeds the set threshold, the redundancy level preparation mode is activated in advance even if the set threshold is not reached.

[0017] Furthermore, the opposite state between the user behavior direction and the control output direction is calculated by the unit change rate of the spatial direction difference vector. If the rate exceeds a threshold, it is considered a high mismatch risk segment. If the overlap rate oscillates rapidly, the system adds a change rate threshold to the linear interpolation output value to avoid unexpected jumps in the output within a continuous period.

[0018] Based on the above design, a discriminant function integrating directional vector difference rate analysis and overlap rate change sensitivity identification is used to determine whether the control system should perform strong intervention or limit output jump behavior. Its calculation formula is as follows:

[0019]

[0020] in:

[0021] Current moment The intervention sensitivity function value is used to determine whether redundant control transitions are triggered or interpolation jumps are suppressed. The unit time interval of the system control cycle; The magnitude of the spatial unit difference vector between the user's intention direction and the system control direction at the current moment, i.e., the degree of mismatch; The first time derivative of the unit difference vector in the direction of user intent; The normalized standard deviation of the current overlap rate fluctuation; Overlap rate The second derivative over time represents the acceleration characteristic of the oscillation trend;

[0022] function The derivation process includes:

[0023] Define the user intent direction vector and the system control output direction vector as unit vectors. Then, calculate the difference vector between them and take its modulus, denoted as [equation missing]. This is used to characterize the degree of orientation mismatch; simultaneously, to determine the evolution rate of this mismatch, the first derivative of the magnitude of the difference vector over time is calculated. , representing the rate of directional mismatch, thus the first part of the risk metric can be derived as follows: This item will increase rapidly when the deviation between user and system behavior intensifies rapidly, serving as an important dynamic indicator for redundancy control.

[0024] On the other hand, to reflect the stability of the matching between the actual trajectory and the target trajectory during the system control process, a definition is made. The current trajectory overlap rate (given by the sovereign method); if the overlap rate fluctuates drastically over multiple periods, it will cause abnormal jumps in the output interpolation, therefore... Find its second derivative This is used to measure the acceleration of the trajectory matching trend; to increase the relative weight of this term under different states, a normalized standard deviation term for overlap rate fluctuation is introduced. This term reflects the magnitude of the perturbation in the trajectory matching curve during the most recent period; multiplying the two yields the second metric. This indicates the risk of system control instability caused by trajectory fluctuations;

[0025] In summary, the two risk terms described above respectively characterize directional antagonistic risk and trajectory disturbance risk. By linearly superimposing them and normalizing them over a unit time period, a complete system intervention sensitivity function can be constructed. Finally, when the function value Exceeding the pre-set system security intervention threshold When the system determines that the current state is a high-risk segment, it can trigger the highest level of redundancy control intervention or start the output change rate suppression mechanism in any case where there is obvious direction mismatch or severe trajectory fluctuation, so as to ensure the safety of system operation and the stability of human-machine collaboration.

[0026] The beneficial effects of this invention are as follows: By introducing a safety redundancy strategy and switching method based on anomaly detection into the exoskeleton control system, this invention achieves dynamic safety assurance and control continuity in the human-computer interaction process. Through real-time acquisition and analysis of the tension between the user's active control intentions and the system's execution behavior, this method can identify potential anomalies at an early stage, including orientation mismatch, decreased action consistency, and deterioration of trajectory overlap. Therefore, the system can not only establish preconditions for redundant control intervention before anomalies occur, but also perform graded responses based on the stability indicators of user behavior trends and historical behavior data, enabling a smooth transition from low intervention to high control, significantly improving the safety and reliability of the exoskeleton in complex motion environments.

[0027] Furthermore, by comparing trajectory projections within the time synchronization window, calculating the direction mismatch rate, and dynamically interpolating and adjusting the trajectory overlap rate, the system avoids sudden jumps or unnecessary oscillations during control intensity adjustments. When there is an irreconcilable conflict between user behavior and system control direction, the system can promptly enter a freeze or conservative mode to prevent secondary risks caused by excessive intervention; and when the user regains synchronization or relinquishes posture, the control level can be gradually restored, achieving a dual balance between safety and flexibility. This invention not only improves the sensitivity and response speed of the exoskeleton system to abnormal states but also establishes a recoverable human-machine trust loop, which helps to significantly improve comfort and safety during long-term wear. Attached Figure Description

[0028] Figure 1 This is the main flowchart of exoskeleton abnormality detection and redundancy switching in this invention.

[0029] Figure 2 This is the main flowchart of the multi-source signal fusion and redundancy dynamic intervention of the present invention.

[0030] Figure 3 This is a diagram illustrating the trajectory-direction dual-sensitive early redundant intervention relationship of the present invention.

