An unmanned aerial vehicle safety control method based on fault risk learning

Abstract

The present application belongs to the technical field of flying robots, and particularly relates to a UAV safety control method based on fault risk learning, comprising the following steps: firstly, performing actuator fault modeling; then, designing a fixed-time fault observer to quickly and effectively estimate the fault; secondly, using conditional value at risk to quantitatively evaluate the fault risk to obtain the required label true value for learning, and learning the influence of the fault risk on the UAV position uncertainty based on a back propagation neural network; finally, designing an adaptive risk-oriented control compensation strategy to realize smooth adjustment and adaptive safety control of the control strategy. The method can improve the control response speed and trajectory tracking accuracy of the UAV under different fault levels, and is particularly suitable for task scenarios with high requirements for flight safety and reliability, such as inspection operations in complex environments, search and rescue and delivery tasks in disaster areas, and dangerous source monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of flying robot technology, specifically relating to the safety control of unmanned aerial vehicles (UAVs). Background Technology

[0002] As a representative of aerial autonomous systems, unmanned aerial vehicles (UAVs) have been widely used in critical mission scenarios such as complex environment inspection, disaster search and rescue delivery, and hazard source monitoring and early warning. These scenarios place extremely high demands on the flight safety and reliability of UAVs. UAVs not only need to respond quickly to sudden actuator failures, but also need to ensure trajectory tracking accuracy under the constraints of lightweight computing power (adapting to the hardware capabilities of small UAVs) to avoid equipment damage, mission failure, or even secondary disasters caused by control delays or trajectory deviations. According to data from the Global UAV Network, in recent years, actuator failures have accounted for over 35% of UAV malfunction accidents, with delayed fault response and insufficient accuracy under lightweight computing power being the core pain points restricting their application in high-safety-demand scenarios.

[0003] Current technical solutions for UAV fault control suffer from three significant shortcomings: First, traditional fault observers rely heavily on linear models, resulting in slow estimation speeds for nonlinear actuator faults, often requiring several seconds to complete fault identification, which fails to meet the real-time control requirements of highly dynamic mission scenarios. Second, fault risk assessments primarily employ qualitative descriptions (e.g., "high / medium / low risk"), lacking quantitative indicators to support them, making it difficult to provide accurate quantitative data for subsequent control strategy adjustments, leading to weak targeted control compensation. Third, existing adaptive control strategies do not fully integrate the correlation between fault risk and position uncertainty, either ignoring the impact of risk on trajectory deviation (relying solely on fixed control parameters) or requiring complex computing power to support risk modeling (e.g., risk assessment based on Monte Carlo simulation), making it difficult to balance the contradiction between "lightweight computing power" and "control accuracy." For example, in disaster area search and rescue missions, if a UAV experiences a motor failure, traditional observers may cause the UAV to temporarily lose control due to estimation delays, while complex risk assessment algorithms consume excessive computing power, squeezing the operational resources of the trajectory tracking module, ultimately causing the search target to miss the optimal rescue window.

[0004] As drone mission scenarios evolve towards "more complex environments, higher safety standards, and smaller hardware," existing technologies are struggling to meet the integrated requirements of "rapid fault observation, accurate risk quantification, and lightweight adaptive control." There is an urgent need for an integrated solution that can achieve rapid fault estimation, risk quantification learning, and adaptive control compensation with lightweight computing power. This solution would address the industry pain points of "slow response, low accuracy, and poor computing power adaptation" for drones under different fault levels, thereby improving their reliable operation in high-safety-requirement scenarios.

[0005] Patent application No. 202410076157.6 proposes a UAV collision risk perception method based on conditional value at risk, but there are two problems: (1) Risk assessment based on Monte Carlo simulation requires complex computing power to support risk modeling; (2) It does not consider the safety control problem when the UAV fails after the collision. Patent application No. 202510989254.9 proposes a multi-UAV distributed fault-tolerant control method, multi-UAV control method and system, but there are two problems: (1) It does not incorporate the "fault risk quantification and learning" mechanism, and the fault-tolerant control lacks the accuracy of "risk tendency"; (2) It does not consider the "lightweight computing power constraint", and it is difficult to adapt to small UAV / single-machine task scenarios. Patent application No. 202510551308.3 proposes a fault-tolerant control method for fault-tolerant UAVs based on incremental nonlinear dynamic inverse terminal sliding mode control, but it is difficult to adapt to multiple types of actuator failures for the specific severe fault scenario of "single rotor or dual rotor failure".

