Fault-tolerant recovery method and system for unmanned aerial vehicle-unmanned ship in inland water area

Abstract

The invention provides an inland water area unmanned aerial vehicle-unmanned ship fault-tolerant recovery method and system, and relates to the technical field of unmanned autonomous system control, and the method comprises the steps: constructing a water-air cooperative digital twin simulation environment; configuring an unmanned aerial vehicle model and an unmanned ship model; collecting multi-dimensional sensor time sequence data of the unmanned aerial vehicle in a normal flight state; based on a mask noise self-supervision training strategy, training the de-noising auto-encoder network to obtain a trained de-noising auto-encoder network; inputting sensor data collected in real time into the trained de-noising auto-encoder network, and outputting reconstructed data; calculating a reconstruction error between the sensor data and the reconstruction data; obtaining effective state data based on the reconstruction error and a self-adaptive threshold value; training a fault-tolerant control strategy network by adopting a near-end strategy optimization algorithm; and inputting the real-time effective state data into the trained fault-tolerant control strategy network, outputting a control instruction, and completing collaborative recovery to the unmanned ship model.

Description

Technical Field

[0001] This invention relates to the field of unmanned autonomous system control technology, and in particular to a fault-tolerant recovery method and system for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waterways. Background Technology

[0002] With the in-depth development of unmanned systems technology, cross-domain air-water collaboration has shown great potential in inland water monitoring, emergency search and rescue and other fields. The complex terrain and flow field characteristics of inland waters have put forward higher requirements for the stability and reliability of UAV-unmanned vessel collaborative recovery. Digital twin simulation platform and intelligent fault-tolerant control technology have become key directions to support the stable operation of the system in complex environments.

[0003] In existing technologies, linear control, PID controllers and other solutions provide basic support for the control of unmanned systems, and physical testing provides direct evidence for technology verification. These technologies ensure the basic operation of the system in structured environments and fault-free conditions, and accumulate practical experience for the technological development of collaborative systems. At the same time, traditional modeling methods also provide a theoretical basis for describing the basic dynamic characteristics of the system.

[0004] However, traditional linear models often ignore higher-order nonlinear terms of the system, making it difficult to adapt to extreme operating conditions. Physical testing is costly and carries significant risks, and obtaining massive amounts of fault data is difficult, making it hard to meet the fault-tolerant control requirements in complex environments. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waters. This method can solve the technical problems of traditional linear models often ignoring higher-order nonlinear terms of the system, making it difficult to adapt to extreme working conditions, resulting in high costs and huge risks in physical testing, difficulty in obtaining massive amounts of fault data, and difficulty in meeting the fault-tolerant control requirements in complex environments.

[0006] A first aspect of this invention provides a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned surface vessels (USVs) in inland waterways, comprising: S1: Construct a simulation environment for a water-air collaborative digital twin; S2: Based on the simulation environment, configure the drone model and the unmanned vessel model; S3: Collect multi-dimensional sensor time-series data under normal flight conditions of the UAV, and construct a training dataset based on the multi-dimensional sensor time-series data; S4: Based on the mask noise self-supervised training strategy, the training dataset is input into the denoising autoencoder network to train the denoising autoencoder network and obtain the trained denoising autoencoder network. S5: Input the real-time sensor data into the trained denoising autoencoder network and output the reconstructed data; S6: Calculate the reconstruction error between the sensor data and the reconstructed data; S7: Based on the reconstruction error and adaptive threshold, obtain effective state data; S8: Based on valid state data, a near-end policy optimization algorithm is used to train the fault-tolerant control policy network; S9: Obtain real-time valid status data; S10: Input the real-time valid state data into the trained fault-tolerant control strategy network, output control commands, and drive the UAV model to complete the cooperative recovery to the unmanned vessel model according to the control commands.

[0007] A second aspect of the present invention provides a fault-tolerant recovery system for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waterways, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method as described in the first aspect.

[0008] A third aspect of the present invention provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method as described in the first aspect.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, to address the shortcomings of traditional linear models that neglect higher-order nonlinear terms, have high costs and risks associated with physical testing, and suffer from a lack of fault data, a water-air collaborative digital twin simulation environment is constructed and a high-fidelity model is configured to accurately reproduce the system's dynamic characteristics and fault mechanisms, adapting to extreme operating conditions. A denoising autoencoder network is trained based on a masked noise self-supervised training strategy, enabling the acquisition of training data without physical testing, thus facilitating sensor fault diagnosis and failure data reconstruction. Combining domain randomization technology, a fault-tolerant control strategy network is trained using a near-end policy optimization algorithm. Through multi-objective optimization guidance, the system can still fly stably and recover accurately under extreme fault conditions, significantly improving the robustness and safety of the air-water cross-domain collaborative system in complex inland environments. Attached Figure Description

[0010] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0011] Figure 1 This is a schematic flowchart of a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waters provided by an embodiment of the present invention.

[0012] Figure 2 This is an overall flowchart of a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waters provided by an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of a noise reduction autoencoder structure provided in an embodiment of the present invention.

[0014] Figure 4 This is a convergence curve of the reconstruction error of a fault diagnosis network provided in an embodiment of the present invention.

[0015] Figure 5 This is a cumulative reward curve for training a fault-tolerant control strategy network provided in an embodiment of the present invention.

[0016] Figure 6 This is a comparison diagram of flight trajectories under sudden sensor failure provided by an embodiment of the present invention.

[0017] Figure 7 This is a schematic diagram of the structure of an inland waterway unmanned aerial vehicle (UAV)-unmanned vessel fault-tolerant recovery system provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] Reference manual attached Figure 2 The diagram shows an overall flowchart of a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waters provided by an embodiment of the present invention.

[0020] Specifically, Figure 2 It comprises three core layers: simulation environment layer, perception and diagnosis layer, and decision and control layer. The simulation environment layer serves as the carrier of physical and fault scenarios, integrating dynamic models, environmental disturbances, and fault injection modules. The dynamic models cover 3-DOF models for unmanned surface vessels and 6-DOF models for unmanned aerial vehicles, environmental disturbances include water flow and wind field models, and the fault injection module covers three types of fault models: actuators, sensors, and water flow disturbances.

