Fault Tolerant Control Method for Delta Wing Aircraft Based on Concurrent Learning Neural Network

Abstract

The present invention discloses a fault-tolerant control method for a delta-wing aircraft. In this method, by establishing a concurrent learning neural network and introducing a fault-tolerant mechanism on the basis of a basic controller, after a fault occurs in the unmanned aircraft, the aircraft can process and solve it online, so as to continue to work in a certain acceptable state. At the same time, concurrent learning can record historical data, ensure the convergence of the neural network weights without continuous excitation, and thus can perform tasks with higher requirements.

Description

Technical Field

[0001] The present invention relates to a fault-tolerant control method for a delta-wing aircraft, and more particularly to a fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network. Background Art

[0002] Unmanned aerial vehicles (UAVs) are widely used in all aspects of national defense and social and economic construction. Compared with manned aircraft, UAVs have the characteristics of flexibility, rapid deployment, low cost, and no risk of personnel casualties. In the fields of national defense and social economy, UAVs have shown extremely broad application potential and development value.

[0003] However, while UAVs are widely used, they also have many problems of their own. Although a certain degree of autonomous flight has been achieved, there is still a gap between their fault handling ability and human intelligence. During flight, UAVs are affected by factors such as wind and engine vibration. And due to the lack of human intervention and environmental uncertainty, UAVs are more likely to lose control after a fault occurs. Nowadays, the automation level of UAV systems is getting higher and higher, the scale is getting larger and larger, and the investment in unmanned systems is also increasing. If a fault occurs during the flight of an aircraft and the control system cannot eliminate or "tolerate" the fault, it may cause the aircraft to become paralyzed and ineffective, and further result in heavy losses of personnel casualties and property. To solve these problems, this patent proposes a fault-tolerant control method for unmanned aerial vehicles, expecting that after a fault occurs in the UAV, the aircraft can solve the problem online and continue to work in an acceptable state.

[0004] During the operation of a UAV, various emergencies may occur, such as aerodynamic nonlinearity generated under large maneuvers or actuator failures. The concurrent learning neural network can adjust the control law online without knowing the fault information to solve unknown nonlinear disturbances and achieve fault-tolerant control of the aircraft. At the same time, concurrent learning can estimate the aerodynamic parameters of the aircraft and the control efficiency of the actuator to solve some faults such as partial actuator failure and change of aerodynamic parameters.

[0005] For the above reasons, the inventor of the present invention has conducted in-depth research on the existing fault-tolerant control methods for delta-wing aircraft, hoping to design a fault-tolerant control method for delta-wing aircraft based on a concurrent learning neural network that can solve the above problems. Summary of the Invention

[0006] To overcome the above problems, the present inventors have conducted intensive research and designed a fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network. In this method, by establishing a concurrent learning neural network and introducing a fault-tolerant mechanism on the basis of a basic controller, after a fault occurs in the unmanned aircraft, the aircraft can process and solve the problem online and continue to work in a certain acceptable state. At the same time, concurrent learning can record historical data, ensure the convergence of the neural network weights without continuous excitation, and then can perform tasks with higher requirements, thus completing the present invention.

[0007] Specifically, the object of the present invention is to provide a fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network, and this method includes the following steps:

[0008] Step 1, perform fault tolerance processing on the output value of the aircraft basic controller to obtain a pseudo-control quantity u;

[0009] The output value of the aircraft basic controller includes the output value of the concurrent learning aerodynamic parameters and the output value of the concurrent control efficiency processing;

[0010] Step 2, output the accurate pseudo-control quantity u to the actuator, and then control the attitude angle of the delta-wing aircraft.