[0006] Therefore, none of the above methods comprehensively consider the issues of rapid fault estimation, risk quantification learning, and adaptive control compensation for UAVs under the constraint of lightweight computing power, so as to complete UAV safe operation tasks with high requirements for flight safety and reliability. Summary of the Invention

[0007] To overcome the shortcomings of existing UAV fault control methods, such as slow fault estimation, lack of risk quantification, and difficulty in balancing computing power and accuracy, this paper proposes a UAV safety control method based on fault risk learning for UAV operation scenarios with high requirements for flight safety and reliability (such as complex environment inspection, disaster area search and rescue and delivery, and hazard source monitoring and early warning). This method aims to achieve the integration of rapid fault estimation, accurate risk quantification, and adaptive control compensation for UAVs under different fault levels with lightweight computing power, thereby improving the control response speed and trajectory tracking accuracy of UAVs under different fault levels.

[0008] The technical solution of this invention is as follows:

[0009] A method for safe control of unmanned aerial vehicles (UAVs) based on fault risk learning includes the following steps:

[0010] S1. Establish an actuator fault model;

[0011] S2. Design a fixed-time fault observer;

[0012] S3. Use conditional value at risk to quantitatively assess fault risk and obtain the true value of the label required for learning. Then, use backpropagation neural network to learn the impact of fault risk on the uncertainty of UAV position.

[0013] S4. Design an adaptive risk-prone control compensation strategy. By constructing a dynamic correlation mechanism between fault risk and control input, the control strategy can be continuously adjusted without chattering and adaptive safety control can be achieved.

[0014] The specific steps of S1 are as follows:

[0015] The quadcopter drone is considered to have a perfectly rigid and axisymmetric structure with uniform mass distribution. Without considering uncertainties, its axis and center of mass are perfectly aligned, and its thrust and drag are proportional to the square of the angular velocity of each brushless motor. Input control is defined based on these conditions. The total lift is generated by the drone's motors. Triaxial torque in the combined system Given this configuration, the mapping relationship between the control input and the motor lift can be expressed as:

[0016] ;

[0017] in This represents the lift generated by the four brushless motors, and the coefficient representing the relationship between the blade lift and the yaw torque. , and These represent the roll arm and pitch arm of the lift generated by a single motor relative to the UAV, respectively.

[0018] Because motors operate in similar environments and under similar loads, the likelihood of multiple motors failing simultaneously is greater. Mathematically, the decrease in lift and torque caused by blade damage and reduced motor speed can be represented by a multiplicative efficiency loss, i.e.:

[0019] ;

[0020] in and Representing actual thrust and commanded thrust respectively, the fault matrix... From the ability coefficient Determined, capability coefficient Indicates the first The degree of failure of each motor.

[0021] Step S2 is specifically as follows: [The text abruptly ends here, likely due to an incomplete sentence or a missing section. Axial velocity Triaxial angular velocity under the combined machine system Constructing the system state vector The state equation for a drone in the event of actuator failure can be derived as follows:

[0022] ;

[0023] Where the invertible matrix Represents the control coefficient matrix. Indicates the quality of the drone. and These represent the roll angle and pitch angle, respectively. The moment of inertia matrix of the machine system. This represents the aggregate of actuator faults. Represents a four-dimensional identity matrix, with known nonlinear terms. It can be represented as:

[0024] ;

[0025] in It represents the acceleration due to gravity.

[0026] Fixed-time fault observers are used to estimate state vectors. and fault set items The design is as follows:

[0027] ;

[0028] in and These are the state vectors. and fault set items The estimation results, It is the difference between the actual value and the estimated value. middle and This is the observer gain. According to the definition of the fault lumped term, the motor capability coefficient can be expressed as:

[0029] ;

[0030] in yes The elements in the item.

