A robust atmospheric data fusion method based on adaptive fault detection and removal strategy

Abstract

The present invention belongs to the field of aerospace technology, and specifically relates to a robust atmospheric data fusion method based on a fault adaptive detection and removal strategy. Traditional solutions for redundant sensors use data processing methods such as "voting system", averaging, and measurement (primary sensor)-monitoring (secondary sensor) systems, which cannot effectively handle sensor failures in the ADS system. Therefore, the present invention introduces a mechanism that can automatically detect and remove sensor faults without human intervention, and is committed to designing an atmospheric data estimation algorithm that is robust to sensor failures. Simulation results verify the feasibility of the proposed robust atmospheric data estimation scheme in a real flight environment.

Description

Technical Field

[0001] The present invention belongs to the field of aerospace technology, and is particularly applicable to redundant atmospheric data systems of aircraft, and relates to a robust atmospheric data fusion method based on a fault adaptive detection-removal strategy. Background Art

[0002] Since the advent of aircraft, flight safety has been a fundamental requirement. Whether autonomous or manned, atmospheric data such as angle of attack, sideslip angle, and airspeed influence aircraft flight control and guidance and are crucial to safe flight. This atmospheric data is typically measured and acquired by the aircraft's air data sensor (ADS) system. Because ADS systems are directly exposed to the flight environment, they are prone to failure in harsh weather conditions due to factors such as icing and water accumulation, resulting in erroneous atmospheric data and posing a significant challenge to aircraft flight safety. Providing consistent and reliable atmospheric data in complex environments to ensure flight safety is a relentless goal within the aviation industry. To mitigate the potential risks posed by ADS system failures, engineers, based on the concept of redundancy, have incorporated multiple sets of the same type of sensors into aircraft ADS systems. However, common data processing methods for redundant sensors, such as voting, averaging, and measurement (primary sensor)-monitoring (secondary sensor) systems, are ineffective in handling sensor failures within the ADS system. Summary of the Invention

[0003] To address the above issues, the present invention adds an atmospheric data sensor fault adaptive detection and removal module to the traditional ADS system, realizing self-detection and self-diagnosis of redundant atmospheric data sensors, as well as information fusion of redundant atmospheric data sensor measurements, thereby significantly improving the reliability of the ADS system and the accuracy of the atmospheric data output, which is of vital importance to ensuring safe flight.

[0004] The purpose of this invention is to design a robust atmospheric data estimation algorithm suitable for redundant ADS systems with faulty sensors, achieving accurate atmospheric data estimation with low computational complexity. Specifically, by leveraging the concepts and algorithms for solving classification problems in the field of artificial intelligence, a novel, simple, and direct atmospheric data estimation method is proposed for redundant aircraft ADS systems. This method implements self-detection of atmospheric sensor faults, thereby meeting the aircraft's demand for reliable and continuous atmospheric data estimates.

[0005] The technical solution of the present invention is:

[0006] An atmospheric data fusion method based on a fault adaptive detection and removal strategy for a redundant atmospheric data system includes:

[0007] The fault-resistant atmospheric data estimation model is established as:

[0008] y t =Hx t +w t

[0009] in, Contains all measurements of the aircraft's dynamic system, is the measurement matrix composed of H1 and H2, w t is the noise vector composed of all sensor noises, y 1t It is obtained from the measurement value of the aircraft inertial navigation system:

[0010] y 1t =H1x t +w 1t

[0011]

[0012] where x t is the aircraft system status, including atmospheric data and Euler angles, w 1t is the inertial navigation system sensor noise; y 2t It is obtained by measuring the value of the aircraft's atmospheric data sensor. The atmospheric data sensor is defined as K1 airspeed sensor, K2 angle of attack sensor, and K3 sideslip angle sensor. The measurement equation is:

[0013]

[0014] in, is the k1th airspeed sensor measurement value, H as is the airspeed sensor measurement matrix, The k1th airspeed sensor measurement noise, is the measurement value of the k1th angle of attack sensor, H aoa is the angle of attack sensor measurement matrix, is the k1 angle of attack sensor measurement noise, is the measured value of the k1th sideslip angle sensor, H ssa is the sideslip angle sensor measurement matrix, For the k1th sideslip angle sensor measurement noise, define We can get:

[0015] y 2t =H2x t +w 2t

[0016] Among them, y 2t ∈R KContains all sensor measurements in the redundant ADS system, K = K1 + K2 + K3, H2 is the measurement matrix formed according to the sources of different components;