[0021] Furthermore, the perception and diagnosis layer is the core of fault diagnosis and data repair. Taking 14-dimensional sensor time-series data as input, it generates reconstructed data through a denoising autoencoder (including encoder, bottleneck layer, and decoder). Faults are then determined by comparing residuals with thresholds, ultimately outputting the original or reconstructed valid state data. The decision and control layer is responsible for outputting control commands. Its core is the PPO reinforcement learning controller, whose policy network receives valid state data and outputs 4-dimensional control quantities. A value network assists in optimization, ultimately converting the control quantities into PWM commands to drive the actuators. In terms of inclusion, each layer integrates corresponding functional modules, forming a complete technical support system. In terms of connectivity, the simulation environment layer transmits sensor data to the perception and diagnosis layer, the perception and diagnosis layer sends valid data to the decision and control layer, and the PWM commands output by the decision and control layer are fed back to the simulation environment layer, forming a closed-loop control link.

[0022] It should be noted that, Figure 2 The three-layer closed-loop architecture clearly presents the collaborative logic of high-fidelity simulation, fault diagnosis and repair, and robust control, providing complete architectural support for the stable implementation of fault-tolerant recovery of unmanned aerial vehicles and unmanned vessels in inland waters.

[0023] Reference manual attached Figure 1 The diagram illustrates a process flow diagram of a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned vessels in inland waters provided by an embodiment of the present invention.

[0024] This invention provides a fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned surface vessels (USVs) in inland waters, which may include the following steps: S1: Construct a simulation environment for a water-air collaborative digital twin.

[0025] Among them, the simulation environment of water-air collaborative digital twin refers to a virtual verification platform that integrates the physical characteristics and fault injection mechanisms of inland water and airspace to reproduce the dynamic characteristics and complex working conditions of the system with high fidelity.

[0026] In one possible implementation, S1 specifically includes sub-steps S101 to S105: S101: Define the inertial coordinate system and the body coordinate system.

[0027] Among them, the inertial coordinate system refers to the reference system that describes the absolute position, linear velocity and gravitational acceleration of the system, while the body coordinate system refers to the reference system that describes the thrust vector, angular velocity and airborne sensor data.

[0028] For example, defining an inertial coordinate system Used to describe the absolute position of the system linear velocity and gravitational acceleration Define the body coordinate system of UAVs and unmanned surface vessels. Used to describe thrust vector and angular velocity And airborne sensor data.

[0029] S102: Determine the yaw angle, pitch angle, and roll angle of the UAV based on the inertial coordinate system and the body coordinate system.

[0030] Among them, the yaw angle is the angle around the Z-axis that determines the direction of the nose, the pitch angle is the angle around the Y-axis that determines the nose lift, and the roll angle is the angle around the X-axis that determines the fuselage tilt.

[0031] S103: Based on yaw, pitch, and roll angles, construct the three basic rotation matrices corresponding to the UAV according to the ZYX rotation order: in, Indicates circling X Axis rotation Rotation matrix of angle, Indicates circling Y Axis rotation Rotation matrix of angle, Indicates circling Z Axis rotation Rotation matrix of angle, sin Represents the sine function. cos This represents the cosine function.

[0032] Among them, the basic rotation matrix is ​​the coordinate transformation matrix corresponding to a single Euler angle, and the coordinate transformation matrix is ​​the mapping matrix between the unified inertial coordinate system and the body coordinate system.

[0033] S104: Perform matrix multiplication on each basic rotation matrix to obtain the coordinate transformation matrix of the UAV.

[0034] S105: Based on the coordinate transformation matrix, construct a water-air collaborative digital twin simulation environment that includes rigid body dynamics, actuator failure mechanisms, and environmental interference effects.

[0035] Specifically, the problem of inconsistency between the actuator reference system and the motion trajectory reference system is first solved by defining two types of coordinate systems. Then, coordinate transformation relationships are constructed through Euler angles and rotation matrices. Finally, a high-fidelity simulation environment that integrates physical characteristics, fault mechanisms and environmental interference is constructed based on this transformation matrix, providing a unified and realistic reference system foundation for subsequent model configuration and algorithm training.

[0036] For example, an inertial coordinate system can be fixed at a point on the Earth's surface to describe the UAV's absolute position and gravitational acceleration. A body coordinate system can be fixed at the UAV's center of mass to describe the thrust vector generated by the four propellers. When constructing the rotation matrix, the yaw angle can be set as the angle by which the nose deviates from true north, the pitch angle can be set as the angle between the nose and the horizontal plane, and the roll angle can be set as the tilt angle between the fuselage and the vertical plane. After constructing the corresponding rotation matrices in the ZYX order, the final coordinate transformation matrix is ​​obtained through matrix multiplication.

[0037] Specifically, to accurately reproduce fault conditions and environmental disturbances, a programmable fault injection model is introduced into the simulation environment, covering three typical fault types: actuators, sensors, and external water flow environment. The mathematical model is as follows: 1. Actuator Failure Model: Used to simulate power output attenuation caused by propeller entanglement with aquatic plants, motor aging, or battery voltage drop in inland waters. The formula is: in, Indicates actual power output. Represents the fault severity matrix. This represents the ideal PWM instruction output by the controller. . Indicates the first i The battery lost 30% of its power.

[0038] 2. Sensor Fault Model Used to simulate visual failures caused by river bridges blocking GPS signals or water surface reflections: in, This represents the sensor's measured value. Indicates the signal loss factor. This represents the actual value corresponding to the sensor. This represents Gaussian white noise. This represents a constant deviation, simulating positioning drift caused by GPS multipath effects.

[0039] Furthermore, This indicates that the sensor is working properly. This simulates a fault where "the sensor is stuck / the data returns to zero".

[0040] 3. Inland water flow disturbance model This describes the impact of continuous river flow velocity on the motion of unmanned surface vessels (USVs). The actual velocity of the USV relative to the ground is a vector sum of the vessel's velocity relative to the water and the water flow velocity, expressed by the following formula: in, This indicates the actual speed of the unmanned vessel relative to the ground. This represents the coordinate transformation matrix from the unmanned vessel's body coordinate system to the Earth coordinate system. This indicates the speed of the unmanned vessel relative to the water. This indicates the speed of water flow in a river.

