Unmanned aerial vehicle nonparametric data fault detection and recovery method based on digital twinning

Abstract

The invention discloses an unmanned aerial vehicle non-parametric data fault detection and recovery method based on digital twinborn, and the method comprises the steps: collecting multi-source data, carrying out the preprocessing, generating a standard data set, and constructing a digital twinborn body; mapping the data to the twinborn body, obtaining a prediction state, and differentiating the prediction state from an actual measurement state; dividing a multi-scale residual error, constructing a nested manifold, and outputting an abnormal geometric position; constructing a reversible envelope, positioning a hierarchy, and determining a target recovery state; calculating a local response relationship, constructing an inversion operator, and solving a recovery control quantity; and synchronizing the control quantity to the twin body, executing verification, and issuing to unmanned aerial vehicle control. According to the method, the digital twin is constructed, and multi-scale residual geometric recognition and partial derivative inversion control are fused, so that accurate detection, reversible recovery and closed-loop control of the unmanned aerial vehicle under the non-parametric data abnormal working condition are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and operational safety assurance technology for unmanned aerial vehicles (UAVs), and in particular to a method for detecting and recovering non-parametric data faults in UAVs based on digital twins. Background Technology

[0002] With the widespread application of drones in fields such as inspection, surveying, logistics, and emergency rescue, their operating environment is becoming increasingly complex, and the duration of flight missions is constantly extending, significantly increasing the risk of malfunctions faced by drones during actual operation. Existing drone operation safety assurance technologies mainly rely on sensor threshold monitoring, state estimation based on physical models, or anomaly judgment methods based on empirical rules to monitor and diagnose flight attitude, power systems, and control signals. These methods can play a certain role under conditions of clear structure and stable operating conditions, but they lack effective adaptability to complex environmental disturbances, system aging, and abnormal responses caused by multi-factor coupling.

[0003] On the other hand, some existing technologies have introduced data-driven or model-predictive methods to analyze and predict the operational status of UAVs, but they still heavily rely on preset parameter models or specific fault samples. Due to the strong nonlinearity, strong coupling, and time-varying characteristics of UAV systems, non-parametric anomalies or unknown types of faults that are difficult to characterize with fixed models often occur in actual operation, leading to misjudgments or missed detections in existing methods during the anomaly detection phase. After detecting anomalies, traditional methods often employ conservative strategies such as degraded control, return to home, or shutdown, lacking specificity in the recovery process and making it difficult to achieve reversible state recovery while ensuring safety.

[0004] In recent years, digital twin technology has begun to be applied to the operation monitoring and simulation analysis of unmanned aerial vehicle (UAV) systems. However, existing solutions mostly focus on state mapping or offline simulation verification, and the coupling between the digital twin and actual control decisions is limited. A closed-loop fault detection and recovery mechanism based on digital twins has not yet been formed. Existing digital twin solutions generally adopt simple error comparison or single-scale analysis methods, which are difficult to characterize complex residual structures and their evolution characteristics, and cannot support accurate location and recovery control of unknown faults.

[0005] Therefore, how to provide a method for detecting and recovering non-parametric data faults in UAVs based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for nonparametric data fault detection and recovery of unmanned aerial vehicles (UAVs) based on digital twins. This invention constructs a digital twin that corresponds to the UAV's operational state in real time, integrating techniques such as multifractal residual structure analysis, reversible steady-state envelope construction, and partial derivative response inversion control. This method detects, locates, and recovers nonparametric anomalies and unknown types of faults generated by UAVs in complex environments. The invention utilizes the digital twin to map and virtually execute verification of multi-source operational data, characterizes the anomaly evolution law through multi-scale residual geometric features, plans reversible recovery states in the digital twin space, and then generates and executes verified recovery control quantities. This method effectively improves the accuracy and stability of UAV fault detection, enhances adaptability to unknown faults, and achieves controllability and reversibility of the recovery process. It has the advantages of strong robustness, high security, and high recovery efficiency.

[0007] A method for detecting and recovering nonparametric data faults in unmanned aerial vehicles based on digital twins according to an embodiment of the present invention includes:

[0008] Collect multi-source operational data generated by drones during flight, preprocess the multi-source operational data to form a standardized non-parametric operational dataset, and construct a digital twin;

[0009] The standardized nonparametric operational dataset is mapped to a digital twin, and the twin predicted state data at the corresponding time point is obtained and differentially processed with the measured operational state data of the UAV to generate a residual data sequence.

[0010] Based on the residual data sequence, multi-scale division is performed according to a preset time scale to form a multi-scale residual segment set. Non-parametric manifold embedding processing is then performed to construct a multi-fractal nested twin residual manifold, and anomaly state identifiers and the geometric positions of the anomaly states in the digital twin space are output.

[0011] In the digital twin space, a reversible steady envelope containing a steady-state kernel, a reversible attraction band, and a degenerate envelope layer is constructed. Based on the geometric position of the abnormal state in the digital twin space, the steady-state envelope level and the target recovery state corresponding to the abnormal state are determined.

[0012] Based on the digital twin, the local partial derivative response relationship of the control input change to the residual data change is calculated. Combined with the control constraints of the UAV physical actuator, the partial derivative response inversion operator is constructed, and the corresponding recovery control quantity is calculated according to the target recovery state.

[0013] The recovery control quantity is synchronized to the digital twin for virtual execution verification. After successful verification, the recovery control quantity is sent to the drone to execute recovery control, and the operational data collected during the recovery execution process is fed back to the digital twin.

[0014] Optionally, the multi-source operating data includes attitude and motion data, power system operating data, control command data, and environmental disturbance data.

[0015] Optionally, the preprocessing of multi-source operating data includes time synchronization, noise filtering, data alignment, and non-parametric normalization of the multi-source operating data.

[0016] Optionally, constructing a digital twin includes:

[0017] Based on a standardized non-parametric operation dataset, the airframe structural parameters, power system parameters, sensor configuration parameters, and actuator connection parameters of the UAV are digitally mapped to form a structural mapping relationship.