[0031] Figure 4 This is the main flowchart of the redundancy hierarchical control for exoskeleton rehabilitation in Embodiment 1 of the present invention.

[0032] Figure 5 This is a flowchart of the step frequency adaptive and trajectory-sensitive redundancy adjustment function in Embodiment 2 of the present invention. Detailed Implementation

[0033] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] Combined with appendix Figure 1This invention relates to an exoskeleton safety redundancy strategy and switching method based on anomaly detection. During the operation of the exoskeleton system, sensors deployed on the user's body and key joints or drive units of the exoskeleton collect the user's active control intention signals in real time. These intention signals can originate from muscle activation patterns detected by electromyography (EMG) sensors, movement trends acquired by joint angle sensors, and force sensor feedback generated by the user during the execution of actions. Simultaneously, the system acquires the response signals of the exoskeleton execution units, including the angular position of the execution joints, movement speed, output torque, and overall dynamic state, thereby establishing a coupling relationship between the user's active control intention and the exoskeleton system's execution behavior. Based on the above coupling relationship, the tension state between the two is calculated, that is, whether there is a difference between the force or displacement trend corresponding to the user's expected control direction and the inertial response direction of the exoskeleton system. When this difference is manifested as asynchrony within a continuous time window and the tension continues to accumulate, the system determines it as a potential abnormal state. This potential abnormal state serves as a prerequisite for triggering subsequent redundancy strategies, ensuring that the system promptly enters the anomaly management process when there is a disconnect between the user's intention and the exoskeleton response, thus providing a reliable basis for the intervention of redundant control strategies and avoiding control failures and safety risks caused by delayed detection or misjudgment.

[0035] Upon triggering of an abnormal state, the system enters a dual-controller parallel operation mode. This means the primary controller and redundant controller simultaneously generate control suggestions within the same control cycle. The primary controller derives its output based on the user's intended movement, while the redundant controller generates corrective or safety-priority control schemes based on the anomaly detection mechanism. Subsequently, the system comprehensively assesses the user's real-time reverse force signals, the overall attitude deviation trend, and the consistency of the user's continuous actions to determine whether control should be taken over by the redundant strategy, whether primary control should continue, or whether a forced freeze state should be entered. During this assessment, the system analyzes the clarity of the user's actions (e.g., the consistency between the direction and speed of the action) and the system's... The response matching degree of the intervention output establishes a coupling relationship between behavioral tendency and the degree of control intervention. This coupling relationship is defined as a dynamic weight mapping model that can adjust the intervention level of redundant control strategies according to the stability of user behavior and the effect of system output. This achieves a balance between maintaining user initiative and ensuring system security. When the user exhibits clear compliant actions, the system maintains a low-intrusion state, and the redundant controller only makes slight corrections. When the user's actions become uncertain or continuously conflict with the system control, the intervention level of the redundant strategy gradually increases, transitioning from a low-intrusion state to a high-takeover state, forming a continuous control evolution mechanism that avoids control abruptness caused by a single jump.

[0036] During the control evolution process, the system continuously monitors the matching relationship between the user's intention input signal and the output of the exoskeleton execution unit. When it detects that the user's reverse intention input and the system control output are consistently directionally opposite or mechanically cancel each other out over multiple consecutive control cycles, and comprehensive calculations reveal that this antagonistic trend is irreconcilable, manifested as a gradual imbalance in the system intervention ratio and a continuous widening of the difference between the user's input and the system output, the system will proactively trigger the control withdrawal mechanism. This terminates further intervention of redundant control strategies, causing the exoskeleton system to transition from normal operating mode to a soft-freeze state or a conservative posture mode. In this state, the controller stops applying new drive commands to the execution joints, only maintaining the most... The system provides necessary posture support to prevent users from falling or causing danger, while reducing joint output torque or movement speed to ensure safety. In the frozen state, the system does not immediately resume normal operation. Instead, it continuously identifies whether the user has resumed active synchronous movements in the same direction as the system output, or whether the user has re-accepted control of the exoskeleton through natural body posture. When the above conditions are met, the system determines that the human-machine trust relationship has been re-established, and then gradually unfreezes the state, restoring the normal operation of the control strategy. This ensures that risks can be reduced in time when there is an uncoordinated conflict between the user and the exoskeleton, and smoothly transitions back to normal interactive control after the trust relationship is rebuilt.