[0041] It should be noted that this method achieves precise unification of the force and motion reference frame of the UAV through the definition of a standardized coordinate system and the construction logic of the rotation matrix, providing a rigorous mathematical foundation for subsequent dynamic modeling. The constructed simulation environment integrates rigid body characteristics, fault mechanisms, and environmental disturbances, and can reproduce complex working conditions such as unsteady gusts in inland waters and GPS multipath effects with high fidelity. It can simulate extreme scenarios without relying on physical testing, which avoids the high cost and risks of physical testing, and provides a reliable virtual verification platform for subsequent model configuration, data acquisition, and algorithm training.

[0042] In this embodiment of the invention, this step does not rely on physical testing and can safely simulate complex scenarios such as valley gusts and sensor failures, solving the problems of high cost and high risk of physical testing, and providing an efficient and reliable basic environment for subsequent model configuration, data collection and algorithm training.

[0043] S2: Based on the simulation environment, configure the drone model and the unmanned vessel model.

[0044] Among them, the UAV model is a mathematical model describing its six-degree-of-freedom dynamic characteristics, and the unmanned ship model is a three-degree-of-freedom motion model considering fluid coupling effects.

[0045] In one possible implementation, S2 specifically includes sub-steps S201 and S202: S201: Configure the drone model based on the simulation environment.

[0046] In one possible implementation, S201 specifically includes sub-steps S2011 to S2015: S2011: Based on the simulation environment, determine the types of forces experienced by the UAV during its translational motion.

[0047] Among them, the force type refers to the gravity, airframe thrust, and air resistance that the drone experiences during flight.

[0048] S2012: Based on Newton's second law, and combining the force type and coordinate transformation matrix, the translational characteristic equation in the inertial coordinate system is constructed as follows: in, m Indicates the quality of the drone. P This represents the translational acceleration of the UAV in the inertial coordinate system. This represents the projection of the UAV's thrust in the inertial coordinate system. mg This indicates the gravitational force acting on the drone. This indicates the air resistance experienced by the drone during its translational motion.

[0049] Among them, the translational characteristic equation is a mathematical expression based on Newton's second law that describes the translational motion of the UAV.

[0050] S2013: Construct the rotational inertia matrix of the UAV in the body coordinate system: in, The moment of inertia matrix of the UAV is a symmetric matrix. These represent the moments of inertia about the X, Y, and Z axes of the body coordinate system, respectively.

[0051] The moment of inertia matrix refers to the symmetric matrix that describes the rotational inertia of the UAV, and its off-diagonal elements are 0.

[0052] S2014: Based on Euler's equations, combined with the moment of inertia matrix and core torque, the rotational characteristic equations of the UAV in the body coordinate system are constructed as follows: in, This represents the moment of inertia matrix of the UAV in the body coordinate system. This indicates the angular velocity of the drone. This represents the total torque during the rotation of the drone. This represents the nonlinear Coriolis torque. Indicates control torque. This indicates the disturbance torque experienced by the drone during its rotation.

[0053] Furthermore, the disturbance torque experienced by the drone during its rotation also includes: in, This represents the gyroscopic torque generated by the high-speed rotation of the propeller itself. This represents the air drag torque.

[0054] Among them, the core torque refers to the control torque, disturbance torque and nonlinear Coriolis torque that affect the attitude of the UAV, and the rotation characteristic equation is a mathematical expression based on the Euler equation to describe the rotation law of the UAV.

[0055] S2015: Based on the translational and rotational characteristic equations, complete the configuration of the UAV model.

[0056] For example, air resistance in translational forces can simulate aerodynamic losses in a valley gust environment; the rotational inertia matrix can be set to a symmetrical matrix form according to the UAV's structural design. The Coriolis torque in the core torque is close to zero when the UAV is hovering smoothly, but when a motor stops, causing the UAV to spin at high speed, this term increases exponentially, becoming a key factor affecting attitude stability. Gyroscopic torque is the torque generated by the high-speed rotation of the propeller itself, and air resistance torque is the opposing torque of the airflow on the aircraft during flight.

[0057] It should be noted that by accurately identifying the types of translational forces and integrating coordinate transformation matrices to construct the equations of motion, higher-order terms such as nonlinear Coriolis torques are fully preserved, thus solving the problem of extreme condition adaptation failure caused by the neglect of nonlinear characteristics in traditional linear models. The constructed six-DOF UAV model can accurately reproduce the attitude instability characteristics under motor failure and valley gust disturbances, providing high-fidelity physical model support for subsequent fault data acquisition, denoising autoencoder training, and fault-tolerant control strategy verification, ensuring the effectiveness and reliability of algorithm training.

[0058] S202: Based on the characteristics of the inland water flow field, an unmanned vessel model is configured.

[0059] Among them, the UAV model refers to the mathematical model that describes its six-degree-of-freedom translational and rotational dynamics characteristics, while the unmanned vessel model refers to the three-degree-of-freedom mathematical model that adapts to the flow field characteristics of inland waters and describes its planar motion laws.

[0060] Specifically, this step involves configuring the models in stages, taking into account the different motion characteristics and application scenarios of UAVs and unmanned vessels: the UAV model needs to fully reproduce its rigid body dynamics and nonlinear coupling terms, while the unmanned vessel model needs to be adapted to the fluid coupling effect and inertial delay characteristics of inland waters, ensuring that both types of models can fit the actual working conditions.

[0061] For example, when configuring a drone model, it is necessary to consider the types of forces such as gravity, thrust, and air resistance, and retain nonlinear terms such as Coriolis torque to reproduce the attitude instability characteristics when a single propeller stops turning; when configuring an unmanned vessel model, the additional mass of the water surrounding the hull needs to be included in the inertia matrix to reflect the inertial delay characteristics of the unmanned vessel's motion in inland waters.

[0062] In one possible implementation, S202 specifically includes sub-steps S2021 to S2024: S2021: Based on the characteristics of the inland water flow field, determine the inertial delay characteristics of unmanned vessel motion.

[0063] Among them, the inertial delay characteristic refers to the motion lag caused by the unmanned vessel propelling itself and the water layer around its hull.

[0064] S2022: Based on the planar motion law of the unmanned vessel, determine the matrix terms in the motion equation of the unmanned vessel.

[0065] Among them, the matrix terms refer to the core matrices in the unmanned vessel's motion equations, such as the inertia matrix, the Coriolis centripetal force matrix, and the hydrodynamic drag matrix.