[0011] In Step 1, the pseudo-control quantity u is obtained through the following formula (I):

[0012]

[0013] Wherein, represents the flight state value of the ideal model, which is the matrix of the ideal roll angle and the ideal roll rate measured in real time by the aircraft in the ideal model, represents taking the value of the second row of the matrix, and Ax(2) represents the value of the second row of the matrix Ax;

[0014] x represents the actual flight state value, where φ represents the aircraft roll angle and p represents the aircraft roll rate;

[0015] K P represents the proportional gain, K P = [-1 -1];

[0016] e represents the difference between the flight state value x ref of the ideal model and the actual state value x;

[0017] v ad represents the output value processed by the concurrent learning neural network;

[0018] A represents the output value of the concurrent learning aerodynamic parameters,

[0019] Represents the output value of the concurrent control efficiency processing, which is calculated from the control efficiency Λ of the actuator.

[0020] The concurrent learning neural network processing uses a single-hidden-layer radial basis function (RBF) neural network, and the output value v of the neural network ad is obtained through the following formula (II):

[0021]

[0022] where V represents the weight between the input layer and the hidden layer, and W represents the weight between the output layer and the hidden layer;

[0023] σ represents the non-linear activation function of the hidden layer. Preferably, the non-linear activation function is a Gaussian basis function;

[0024] represents the input layer of the neural network;

[0025] The is obtained through the following formula (III):

[0026]

[0027] where b v represents the bias, which is set to the constant 1;

[0028] x in is the input value of the input layer. Here, there are two input values, which are φ and p actually measured by the aircraft.

[0029] The V and W are obtained through the following formulas (V) and (VI):

[0030]

[0031]

[0032] where, represents the first derivative of W, represents the first derivative of v;

[0033]

[0034] I represents the identity matrix of an appropriate dimension, which is set to a 2-dimensional identity matrix;

[0035] Γ W represents the gain of w, which is set to the constant 3;

[0036] Γ V represents the gain of V, which is set to the constant 1;

[0037] Γ W1 represents the concurrent learning gain of W, which is set to the constant 1;

[0038] Γ V1 represents the concurrent learning gain of V, which is set to the constant 1;

[0039] k represents the neural network damping gain, which is set to a constant;

[0040] ||e|| represents the norm of e, that is, the norm of the difference between the flight state value of the ideal model and the actual flight state value;

[0041] represents the residual signal;

[0042] σ′ represents the derivative of the non-linear activation function σ of the hidden layer;

[0043] represents the recorded neural network input layer value;

[0044] V c represents an intermediate variable.

[0045] In this method, while performing the said step 1, step 1′ is also performed, and the said step 1′ includes the following sub-steps:

[0046] Step 1′-1, start reading data points synchronously when the aircraft starts to run, read once every 0.005 seconds, and store the read data points in the database;

[0047] Step 1′-2, when the database is full, determine the data points read again, and list the data points that meet the determination conditions as alternative data points;

[0048] Step 1′-3, perform singular value maximization processing on the alternative data points and the data points in the database, and update the data points in the database;

[0049] Step 1′-4, obtain the residual signal in real time through the database

[0050] Preferably, the said step 1′-3 includes the following sub-steps,

[0051] Sub-step a, form a data set with the alternative data points and 5 data points in the database;

[0052] Sub-step b, randomly select 5 data points from the data set and calculate their singular values;

[0053] Sub-step c, repeat sub-step b 5 times, and the data points selected each time are not exactly the same;

[0054] Sub-step d: Sort the obtained 6 singular values in descending order and select the largest one.

[0055] Sub-step e: Retrieve the 5 data points corresponding to the largest singular value and update the database with the retrieved 5 data points.

[0056] In step 1'-2, the determination condition for the data points is:

[0057] where x(t) represents the flight state value of the data point read at time t, and x p represents the flight state value of the last data point stored in the database;

[0058] represents the set determination condition, set as the constant 0.3.

[0059] When the database is full, 5 to 10 data points are recorded. Preferably, 5 data points are recorded when the database is full.

[0060] When the data points in the database are updated, the residual signal is obtained by the following formula (VII):

[0061]

[0062] where Δ represents the calculated value of the model error,

[0063] represents the second derivative of x i ;

[0064] U i represents the value of the pseudo-control quantity u recorded by the concurrent learning data record processing;

[0065] When the data points in the database are not updated, the residual signal takes the value of 0.