[0037] Combined with appendix Figure 2 The tension state determination process is completed through multi-source signal fusion analysis. The system collects the user's active force signals in real time through electromyography sensors deployed at the user's muscle groups, and simultaneously records the execution response signals of the exoskeleton joint drive unit. When the response time difference between the two exceeds the dynamic response threshold set by the system, it is determined that there is a potential abnormal state. To avoid misjudgment caused by a single indicator, the system further introduces at least three joint judgment indicators for comprehensive analysis: one is the joint angle deviation indicator, which compares the difference between the real-time angle of the exoskeleton's execution joint and the user's expected angle or action template. If it exceeds the tolerance range, it is counted as an abnormal state. The system employs a multi-factor approach, which considers several factors. First, it considers the change in electromyographic activation patterns. Second, it compares changes in amplitude distribution, activation timing, and frequency domain characteristics of current electromyographic signals to identify any significant deviations from normal movement patterns. Third, it considers the consistency of movement response. The system calculates the degree of overlap between the user's expected movement trajectory and the exoskeleton's actual execution trajectory in time and space. If the consistency significantly decreases, it is considered a potential risk. By combining these multiple indicators, the system uses a weighted judgment method. When the response time difference exceeds a threshold and two or more of the three indicators meet the abnormal conditions, it is ultimately confirmed to be in a potential abnormal state, thus providing a reliable basis for triggering redundant control strategies.

[0038] During the parallel output phase of the main controller and redundant controller, the system dynamically monitors the user's action execution trend to allocate the control output weights of the two controllers. Within the same control cycle, the main controller generates a normal output signal based on the user's expected action, while the redundant controller generates a compensation or correction signal based on safety judgment. The system uses the user's action trend as the basis for weight adjustment. When the user's action is detected to have clear continuity and directionality, the weight of the main controller is increased to ensure the user's intention takes precedence. When the user's action is detected to have uncertainty or a significant deviation from the expected trajectory, the weight of the redundant controller is increased to enhance safety intervention capabilities. In the process of determining control ownership, the system further introduces user action compliance as an auxiliary reference. The compliance level is determined by collecting the user's electromyographic signal pattern, joint angle changes, and the matching degree between the execution trajectory and the main controller's output direction within a unit time window. If the user continuously executes actions consistent with the main controller's output direction within this time window, the main controller's priority remains unchanged, and the redundant controller is in a low-weight state, only making slight corrections when necessary. This maximizes user dominance while ensuring safety redundancy, achieving reasonable allocation and dynamic switching of control rights.

[0039] The intervention level of the redundancy control strategy is not fixed, but dynamically adjusted based on stability indicators of user movement trends. These stability indicators include at least the rate of change of movement speed, trajectory repetition rate, and behavioral direction consistency. The rate of change of movement speed is obtained by calculating the standard deviation of gait or limb movement speed within a preset time window by collecting real-time user joint motion data; this indicator reflects whether user movements are volatile. The trajectory repetition rate is obtained by comparing the movement trajectories of multiple consecutive movement cycles and calculating their overlap area and path difference; this indicator characterizes the repeatability and regularity of user movement execution. Behavioral direction consistency is calculated by detecting the change in the angle between the expected and actual movement directions and taking the average value; this indicator reflects… Whether the user's control intention and actual execution remain consistent in direction, when any of the above indicators show a decline in stability, the system gradually increases the intervention level of the redundant control strategy according to the set hierarchical threshold, ensuring that the system can promptly enhance assistance or safety intervention when the user's actions become unstable; at the same time, during the redundant control strategy intervention stage, the system also dynamically adjusts the output intervention intensity according to the overlap rate between the user's actual execution trajectory and the target trajectory. When the overlap rate is high, the system output maintains a low intensity correction to avoid excessive intervention, while when the overlap rate continues to decrease and the deviation expands, the system gradually increases the control output intensity, ensuring that the redundant control smoothly transitions from mild correction to deep takeover, based on hierarchical intervention control under the dual judgment of user behavior stability and trajectory deviation.

[0040] Among these, the rate of change in movement speed and the trajectory repetition rate are used as important criteria for dynamically adjusting the intervention level of the redundancy control strategy. During system operation, the user's gait speed data is first statistically processed within a preset time window. The rate of change in movement speed is obtained by calculating the standard deviation of the speed values ​​within this time window. When the standard deviation continuously exceeds the fluctuation threshold set by the system in multiple consecutive time windows, it indicates that the user's gait speed has large fluctuations and instability. The system then raises the intervention level of the redundancy control strategy by one level to enhance the auxiliary control strength, thereby intervening in a timely manner during the unstable phase of the user's movement to ensure safety. At the same time, the system also calculates the trajectory repetition rate by comparing the user's movement trajectory over two consecutive movement cycles. This repetition rate is composed of the trajectory overlap area and the path difference. If the trajectory repetition rate shows a downward trend in multiple consecutive movement cycles, it is determined that the regularity of the user's movement execution has decreased and the stability has weakened. Based on this, the system raises the level of the redundancy strategy, so that the exoskeleton control strategy gradually transitions from mild correction to a higher level of deep intervention, in order to provide stronger movement assistance and safety redundancy guarantee when the movement repetition is insufficient. The degree of redundancy intervention is dynamically adjusted according to the user's actual movement stability.