[0066] S2023: Based on matrix terms, and combining Fossen's ship motion equations with the fluid coupling effect of water, the motion equations of unmanned vessels are constructed: in, M The inertial matrix represents the unmanned surface vessel. The time derivative of the unmanned surface vessel's motion state vector. This represents the planar motion state vector of the unmanned surface vessel. Represents the Coriolis centripetal force matrix. Represents the hydrodynamic resistance matrix. This indicates the propulsion torque of the unmanned vessel. This represents the generalized force vector of environmental disturbances experienced by the unmanned vessel.

[0067] Among them, the planar motion state vector of the unmanned vessel These represent the sway velocity, roll velocity, and bow roll rate, respectively.

[0068] Furthermore, for the unmanned vessel subsystem, given the relatively small hull tilt angle during inland water operations, a simplified three-degree-of-freedom kinematic equation is adopted, considering only the transformations of planar position and heading angle, resulting in the following coordinate transformation matrix for the unmanned vessel: in, This represents the coordinate transformation matrix of the unmanned surface vessel. Indicates the heading angle of the unmanned vessel. Indicates heading angle The sine value, Indicates heading angle The cosine value.

[0069] Among them, Fossen's equations of motion refer to the classical dynamic equations adapted to the planar motion of ships; fluid coupling effect refers to the influence of the water medium on the motion of unmanned ships.

[0070] S2024: Based on the equations of motion of the unmanned vessel, complete the configuration of the unmanned vessel model.

[0071] It should be noted that, considering the high water density and complex flow field characteristics of inland waters, a model was constructed by clearly defining the inertial delay characteristics, precisely defining matrix terms, and combining them with Fossen's ship motion equations. The inertial matrix includes added mass, and the hydrodynamic resistance matrix adapts to different sailing speeds, accurately reproducing the real motion law of the unmanned surface vessel (USV). The configured USV model can effectively simulate the flow field disturbances and motion delays in inland waters, and works in conjunction with the UAV model to support collaborative recovery simulation. This provides a virtual platform that fits the actual working conditions for subsequent algorithm training, ensuring the practicality of the collaborative recovery strategy.

[0072] In this embodiment of the invention, this step accurately reproduces nonlinear terms such as the Coriolis torque of the UAV and the inertial delay characteristics of the unmanned vessel, avoiding the adaptation limitations of traditional linear models and providing a realistic physical model support for subsequent data acquisition and control training.

[0073] S3: Collect multi-dimensional sensor time-series data under normal flight conditions of the UAV, and construct a training dataset based on the multi-dimensional sensor time-series data.

[0074] Among them, multidimensional sensor time-series data refers to 14-dimensional state vector time-series data collected by sensors such as UAV IMU, vision, and GPS. It includes three-dimensional position, three-dimensional velocity, three-dimensional Euler angles, visual relative position, and system flags. The training dataset is a sample set that integrates normal data and fault simulation data.

[0075] In one possible implementation, S3 specifically includes sub-steps S301 to S303: S301: Collects multi-dimensional sensor time-series data during the flight of the UAV.

[0076] For example, it is possible to control a drone to complete 2,000 fault-free flight paths in a simulated environment.

[0077] Among them, multidimensional sensor time-series data refers to 14-dimensional state time-series data collected by sensors such as UAV IMU and vision.

[0078] For example, in the 14-dimensional sensor data collected, the first 9 dimensions are IMU and odometer data, the middle 3 dimensions are visual relative positioning data, and the last 2 dimensions are system flag bits.

[0079] S302: Perform masking noise operation on the multidimensional sensor time series data to obtain corrupted data.

[0080] Masking noise operation refers to the process of randomly setting some data dimensions to zero to simulate sensor failure. Corrupted data refers to sensor data that has been processed by masking noise operation to simulate failure; training dataset refers to the sample set that integrates normal data and corrupted data for network training.

[0081] For example, a mask matrix can be randomly generated with a 20% probability, and some dimensions of the visual relative positioning data can be set to zero to simulate GPS signal loss caused by river bridge obstruction.

[0082] S303: Integrate multidimensional sensor time-series data and damage data to obtain a training dataset.

[0083] For example, 2,000 original data entries can be matched one-to-one with the corresponding masked corrupted data entries to form a training dataset of 4,000 samples.

[0084] It should be noted that by collecting fault-free data first and then simulating faults, sufficient training samples can be obtained without physical fault testing, solving the problem of the scarcity of massive fault data. The masking noise operation accurately simulates sensor failure scenarios, and the integrated training dataset takes into account both normal and fault conditions, providing a realistic sample basis for the subsequent self-supervised training of the denoising autoencoder network, ensuring that the network can learn the physical constraints between data.

[0085] In this embodiment of the invention, this step can obtain sufficient training data without actual fault testing, which solves the problem of the scarcity of massive fault data, provides a sample basis that fits the working conditions for the training of the denoising autoencoder network, and ensures the learning effect of the network.

[0086] S4: Based on the mask noise self-supervised training strategy, the training dataset is input into the denoising autoencoder network to train the denoising autoencoder network and obtain the trained denoising autoencoder network.

[0087] Among them, the mask noise self-supervised training strategy refers to the unsupervised training method that simulates faults by randomly masking part of the data, and the denoising autoencoder network is a fault diagnosis and data reconstruction network containing encoder and decoder.

[0088] In one possible implementation, S4 specifically includes sub-steps S401 to S403: S401: Construct a denoising autoencoder network that includes an encoder and a decoder.

[0089] For example, the number of nodes in the encoder layer can be set to [14→32→16→8]. By reducing the dimensionality layer by layer and removing redundant noise, the number of nodes in the decoder layer can be set to [8→16→32→14]. The low-dimensional features are mapped back to the 14-dimensional original data space. The hidden layer uses the ReLU activation function, and the output layer uses a linear function.

[0090] Reference manual attached Figure 3 The diagram shows a denoising autoencoder structure provided by an embodiment of the present invention.

[0091] Specifically, Figure 3 In this framework, the original sensor data is a 14-dimensional vector (formatted as 2×7), representing the raw state data acquired by the unmanned system's sensors. The random mask module simulates sensor failure, partially obscuring the original sensor data. The corrupted data is a 14-dimensional vector (formatted as 2×7dim) obtained after processing with the random mask, corresponding to the failed data after the sensor failure. The encoder is the compression component of the denoising autoencoder (DAE), used to compress the corrupted data into low-dimensional core features. The latent variable z is the compressed feature vector output by the encoder, representing the key form of the data. The decoder is the restoration component of the DAE, used to restore the latent variable z to complete data. The reconstructed data is a 14-dimensional vector (formatted as 2×7dim) output by the decoder, representing the repaired sensor data. The optimization objective, "Minimizing the Reconstruction Error (MSE Loss)," is the network's training criterion, used to measure the degree of difference between the reconstructed data and the original sensor data.