[0031] Step S3 is as follows:

[0032] In malfunction situations, the drone exhibits phenomena such as altitude loss and attitude turbulence, and its flight trajectory shows greater volatility compared to normal conditions. Under these circumstances, the drone's position... Estimation can be done using distributions It indicates. Among them. Let represent the covariance matrix, which describes the positional uncertainty in three dimensions. Considering the combined effects of actuator capability coefficients, the number of actuator failures, and UAV flight speed on the UAV's positional uncertainty, the safety loss function is established as follows:

[0033] ;

[0034] in This represents the element-wise Hadamard product. Indicates the actual location of the drone and desired location The deviation. For ease of description, the subsequent safety loss function... Abbreviated as .

[0035] Before using conditional value at risk to measure failure risk, we first introduce the value at risk measurement of the safety loss function. , is represented as:

[0036] ;

[0037] in The first term represents the security loss function. One element, For the first The nth standard basis vector, whose nth... One component is 1, and the rest are 0. Indicates an event The probability of occurrence Indicates the confidence level. This indicates the potential safety loss value.

[0038] Based on this The conditions under risk value Defined as:

[0039] ;

[0040] in express , The operator represents the mathematical expectation of a random variable.

[0041] After quantifying the fault risk based on conditional risk value and obtaining the true value of the label, a backpropagation neural network is used to learn the impact of fault risk on the uncertainty of the UAV's position. Let the current time be... Then in The dimensional features of the model input within each time step include the motor capability coefficient sequence. Fault risk confidence sequence and drone speed sequence ,in Therefore, define the input variables. and corresponding fault risk quantification output They are respectively:

[0042] ;

[0043] To achieve this mapping relationship, a multi-layer backpropagation neural network is used to model the temporal input. The network consists of an input layer, hidden layers, and an output layer, with each layer passing through a hyperbolic tangent function. A nonlinear transformation is performed on the output of the previous layer. The forward propagation process is as follows:

[0044] ;

[0045] in This represents the feature state vector of the hidden layer. This represents a single-step input vector within a time window. and These represent the input and output weight matrices, respectively. and These represent the biases of the hidden layer and the output layer, respectively. The number of neurons in the hidden layer is represented by the mean squared error loss function used for the regression task, which can be expressed as:

[0046] ;

[0047] in The number of training samples was optimized using the Levenberg-Marquardt algorithm.

[0048] Step S4 is as follows:

[0049] ① To address the fault risks in each direction of three-dimensional space, a quantitative assessment model based on the sensitivity of risk direction is established;

[0050] ②Based on the sensitivity weights of each direction output by the model, a mapping relationship between three-dimensional fault risk and multiple control input channels is constructed, and the risk level of different directions is converted into the compensation intensity coefficient of the corresponding control channel.

[0051] ③ This mapping mechanism enables precise matching between risk direction and control compensation, thereby improving the controller's risk response capability and directional adaptive capability.

[0052] Constructing a risk propensity weight vector for:

[0053] ;

[0054] in It is a risk allocation mapping matrix, which transforms the failure risks in each direction of three-dimensional space into the influence weights of the four control channels of the UAV, thereby quantifying and establishing the mapping relationship from three-dimensional risks to multiple control channels.

[0055] Based on this, the safety control law is designed as follows:

[0056] ;

[0057] in It is a cascaded control baseline controller. Represents the vector The diagonal matrix formed by the components is used to weight the risk compensation intensity on the control channel.

[0058] The advantages of this invention compared to the prior art are:

[0059] This invention relates to a UAV safety control method based on fault risk learning, primarily targeting high-safety-requirement UAV mission scenarios such as complex environment inspection, disaster area search and rescue delivery, and hazard source monitoring and early warning. First, actuator fault modeling is performed; then, a fixed-time fault observer is designed to quickly and effectively estimate the fault; next, conditional value at risk (CWH) is used to quantify the fault risk and obtain the ground truth label required for learning, and the impact of fault risk on UAV position uncertainty is learned based on a backpropagation neural network; finally, an adaptive risk-prone control compensation strategy is designed to achieve smooth adjustment of the control strategy and adaptive safety control.