[0018] Based on a standardized non-parametric operation dataset, state correlation processing is performed on the attitude data, motion data, and dynamic data of UAVs under different control commands and environmental disturbances to form a state evolution mapping relationship.

[0019] Based on the standardized non-parametric operation dataset and the collected control command data, a joint mapping is performed on the correspondence between changes in control input and changes in UAV state to form a control response mapping relationship.

[0020] Consistency constraints are fused to integrate the structural mapping relationship, state evolution mapping relationship, and control response mapping relationship. The mapping relationship is synchronously updated based on real-time collected multi-source operation data to form a digital twin that corresponds to the real-time operation status of the UAV.

[0021] Optionally, generating the residual data sequence includes:

[0022] The standardized non-parametric running dataset is loaded into the digital twin according to the data time stamp and feature correspondence, so that each running state variable in the digital twin and the corresponding data in the standardized non-parametric running dataset are mapped one-to-one.

[0023] In a digital twin, standardized non-parametric operating data is synchronized based on a mapping relationship to generate a twin operating state corresponding to the current moment, and twin predictive state data is formed from the twin operating state.

[0024] Under the same time marker, the twin predicted state data and the measured operating state data collected by the UAV are subjected to feature alignment processing in the digital twin;

[0025] The twin prediction state data and the measured operating state data that have completed feature alignment are differentially processed item by item according to the corresponding features to obtain residual data that characterizes the deviation relationship between the twin prediction state and the measured operating state.

[0026] The residual data obtained from multiple consecutive time markers are integrated and processed in chronological order to form a residual data sequence.

[0027] Optionally, the construction of the multifractal nested twin residual manifold, and the output of the abnormal state identifier and the geometric position of the abnormal state in the digital twin space, includes:

[0028] Obtain the residual data sequence, organize the residual data sequence in order according to the time stamp, and bind the residual data at each time step with the corresponding digital twin space state stamp to form a residual sample sequence with state stamp;

[0029] Based on a preset time scale, the residual sample sequence with state labels is divided into multiple scales. At each time scale, continuous residual segments are extracted in a sliding window manner to form a set of multi-scale residual segments. Feature dimension consistency processing is performed on the set of multi-scale residual segments.

[0030] At each time scale, nonparametric manifold embedding is performed on the corresponding set of residual segments to obtain the residual manifold representation at the time scale. During the embedding process, twin space state labels are introduced as neighborhood constraints to form a state-constrained residual manifold.

[0031] A cross-scale nested construction process is performed on the state-constrained residual manifolds formed at different time scales. The cross-scale nested construction process includes determining the correspondence between residual manifolds at each scale, establishing a cross-scale alignment mapping, generating a cross-scale nested boundary, and combining the residual manifolds at each scale into a multifractal nested twin residual manifold based on the cross-scale nested boundary.

[0032] Embedding and locating the current residual segment in a multifractal nested twin residual manifold, calculating the positional relationship of the current embedding point relative to the cross-scale nesting boundary, outputting an anomaly status identifier based on the positional relationship, and outputting the geometric position of the anomaly status in the digital twin space.

[0033] Optionally, determining the steady-state envelope level and target recovery state corresponding to the abnormal state based on its geometric location in the digital twin space includes:

[0034] Based on multifractal nested twin residual manifolds, residual embedding samples corresponding to normal operation periods are selected in the digital twin space, and organized according to time markers and feature dimensions to determine the sample set that simultaneously satisfies residual stability and state evolution consistency at each scale. The minimum enclosing domain of the sample set is used as the initial region boundary of the steady-state kernel.

[0035] In a digital twin, a bidirectional rolling simulation is performed around the initial region boundary of the steady-state kernel. Amplitude-limited control perturbations and environmental perturbations are applied to the state of the boundary neighborhood. The set of states that can revert to the steady-state kernel within a finite number of steps without external reconfiguration is recorded. The continuous region between the state set and the steady-state kernel is defined as a reversible attraction zone.

[0036] For the remaining states that do not meet the conditions for reversible attraction bands, a local reversible geometric transformation generation process is executed in the digital twin. Multiple sets of candidate geometric transformations are generated through local state reparameterization, control law fine-tuning, and constraint unification. States that can enter the reversible attraction band within a finite number of steps are uniformly collected. The collected state set is determined as the degenerate envelope layer, and the corresponding local reversible geometric transformation label is recorded for each type of state.

[0037] In the digital twin, the boundaries of the steady-state kernel, reversible attraction band and degenerate envelope are self-consistently calibrated. Antifactual perturbation experiments and time-symmetric replay verifications are performed in the boundary neighborhood. Based on the consistency of the cross-scale residual structure and the consistency of state evolution, the boundaries of the three types of regions are synchronously updated to form a reversible steady-state envelope.

[0038] Based on the geometric location of the abnormal state in the digital twin space, determine its corresponding level:

[0039] When the position is within the steady-state kernel, the target recovery state is determined to be the current corresponding steady-state point;

[0040] When the location is within the reversible attraction zone, the target recovery state is determined to be the steady state point with the shortest time to reach steady-state nuclear evolution.

[0041] When the location is within the degenerate envelope, the intermediate target entering the reversible attraction zone is determined based on the corresponding local reversible geometric transformation marker, and the final target recovery state is determined.

[0042] Optionally, the step of constructing a partial derivative response inversion operator based on the control constraints of the UAV's physical actuators, and calculating the corresponding recovery control quantity according to the target recovery state, includes:

[0043] Obtain the residual data corresponding to the current abnormal state from the digital twin, and obtain the current control input corresponding to the current abnormal state, where the residual data comes from the residual data sequence and the current control input comes from the control command data;

[0044] In the digital twin, multiple sets of constrained control disturbance sequences are generated based on the current control input. The constrained control disturbance sequences are synchronized to the digital twin for virtual execution. The residual change data corresponding to each set of constrained control disturbance sequences are recorded to form a corresponding sample set of control disturbance and residual change.