[0092] Furthermore, in terms of inclusion relationships: this DAE network architecture integrates a random mask module, an encoder, and a decoder, while also linking the original sensor data, corrupted data, latent variable z, reconstructed data, and the "Minimize Reconstruction Error (MSELoss)" optimization objective used for training. In terms of connectivity relationships: the original sensor data is connected to the random mask module, which partially masks the data to generate corrupted data. The corrupted data is connected to the encoder, which compresses it into latent variable z. The latent variable z is connected to the decoder, which restores and outputs the reconstructed data. The reconstructed data and the original sensor data are linked through the "Minimize Reconstruction Error (MSELoss)" optimization objective, which guides the optimization and adjustment of network parameters.

[0093] It should be noted that, Figure 3 The denoising autoencoder (DAE) enables accurate repair of sensor fault data, improves the reliability of perception information of unmanned systems in fault scenarios, and provides stable perception data support for subsequent collaborative recovery control of UAVs and unmanned vessels.

[0094] S402: Based on a mask noise self-supervised training strategy, the denoising autoencoder network is optimized through a loss function. in, Represents the loss function. Indicates to N Summing the training data, Represents the square of the L2 norm. Represents the decoder function. Represents the encoder function. This represents the reconstructed output data of the network. Indicates the first i Original, undamaged sensor data.

[0095] The loss function refers to the mean squared error function that measures the error between the reconstructed data and the original data.

[0096] It should be noted that the sensor data is first deliberately "damaged" (simulating a fault), and then the network is trained to repair the "bad data" to be close to the original data, so that the network can learn to use complementary information between sensors to deal with faults.

[0097] For example, corrupted data from the training dataset can be input into the network, with the mean square error between the reconstructed output and the original uncorrupted data as the optimization objective.

[0098] S403: Iterate and optimize until the loss value of the loss function is less than the preset loss value, and obtain the trained denoising autoencoder network.

[0099] The preset loss value refers to the critical error value used to determine whether the network training has reached the target and can accurately reconstruct the data.

[0100] It should be noted that those skilled in the art can set the preset loss value according to actual needs, and this invention does not limit that.

[0101] Specifically, this step first constructs a symmetric network architecture adapted to high-dimensional sensor data, then adopts a mask noise self-supervised training strategy to optimize network parameters with the goal of minimizing reconstruction error, and continuously iterates until the loss value is lower than the preset standard to obtain a denoising autoencoder network that can accurately learn the physical manifold of the system.

[0102] For example, the preset loss value can be set to the order of 10⁻³. When the network training is iterated until the loss value is stably lower than this value, the training is considered complete.

[0103] Reference manual attached Figure 4 The figure shows a convergence curve of reconstruction error of a fault diagnosis network provided by an embodiment of the present invention.

[0104] Specifically, Figure 4The horizontal axis represents the number of training epochs, indicating the number of training iterations for the fault diagnosis network (denoising autoencoder). The vertical axis represents the mean squared error (MSE Loss), used to measure the deviation between the network's reconstructed data and the original data. The solid line corresponds to the training set error, and the dashed line corresponds to the validation set error. The annotation "converged to the order of 10⁻³ (indicating the acquisition of the physical manifold)" means that when the error stabilizes within this range, the network has mastered the physical constraints for the normal operation of the system.

[0105] Furthermore, in terms of inclusion relationships, this curve represents the results of the fault diagnosis network training process, integrating error data from the training and validation sets, as well as convergence status annotations. Regarding connectivity, as the number of training rounds increases, the errors in both the training and validation sets decrease synchronously, eventually converging to the order of 10⁻³, demonstrating the stability and effectiveness of the network training.

[0106] It should be noted that, Figure 4 By visually demonstrating the error convergence process, the fault diagnosis network was verified to effectively learn the physical constraints of the system, providing reliable model support for the accurate diagnosis and data reconstruction of subsequent sensor faults.

[0107] It should be noted that the constructed symmetric network architecture is adapted to the characteristics of high-dimensional sensor data, and the self-supervised training strategy does not require labeling fault samples, thus reducing data dependence. Using the loss value being less than a preset value as the stopping condition for iteration ensures that the network can accurately learn the physical manifold of the system, providing a high-performance core model for subsequent sensor fault diagnosis and failure data reconstruction, and avoiding the dependence of traditional methods on precise mathematical observers.

[0108] In this embodiment of the invention, this step does not require labeling fault samples and can be trained using only normal data, reducing data dependence. The network can learn the physical constraints between sensor data, laying the core model foundation for subsequent fault diagnosis and data repair.

[0109] S5: Input the real-time sensor data into the trained denoising autoencoder network and output the reconstructed data.

[0110] Among them, real-time sensor data refers to the raw sensor data acquired in real time during UAV operations, and reconstructed data refers to reliable state data that replaces fault data generated by the network based on physical constraints.

[0111] In this embodiment of the invention, this step enables real-time virtual repair of fault data without the need for a precise mathematical observer, thus compensating for data failure caused by sensor malfunctions, ensuring the continuity of state data, and providing support for the stability of the control loop.

[0112] S6: Calculate the reconstruction error between the sensor data and the reconstructed data.

[0113] Among them, reconstruction error refers to the residual between real-time sensor data and reconstructed data, reflecting the degree to which the data deviates from the normal physical manifold of the system.

[0114] In this embodiment of the invention, this step provides a quantitative basis for sensor fault diagnosis, can accurately capture data anomalies, avoid the one-sidedness of traditional threshold judgment, ensure the agility and accuracy of fault detection, and provide a key reference for subsequent acquisition of effective status data.

[0115] S7: Based on the reconstruction error and the adaptive threshold, obtain the effective state data.

[0116] Among them, the adaptive threshold refers to the critical value set based on the statistical characteristics of the reconstruction error when the system is running normally, and the effective state data refers to the reliable state data used for control training after fault determination.

[0117] It should be noted that those skilled in the art can set the size of the adaptive threshold according to actual needs, and this invention does not limit this.

[0118] In one possible implementation, S7 specifically includes sub-steps S701 and S702: S701: Set an adaptive threshold based on reconstruction error.