[0060] Compared to existing UAV fault control technologies, traditional methods may suffer from delayed fault estimation, lack of risk quantification capabilities, or difficulty in balancing computing power and accuracy. As a result, under the constraint of lightweight computing power, UAVs cannot simultaneously achieve fault response speed, trajectory tracking accuracy, and adaptability to multiple fault scenarios, making it difficult to meet the operational requirements of high-safety-demand missions. This method addresses the problems of slow control compensation response and low trajectory tracking accuracy of UAVs under the influence of actuator failures. It can improve the control response speed and trajectory tracking accuracy of UAVs under different fault levels and is suitable for scenarios with high requirements for flight safety and reliability. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the design of the UAV safety control method based on fault risk learning according to the present invention.

[0062] Figure 2 The drone used for method verification in the specific embodiments.

[0063] Figure 3 This is a drone trajectory tracking diagram from a specific embodiment.

[0064] Figure 4 This is a diagram of the UAV thrust command in a specific embodiment.

[0065] Figure 5 The image shows the estimation effect of the fixed-time fault observer in a specific embodiment.

[0066] Figure 6This is a trajectory diagram comparing the control methods in a specific embodiment. Detailed Implementation

[0067] Taking a general-purpose UAV flight platform as an example to illustrate the specific implementation of the system and method, when a UAV performs high-precision tasks in the event of an actuator failure, it has very high requirements for its own safety.

[0068] like Figure 1 As shown, the specific implementation steps of this invention are as follows:

[0069] S1. Establish the actuator failure model. The quadcopter UAV is considered to have a perfectly rigid and axisymmetric structure with uniform mass distribution. Without considering uncertainties, its axis and center of mass are perfectly aligned. Thrust and drag are proportional to the square of the angular velocity of each brushless motor. Based on these conditions, define the input control... The total lift is generated by the drone's motors. Triaxial torque in the combined system Given this configuration, the mapping relationship between the control input and the motor lift can be expressed as:

[0070] ;

[0071] in This represents the lift generated by the four brushless motors, and the coefficient representing the relationship between the blade lift and the yaw torque. , and These represent the roll arm and pitch arm of the lift generated by a single motor relative to the UAV, respectively.

[0072] Because motors operate in similar environments and under similar loads, the likelihood of multiple motors failing simultaneously is greater. Mathematically, the decrease in lift and torque caused by blade damage and reduced motor speed can be represented by a multiplicative efficiency loss, i.e.:

[0073] ;

[0074] in and Representing actual thrust and commanded thrust respectively, the fault matrix... From the ability coefficient Determined, capability coefficient Indicates the first The degree of failure of each motor.

[0075] S2. Design a fixed-time fault observer. (By...) Axial velocity Triaxial angular velocity under the combined machine system Constructing the system state vector The state equation for a drone in the event of actuator failure can be derived as follows:

[0076] ;

[0077] Where the invertible matrix Represents the control coefficient matrix. Indicates the quality of the drone. and These represent the roll angle and pitch angle, respectively. The moment of inertia matrix of the machine system. This represents the aggregate of actuator faults. Represents a four-dimensional identity matrix, with known nonlinear terms. It can be represented as:

[0078] ;

[0079] in It represents the acceleration due to gravity.

[0080] Fixed-time fault observers are used to estimate state vectors. and fault set items The design is as follows:

[0081] ;

[0082] in and These are the state vectors. and fault set items The estimation results, It is the difference between the actual value and the estimated value. middle and This is the observer gain. According to the definition of the fault lumped term, the motor capability coefficient can be expressed as:

[0083] ;

[0084] in yes The elements in the item.

[0085] S3. Use Conditional Value at Risk (CVO) to quantify the failure risk and obtain the true value of the label required for learning. Then, use a backpropagation neural network to learn the impact of failure risk on the uncertainty of the UAV's position.

[0086] In malfunction situations, the drone exhibits phenomena such as altitude loss and attitude turbulence, and its flight trajectory shows greater volatility compared to normal conditions. Under these circumstances, the drone's position... Estimation can be done using distributions It indicates. Among them. Let represent the covariance matrix, which describes the positional uncertainty in three dimensions. Considering the combined effects of actuator capability coefficients, the number of actuator failures, and UAV flight speed on the UAV's positional uncertainty, the safety loss function is established as follows:

[0087] ;

[0088] in This represents the element-wise Hadamard product. Indicates the actual location of the drone and desired location The deviation. For ease of description, the subsequent safety loss function... Abbreviated as .