[0066] The output value of the concurrent control efficiency processing is obtained by the following formula (VIII):

[0067]

[0068] where Γ2 represents the gain term, set as the constant 1;

[0069] U i represents the value of the pseudo-control quantity u recorded by the concurrent learning data record processing; B represents the matrix

[0070] represents The first derivative of;

[0071] Y i represents the Y value of the concurrent learning data record processing record, where

[0072] Preferably, the a 11 、a 12 、a 21 and a 22 are obtained by the following formula (IX):

[0073]

[0074] where k CL represents the concurrent learning gain, set to the constant 0.01;

[0075] Γ1 represents the gain matrix, set to a 4-dimensional identity matrix;

[0076] Y(x) represents the state matrix,

[0077] y represents the integral of Y(x), y i represents the y of the concurrent learning data record processing record;

[0078] θ represents the parameter of the aircraft model,

[0079] represents the first derivative of θ.

[0080] The beneficial effects of the present invention include:

[0081] (1) According to the fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network provided by the present invention, the control law can be online reconstructed in the case of unknown fault information, the unknown non-linear fault can be cancelled, and thus the controller performance can be restored to a certain extent to achieve fault-tolerant control;

[0082] (2) According to the fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network provided by the present invention, it can ensure the convergence of the neural network weights to the true value without persistent excitation, improve the fault-tolerant control efficiency of the neural network, and when the fault does not change, the neural network adaptive control based on concurrent learning does not need to learn every time there is an excitation;

[0083] (3) According to the fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network provided by the present invention, it does not require a large amount of real-time calculation and is easy to implement in the case of limited computing resources of the aircraft;

[0084] (4) According to the fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network provided by the present invention, the efficiency of the actuator and the aerodynamic parameters can be estimated by using concurrent learning, and the estimation can be carried out without persistent excitation, so as to better solve the faults of the reduction of the actuator control efficiency and the change of the aerodynamic parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 Shows the framework of the fault-tolerant control method for a delta-wing aircraft with a concurrent learning neural network according to the present invention;

[0086] Figure 2 Shows the actual roll angle tracking effect of the aircraft with the fault-tolerant control method set in the experimental example of the present invention after an error occurs;

[0087] Figure 3 Shows the output value v of the concurrent learning neural network of the aircraft with the fault-tolerant control method set in the experimental example of the present invention ad Curve;

[0088] Figure 4 Shows the fitting curve of the neural network to the unknown non-linear interference in the experimental example of the present invention;

[0089] Figure 5 Shows the actual roll angle tracking effect of the aircraft without the fault-tolerant control method set in the experimental example of the present invention after an error occurs. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0090] The present invention will be further described in detail below with reference to the drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become more clearly defined.

[0091] The special term "exemplary" here means "serving as an example, embodiment or illustration". Any embodiment described as "exemplary" here does not have to be construed as superior to or better than other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings do not have to be drawn to scale unless otherwise specified.

[0092] According to the fault-tolerant control method for a delta-wing aircraft provided by the present invention, in this method,

[0093] This method includes the following steps:

[0094] Step 1, perform fault-tolerant processing on the output value of the aircraft basic controller to obtain an accurate pseudo-control quantity u;

[0095] The output value of the aircraft basic controller includes the output value of the concurrent learning aerodynamic parameters and the output value of the concurrent control efficiency processing;

[0096] Step 2, output the accurate pseudo-control quantity u to the actuator, and then control the attitude angle of the delta-wing aircraft.

[0097] In step 1, the pseudo-control quantity is obtained by the following formula (1):

[0098]

[0099] Wherein, represents the flight state value of the ideal model, which is the matrix of the ideal roll angle and the ideal roll rate measured in real time by the aircraft in the ideal model; represents taking the value of the second row of the matrix Ax(2) represents the value of the second row of the matrix Ax;

[0100] The ideal model is generally referred to as the reference model, and this model can make the expected response to the command;

[0101] x represents the actual flight state value, which can be expressed as where φ represents the aircraft roll angle (rad), and p represents the aircraft roll rate (rad / s);

[0102] Both φ and p are obtained by real-time reading of the sensor. The sensor is a rate gyro sensor, and the working frequency of the sensor is 200 Hz.