[0119] For example, by statistically analyzing the maximum value and distribution pattern of 1000 sets of reconstruction errors under fault-free conditions, the adaptive threshold can be set to 1.2 times the maximum value.

[0120] S702: Determine if the reconstruction error is less than or equal to the adaptive threshold. If yes, use the real-time acquired sensor data as valid status data. Otherwise, determine that the sensor is faulty and use the reconstructed data as valid status data. in, r This represents the residual vector between the real-time acquired sensor data and the reconstructed data, used to quantify the degree of deviation between the two. This represents the raw sensor data collected in real time. This represents the reconstructed data output by the denoising autoencoder network. This represents the composite reward function, used to guide the training direction of the fault-tolerant control policy network. Indicates distance reward. This represents a penalty for attitude stability. This indicates a smoothing penalty for the control quantity. This indicates a sparse success reward. This indicates the operation of taking the absolute value.

[0121] Among them, reconstructed data refers to reliable data that replaces fault data generated by the denoising autoencoder network.

[0122] Specifically, this step first analyzes the reconstruction error distribution characteristics during normal system operation, sets a reasonable adaptive threshold as a fault judgment criterion, and then automatically identifies whether the sensor is faulty by comparing the magnitude of the reconstruction error with the threshold in real time. In turn, reliable status data is selected as the input for subsequent control to ensure the stability of the control loop.

[0123] For example, when the visual sensor's data drops to zero due to mirror reflection from the water surface, the calculated reconstruction error will far exceed the adaptive threshold, indicating a sensor malfunction. In this case, the reconstructed data output by the network is directly used as valid state data. If the sensor data is normal and the reconstruction error is less than the adaptive threshold, then the real-time sensor data is used directly.

[0124] It should be noted that the threshold set based on the reconstruction error is in line with the actual operating characteristics of the system, and the judgment logic is accurate and efficient. By comparing errors, automatic fault identification and adaptive data replacement are achieved, which not only ensures the direct use of the original data when there is no fault, but also solves the problem of data failure caused by sensor failure. This provides continuous and reliable state input for the subsequent fault-tolerant control strategy network and avoids control divergence.

[0125] In this embodiment of the invention, this step enables automatic determination of sensor faults and data adaptation, ensuring the reliability of the state data input to the control network, avoiding control divergence caused by fault data, and providing a stable input for fault-tolerant control.

[0126] S8: Based on valid state data, a near-end policy optimization algorithm is used to train the fault-tolerant control policy network.

[0127] Among them, the proximal policy optimization algorithm refers to a reinforcement learning algorithm that limits the policy update step size and ensures training convergence, and the fault-tolerant control policy network is a robust control instruction output network with an Actor-Critic architecture.

[0128] In one possible implementation, S8 specifically includes sub-steps S801 to S804: S801: Construct a fault-tolerant control strategy network based on valid state data using an Actor-Critic architecture.

[0129] Among them, the Actor-Critic architecture refers to a dual-network structure containing an Actor network and a Critic network.

[0130] For example, the Actor network can take 14-dimensional effective state data as input and output 4-dimensional normalized control quantities corresponding to throttle, roll, pitch, and yaw, while the Critic network outputs a 1-dimensional scalar reward estimate.

[0131] S802: Based on the multi-objective optimization requirements of collaborative recovery, the training direction of the fault-tolerant control policy network is determined through a composite reward function. in, This represents the composite reward function. Indicates distance reward. This represents a penalty for attitude stability. This indicates a smoothing penalty for the control quantity. This indicates a sparse success reward.

[0132] Among them, the multi-objective optimization requirement for collaborative recovery refers to the optimization objective that takes into account the flight stability, control smoothness, and recovery accuracy of the UAV, while the composite reward function refers to the reward calculation function that integrates the multi-objective optimization requirements.

[0133] Furthermore, the distance reward increases because the drone needs to be close to the origin of the unmanned vessel's coordinates; the closer the drone is, the greater the reward. in, Indicates distance reward. This represents the weighting coefficient for distance rewards. Indicates the location coordinates of the drone. Indicates the location coordinates of the unmanned vessel. This represents the square of the L2 norm.

[0134] When a motor fails, the most common fatal problem for a drone is violent tumbling, so it is necessary to limit the Euler angle (tilt angle) and angular velocity (rotation speed) of the drone: in, Indicates a gesture reward. This represents the weighting coefficient for posture rewards. This indicates the angular velocity of the drone. Represents the Euler angles of the drone.

[0135] Because the signal output by the neural network may have high-frequency jitter, and in a real-world motor, the PWM signal jumps wildly from 10% to 90% and then back to 20% within 0.1 seconds, the motor and ESC will overheat and burn out, and the mechanical structure will vibrate and break. Therefore, operations at two points with larger differences should be penalized. in, This indicates that the reward is controlled to be smooth. This represents the weighting coefficients used to control the smoothing of rewards. Indicates the first t Timing control signals, Indicates the first t- Control signal at time 1.

[0136] For example, the distance reward can be set to increase as the drone and unmanned surface vessel are closer to the target position; the attitude stability penalty can limit the range of Euler angles and angular velocities; the control quantity smoothing penalty can suppress drastic fluctuations in control quantities at adjacent moments; and the sparse success reward can give a positive reward of +100 when the drone's landing error is less than 0.2m and the speed meets the target.

[0137] S803: Based on the domain randomization strategy, the physical parameters of the environment are dynamically randomized during the training process, and the fault injection module in the simulation environment is activated to generate training samples containing fault states.

[0138] Among them, the domain randomization strategy refers to the training method of dynamically adjusting environmental parameters to improve the network's generalization ability; environmental physical parameters refer to parameters that affect motion characteristics, such as UAV mass and drag coefficient; the fault injection module refers to a programmable module that simulates faults such as sensor failure and power attenuation; and the training samples refer to a sample set of system state-control commands-rewards that includes normal and fault states.

[0139] For example, at the beginning of each training round, the drone mass can be randomly adjusted to 0.9 to 1.1 times, and every 1000 steps, there is a 30% probability of triggering a failure where the visual observation value is set to zero or the motor thrust coefficient decreases to 0.5 to 0.8 times.

[0140] S804: Based on the composite reward function, the fault-tolerant control policy network is iteratively trained using training samples through the near-end policy optimization algorithm.