[0089] Before using conditional value at risk to measure failure risk, we first introduce the value at risk measurement of the safety loss function. , is represented as:

[0090] ;

[0091] in The first term represents the security loss function. One element, For the first The nth standard basis vector, whose nth... One component is 1, and the rest are 0. Indicates an event The probability of occurrence Indicates the confidence level. This indicates the potential safety loss value.

[0092] Based on this The conditions under risk value Defined as:

[0093] ;

[0094] in express , The operator represents the mathematical expectation of a random variable.

[0095] After quantifying the fault risk based on conditional risk value and obtaining the true value of the label, a backpropagation neural network is used to learn the impact of fault risk on the uncertainty of the UAV's position. Let the current time be... Then in The dimensional features of the model input within each time step include the motor capability coefficient sequence. Fault risk confidence sequence and drone speed sequence ,in Therefore, define the input variables. and corresponding fault risk quantification output They are respectively:

[0096] ;

[0097] To achieve this mapping relationship, a multi-layer backpropagation neural network is used to model the temporal input. The network consists of an input layer, hidden layers, and an output layer, with each layer passing through a hyperbolic tangent function. A nonlinear transformation is performed on the output of the previous layer. The forward propagation process is as follows:

[0098] ;

[0099] in This represents the feature state vector of the hidden layer. This represents a single-step input vector within a time window. and These represent the input and output weight matrices, respectively. and These represent the biases of the hidden layer and the output layer, respectively. The number of neurons in the hidden layer is represented by the mean squared error loss function used for the regression task, which can be expressed as:

[0100] ;

[0101] in The number of training samples was optimized using the Levenberg-Marquardt algorithm.

[0102] S4. Design an adaptive risk-prone control compensation strategy. By constructing a dynamic correlation mechanism between fault risk and control input, continuous chatter-free adjustment and adaptive safety control of the control strategy are achieved. Specifically, firstly, a quantitative evaluation model based on risk direction sensitivity is established for fault risk in each direction of three-dimensional space. Then, based on the sensitivity weights of each direction output by this model, a mapping relationship between three-dimensional fault risk and multiple control input channels is constructed, converting the risk level in different directions into the compensation intensity coefficient of the corresponding control channel. Finally, through this mapping mechanism, precise matching between risk direction and control compensation is achieved, thereby improving the controller's risk response capability and direction adaptation capability. A risk propensity weight vector is constructed. for:

[0103] ;

[0104] in It is a risk allocation mapping matrix, which transforms the failure risks in each direction of three-dimensional space into the influence weights of the four control channels of the UAV, thereby quantifying and establishing the mapping relationship from three-dimensional risks to multiple control channels.

[0105] Based on this, the safety control law is designed as follows:

[0106] ;

[0107] in It is a cascaded control baseline controller. Represents the vector The diagonal matrix formed by the components is used to weight the risk compensation intensity on the control channel.

[0108] The following specific embodiments illustrate the implementation principle of the present invention:

[0109] use Figure 2 The drone shown was used in an experiment, and the experimental procedure is as follows: The mass of the drone... The lift generated by a single motor relative to the roll arm of the drone is 1.35 kg. and pitch lever arm All are 0.23m, the moments of inertia of the three axes in the machine system. for kg·m 2 The coefficient relating blade lift to yaw moment. The value is 0.01. In the experiment, the failure rate of the UAV actuators covered a continuous gradient range of 5%-40% (in 5% intervals). To demonstrate the effectiveness of the proposed fixed-time fault observer and adaptive control method, we selected the experimental results under severe fault conditions of 30%, 20%, 40%, and 10% failure rates for motors 1-4, respectively, i.e., the capability coefficients of the four motors. The values ​​are (0.7, 0.8, 0.6, 0.9). The test results are as follows: Figures 3 to 6 As shown.