[0041] Among them, behavioral direction consistency is an important indicator for judging the reliability of user control intention. It is achieved by detecting the angle formed between the user's expected movement direction and the actual movement direction of the exoskeleton system in each action cycle and calculating the average value of the angle over consecutive cycles as a reference for judging directional consistency. When the average angle value is small, it indicates that the user's intention and the system execution are highly consistent, and the system maintains a low intervention control state. When the angle continues to increase, it indicates that there is a deviation trend between the user's control intention and the system response. The system will then increase the intervention level of the redundancy strategy to enhance the assistance or correction capabilities. In addition, to further improve the system's adaptability and individualization capabilities in the initialization phase, the initial level of the redundancy control strategy is determined by the user's historical action trend stability score. The score is calculated by weighted averaging of multiple action stability indicators recorded by the user in the past control cycles, including but not limited to the rate of change of action speed, trajectory repetition rate, and behavioral direction consistency. The system presets the redundancy intervention level based on this stability score. If the score is high, the system's initial intervention level is low to maintain user dominance. If the score is low, the system initially increases the proportion of redundancy control to enhance safety.

[0042] When calculating the overlap rate between the user's execution trajectory and the system's target trajectory, to ensure consistency in the comparison process across time, the system sets a time synchronization window. The user's actual trajectory within this window is projected onto the corresponding time period of the target trajectory according to a time index for matching and comparison. This time alignment method avoids trajectory mismatch caused by differences in action rhythm, thus more accurately reflecting the degree of agreement between the user's actions and the target expectation. Based on this, the system uses the overlap rate as a reference signal for the output intervention intensity, employing a proportional adjustment method to linearly interpolate the output value. Specifically, when the overlap rate is high, the interpolation result corresponds to a lower output intensity, keeping redundant control with low intervention; when the overlap rate is high... As the system descends, the interpolation result gradually increases, and the output intensity increases accordingly. This ensures a smooth transition in the controller's output changes over multiple consecutive cycles, avoiding discomfort or danger to the user caused by sudden mechanical jumps. Simultaneously, the system performs joint judgment on abnormal states. When the trajectory overlap rate is consistently below a set threshold, and the user's behavior direction is detected to be opposite to the system's control output direction for two or more consecutive action cycles, the system immediately identifies it as a high-risk state and triggers the highest level of redundant control intervention. At this level, the redundant controller completely takes over the main output, while the main controller retains only the monitoring function. This ensures that safety protection can be quickly established and risks minimized when there is a serious confrontation between the user and the system.

[0043] Combined with appendix Figure 3 During the matching process between the execution trajectory and the target trajectory, an asymmetric time synchronization window is adopted. This time window is designed to perform forward multi-frame extension and backward few-frame backtracking on the user trajectory data. That is, based on the current matching time point, the system prioritizes the introduction of predictive trajectory data of the user in the next few frames, while retaining a small amount of historical frames for correction reference. This constructs an early capture mechanism for user behavior trends. This structure makes trajectory projection matching more focused on the early response capability of user actions, thereby more effectively identifying potential intention deviations or action anomalies. During the continuous monitoring of trajectory overlap rate, the system further calculates the first derivative of the overlap rate change and captures the trend of its rate of change in real time. When the system detects a sudden change in the first derivative value, which is manifested as a reversal (from decrease to increase) or a sharp steepening (rapid decrease in a short period of time) of the overlap rate decline trend in a short period of time, it determines that there is a potential risk of unstable fluctuation in trajectory matching. The system regards this as a precursor to entering the action mismatch state, and immediately triggers the redundancy control strategy to increase the intervention level in advance, so that the system can complete the transfer of control weights and upgrade of auxiliary mechanisms before the mismatch develops into a serious anomaly.

[0044] Before the actual trajectory is projected onto the target trajectory, the system detects the user's movement rhythm characteristics. It obtains the user's current step frequency through periodic analysis of cadence sensors or joint angular velocities and uses this as an adaptive adjustment parameter to dynamically compress or expand the time scale of the target trajectory. When the user is in a high-frequency movement state, the target trajectory is appropriately compressed to maintain consistency with the user's movement rhythm, thereby ensuring improved time alignment accuracy between the two parties during trajectory matching and avoiding matching deviations caused by frequency differences. On this basis, in the overlap rate monitoring stage, the system not only judges whether the current overlap rate is lower than the set threshold, but also introduces the historical average value and the rate of decline as joint judgment conditions. When the overlap rate has not reached the minimum threshold but is lower than the historical average value and its rate of decline exceeds the system's preset threshold, the system considers that the deviation between the user trajectory and the target trajectory has a developing trend. Therefore, it activates the redundancy level preparation mode in advance. In this mode, the redundant controller gradually increases the output ratio and enters the takeover state, thereby completing the defensive deployment before a real serious mismatch occurs.