[0103] K P represents the proportional gain, K P = [-1 -1];

[0104] e represents the difference between the flight state value of the ideal model and the actual flight state value x;

[0105] v ad represents the output value processed by the concurrent learning neural network;

[0106] A represents the output value of the concurrent learning aerodynamic parameters,

[0107] represents the output value processed by the concurrent control efficiency, and is calculated from the control efficiency Λ of the actuator.

[0108] The basic controller of the aircraft is a simple proportional-integral-derivative controller (PID). The roll dynamic model of the delta-wing aircraft is expressed as

[0109] Wherein, is the first derivative of x, and is obtained by calculation; is directly measured by the sensor, is obtained by taking the first derivative of p.

[0110] The delta-wing aircraft is open-loop unstable during large-angle-of-attack rolling maneuvers, and this instability is known as the "wing rock phenomenon". This is caused by asymmetric aerodynamic effects acting on the delta wing, resulting in unstable rolling motion of the aircraft. Therefore, it is necessary to calculate and control the pseudo-control quantity u to actively control the delta-wing aircraft.

[0111] In the present invention, the framework of the fault-tolerant control method for a delta-wing aircraft using a concurrent learning neural network is as Figure 1 shown. Through fault-tolerant processing, the output value of the aircraft's basic controller is cancelled out, and after cancellation, the accurate pseudo-control quantity u is output. Then, the accurate pseudo-control quantity u is output to the actuator to control the unmanned aircraft, reducing the impact of sudden errors on the flight of the aircraft.

[0112] According to a preferred embodiment of the present invention, the concurrent learning neural network processing uses a single-hidden-layer radial basis function (RBF) neural network, and the output value v of the concurrent learning neural network processing ad is obtained through the following formula (2):

[0113]

[0114] where V represents the weight between the input layer and the hidden layer, and W represents the weight between the output layer and the hidden layer;

[0115] σ represents the non-linear activation function of the hidden layer. Preferably, the non-linear activation function is a Gaussian basis function;

[0116] represents the input layer of the neural network;

[0117] Preferably, the is obtained through the following formula (3):

[0118]

[0119] where b v represents the bias, which is set to the constant 1;

[0120] x in is the input value of the input layer. Here, there are two input values, which are φ and p actually measured by the aircraft.

[0121] The single-hidden-layer radial basis function (RBF) neural network means that there is a hidden layer between the input and the output, that is, the output of the input layer is the input of the hidden layer, and the product of the output of the hidden layer and the corresponding weight is the input of the output layer, and the output of the output layer is the final output.

[0122] Nonlinear activation functions play a very important role for artificial neural network models to learn and understand very complex and nonlinear functions. They introduce nonlinear characteristics into our network. In a neuron, after the input values are summed up with weights, a function is applied to them, and this function is the nonlinear activation function. The introduction of the nonlinear activation function is to increase the nonlinearity of the neural network model, enabling the neural network to approximate any nonlinear function, so that the neural network can be applied to numerous nonlinear models. The Gaussian function, also known as the radial basis function, is a commonly used kernel function, usually defined as a monotonic function of the Euclidean distance between any point in space and a certain center.

[0123] In this application, the nonlinear activation function is expressed as the following formula (IV):

[0124]

[0125] where j = 1,..., n2, b j is a positive scalar representing the width of the Gaussian function, and c j is the center vector, having the same dimension as the input parameter vector x, and the Euclidean distance between them is defined as ||x - c j ||.

[0126] According to the present invention, obtaining the weight value is an important link to ensure that the combination of concurrent learning and the neural network weight adjustment algorithm can ensure that the weights converge to the true values. Based on the estimated parameters of historical data and instantaneous data, even without external excitation, when the historical data contains sufficient information, the result can be ensured to converge to the true value.