[0110] Figure 3 The figure shows the trajectory tracking performance of the UAV along the x, y, and z axes over time when using a baseline controller under this severe fault condition. From the three curves in Figure 3, we can further observe that: in the x and y axes, the actual flight trajectory initially deviates slightly from the expected trajectory but generally matches it; however, during the fault-affected phase (approximately 11 seconds later), the difference in fluctuation between the two gradually increases; and in the z-axis direction, in addition to the "altitude drop" phenomenon, a short-term, sharp fluctuation can be observed in the actual trajectory after the fault occurs, creating a strong contrast with the stable state of the expected trajectory. This fluctuation further exacerbates the risk of loss of position control in the z-axis direction.

[0111] Figure 4 The figure shows the thrust command curves of the four motors of the UAV as a function of time when using the baseline controller under this severe failure condition. Figure 4 As can be seen, before the fault (11 seconds ago), the thrust commands of the four motors fluctuated smoothly and the values ​​were close; after 11 seconds, the fault was triggered, and all four motors failed to varying degrees. The thrust commands of each motor increased sharply and the high-frequency oscillation intensified, and the differences widened significantly, reflecting the instability of the baseline controller under the fault.

[0112] Figure 5 The estimation performance of the fixed-time fault observer required for model training is presented. It can be seen that before the fault injection, the capability coefficients of each motor are close to 1 and fluctuate little; after the sudden injection of the failure fault, the coefficients drop rapidly, and the fixed-time observer converges within about 1 second, accurately tracking the differentiated decay trend of the capability coefficients of different actuators, demonstrating the observer's fast response and estimation accuracy.

[0113] A comparison of the proposed method with the overall sliding mode control method based on a fixed-time fault observer in terms of controller response speed and trajectory tracking accuracy under severe fault conditions is shown below. Figure 6 As shown in the figure, the proposed method directly utilizes the learned fault risk information for adaptive safety control, thus achieving a faster control response speed compared to the overall sliding mode control method. Furthermore, after fault injection, the proposed method significantly improves the height drop phenomenon in the z-direction compared to the overall sliding mode control method, while exhibiting smaller tracking errors in the x and y directions and faster error convergence.

[0114] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for unmanned aerial vehicle (UAV) safety control based on fault risk learning, characterized in that, Includes the following steps: S1. Establish an actuator fault model; S2. Design a fixed-time fault observer; S3. Use conditional value at risk to quantitatively assess fault risk and obtain the true value of the label required for learning. Then, use backpropagation neural network to learn the impact of fault risk on the uncertainty of UAV position. S4. Design an adaptive risk-prone control compensation strategy. By constructing a dynamic correlation mechanism between fault risk and control input, the control strategy can be continuously adjusted without chattering and adaptive safety control can be achieved. Step S3 is as follows: In the event of a malfunction, the drone exhibits phenomena such as altitude loss and attitude turbulence, and its flight trajectory shows greater volatility compared to its normal state; under these circumstances, the drone's position... Estimation can be done using distributions Indicates; among which Indicates a multivariate Gaussian distribution. This is a location estimate. The covariance matrix represents the positional uncertainty in three dimensions. Considering the influence of actuator capability coefficients, the number of actuator failures, and the drone's flight speed on its positional uncertainty, the safety loss function is established as follows: ； in This represents the element-wise Hadamard product. Indicates the actual location of the drone and desired location The deviation; for ease of description, the subsequent safety loss function Abbreviated as ; Before using conditional value at risk to measure failure risk, we first introduce the value at risk measurement of the safety loss function. , represented as: ； in The first term represents the security loss function. One element, For the first The nth standard basis vector, whose nth... One component is 1, and the rest are 0. Indicates an event The probability of occurrence Indicates the confidence level. Indicates the potential safety loss value; Based on this The conditions under risk value Defined as: ； in express , Operators that express the mathematical expectation of a random variable; After quantifying the fault risk based on the conditional value of risk and obtaining the true value of the label, a backpropagation neural network is used to learn the impact of fault risk on the uncertainty of the UAV's position; let the current time be... Then in The dimensional features of the model input within each time step include the motor capability coefficient sequence. Fault risk confidence sequence and drone speed sequence ,in Therefore, define the input variables. and corresponding fault risk quantification output They are respectively: ； To achieve this mapping relationship, a multi-layer backpropagation neural network is used to model the temporal input. The network consists of an input layer, hidden layers, and an output layer, with each layer passing through a hyperbolic tangent function. A nonlinear transformation is performed on the output of the previous layer; the forward propagation process is as follows: ； in This represents the feature state vector of the hidden layer. This represents a single-step input vector within a time window. and These represent the input and output weight matrices, respectively. and These represent the biases of the hidden layer and the output layer, respectively. The number of neurons in the hidden layer is represented by the mean squared error loss function used for the regression task, which can be expressed as: ； in The number of training samples was optimized using the Levenberg-Marquardt algorithm.