[0045] When the user's behavior direction conflicts with the system's control output direction, the system calculates the unit rate of change of the spatial direction difference vector. If this rate exceeds a preset threshold, it is identified as a high-risk mismatch segment, triggering the early intervention of redundancy strategies. Simultaneously, in trajectory overlap rate monitoring, when the overlap rate oscillates rapidly over multiple consecutive cycles, the system adds a rate of change threshold to the original linear interpolation output value to prevent unexpected large jumps in the output control quantity within a short period. Based on the above design, the system constructs a discriminant function that integrates direction vector difference rate analysis and overlap rate change sensitivity identification. This function uniformly measures whether strong intervention control should be triggered or the jump in output interpolation should be limited in the current state. The specific calculation formula is as follows:

[0046]

[0047] in,

[0048] For the current moment

[0049] The intervention-sensitive function value is used to determine whether redundant control transitions or interpolation suppression need to be triggered. The unit time interval of the system control cycle; It represents the magnitude of the spatial unit difference vector between the user's intended direction and the system's control direction, i.e., the degree of direction mismatch; The first derivative of the difference vector in time reflects the rate of directional mismatch. This is the normalized standard deviation of the fluctuation in trajectory overlap rate, used to characterize trajectory stability; Trajectory overlap rate The second derivative over time represents the acceleration characteristic of the trajectory oscillation trend. Its derivation process includes: first, normalizing both the user's intended direction vector and the system output direction vector, calculating their difference vector, and then taking the modulus to obtain the result. And combined with its time derivative The first part of the risk items was obtained. This is used to reflect the worsening trend of orientation mismatch over time; secondly, in the trajectory matching dimension, it defines... To calculate the real-time overlap rate between the execution trajectory and the target trajectory, and using its second derivative... Oscillatory acceleration reflecting changes in overlap rate, while simultaneously expressed as the normalized standard deviation of trajectory fluctuations. As a weighting adjustment term, it forms the second part of the risk measurement. Finally, the two risk measures mentioned above are normalized and superimposed according to the time step to form a complete intervention sensitivity function. ,when Exceeding the set safety intervention threshold When the system determines that the current segment is high-risk, it will trigger the highest level of redundancy control strategy to take over or apply a rate suppression mechanism to the interpolation output, thereby achieving dual monitoring and safety control of direction mismatch and trajectory fluctuation.

[0050] Example 1:

[0051] Combined with appendix Figure 4In this embodiment, during lower limb exoskeleton-assisted walking training for a rehabilitation patient, the system collects electromyographic signals of the quadriceps and hamstring muscles in real time and monitors lower limb movements through joint angle sensors and an inertial measurement unit. When the patient actively generates a force signal during stepping, the system detects that the electromyographic amplitude of the quadriceps muscle rapidly increases within 200ms, indicating a clear intention to initiate movement. However, the driving response of the exoskeleton's knee joint only produces a corresponding angle change after 450ms, resulting in a response time difference of 250ms, exceeding the system's preset 200ms dynamic response threshold. At this point, the system determines that there is a potential abnormality. Furthermore, the system combined three indicators for joint judgment: First, regarding joint angle deviation, the patient expected a knee flexion angle of 30°, while the exoskeleton actually executed only 22°, a deviation of 8°, exceeding the set tolerance range of 5°; Second, regarding electromyographic activation pattern, the patient's hamstring signal showed an abnormally early peak during the flexion-extension coordination phase, causing the timing to deviate by 18% compared to the normal gait database, indicating an abnormal activation pattern; Third, regarding movement response consistency, the patient expected a stepping direction of 10° forward, while the system actually executed a forward direction of 2°, with an average deviation of 8°, and the consistency judgment was below the standard lower limit of 85%. Since two of the three indicators exceeded the threshold, the system ultimately confirmed that the system had entered an abnormal state. At this point, the system enters the parallel output phase of the main controller and the redundant controller. The main controller generates a joint output torque of 12 Nm based on the patient's intention, while the redundant controller generates a conservative output torque of 16 Nm based on anomaly detection. The system dynamically allocates weights according to the user's movement execution trend. In the initial stage, the weight of the main controller is set to 0.6, the weight of the redundant controller is 0.4, and the synthesized output is 13.6 Nm. However, in the following three control cycles, the system observes that the user's electromyographic signals and the actual joint movement direction are consistent again. For example, within a unit time window of 1 second, the consistency between the patient's expected knee flexion direction and the output direction of the main controller remains above 95%, and the electromyographic activation mode recovers to a 92% matching degree with the normal gait database. The joint angle deviation gradually decreases to 3°, indicating that the user's movement compliance is good. In this case, the system maintains the priority of the main controller, increasing its weight to 0.8, while reducing the weight of the redundant controller to 0.2. The final synthesized output value returns to 12.8 Nm to ensure the dominance of the user's intention and reduce unnecessary intervention from redundant control.