[0127] The V and W are obtained through the following formulas (V) and (VI):

[0128]

[0129] where represents the first derivative of W, represents the first derivative of V;

[0130]

[0131] I represents the identity matrix of an appropriate dimension, set as the identity matrix of dimension ;

[0132] Γ W represents the gain of W, set as the constant 3;

[0133] Γ V represents the gain of V, set as the constant 1;

[0134] Γ W1 represents the concurrent learning gain of W, set as the constant 1;

[0135] Γ V1 represents the concurrent learning gain of V, which is set to the constant 1;

[0136] k represents the neural network damping gain, which is set to the constant 0.01;

[0137] ||e|| represents the norm of e, that is, the norm of the difference between the flight state value of the ideal model and the actual flight state value;

[0138] represents the residual signal;

[0139] σ′ represents the derivative of the non-linear activation function σ of the hidden layer;

[0140] represents the recorded neural network input layer value;

[0141] V c represents an intermediate variable, which has no specific physical meaning and can be calculated according to the expression.

[0142] According to a preferred embodiment of the present invention, the concurrent learning neural network processing further includes concurrent learning data recording processing, which is to record and calculate data points, determine which data points can be recorded, and can play a judgment value for subsequent data.

[0143] In this method, while performing step 1, step 1′ is also executed, and step 1′ includes the following sub-steps:

[0144] Step 1′-1, start reading data points synchronously when the aircraft starts to run, read once every 0.005 seconds, and store the read data points in the database;

[0145] Step 1′-2, when the database is full, determine the data points read again, and list the data points that meet the determination conditions as alternative data points;

[0146] Step 1′-3, perform singular value maximization processing on the alternative data points and the data points in the database, and update the data points in the database;

[0147] Step 1′-4, obtain the residual signal in real time through the database

[0148] The data points are the aircraft roll angle φ and the aircraft roll rate p read by the sensor at the current moment.

[0149] Preferably, step 1′-3 includes the following sub-steps,

[0150] Sub-step a: Combine the alternative data points and 5 data points in the database to form a data set;

[0151] Sub-step b: Arbitrarily select 5 data points from the data set and calculate their singular values;

[0152] Sub-step c: Repeat sub-step b 5 times, and the data points selected each time are not completely the same;

[0153] Sub-step d: Sort the 6 obtained singular values in ascending order and select the largest singular value among them;

[0154] Sub-step e: Retrieve the 5 data points corresponding to the largest singular value and update the database with the retrieved 5 data points.

[0155] According to the present invention, 5 to 10 data points are recorded in the database at full load. Preferably, 5 data points are recorded in the database at full load.

[0156] Since the convergence speed of the concurrent learning data record processing is proportional to the minimum singular value of the historical data stack, in order to obtain better results, it is necessary to ensure that: 1) the rank condition is satisfied, and 2) the minimum singular value is maximized.

[0157] The rank condition refers to that the rank of the historical data stack Z is equal to the number of rows of the true value weight matrix. That is, let the weight true value W * ∈R m ×n , and let rank(Z) = m. Ensuring that the rank condition is satisfied can ensure that the tracking error converges to 0, and maximizing the minimum singular value can achieve a faster convergence speed.

[0158] Although the more recording points in the database, the better, considering that the calculation time of the calculation program is too long and it is necessary to ensure that the aircraft responds and controls within a reasonable time, the number of recording points in the database cannot be set too many; if the number of data points is too small, it will lead to too little data stored in the database, unable to accurately judge the data value, which may lead to misjudgment and unable to achieve accurate control of the aircraft.

[0159] In step 1'-2, the determination condition for recording data points is:

[0160] where x(t) represents the flight state value of the data point read at time t, and x p represents the flight state value of the last data point stored in the database;

[0161] represents the set determination condition, which is set as a constant 0.3.

[0162] If the data point at the current moment meets the determination condition, the data point at time t is listed as an alternative data point, and steps 1'-3 are performed. The alternative data is processed to maximize the matrix singular value. At the same time, it is necessary to meet the maximization of the matrix singular value of the database, and the data point at time t will be recorded in the database.