[0045] Based on the corresponding sample set, the local partial derivative response relationship of control input change to residual data change is constructed. By performing consistency screening on the residual change results corresponding to control disturbances of different amplitudes, the control disturbance results that maintain a stable change trend in the geometric neighborhood of multi-scale residual structure and abnormal state are retained, forming the local partial derivative response relationship of structural constraint.

[0046] A partial derivative response inversion operator is constructed based on the structural constraint local partial derivative response relationship. The partial derivative response inversion operator consists of control constraint projection processing, residual target alignment processing, and inversion stabilization processing, wherein:

[0047] Control constraint projection processing maps candidate control input variations to feasible control domains that satisfy physical actuator constraints;

[0048] The residual target alignment process aligns the current residual change direction with the residual change direction corresponding to the target recovery state.

[0049] The inversion stabilization process suppresses ill-posed directions during the inversion process and preferentially preserves inversion solutions in the neighborhood of the reversible attraction band.

[0050] The target residual change corresponding to the target recovery state is used as the inversion input, and the current control input is used as the control reference. The recovery control quantity is calculated by the partial derivative response inversion operator.

[0051] Optionally, the step of synchronizing the recovery control quantity to the digital twin for virtual execution verification, and then issuing the recovery control quantity to the drone for recovery control after successful verification, includes:

[0052] Obtain the recovery control quantity, synchronize the recovery control quantity to the digital twin, and enable the digital twin to enter a virtual execution state with the recovery control quantity as the control input;

[0053] The operation process of the UAV is virtually executed in the digital twin according to the recovery control quantity, generating a virtual operation state sequence corresponding to the recovery control process. The virtual operation state sequence is consistent with the actual operation state of the UAV in terms of time sequence and state dimension.

[0054] Virtual residual data is generated based on the virtual running state sequence in the digital twin. The virtual residual data is compared with the residual data sequence to complete the virtual execution verification and judgment.

[0055] If the virtual execution verification is successful, the recovery control quantity will be sent to the UAV flight control system, so that the UAV will perform the recovery control operation according to the recovery control quantity;

[0056] During the process of the drone performing recovery control operations, real-time operation data of the drone is collected and synchronously fed back to the digital twin for updates.

[0057] The beneficial effects of this invention are:

[0058] This invention achieves unified mapping and synchronous evolution of multi-source UAV operational data by introducing a tight coupling between a digital twin and the actual operation of the UAV. This allows fault detection to move beyond fixed physical models or empirical thresholds and instead rely on analysis based on non-parametric data characteristics. By constructing multi-fractal nested twin residual manifolds in the digital twin space, the evolution of residual structures can be characterized from multiple time scales and multi-dimensional feature perspectives. This effectively identifies non-parametric anomalies and unknown types of faults that are difficult to detect using traditional methods, reducing the risk of misjudgment and missed detection.

[0059] In the fault location and recovery planning stage, this invention constructs a reversible steady-state envelope containing a steady-state core, a reversible attraction band, and a degenerate envelope layer. This allows the geometric location of the UAV's abnormal state in the digital twin space to correspond with the recovery strategy, avoiding the limitations of relying on human experience or fixed strategies for handling. It can automatically determine the target recovery state based on the severity and evolution characteristics of the abnormal state, giving the recovery process clear spatial constraints and a reversible path. This improves the targeting and reliability of the recovery while ensuring operational safety.

[0060] During the recovery control execution phase, this invention utilizes a digital twin to virtually verify the recovery control quantities and generates feasible recovery control inputs by combining the constraints of the physical actuators, thus achieving safety verification and optimization of the recovery control before its issuance. By continuously feeding operational data during the recovery execution process back to the digital twin, a closed-loop update mechanism is formed, continuously improving the adaptability to complex operating conditions and new types of faults. Overall, this invention enhances the accuracy of UAV fault detection, the stability of recovery control, and the robustness of system operation in complex environments. Attached Figure Description

[0061] 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:

[0062] Figure 1 This is a flowchart of a method for detecting and recovering nonparametric data faults in unmanned aerial vehicles based on digital twins, as proposed in this invention.

[0063] Figure 2 This is a schematic diagram of the reversible steady-state envelope structure of a nonparametric data fault detection and recovery method for unmanned aerial vehicles based on digital twins proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figure 1 and Figure 2 A method for nonparametric data fault detection and recovery of unmanned aerial vehicles based on digital twins, comprising:

[0066] Collect multi-source operational data generated by drones during flight, preprocess the multi-source operational data to form a standardized non-parametric operational dataset, and construct a digital twin;

[0067] The standardized nonparametric operational dataset is mapped to a digital twin, and the twin predicted state data at the corresponding time point is obtained and differentially processed with the measured operational state data of the UAV to generate a residual data sequence.

[0068] Based on the residual data sequence, multi-scale division is performed according to a preset time scale to form a multi-scale residual segment set. Non-parametric manifold embedding processing is then performed to construct a multi-fractal nested twin residual manifold, and anomaly state identifiers and the geometric positions of the anomaly states in the digital twin space are output.

[0069] In the digital twin space, a reversible steady envelope containing a steady-state kernel, a reversible attraction band, and a degenerate envelope layer is constructed. Based on the geometric position of the abnormal state in the digital twin space, the steady-state envelope level and the target recovery state corresponding to the abnormal state are determined.

[0070] Based on the digital twin, the local partial derivative response relationship of the control input change to the residual data change is calculated. Combined with the control constraints of the UAV physical actuator, the partial derivative response inversion operator is constructed, and the corresponding recovery control quantity is calculated according to the target recovery state.

[0071] The recovery control quantity is synchronized to the digital twin for virtual execution verification. After successful verification, the recovery control quantity is sent to the drone to execute recovery control, and the operational data collected during the recovery execution process is fed back to the digital twin.

[0072] In this embodiment, the multi-source operating data includes attitude and motion data, power system operating data, control command data, and environmental disturbance data.

[0073] In this embodiment, the preprocessing of multi-source operating data includes time synchronization, noise filtering, data alignment, and non-parametric normalization of the multi-source operating data.