[0141] Among them, the proximal policy optimization algorithm refers to a reinforcement learning algorithm that restricts the policy update step size and ensures stable convergence of training.

[0142] Specifically, this step first constructs a dual-network architecture adapted to continuous control requirements, then clarifies the multi-objective optimization direction through a composite reward function, subsequently generates rich training samples through a domain randomization strategy to bridge the gap between simulation and reality, and finally uses a stable and efficient proximal policy optimization algorithm to iteratively train the network so that the network learns robust control laws under fault conditions.

[0143] For example, 50,000 iterative training steps can be performed, and the pruning mechanism of the near-end policy optimization algorithm ensures stable policy updates.

[0144] Reference manual attached Figure 5 The diagram shows the cumulative reward curve for training a fault-tolerant control strategy network provided in an embodiment of the present invention.

[0145] Specifically, Figure 5In the graph, the horizontal axis "Episodes" represents the number of training iterations for the fault-tolerant control strategy. The vertical axis "Total Reward" represents the total reward obtained by the agent (e.g., a drone control strategy) in each training round. "Original Round Reward" corresponds to the actual reward data for each training round. "MovingAvg" is the smoothed average of the original rewards over multiple rounds, used to reflect the overall direction of reward change. "Convergence Target Threshold" is a pre-set baseline, representing the reward standard for achieving stable policy training. "Fluctuations stem from random fault injection (prompting the agent to learn adaptability)" explains the fluctuations in the curve, i.e., randomly adding fault scenarios during training to improve the agent's adaptability.

[0146] Furthermore, regarding the inclusion relationship: Figure 5 As a visualization of the training process, it includes a horizontal axis (training rounds), a vertical axis (cumulative reward), a curve showing the original round reward data, an average reward trend curve, a convergence target threshold, and annotations explaining the causes of fluctuations. In terms of connectivity: the horizontal axis (training rounds) and the vertical axis (cumulative reward) together form a coordinate plane; the original round reward data corresponds to the reward value of each training round and forms a curve. The average reward trend is calculated based on the original round rewards and displays the overall change in reward as a smooth curve. The convergence target threshold serves as a benchmark, compared with the average reward trend curve, showing whether the training is approaching the expected results. The annotations for the causes of fluctuations correspond to the fluctuating areas of the curve, explaining the triggering factors for reward changes.

[0147] It should be noted that, Figure 5 By demonstrating the reward change process under domain randomization, the learning adaptability of the fault-tolerant control strategy is intuitively demonstrated, and the convergence of the strategy under random failure scenarios is verified, providing visual support for the effective training of fault-tolerant control strategies for unmanned systems.

[0148] Reference manual attached Figure 6 The diagram shows a comparison of flight trajectories under sudden sensor failure, as provided in an embodiment of the present invention.

[0149] Specifically, Figure 6 The horizontal axis represents horizontal displacement (m), indicating the drone's horizontal position during recovery. The vertical axis represents flight altitude (m), reflecting the drone's vertical position; together, they constitute the recovery trajectory. The dashed line corresponds to the flight trajectory of traditional PID control (showing a divergent trend), while the solid line corresponds to the flight trajectory of the fault-tolerant control of this invention (showing a convergent trend). The dotted line represents the normal flight phase, and the diamond mark indicates the target landing pad. "Sudden sensor failure (t=3s)" is the fault trigger node marker (corresponding to a horizontal displacement of approximately 3m), and "Target landing pad" is the recovery endpoint marker (horizontal displacement 10m, flight altitude 0m).

[0150] Furthermore, in terms of inclusion relationship, this Figure 6 This presentation compares the recovery performance under complex sea conditions, integrating flight trajectory data, fault triggering nodes, and recovery target nodes from different control methods. In terms of connectivity, during the normal flight phase (dotted line), the flight trajectories (changes in altitude and horizontal displacement) of the two control methods remain synchronized. After a sudden sensor failure at t=3s, the trajectory curve of the traditional PID control (dashed line) shows a continuous increase in flight altitude (trajectory divergence), while the trajectory curve of the fault-tolerant control of this invention (solid line) remains stable and gradually converges, ultimately reaching the target landing pad, demonstrating the difference in effectiveness between different control schemes under fault conditions.

[0151] It should be noted that, Figure 6 By comparing the trajectory formed by horizontal displacement and flight altitude, the stability and convergence of the fault-tolerant control of the present invention under sudden failure are clearly demonstrated, providing intuitive effect support for the robustness of the method in complex inland water recovery scenarios.

[0152] It should be noted that the constructed dual-network architecture is adapted to continuous control requirements, and the composite reward function accurately guides multi-objective optimization. Domain randomization and fault injection generate rich training samples, bridging the gap between simulation and reality. The proximal policy optimization algorithm ensures training stability, enabling the network to learn robust control laws under fault conditions, effectively overcoming the failure problem of traditional linear control under nonlinear faults, and providing reliable control support for collaborative recovery.

[0153] In this embodiment of the invention, this step combines effective state data with domain randomization training, enabling the network to adapt to extreme faults and complex environments, avoiding the failure problem of traditional linear control, and significantly improving the robustness and generalization ability of the control strategy.

[0154] S9: Obtain real-time valid status data.

[0155] S10: Input the real-time valid state data into the trained fault-tolerant control strategy network, output control commands, and drive the UAV model to complete the cooperative recovery to the unmanned vessel model according to the control commands.

[0156] Collaborative recovery refers to the process of a drone accurately landing in the target area of ​​an unmanned vessel in the event of a malfunction or in a complex environment.

[0157] In this embodiment of the invention, this step directly outputs the appropriate control commands, enabling stable flight and precise landing under fault conditions. It overcomes the recovery challenges caused by complex interference in inland waters and system failures, and significantly improves the recovery success rate and safety of the air-water cross-domain collaborative system.

[0158] Reference manual attached Figure 7The diagram shows a structural schematic of an inland waterway unmanned aerial vehicle (UAV)-unmanned vessel fault-tolerant recovery system provided by an embodiment of the present invention.