[0163] If the data point at the current moment does not meet the determination condition or meets the determination condition but does not meet the maximization of the matrix singular value of the database, the current data point will not be recorded in the database.

[0164] According to the present application, maximizing the singular value can ensure that the data in the database is sufficiently different, and maximizing the minimum singular value of the database matrix can better reflect the data situation during the actual flight of the aircraft.

[0165] In steps 1'-3, the process of maximizing the matrix singular value is as follows. First, record the points that are sufficiently different from the previous data. If the number of stored data points exceeds the maximum allowable number, the algorithm will merge the new data points in a way that maximizes the minimum singular value of the data stack. To achieve this, the algorithm sequentially replaces the data points in the historical data stack with the current data point and stores the resulting minimum singular value in a variable, finds the maximum minimum singular value, and replaces the original data point with the new data point. The specific algorithm is as follows:

[0166] According to a preferred embodiment of the present invention, the residual signal represents the difference between the instantaneous estimate of the model error and the stored estimate of the model error, and can be directly used as the training information for the adaptive update law.

[0167] When the data points in the database are updated, the residual signal is obtained by the following formula (VII):

[0168]

[0169] where Δ represents the calculated value of the model error,

[0170] represents the second derivative of x i ;

[0171] x i represents the flight state value of the i-th data point of the latest record of the data point update, which is obtained by real-time measurement by the sensor;

[0172] U i represents the value of the pseudo-control quantity u recorded by the concurrent learning data record processing;

[0173] When the data points in the database are not updated, the residual signal takes the value of 0.

[0174] Through the above method, the roll angle and roll rate of the aircraft can be obtained in real time by the delta-wing aircraft, the weights V and W of the concurrent learning neural network can be obtained, and further the output value processed by the concurrent learning neural network can be obtained, which is offset against the error data in the aircraft basic controller to achieve the effect of fault tolerance.

[0175] According to a preferred embodiment of the present invention, the concurrent learning control efficiency processing is further included in the fault tolerance processing, and the calculated value of the control efficiency during the real-time flight of the aircraft can be obtained through the concurrent learning control efficiency processing

[0176] The calculated value of the control efficiency is obtained through the following formula (VIII):

[0177]

[0178] where Γ2 represents the gain term, which is set to the constant 1;

[0179] U i represents the value of the pseudo control quantity u recorded by the concurrent learning data recording process; B represents the matrix

[0180] represents the first derivative of

[0181] Y i represents the Y value recorded by the concurrent learning data recording process, where

[0182] Through the concurrent learning estimated control efficiency processing, the real-time control efficiency of the aircraft can be calculated and output. Using the output value of the concurrent learning control efficiency processing for the calculation of the pseudo control quantity u, a pseudo control quantity u closer to the actual situation can be obtained, so that when the aircraft encounters a fault, precise control of the aircraft can be achieved through this method, avoiding the crash of the unmanned aircraft and completing the basic work tasks.

[0183] Preferably, the a 11 、a 12 、a 21 and a 22 are obtained through the following formula (IX):

[0184]

[0185] where k CL represents the concurrent learning gain, which is set to the constant 0.01;

[0186] Γ1 represents the gain matrix, which is set as a constant 4-dimensional identity matrix;

[0187] Y(x) represents the state matrix,

[0188] y represents the integral of Y(x), y i represents the y processed from the concurrent learning data record;

[0189] θ represents the parameter of the aircraft model,

[0190] represents the first derivative of θ.

[0191] By solving the aerodynamic parameter θ, the output value A of the concurrent learning aerodynamic parameter processing can be obtained.

[0192] Through the concurrent learning aerodynamic parameter processing, the output value A of the concurrent learning aerodynamic parameter can be obtained in real time. When there is an aerodynamic effect on the flight of the aircraft, the aerodynamic parameter is adjusted through the output value A to reduce the influence caused by environmental factors on the aircraft and reduce the probability of the aircraft malfunctioning.