[0074] In this embodiment, constructing a digital twin includes:

[0075] Based on a standardized non-parametric operational dataset, the airframe structural parameters, power system parameters, sensor configuration parameters, and actuator connection parameters of the UAV are digitally mapped to form a structural mapping relationship, which is as follows:

[0076] Using the physical components of the UAV and their connection relationships as mapping objects, a one-to-one correspondence is established in the digital twin between the airframe configuration, power transmission path, sensor spatial arrangement, and actuator action link, so that the structural parameters in the digital space are consistent with the topological, assembly, and constraint relationships of the physical UAV, and can be adjusted synchronously as the structural parameters are updated.

[0077] Based on a standardized nonparametric operational dataset, state correlation processing is performed on the attitude, motion, and dynamic data of UAVs under different control commands and environmental disturbances to form a state evolution mapping relationship, which is as follows:

[0078] Under continuous time conditions, the control command sequence and environmental disturbance parameters are associated with the current attitude, motion and dynamic state of the UAV, and the correspondence between the attitude change, motion change and dynamic response of the UAV in the next moment under the action of control commands and environmental disturbances is further determined, so that the multi-dimensional operating state of the UAV can evolve continuously according to the change sequence of control input and external disturbances.

[0079] Based on the standardized non-parametric operational dataset and the collected control command data, a joint mapping is performed on the correspondence between changes in control input and changes in UAV state to form a control response mapping relationship, wherein the control response mapping relationship is as follows:

[0080] Under continuous time conditions, the changes of each control quantity in the control command are correlated with the actual response changes of the UAV's attitude state, motion state and dynamic state to determine the corresponding response results of the UAV's multi-dimensional operating state under different control input change amplitudes and change sequences.

[0081] Consistency constraints are fused to integrate the structural mapping relationship, state evolution mapping relationship, and control response mapping relationship. The mapping relationship is synchronously updated based on real-time collected multi-source operation data to form a digital twin that corresponds to the real-time operation status of the UAV.

[0082] In this embodiment, generating the residual data sequence includes:

[0083] The standardized non-parametric running dataset is loaded into the digital twin according to the data time stamp and feature correspondence, so that each running state variable in the digital twin and the corresponding data in the standardized non-parametric running dataset are mapped one-to-one.

[0084] In a digital twin, standardized non-parametric operating data is synchronized based on a mapping relationship to generate a twin operating state corresponding to the current moment, and twin predictive state data is formed from the twin operating state.

[0085] Under the same time marker, the twin predicted state data and the measured operating state data collected by the UAV are subjected to feature alignment processing in the digital twin;

[0086] The twin prediction state data and the measured operating state data that have completed feature alignment are differentially processed item by item according to the corresponding features to obtain residual data that characterizes the deviation relationship between the twin prediction state and the measured operating state.

[0087] The residual data obtained from multiple consecutive time markers are integrated and processed in chronological order to form a residual data sequence.

[0088] In this embodiment, the construction of a multifractal nested twin residual manifold and the output of anomaly state identifiers and the geometric positions of the anomaly states in the digital twin space include:

[0089] Obtain the residual data sequence, organize the residual data sequence in order according to the time stamp, and bind the residual data at each time step with the corresponding digital twin space state stamp to form a residual sample sequence with state stamp;

[0090] Based on a preset time scale, the residual sample sequence with state labels is divided into multiple scales. At each time scale, continuous residual segments are extracted in a sliding window manner to form a set of multi-scale residual segments. Feature dimension consistency processing is performed on the set of multi-scale residual segments.

[0091] At each time scale, a nonparametric manifold embedding process is performed on the corresponding set of residual fragments to obtain the time-scale residual manifold representation. A twin space state label is introduced as a neighborhood constraint during the embedding process to form a state-constrained residual manifold. Specifically, the nonparametric manifold embedding process on the corresponding set of residual fragments is as follows:

[0092] The residual segments at the same time scale are uniformly organized according to feature dimensions and time order, and local neighborhood relationships are constructed based on the similarity between residual segments, so that residual segments with similar residual evolution features are preferentially established in neighborhood association during the embedding process;

[0093] Based on the construction of local neighborhood relations, a low-dimensional embedding transformation is performed on the residual fragments to map the high-dimensional residual fragments to the low-dimensional embedding space, while maintaining the local neighborhood structure between the residual fragments without destruction during the mapping process;

[0094] In the low-dimensional embedding process, twin space state labels corresponding to residual segments are introduced to constrain the embedding neighborhood, so that residual segments with the same or adjacent twin space state labels are kept in close proximity in the embedding space, forming a residual manifold representation constrained by twin space state.

[0095] A cross-scale nested construction process is performed on state-constrained residual manifolds formed at different time scales. This process includes determining the correspondence between residual manifolds at each scale, establishing cross-scale alignment mappings, generating cross-scale nested boundaries, and combining residual manifolds at each scale into a multifractal nested twin residual manifold based on these boundaries.

[0096] To determine the correspondence between residual manifolds at different scales, specifically: at different time scales, according to the overlap relationship of the time intervals corresponding to the residual segments and the consistency of the twin space state labels, the embedding points in the residual manifolds at each scale are associated and matched, and the embedding points that come from the same or adjacent time intervals and have the same twin space state labels are determined as cross-scale corresponding points, thus establishing the correspondence between residual manifolds at different time scales.

[0097] Establishing a cross-scale alignment mapping involves: based on the cross-scale correspondence, performing unified alignment processing on residual manifolds at different time scales, so that the residual manifold at a longer time scale serves as the global structural reference, and the residual manifold at a shorter time scale is embedded into the global structure according to the correspondence, and performing position correction and boundary adjustment on the scale differences generated during the embedding process, forming a cross-scale residual manifold representation that can be aligned under the same twin space coordinate system;

[0098] The cross-scale nesting boundary is jointly determined by the statistical consistency constraint of the residual fragments output by the digital twin at different time scales and the consistency constraint of the state evolution of the twin space;

[0099] Embedding and locating the current residual segment in a multifractal nested twin residual manifold, calculating the positional relationship of the current embedding point relative to the cross-scale nesting boundary, outputting an anomaly state identifier based on the positional relationship, and outputting the geometric position of the anomaly state in the digital twin space. Specifically, calculating the positional relationship of the current embedding point relative to the cross-scale nesting boundary involves:

[0100] In a multifractal nested twin residual manifold, the cross-scale nested boundary representation corresponding to the time scale of the current embedding point is extracted, and the cross-scale nested boundary representation is discretized into several continuous boundary point sets.