[0052] As the patient gradually progresses to the continuous walking stage, the exoskeleton system begins to monitor the stability indicators of the user's movement trends in real time, in order to dynamically adjust the intervention level of the redundancy control strategy. Regarding the rate of change in movement speed, the system records the patient's knee flexion and extension speed data in a preset time window of 2 seconds, calculating a gait speed standard deviation of 0.42 m / s. The set fluctuation threshold is 0.35 m / s. Since this standard deviation exceeds the threshold for three consecutive time windows, it indicates significant fluctuations in the patient's gait rhythm. Therefore, the system upgrades the redundancy strategy level from level 1 to level 2, increasing the proportion of the redundant controller in the output weight from 20% to 40%, thus enhancing intervention for gait instability. Secondly, regarding the trajectory repetition rate, the system matched the hip, knee, and ankle trajectories of the patient for two consecutive gait cycles, calculated the overlap area to be 142 cm², the path difference to be 36 cm, and the overall repetition rate to be 74%, while the repetition rate of the previous cycle was 82%. The repetition rate showed a decreasing trend over three consecutive cycles, with a cumulative decrease of 12%. As a result, the system raised the redundancy strategy level from level 2 to level 3, and further adjusted the corresponding output weight ratio to 50% for the main controller and 50% for the redundant controller, in order to significantly increase the protective effect of the redundancy control.

[0053] Finally, regarding consistency in behavioral direction, the system detected the change in the angle between the patient's expected walking direction and the actual execution direction within a single step cycle. The calculated average angle was 11°, lower than the set consistency reference threshold of 15°, indicating acceptable directional consistency. Therefore, no additional level enhancement was triggered. During the intervention of the redundancy control strategy, the system also dynamically adjusted the output intervention intensity based on the overlap rate between the actual execution trajectory and the target trajectory. For example, if the planned step length for the target trajectory is 0.65m, while the patient's actual step length is 0.58m, the overlap rate is 89%, and the corresponding output intervention intensity is adjusted to 0.3 times the redundancy correction. When the overlap rate drops to 76% in the next step, the output intervention intensity increases to 0.6 times the redundancy correction, thus achieving a smooth increase in intervention intensity.

[0054] The exoskeleton system further optimizes the dynamic response of the redundant control strategy by combining behavioral orientation consistency and historical stability scores. In the initial training phase, the system statistically analyzes the patient's movement stability based on control cycle data from the past 5 minutes. The average rate of change in movement speed is calculated to be 0.28 m / s (threshold 0.35 m / s), the average trajectory repetition rate is 81% (threshold 75%), and the average angle of behavioral orientation consistency is 9° (threshold 15°). Using a weighted average formula, a speed weight of 0.4, a trajectory weight of 0.35, and a direction weight of 0.25 are assigned, resulting in a stability score of 0.86 (out of 1.0). Based on this, the system sets the initial level of the redundant control strategy to Level 1, meaning the redundant controller output accounts for only 20%, with the main controller maintaining dominance to ensure user autonomy. In subsequent single-step movement cycles, the system monitors the angle between the user's desired direction of movement and the actual direction of execution in real time. For example, within a 1.2-second gait cycle, if the desired direction is 12° forward and the actual direction fluctuates between 14° and 17°, the calculated average angle deviation is 4.2°, which is less than the reference threshold of 15°, indicating good directional consistency. Therefore, the current redundancy level remains unchanged. However, in the next cycle, due to fatigue, the patient's desired direction shifts from 10° forward to -5° forward, resulting in an angle change of 15°. This deviation exceeds the threshold for two consecutive cycles, indicating decreased directional consistency. The system immediately triggers a level upgrade, raising the redundancy control level to level two and adjusting the proportion to 40%. Simultaneously, the system enters the process of calculating the overlap rate between the execution trajectory and the target trajectory. Within a 3-second time synchronization window, the patient's actual trajectory points are projected onto the corresponding time period of the target trajectory for matching and comparison. The current overlap rate is found to be 72%, which is lower than the set threshold of 75%. According to the proportional adjustment method, the output intervention intensity is increased from the original 0.3 times correction to 0.6 times correction. To ensure continuity, the system uses a linear interpolation method to gradually increase the intervention intensity over the next three cycles, rather than a one-time jump. For example, it increases to 0.4 times in the first cycle, 0.5 times in the second cycle, and 0.6 times in the third cycle, thereby avoiding unexpected large jumps in output torque within a short period of time. If the user's behavior direction and the system control output direction remain opposite for two consecutive cycles, for example, if the user continuously intends to retreat while the exoskeleton control output continues to push forward, the system determines that it has entered a severe adversarial state. When both the trajectory overlap rate and the opposite direction conditions are met, the redundant control strategy is switched to the highest level, the output weight of the redundant controller is increased to 100%, and the action control is forcibly taken over. The exoskeleton is locked in a safe and conservative mode to ensure that the patient will not fall or suffer joint damage due to human-machine conflict.