[0193] According to the fault-tolerant control method of the delta-wing aircraft provided by the present invention, the control law can be reconfigured online in the case of unknown fault information, and the unknown non-linear fault can be cancelled, so as to restore the controller performance to a certain extent and achieve fault-tolerant control. By using concurrent learning to estimate the actuator efficiency and aerodynamic parameters, the estimation can be carried out under the condition of no persistent excitation to better solve the faults of the reduction of the actuator control efficiency and the change of the aerodynamic parameters.

[0194] Embodiment:

[0195] Select the following rolling dynamic model of the delta-wing aircraft:

[0196]

[0197] Start calculating the output value v of the concurrent learning neural network from the flight start moment, ad At this time, v ad is 0, and a fault occurs 10 s after the flight starts due to the influence of strong wind.

[0198] There is a fault-tolerant control method set in the aircraft. The aircraft obtains its own φ and p values through detection and obtains the output value v of the concurrent learning neural network. ad The curve of Figure 3 is as shown, and the fitting curve of the neural network for unknown non-linear interference is as shown in Figure 4 The output value of the concurrent control efficiency processing is 0.75. K p = [-1 -1],

[0199] Through the pseudo-control quantity u is obtained in real time.

[0200] Through the fault-tolerant control method in the aircraft, after the aircraft fails, the fault can be processed, the tracking of the roll angle can be ensured, and the actual roll angle tracking effect is as Figure 2 shown.

[0201] Comparative example:

[0202] Select the same roll dynamic model of the general delta-wing aircraft as in the embodiment, and calculate the output value v of the concurrent learning neural network starting from the flight start moment. ad At this time, v ad is 0, and a fault occurs 10 s after the flight starts due to the influence of strong wind.

[0203] In the aircraft, no fault-tolerant control method is set. Only through the basic controller, after a fault occurs, the fault cannot be processed, and the actual roll angle tracking effect is as Figure 3 shown.

[0204] In Figures 2 to 5 , r represents the roll angle command, ref represents the roll angle in the ideal model, plant represents the roll angle during the actual flight of the aircraft, and nonlinear represents the non-linear fitting curve.

[0205] From Figure 2 it can be seen that through the fault-tolerant control method in this application, problem data can be processed in a timely manner, the pseudo-control quantity of the aircraft can be corrected, the roll angle tracking can be ensured in case of emergencies, and further, it can be ensured that the UAV does not crash in case of a fault and can continue to complete the task.

[0206] From Figure 5 it can be seen that when an emergency occurs in an aircraft without a fault-tolerant control method, fault tolerance cannot be processed, resulting in the UAV being unable to track the roll angle and possibly causing the UAV to crash.

[0207] In summary, through Figures 2 to 5 it can be seen that the fault-tolerant control method in this application can realize the online adjustment of the control law to solve unknown non-linear interference and realize the fault-tolerant control of the aircraft.

[0208] The present invention has been described above in combination with preferred embodiments, but these embodiments are only exemplary and only play an illustrative role. On this basis, various substitutions and improvements can be made to the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A fault-tolerant control method for a delta-wing aircraft based on a concurrent learning neural network, characterized in that, The method includes the following steps: Step 1: Perform fault tolerance processing on the output value of the aircraft basic controller to obtain the pseudo control quantity u; Step 2: Output the accurate pseudo control quantity u to the actuator, and then control the attitude angle of the delta wing aircraft; In Step 1, the pseudo control quantity u is obtained through the following formula (1): Among them, represents the flight state value of the ideal model, which is the matrix of the ideal roll angle and the ideal roll rate measured in real time by the aircraft in the ideal model represents taking the matrix The value of the second row, Ax(2) represents the value of the second row of the matrix Ax; x represents the actual flight state value, where φ represents the aircraft roll angle and p represents the aircraft roll rate; K P represents the proportional gain, K P = [-1 -1]; e represents the difference between the flight state value of the ideal model and the actual state value x; v ad represents the output value processed by the concurrent learning neural network; A represents the output value of the concurrent learning aerodynamic parameters, Represents the output value of concurrent control efficiency processing, which is calculated from the control efficiency Λ of the actuator.