[0101] Using the current embedding point as a reference point, calculate the geometric distance distribution between the current embedding point and each boundary point in the set of boundary points, and determine the position of the boundary point corresponding to the minimum distance;

[0102] Based on the boundary point position corresponding to the minimum distance, the spatial lateral relationship of the current embedding point relative to the cross-scale nested boundary is determined. Combined with the position changes of the previous and next adjacent embedding points in the time evolution direction, the positional relationship of the current embedding point relative to the inner, neighborhood or outer side of the cross-scale nested boundary is determined.

[0103] In this embodiment, determining the steady-state envelope level and target recovery state corresponding to the abnormal state based on its geometric location in the digital twin space includes:

[0104] Based on multifractal nested twin residual manifolds, residual embedding samples corresponding to normal operation periods are selected in the digital twin space, and organized according to time markers and feature dimensions to determine the sample set that simultaneously satisfies residual stability and state evolution consistency at each scale. The minimum enclosing domain of the sample set is used as the initial region boundary of the steady-state kernel.

[0105] In a digital twin, a bidirectional rolling simulation is performed around the initial region boundary of the steady-state kernel. Amplitude-limited control perturbations and environmental perturbations are applied to the state of the boundary neighborhood. The set of states that can revert to the steady-state kernel within a finite number of steps without external reconfiguration is recorded. The continuous region between the state set and the steady-state kernel is defined as a reversible attraction zone.

[0106] For the remaining states that do not satisfy the reversible attraction band condition, a local reversible geometric transformation generation process is executed in the digital twin. Multiple sets of candidate geometric transformations are generated through local state reparameterization, control law fine-tuning, and constraint unification. States that can enter the reversible attraction band within a finite number of steps are uniformly grouped. The grouped state set is determined as the degenerate envelope layer. A corresponding local reversible geometric transformation label is recorded for each state type, where:

[0107] Local state reparameterization: In the digital twin, with the current abnormal state as a reference, the local coordinates of the attitude state, motion state and dynamic state corresponding to the abnormal state are reconstructed, and the original state description is converted into a local state expression form centered on the abnormal state. Under the local state expression form, the direction and magnitude of state change are recalibrated to eliminate the influence of the difference in the dimensions of different states on the geometric transformation generation process.

[0108] Control law fine-tuning: After completing the local state reparameterization, the control command is adjusted in a small and continuous manner in the digital twin based on the current control command. Virtual evolution is performed after each adjustment to record the state change trajectory. Control adjustment methods that can guide the local state to approach the reversible attraction zone along the stable evolution direction are retained to form candidate control adjustment combinations that match the local state expression.

[0109] Constraint Consistency: During the generation of candidate control adjustment combinations, the amplitude limits, rate of change limits, and channel coupling relationships of the UAV physical actuators are synchronously introduced into the digital twin. Candidate geometric transformations that do not meet physical constraints or cause discontinuous state evolution are eliminated, while candidate geometric transformations that meet the constraints and can maintain the continuity of state evolution are retained. The retained geometric transformations are recorded as locally reversible geometric transformations.

[0110] In the digital twin, the boundaries of the steady-state kernel, reversible attraction band and degenerate envelope are self-consistently calibrated. Antifactual perturbation experiments and time-symmetric replay verifications are performed in the boundary neighborhood. Based on the consistency of the cross-scale residual structure and the consistency of state evolution, the boundaries of the three types of regions are synchronously updated to form a reversible steady-state envelope.

[0111] Based on the geometric location of the abnormal state in the digital twin space, determine its corresponding level:

[0112] When the position is within the steady-state kernel, the target recovery state is determined to be the current corresponding steady-state point;

[0113] When the location is within the reversible attraction zone, the target recovery state is determined to be the steady state point with the shortest time to reach steady-state nuclear evolution.

[0114] When the location is within the degenerate envelope, the intermediate target entering the reversible attraction zone is determined based on the corresponding local reversible geometric transformation marker, and the final target recovery state is determined.

[0115] In this embodiment, the step of constructing a partial derivative response inversion operator based on the control constraints of the UAV's physical actuators and calculating the corresponding recovery control quantity according to the target recovery state includes:

[0116] Obtain the residual data corresponding to the current abnormal state from the digital twin, and obtain the current control input corresponding to the current abnormal state, where the residual data comes from the residual data sequence and the current control input comes from the control command data;

[0117] In the digital twin, multiple sets of constrained control disturbance sequences are generated based on the current control input. The constrained control disturbance sequences are synchronized to the digital twin for virtual execution. The residual change data corresponding to each set of constrained control disturbance sequences are recorded to form a corresponding sample set of control disturbance and residual change.

[0118] Based on the corresponding sample set, the local partial derivative response relationship of control input change to residual data change is constructed. By performing consistency screening on the residual change results corresponding to control disturbances of different amplitudes, the control disturbance results that maintain a stable change trend in the geometric neighborhood of multi-scale residual structure and abnormal state are retained, forming the local partial derivative response relationship of structural constraint.

[0119] A partial derivative response inversion operator is constructed based on the structural constraint local partial derivative response relationship. The partial derivative response inversion operator consists of control constraint projection processing, residual target alignment processing, and inversion stabilization processing, wherein:

[0120] Control constraint projection processing maps candidate control input variations to feasible control domains that satisfy physical actuator constraints. Specifically, this mapping process involves:

[0121] The control constraint parameters corresponding to the physical actuators of the UAV are obtained in the digital twin. The control constraint parameters are uniformly applied to the changes in candidate control inputs. The part of the changes in candidate control inputs that exceeds the allowable amplitude range of the actuators is pruned so that the changes in candidate control inputs meet the amplitude constraints of each control channel.