[0055] Example 2:

[0056] Combined with appendix Figure 5As patients' rehabilitation training progresses to more complex movements, the exoskeleton system introduces an asymmetric structure of time synchronization windows and an adaptive gait frequency adjustment mechanism to achieve more forward-looking and sensitive redundant control. When a patient is briskly walking, their average gait frequency increases to 1.9 Hz, while the original rhythm of the system's target trajectory is 1.5 Hz. The system first adaptively compresses the time scale of the target trajectory, with a scaling factor calculated as 1.9 ÷ 1.5 = 1.27, meaning the time axis of the target trajectory is compressed by 27%. This ensures that the rhythm planned by the system aligns with the user's high-frequency movements, avoiding system lag during trajectory matching. Subsequently, in the trajectory comparison phase, the system uses an asymmetric time synchronization window, extending the next 5 frames of data forward while only retracing 2 frames for correction, making the trajectory projection more biased towards predicting the patient's advance movement trend. For example, in a certain cycle, if the patient's actual stride trajectory shows a lower limb swing signal 150ms earlier than the target trajectory, the system effectively captures this advance trend by extending forward. The resulting trajectory overlap rate is 84%, which is about 9% higher than that of the traditional symmetrical window, indicating that capturing the advance trend improves matching accuracy. During continuous monitoring, the system calculates the first derivative of the trajectory overlap rate. When the overlap rate continuously decreases from 88% to 73% over three cycles (with a first derivative of -5% / cycle), and then suddenly rebounds to 79% in the fourth cycle (with the derivative turning from negative to positive and the change exceeding 7% / cycle), the system determines this as a reversal of the downward trend and immediately triggers an upgrade of the redundancy intervention level from level 2 to level 3, increasing the redundancy controller weight from 40% to 60% to prevent abnormal amplification of movements in advance. On the other hand, during high-frequency, fast-paced operation, although the overlap rate did not fall below the set threshold of 70%, it dropped sharply from the historical average of 85% to 74%, a decrease rate of 11% per cycle, exceeding the set threshold of 8% per cycle. The system immediately activated the redundancy level preparation mode, increasing the redundancy output ratio from 30% to 50% in the next cycle, and gradually increasing it over two control cycles through linear interpolation to avoid sudden intervention. When the patient showed a significant mismatch due to fatigue, the system had already entered a high-level preparation mode, ensuring smooth transitions and safe takeover of the exoskeleton movements.

[0057] The exoskeleton system enters a high-dynamic monitoring phase. At this point, the patient attempts to increase their cadence to 2.1 Hz, but due to insufficient lower limb muscle strength, movement stability decreases, causing a continuous deviation between the user's intended direction and the system's control direction. (Set at time...) The user expects the direction vector to be The system output direction vector is The difference vector between the two is The modulus is Over three consecutive control cycles, the magnitude of the direction difference vector increased from 0.45 to 0.63, and then to 0.88. The first derivative can be approximated as follows: ,set up ,but Therefore, the first part of the risk item is... .

[0058] On the other hand, the system monitors the trajectory overlap rate. The values ​​for the last five periods were 0.88, 0.81, 0.69, 0.76, and 0.62, indicating significant oscillations. An approximate calculation using its second derivative is performed:

[0059]

[0060] For example in When, substitute ,have to:

[0061]

[0062] Simultaneously calculate the normalized standard deviation of the nearest window. Assume the mean overlap rate of the trajectory is 0.752 and the standard deviation is 0.095. After normalization, take... The second part of the risk items is:

[0063]

[0064] In summary, substituting the two metrics into the intervention sensitivity function:

[0065]