2. The fault-tolerant control method for a delta-wing aircraft according to claim 1, characterized in that, The concurrent learning neural network processing adopts a single-hidden-layer radial basis function (RBF) neural network, and the output value v of the neural network ad is obtained by the following formula (2): Where, V represents the weight between the input layer and the hidden layer, and W represents the weight between the output layer and the hidden layer; σ represents the non-linear activation function of the hidden layer, and the non-linear activation function is a Gaussian basis function; Represents the input layer of the neural network.

3. The fault-tolerant control method for a delta-wing aircraft according to claim 2, characterized in that, The said is obtained by the following formula (III): where b v represents the offset, set to the constant 1; x in Namely, the input values of the input layer. Here, there are two input values, namely φ and p actually measured by the aircraft.

4. The fault-tolerant control method for a delta-wing aircraft according to claim 3, characterized in that, The V and W are obtained through the following formula (5) and formula (6): Among them, represents the first derivative of W, represents the first derivative of V; I represents the identity matrix of an appropriate dimension, set as a 2-dimensional identity matrix; Γ W Represents the gain of W, set to the constant 3; Γ V Represents the gain of V, set to the constant 1; Γ W1 Represents the concurrent learning gain of W, set to the constant 1; Γ V1 Represents the concurrent learning gain of V, set to the constant 1; κ represents the neural network damping gain, set as the constant 0.01; ||e|| represents the norm of e, that is, the norm of the difference between the flight state value of the ideal model and the actual flight state value; Indicates a residual signal; σ′ represents the derivative of the non-linear activation function σ of the hidden layer; Indicates the input layer of the neural network for the record value; V c represents an intermediate variable.

5. The fault-tolerant control method of a delta-wing aircraft according to claim 1, characterized in that, In this method, while performing Step 1, Step 1′ is also performed. The Step 1′ includes the following sub-steps: Step 1′-1: Start reading data points synchronously when the aircraft starts to run, read once every 0.005 seconds, and store the read data points in the database; Step 1′-2: When the database is full, judge the data points read again, and list the data points that meet the judgment conditions as alternative data points; Step 1′-3: Perform singular value maximization processing on the alternative data points and the data points in the database, and update the data points in the database; Steps 1'-4, obtain the residual signal in real time through the database The Step 1′-3 includes the following sub-steps: Sub-step a: Combine the alternative data points and 5 data points in the database to form a data set; Sub-step b: Arbitrarily select 5 data points from the data set and calculate their singular values; Sub-step C: Repeat Sub-step b 5 times, and the data points selected each time are not exactly the same; Sub-step d: Sort the 6 obtained singular values in descending order and select the largest singular value among them; Sub-step e: Retrieve the 5 data points corresponding to the largest singular value, and update the database with the retrieved 5 data points.

6. The fault-tolerant control method of a delta-wing aircraft according to claim 5, characterized in that, In step 1'-2, the determination condition for recording data points is: Among them, x(t) represents the flight state value of the data point read at time t, and x p represents the flight state value of the last data point stored in the database; Indicates the set determination condition, set as the constant 0.

3.

7. The fault-tolerant control method of a delta-wing aircraft according to claim 5, characterized in that, When the database is full, 5 to 10 data points are recorded.

8. The fault-tolerant control method of a delta-wing aircraft according to claim 4, characterized in that, When the data points in the database are updated, the residual signal is obtained by the following formula (VII): where Δ represents the calculated value of the model error, Denote the second derivative of x i ; U i Represents the pseudo-control quantity u value processed by the concurrent learning data record; When the data points in the database are not updated, the residual signal takes a value of 0.

9. The fault-tolerant control method of a delta-wing aircraft according to claim 8, characterized in that, The output value of the concurrent control efficiency processing is obtained by the following formula (VIII): Where, Γ2 represents the gain term, set as the constant 1; B represents a matrix denote the first derivative of; Y i represents the Y value processed by the concurrent learning data record, where

Images