[0122] After amplitude clipping is completed, a control input change rate constraint is applied to the candidate control input changes. Candidate control input changes that do not meet the actuator change rate limit are smoothly adjusted so that the candidate control input changes conform to the response capability of the physical actuator in terms of time continuity.

[0123] Based on satisfying the amplitude constraint and the rate of change constraint, the candidate control input change is subjected to multi-control channel consistency correction, and the candidate control input change that violates the control channel coupling relationship is jointly coordinated and adjusted to obtain the feasible control input change that simultaneously satisfies the amplitude constraint, the rate of change constraint and the channel coupling constraint. The feasible control input change is determined to be the control input change located in the feasible control domain of the physical actuator.

[0124] The residual target alignment process aligns the current residual change direction with the residual change direction corresponding to the target recovery state. Specifically, this process involves:

[0125] Obtain the residual change trajectory corresponding to the current abnormal state in the digital twin, and extract the dominant change trend of the residual change trajectory within the most recent time window as a representation of the current residual change direction.

[0126] Obtain the residual change trajectory corresponding to the target recovery state in the digital twin, extract the dominant change trend of the residual change trajectory in the steady-state evolution process, and use it as a representation of the residual change direction corresponding to the target recovery state.

[0127] Finally, in the digital twin, the current residual change direction is aligned with the residual change direction corresponding to the target recovery state. The change components in the current residual change that deviate from the target recovery direction are weakened, while the change components that are consistent with the target recovery direction are retained and strengthened, so that the current residual change direction converges to the residual change direction corresponding to the target recovery state.

[0128] The inversion stabilization process suppresses ill-posed directions during the inversion process and preferentially preserves inversion solutions in the neighborhood of the reversible attraction band. Specifically, the inversion stabilization process suppresses ill-posed directions during the inversion process and preferentially preserves inversion solutions in the neighborhood of the reversible attraction band.

[0129] In the digital twin, the state evolution test is performed on multiple sets of inversion solutions obtained during the inversion process. The residual change trend caused by each inversion solution during the virtual execution process is recorded. Inversion solutions that cause rapid amplification of residual amplitude, state oscillation or discontinuous evolution are marked as unstable inversion solutions, and the corresponding inversion directions are suppressed.

[0130] After the unstable inversion solutions are screened out, the reversibility of the remaining inversion solutions is evaluated. The inversion solutions are applied to the digital twin and it is observed whether the state can enter or approach the reversible attraction zone region within a finite number of steps. Inversion solutions that cannot enter the reversible attraction zone or cause the state to deviate from the reversible attraction zone are downweighted.

[0131] Among the inversion solutions that satisfy the stability and reversibility conditions, the inversion solutions that make the state evolution trajectory lie in the neighborhood of the reversible attraction band and have a continuous and smooth evolution process are preferentially retained, and the inversion solutions are regarded as effective inversion solutions after stabilization.

[0132] The target residual change corresponding to the target recovery state is used as the inversion input, and the current control input is used as the control reference. The recovery control quantity is calculated by the partial derivative response inversion operator.

[0133] In this embodiment, the step of synchronizing the recovery control quantity to the digital twin for virtual execution verification, and then issuing the recovery control quantity to the drone for recovery control after successful verification, includes:

[0134] Obtain the recovery control quantity, synchronize the recovery control quantity to the digital twin, and enable the digital twin to enter a virtual execution state with the recovery control quantity as the control input;

[0135] The operation process of the UAV is virtually executed in the digital twin according to the recovery control quantity, generating a virtual operation state sequence corresponding to the recovery control process. The virtual operation state sequence is consistent with the actual operation state of the UAV in terms of time sequence and state dimension.

[0136] Virtual residual data is generated based on the virtual running state sequence in the digital twin. The virtual residual data is compared with the residual data sequence to complete the virtual execution verification and judgment.

[0137] If the virtual execution verification is successful, the recovery control quantity will be sent to the UAV flight control system, so that the UAV will perform the recovery control operation according to the recovery control quantity;

[0138] During the process of the drone performing recovery control operations, real-time operation data of the drone is collected and synchronously fed back to the digital twin for updates.

[0139] Example 1:

[0140] To verify the feasibility of this invention in practice, it was applied to the routine operation of a power line inspection drone in a mountainous area. This region is characterized by hills and low mountains with significant topographic relief. The inspection routes follow ridges and valleys, requiring frequent adjustments to the drone's flight attitude and heading during missions. The flight environment also experiences continuous but unstable crosswind disturbances. After prolonged operation in this area, the drone's power system and sensors gradually exhibited performance drift, resulting in a non-linear deviation between flight control commands and actual responses. However, this deviation did not yet reach the traditional threshold alarm conditions, representing a typical non-parametric anomaly.

[0141] In this scenario, when the drone performs routine inspection tasks, the system first synchronously collects attitude data, trajectory data, motor speed and current data, control command data, and environmental wind speed data gathered during flight. It then performs time alignment and non-parametric preprocessing to form a standardized non-parametric operational dataset. This data is continuously mapped to a digital twin, enabling the digital twin to maintain consistency with the physical drone during operation.

[0142] During operation, the digital twin generates a predicted twin state based on standardized nonparametric operational data and performs differential processing with the measured operational state of the UAV to form a residual data sequence. By analyzing the changes in the residual data at multiple time scales, the system constructs a multifractal nested twin residual manifold in the digital twin space. This manifold can reflect the structural changes of the residuals under short-term disturbances and medium-term cumulative offset conditions, enabling the system to identify the geometric location of the anomalous state in the digital twin space before the anomaly is significantly amplified.

[0143] During an actual inspection, a drone exhibited slight attitude fluctuations and a slow increase in motor current after approximately twenty minutes of continuous flight. Traditional methods did not trigger an alarm, while the method of this invention, through a multi-fractal nested twin residual manifold, determined that the abnormal state had deviated from the steady-state core region and was located within the reversible attraction band. The system then constructed a reversible steady-state envelope in the digital twin space, including a steady-state core, a reversible attraction band, and a degenerate envelope layer, and determined the target recovery state accordingly.