2. The UAV safety control method based on fault risk learning according to claim 1, characterized in that, The specific steps of S1 are as follows: The structure of a quadcopter drone is considered to be completely rigid and axisymmetric, with uniform mass distribution. Without considering uncertainties, its axis and center of mass are completely coincident, and its thrust and drag are proportional to the square of the angular velocity of each brushless motor. Input control is defined based on the above conditions. The total lift is generated by the drone's motors. Triaxial torque in the combined system Composition, in which Let represent the torques along the roll, pitch, and yaw axes, respectively. The mapping relationship between the control input and the motor lift is then expressed as: ； in This indicates the lift generated by the four brushless motors. Represents the control allocation matrix. A coefficient representing the correlation between propeller blade lift and yaw moment. and These represent the roll arm and pitch arm of the lift generated by a single motor relative to the UAV, respectively. Because the operating environment and workload of motors are similar, the possibility of multiple motors failing simultaneously is greater. Mathematically, the decrease in lift and torque caused by blade damage and reduced motor speed can be represented by a multiplicative efficiency loss, namely: ； in and Representing actual thrust and commanded thrust respectively, the fault matrix... From the ability coefficient Determined, capability coefficient Indicates the first The degree of failure of each motor.

3. The UAV safety control method based on fault risk learning according to claim 2, characterized in that, Step S2 is as follows: Depend on Axial velocity Triaxial angular velocity under the combined machine system Constructing the system state vector The state equation for the UAV under actuator failure conditions is derived as follows: ； Where the invertible matrix Represents the control coefficient matrix. Indicates the quality of the drone. and These represent the roll angle and pitch angle, respectively. The moment of inertia matrix in the mechanical system is represented. This represents the actuator fault set item. Represents a four-dimensional identity matrix, with known nonlinear terms. Represented as: ;in Represents gravitational acceleration; Fixed-time fault observers are used to estimate state vectors. and fault set items The design is as follows: ； in This represents the observer convergence control coefficient. and These are the state vectors. and fault set items The estimation results It is the difference between the actual value and the estimated value. It is a fixed-time convergent function. middle and It is the observer gain. Represents an element-wise symbolic function; According to the definition of the fault lumped term, the motor capability coefficient is expressed as: ； in yes The elements in the item.

4. The UAV safety control method based on fault risk learning according to claim 3, characterized in that, Step S4 is as follows: ① To address the fault risks in each direction of three-dimensional space, a quantitative assessment model based on the sensitivity of risk direction is established; ②Based on the sensitivity weights of each direction output by the model, a mapping relationship between three-dimensional fault risk and multiple control input channels is constructed, and the risk level in different directions is converted into the compensation intensity coefficient of the corresponding control channel. ③ This mapping mechanism enables precise matching between risk direction and control compensation, thereby improving the controller's risk response capability and directional adaptive capability.

5. The UAV safety control method based on fault risk learning according to claim 4, characterized in that, Constructing a risk propensity weight vector for: ； in It is a risk allocation mapping matrix, which transforms the failure risks in each direction of three-dimensional space into the influence weights of the four control channels of the UAV, thereby quantifying and establishing the mapping relationship from three-dimensional risks to multiple control channels. Based on this, the safety control law is designed as follows: ； in It is a cascaded control baseline controller. Represents the vector The diagonal matrix formed by the components is used to weight the risk compensation intensity on the control channel.

Images