[0159] This invention provides an inland waterway unmanned aerial vehicle (UAV)-unmanned vessel fault-tolerant recovery system 20, comprising: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-described inland waterway UAV-unmanned vessel fault-tolerant recovery method and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0160] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0161] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0162] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0163] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0164] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0165] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0166] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0167] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0168] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0169] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0170] This invention provides a readable storage medium comprising: storing a program or instructions on the readable storage medium, wherein when the program or instructions are executed by a processor, the program or instructions implement the steps of the above-described inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A fault-tolerant recovery method for unmanned aerial vehicles (UAVs) and unmanned surface vessels (USVs) in inland waterways, characterized in that, include: S1: Construct a simulation environment for a water-air collaborative digital twin; S2: Based on the simulation environment, configure the UAV model and the unmanned vessel model; S3: Collect multi-dimensional sensor time-series data under normal flight conditions of the UAV, and construct a training dataset based on the multi-dimensional sensor time-series data; S4: Based on the mask noise self-supervised training strategy, the training dataset is input into the denoising autoencoder network, and the denoising autoencoder network is trained to obtain the trained denoising autoencoder network. S5: Input the real-time collected sensor data into the trained denoising autoencoder network and output the reconstructed data; S6: Calculate the reconstruction error between the sensor data and the reconstructed data; S7: Based on the reconstruction error and the adaptive threshold, obtain the effective state data; S8: Based on the effective state data, a fault-tolerant control strategy network is trained using a near-end policy optimization algorithm; S9: Obtain real-time valid status data; S10: Input the real-time valid state data into the trained fault-tolerant control strategy network, output control commands, and drive the UAV model to complete the cooperative recovery to the unmanned vessel model according to the control commands.

2. The inland waterway UAV-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, S1 specifically includes: S101: Define the inertial coordinate system and the body coordinate system; S102: Based on the inertial coordinate system and the body coordinate system, determine the yaw angle, pitch angle and roll angle of the UAV; S103: Based on the yaw angle, the pitch angle, and the roll angle, construct three basic rotation matrices corresponding to the UAV in the ZYX rotation order; S104: Perform matrix multiplication on each of the basic rotation matrices to obtain the coordinate transformation matrix of the UAV; S105: Based on the coordinate transformation matrix, construct a water-air collaborative digital twin simulation environment that includes rigid body dynamics characteristics, actuator failure mechanisms, and environmental interference effects.

3. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, S2 specifically includes: S201: Configure the UAV model based on the simulation environment; S202: Based on the characteristics of the inland water flow field, configure the unmanned vessel model.

4. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 3, characterized in that, S201 specifically includes: S2011: Based on the simulation environment, determine the type of force experienced by the UAV during its translational motion; S2012: Based on Newton's second law, and combining the force type and coordinate transformation matrix, the translational characteristic equation in the inertial coordinate system is constructed as follows: in, m Indicates the quality of the drone. P This represents the translational acceleration of the UAV in the inertial coordinate system. This represents the projection of the UAV's thrust in the inertial coordinate system. mg This indicates the gravitational force acting on the drone. This indicates the air resistance experienced by the drone during its translational motion; S2013: Construct the rotational inertia matrix of the UAV in the body coordinate system; S2014: Based on Euler's equations, and combining the aforementioned moment of inertia matrix and core torque, the rotational characteristic equations of the UAV in the body coordinate system are constructed as follows: in, This represents the moment of inertia matrix of the UAV in the body coordinate system. This indicates the angular velocity of the drone. This represents the total torque during the rotation of the drone. This represents the nonlinear Coriolis torque. Indicates control torque. This indicates the disturbance torque experienced by the drone during its rotation; S2015: Based on the translational characteristic equation and the rotational characteristic equation, complete the configuration of the UAV model.

5. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 3, characterized in that, S202 specifically includes: S2021: Based on the characteristics of the inland water flow field, determine the inertial delay characteristics of unmanned vessel motion; S2022: Based on the planar motion law of the unmanned vessel, determine the matrix terms in the motion equation of the unmanned vessel; S2023: Based on the aforementioned matrix terms, and combining Fossen's equations of motion for ships with the fluid coupling effect of the water medium, the equations of motion for the unmanned vessel are constructed as follows: in, M The inertial matrix represents the unmanned surface vessel. The time derivative of the unmanned surface vessel's motion state vector. This represents the planar motion state vector of the unmanned surface vessel. Represents the Coriolis centripetal force matrix. Represents the hydrodynamic resistance matrix. This indicates the propulsion torque of the unmanned vessel. This represents the generalized force vector of environmental disturbances experienced by the unmanned surface vessel. S2024: Based on the unmanned vessel's motion equations, complete the configuration of the unmanned vessel model.

6. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, S3 specifically includes: S301: During the flight of the UAV, collect the time-series data of the multi-dimensional sensor; S302: Perform a masking noise operation on the multidimensional sensor time series data to obtain corrupted data; S303: Integrate the multidimensional sensor time-series data with the damaged data to obtain the training dataset.

7. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, S4 specifically includes: S401: Construct a denoising autoencoder network including an encoder and a decoder; S402: Based on the mask noise self-supervised training strategy, the denoising autoencoder network is optimized through a loss function; S403: Iterate and optimize until the loss value of the loss function is less than the preset loss value, and obtain the trained denoising autoencoder network.

8. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, Specifically, S7 includes: S701: Based on the reconstruction error, set the adaptive threshold; S702: Determine whether the reconstruction error is less than or equal to the adaptive threshold; if so, use the real-time acquired sensor data as the valid state data; otherwise, determine that the sensor is faulty and use the reconstructed data as the valid state data.

9. The inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method according to claim 1, characterized in that, S8 specifically includes: S801: Based on the effective state data, construct a fault-tolerant control strategy network with an Actor-Critic architecture; S802: Based on the multi-objective optimization requirements of collaborative recycling, the training direction of the fault-tolerant control strategy network is determined through a composite reward function: in, This represents the composite reward function. Indicates distance reward. This represents a penalty for attitude stability. This indicates a smoothing penalty for the control quantity. Indicates a reward for sparse success; S803: Based on the domain randomization strategy, the physical parameters of the environment are dynamically randomized during the training process, and the fault injection module in the simulation environment is activated to generate training samples containing fault states. S804: Based on the composite reward function, the fault-tolerant control policy network is iteratively trained using the training samples through the near-end policy optimization algorithm.

10. A fault-tolerant recovery system for unmanned aerial vehicles (UAVs) and unmanned surface vessels (USVs) in inland waterways, characterized in that: include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the inland waterway unmanned aerial vehicle-unmanned vessel fault-tolerant recovery method as described in any one of claims 1 to 9.

Images