[0144] During the recovery control phase, the system generates multiple sets of constrained control disturbances that satisfy the physical actuator constraints based on the current control input, and performs virtual execution within the digital twin to establish the correspondence between control disturbances and residual changes. Based on this, a partial derivative response inversion operator is constructed to calculate the recovery control quantity, which is then verified through virtual execution within the digital twin. Verification results show that the recovery control quantity effectively suppresses the residual growth trend. Afterward, the recovery control quantity is issued to the UAV for execution, and the flight attitude and motor load gradually recover to the stable range. Operational data collected during the recovery process is continuously fed back to the digital twin for subsequent updates to the residual structure and control response relationship.

[0145] During the month-long inspection mission, the method was verified multiple times on the same type of drone, covering different wind speeds and load conditions.

[0146] Table 1. Statistical Table of Non-parametric Anomaly Detection and Recovery Results in UAV Inspection Missions

[0147]

[0148] As shown in Table 1, the method of this invention can achieve early detection and effective recovery control of non-parametric anomalies of UAVs under different flight dates and environmental wind speeds. Within the average wind speed range of 2.1 m / s to 5.2 m / s, the anomaly detection lead time is stably distributed between 7 and 12 seconds, indicating that this invention does not rely on fixed thresholds or single-condition assumptions and can maintain relatively stable anomaly perception capabilities even under significant wind speed variations. Even under relatively high wind speeds of 5.2 m / s, the system can still complete the identification before the anomaly is significantly amplified, demonstrating good adaptability to complex environmental disturbances.

[0149] Regarding the recovery control performance, Table 1 shows that the overall recovery control time is controlled between 2.6s and 3.8s, and exhibits a reasonable slight upward trend with increasing wind speed. The recovery control process can dynamically adjust the control response intensity according to the actual severity of the abnormal state, without over-control or response lag. The maximum attitude deviation remains within the range of 2.2° to 3.6°, without drastic attitude fluctuations, reflecting that the control quantity verified based on digital twin virtual execution has good stability and security.

[0150] From the mission execution results, all test flights successfully completed their inspection tasks without any mid-flight return or mission interruption. This demonstrates that the method of this invention, while performing anomaly detection and recovery control, did not interfere with the normal mission flow of the UAV, but rather completed the state recovery in a smooth and controllable manner. In summary, the above analysis shows that this invention can balance detection proactiveness, recovery efficiency, and flight stability under real-world operating conditions, providing reliable technical support for the long-term safe operation of UAVs in complex environments.

[0151] 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 non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins, characterized in that, include: Collect multi-source operational data generated by drones during flight, preprocess the multi-source operational data to form a standardized non-parametric operational dataset, and construct a digital twin; The standardized nonparametric operational dataset is mapped to a digital twin, and the twin predicted state data at the corresponding time point is obtained and differentially processed with the measured operational state data of the UAV to generate a residual data sequence. Based on the residual data sequence, multi-scale division is performed according to a preset time scale to form a multi-scale residual segment set. Non-parametric manifold embedding processing is then performed to construct a multi-fractal nested twin residual manifold, and anomaly state identifiers and the geometric positions of the anomaly states in the digital twin space are output. In the digital twin space, a reversible steady envelope containing a steady-state kernel, a reversible attraction band, and a degenerate envelope layer is constructed. Based on the geometric position of the abnormal state in the digital twin space, the steady-state envelope level and the target recovery state corresponding to the abnormal state are determined. Based on the digital twin, the local partial derivative response relationship of the control input change to the residual data change is calculated. Combined with the control constraints of the UAV physical actuator, the partial derivative response inversion operator is constructed, and the corresponding recovery control quantity is calculated according to the target recovery state. The recovery control quantity is synchronized to the digital twin for virtual execution verification. After successful verification, the recovery control quantity is sent to the drone to execute recovery control, and the operational data collected during the recovery execution process is fed back to the digital twin.

2. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The multi-source operational data includes attitude and motion data, power system operational data, control command data, and environmental disturbance data.

3. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The preprocessing of multi-source operating data includes time synchronization, noise filtering, data alignment, and non-parametric normalization.

4. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The construction of the digital twin includes: Based on a standardized non-parametric operation dataset, the airframe structural parameters, power system parameters, sensor configuration parameters, and actuator connection parameters of the UAV are digitally mapped to form a structural mapping relationship. Based on a standardized non-parametric operation dataset, state correlation processing is performed on the attitude data, motion data, and dynamic data of UAVs under different control commands and environmental disturbances to form a state evolution mapping relationship. Based on the standardized non-parametric operation dataset and the collected control command data, a joint mapping is performed on the correspondence between changes in control input and changes in UAV state to form a control response mapping relationship. Consistency constraints are fused to integrate the structural mapping relationship, state evolution mapping relationship, and control response mapping relationship. The mapping relationship is synchronously updated based on real-time collected multi-source operation data to form a digital twin that corresponds to the real-time operation status of the UAV.

5. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The generation of the residual data sequence includes: The standardized non-parametric running dataset is loaded into the digital twin according to the data time stamp and feature correspondence, so that each running state variable in the digital twin and the corresponding data in the standardized non-parametric running dataset are mapped one-to-one. In a digital twin, standardized non-parametric operating data is synchronized based on a mapping relationship to generate a twin operating state corresponding to the current moment, and twin predictive state data is formed from the twin operating state. Under the same time marker, the twin predicted state data and the measured operating state data collected by the UAV are subjected to feature alignment processing in the digital twin; The twin prediction state data and the measured operating state data that have completed feature alignment are differentially processed item by item according to the corresponding features to obtain residual data that characterizes the deviation relationship between the twin prediction state and the measured operating state. The residual data obtained from multiple consecutive time markers are integrated and processed in chronological order to form a residual data sequence.

6. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The construction of the multifractal nested twin residual manifold, outputting the anomaly state identifier and the geometric location of the anomaly state in the digital twin space, includes: Obtain the residual data sequence, organize the residual data sequence in order according to the time stamp, and bind the residual data at each time step with the corresponding digital twin space state stamp to form a residual sample sequence with state stamp; Based on a preset time scale, the residual sample sequence with state labels is divided into multiple scales. At each time scale, continuous residual segments are extracted in a sliding window manner to form a set of multi-scale residual segments. Feature dimension consistency processing is performed on the set of multi-scale residual segments. At each time scale, nonparametric manifold embedding is performed on the corresponding set of residual segments to obtain the residual manifold representation at the time scale. During the embedding process, twin space state labels are introduced as neighborhood constraints to form a state-constrained residual manifold. A cross-scale nested construction process is performed on the state-constrained residual manifolds formed at different time scales. The cross-scale nested construction process includes determining the correspondence between residual manifolds at each scale, establishing a cross-scale alignment mapping, generating a cross-scale nested boundary, and combining the residual manifolds at each scale into a multifractal nested twin residual manifold based on the cross-scale nested boundary. Embedding and locating the current residual segment in a multifractal nested twin residual manifold, calculating the positional relationship of the current embedding point relative to the cross-scale nesting boundary, outputting an anomaly status identifier based on the positional relationship, and outputting the geometric position of the anomaly status in the digital twin space.

7. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The step of determining the steady-state envelope level and target recovery state corresponding to the abnormal state based on its geometric location in the digital twin space includes: Based on multifractal nested twin residual manifolds, residual embedding samples corresponding to normal operation periods are selected in the digital twin space, and organized according to time markers and feature dimensions to determine the sample set that simultaneously satisfies residual stability and state evolution consistency at each scale. The minimum enclosing domain of the sample set is used as the initial region boundary of the steady-state kernel. In a digital twin, a bidirectional rolling simulation is performed around the initial region boundary of the steady-state kernel. Amplitude-limited control perturbations and environmental perturbations are applied to the state of the boundary neighborhood. The set of states that can revert to the steady-state kernel within a finite number of steps without external reconfiguration is recorded. The continuous region between the state set and the steady-state kernel is defined as a reversible attraction zone. For the remaining states that do not meet the conditions for reversible attraction bands, a local reversible geometric transformation generation process is executed in the digital twin. Multiple sets of candidate geometric transformations are generated through local state reparameterization, control law fine-tuning, and constraint unification. States that can enter the reversible attraction band within a finite number of steps are uniformly collected. The collected state set is determined as the degenerate envelope layer, and the corresponding local reversible geometric transformation label is recorded for each type of state. In the digital twin, the boundaries of the steady-state kernel, reversible attraction band and degenerate envelope are self-consistently calibrated. Antifactual perturbation experiments and time-symmetric replay verifications are performed in the boundary neighborhood. Based on the consistency of the cross-scale residual structure and the consistency of state evolution, the boundaries of the three types of regions are synchronously updated to form a reversible steady-state envelope. Based on the geometric location of the abnormal state in the digital twin space, determine its corresponding level: When the position is within the steady-state kernel, the target recovery state is determined to be the current corresponding steady-state point; When the location is within the reversible attraction zone, the target recovery state is determined to be the steady state point with the shortest time to reach steady-state nuclear evolution. When the location is within the degenerate envelope, the intermediate target entering the reversible attraction zone is determined based on the corresponding local reversible geometric transformation marker, and the final target recovery state is determined.

8. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The partial derivative response inversion operator is constructed by combining the control constraints of the UAV's physical actuators, and the corresponding recovery control quantity is calculated based on the target recovery state, including: Obtain the residual data corresponding to the current abnormal state from the digital twin, and obtain the current control input corresponding to the current abnormal state, where the residual data comes from the residual data sequence and the current control input comes from the control command data; In the digital twin, multiple sets of constrained control disturbance sequences are generated based on the current control input. The constrained control disturbance sequences are synchronized to the digital twin for virtual execution. The residual change data corresponding to each set of constrained control disturbance sequences are recorded to form a corresponding sample set of control disturbance and residual change. Based on the corresponding sample set, the local partial derivative response relationship of control input change to residual data change is constructed. By performing consistency screening on the residual change results corresponding to control disturbances of different amplitudes, the control disturbance results that maintain a stable change trend in the geometric neighborhood of multi-scale residual structure and abnormal state are retained, forming the local partial derivative response relationship of structural constraint. A partial derivative response inversion operator is constructed based on the structural constraint local partial derivative response relationship. The partial derivative response inversion operator consists of control constraint projection processing, residual target alignment processing, and inversion stabilization processing, wherein: Control constraint projection processing maps candidate control input variations to feasible control domains that satisfy physical actuator constraints; The residual target alignment process aligns the current residual change direction with the residual change direction corresponding to the target recovery state. The inversion stabilization process suppresses ill-posed directions during the inversion process and preferentially preserves inversion solutions in the neighborhood of the reversible attraction band. The target residual change corresponding to the target recovery state is used as the inversion input, and the current control input is used as the control reference. The recovery control quantity is calculated by the partial derivative response inversion operator.

9. The method for non-parametric data fault detection and recovery of unmanned aerial vehicles based on digital twins according to claim 1, characterized in that, The process of synchronizing the recovery control quantity to the digital twin for virtual execution verification, and then issuing the recovery control quantity to the drone for recovery control after successful verification, includes: Obtain the recovery control quantity, synchronize the recovery control quantity to the digital twin, and enable the digital twin to enter a virtual execution state with the recovery control quantity as the control input; The operation process of the UAV is virtually executed in the digital twin according to the recovery control quantity, generating a virtual operation state sequence corresponding to the recovery control process. The virtual operation state sequence is consistent with the actual operation state of the UAV in terms of time sequence and state dimension. Virtual residual data is generated based on the virtual running state sequence in the digital twin. The virtual residual data is compared with the residual data sequence to complete the virtual execution verification and judgment. If the virtual execution verification is successful, the recovery control quantity will be sent to the UAV flight control system, so that the UAV will perform the recovery control operation according to the recovery control quantity; During the process of the drone performing recovery control operations, real-time operation data of the drone is collected and synchronously fed back to the digital twin for updates.

Images