[0066] Set the system's security intervention threshold to Then the current calculated value The system immediately identifies this as a high-risk mismatch segment and triggers the highest level of redundant control intervention. At this point, the redundant controller weight is increased from 50% to 100%, forcibly taking over motion output. Simultaneously, an interpolation output rate threshold is activated, limiting the output change rate to within 10% per cycle to prevent unexpected torque jumps caused by overlap rate oscillations. Under this protection, although the patient exhibits significant directional resistance and trajectory fluctuations, the system, through formula-based judgment and real-time intervention, achieves sensitive detection and rapid response to sudden risks, ensuring the patient's safety and stability during high-frequency, fast-paced rehabilitation training.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An exoskeleton safety redundancy strategy and switching method based on anomaly detection, characterized in that... include: The system collects the tension state between the user's active control intention and the exoskeleton system's response execution. When a continuous asynchronous tension is detected between the user's desired control direction and the system's inertial behavior, it is defined as a potential abnormal state and serves as a precondition for triggering redundant control strategies. After the redundant control strategy is triggered, the system allows the main controller and the redundant controller to output control suggestions at the same time. It also determines whether the control should be taken over by the redundant control strategy, continue to maintain the main control, or enter a forced freeze state based on the user's reverse force application and attitude deviation trend behavior. Based on the clarity of the user's current behavior and the responsiveness of the system intervention, a coupling relationship between behavioral tendency and control intervention level is constructed, and the intervention level of redundant control strategies is dynamically adjusted to achieve continuous control evolution from low intrusion to high takeover. If an irreconcilable antagonistic trend is detected between the user's continuous reverse intention input and the system's control output during the control evolution process, and the system intervention ratio imbalance continues to expand, then the system will actively withdraw control and enter a soft-freeze state or conservative posture mode, terminating further control command output. In the soft-freeze state, the system will determine whether the human-machine trust relationship has been rebuilt by identifying whether the user's behavior has re-demonstrated active synchronization or posture surrender actions. The process of determining the tension state is completed by multi-source signal fusion analysis. The system collects the user's active force signals in real time through electromyography sensors deployed at the user's muscle group locations, and simultaneously records the execution response signals of the exoskeleton joint drive unit. When the response time difference between the two exceeds the dynamic response threshold set by the system, it is determined that there is a potential abnormal state. When judging the abnormal state, three indicators are referenced: joint angle deviation index, electromyography activation mode change index, and action response consistency index. The system adopts a weighted judgment method. When the response time difference exceeds the threshold and two or more of the three indicators meet the abnormal conditions, it is finally confirmed that a potential abnormal state has been entered. During the phase when the main controller and the redundant controller are outputting simultaneously, the control output weights of the two controllers are dynamically allocated based on the execution trend of the user's actions. When determining the ownership of control, the user's action compliance performance is taken into account. If the user continuously performs actions in the same direction as the output of the main controller within a unit time window, the main controller is given priority and the redundant controller is in a low-weight state. The intervention level of the redundancy control strategy is adjusted in stages according to the user's action trend stability index, which includes the action speed change rate, trajectory repetition rate, and behavior direction consistency. When the redundancy control strategy intervenes, the output intervention intensity is adjusted according to the overlap rate between the user's actual execution trajectory and the target trajectory.

2. The exoskeleton safety redundancy strategy and switching method based on anomaly detection according to claim 1, characterized in that... The rate of change of movement speed is the standard deviation of the user's gait speed within a preset time window. The redundancy control strategy intervention level is increased by one level when the standard deviation continuously exceeds the set fluctuation threshold. The trajectory repetition rate is calculated by comparing the trajectory overlap area and path difference between two consecutive movement cycles. If the repetition rate shows a decreasing trend in multiple consecutive cycles, the redundancy control strategy level is increased.

3. The exoskeleton safety redundancy strategy and switching method based on anomaly detection according to claim 1, characterized in that... The consistency of the behavioral direction is achieved by detecting the change in the angle between the expected movement direction and the actual movement direction within each action cycle, and using the average value of this angle change as a consistency reference index; the initial level of the redundancy control strategy is determined by the user's historical action trend stability score, which is a weighted average of multiple action trend stability indices over past control cycles.

4. The exoskeleton safety redundancy strategy and switching method based on anomaly detection according to claim 1, characterized in that... When calculating the overlap rate between the execution trajectory and the target trajectory, the actual execution trajectory is projected onto the corresponding time period of the target trajectory for matching and comparison based on a preset time synchronization window; The output intervention intensity adjustment uses the overlap rate as a reference signal and employs a proportional adjustment method to perform linear interpolation of the output value, so that the output intensity change of the redundant controller smoothly transitions over multiple consecutive cycles. When the overlap rate is lower than the set threshold and the user's behavior direction is continuously opposite to the system control output direction for more than two cycles, the redundant control strategy automatically enters the highest intervention level.

5. The exoskeleton safety redundancy strategy and switching method based on anomaly detection according to claim 4, characterized in that... The time synchronization window has an asymmetric structure, which extends the actual execution trajectory forward by multiple frames and back by a few frames, so that the projection matching focuses on the early trend of user behavior.

6. The exoskeleton safety redundancy strategy and switching method based on anomaly detection according to claim 4, characterized in that... Before the actual execution trajectory is projected onto the target trajectory, the time scale of the target trajectory is adaptively adjusted based on the user's current step frequency to align the rhythm in the high-frequency motion state. When the overlap rate is lower than the historical average and the descent rate exceeds the set threshold, the redundancy level preparation mode is activated in advance even if the set threshold is not reached. In this mode, the redundant controller gradually increases the output ratio and enters the takeover state.

Images