[0017] Based on the established estimation model, the estimated value of the atmospheric data is obtained through the fault adaptive detection and removal strategy, specifically:

[0018] The sensor observation value y t The i-th component introduces a binary indicator variable z it To characterize its properties, if y it is a valid observation value, then z it =1, otherwise y it is an outlier, then z it =0, that is:

[0019]

[0020] Among them, h i is the i-th row of the measurement matrix H, when y it is the effective observation value, i.e. z it =1, y it The likelihood function is σ ii,t is the diagonal matrix R t The i-th main diagonal element of ; and when z it =0, y it The likelihood function value is a constant, so y t The likelihood function is unified as:

[0021]

[0022] Where m is the index number of the sensor.

[0023] The indicator variable z it Modeled as a random variable following a Beta-Bernoulli distribution:

[0024]

[0025] Among them, Beta(·,·) represents the beta function, e it and f it is the parameter of the Beta-Bernoulli distribution;

[0026] definition and π t ={π it}, the unknown quantity to be determined is obtained through the following maximum a posteriori probability estimation problem:

[0027]

[0028] The variational Bayesian inference method is used to approximate the solution:

[0029] p(x t ,z t ,π t |y 1:t )≈q(x t ,z t ,π t )=q(x t )q(z t )q(π t )

[0030] According to the VBI theory, the distribution of different factors is solved through the following sequential iterative process:

[0031]

[0032] Where k is the iteration index of the VBI inference process, is the total probability density function, q k (x t ) is solved by the solution of Kalman smoothing theory, q k (z t ) and q k (π t ) Due to the relationship between the conjugate prior distribution, there is a closed solution; after a certain iterative process, the VBI process converges, and at this time q k (x t ) is approximately Gaussian, and its expectation is used as the estimated value of the state, that is,

[0033]

[0034] According to the definition of system state, x t Contains atmospheric data and Euler angles to obtain estimated values ​​of atmospheric data.

[0035] The beneficial effect of the present invention is that the atmospheric data estimation method proposed in the present invention can estimate redundant atmospheric data with relatively low computational complexity, thereby providing a guarantee for the safe flight of an aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The measured and estimated values ​​of the angle of attack in case 1 (S1 angle of attack sensor is stuck);

[0037] Figure 2 are the measured and estimated values ​​of the angle of attack in case 2 (the stuck time of S1 and S2 angle of attack sensors does not overlap);

[0038] Figure 3 These are the measured and estimated values ​​of the angle of attack in scenario three (the stuck time of the S1 and S2 angle of attack sensors overlaps). DETAILED DESCRIPTION

[0039] The present invention is described in detail below with reference to the accompanying drawings and simulation examples to demonstrate the practicability of the present invention.

[0040] The aircraft system kinematic model corresponding to the aircraft VADS can be described by the following differential equations

[0041]

[0042] Among them, the system state x t =(V t ,α t ,β t ,φ t ,θ t ,ψ t ) T It can be mainly divided into two categories, namely, airspeed V t , angle of attack α t , sideslip angle β t Atmospheric data including the roll angle φ t , pitch angle θ t and yaw angle ψ t The Euler angles included, the system input Including three-axis acceleration Roll angular rate p t , pitch angle rate q t and the yaw rate r t , f(·)=(f1,f2,f3,f4,f5,f6) T It is the following nonlinear function

[0043]

[0044] f4=p t +q t sinφ t tanθ t +r t cosφ t tanθ t

[0045] f5=q t cosφ t -r t sinφ t

[0046]

[0047] The system input can be directly measured by the IMU sensor equipped in the aircraft, that is,

[0048] utm =u t +v t

[0049] Among them, u tm is the reading of the IMU sensor, v t is the measurement noise of the IMU sensor. Substituting the IMU reading into the system input, the aircraft dynamics equation based on the IMU data can be obtained:

[0050]

[0051] Among them, the scaling matrix G(x t ) is defined as

[0052]

[0053] Considering that the measurement values ​​of sensors in the ADS system are discrete, it is necessary to discretize the aircraft dynamics model described by the stochastic differential equation. Common discretization methods include the Euler method, that is,

[0054] x t+1 =x t +f(x t ,u tm )△t+G(x t )v t △t

[0055] Where △t is the discrete time interval, which is usually equal to the sampling time of the sensor in the ADS system.

[0056] The measurements of the aircraft system can be divided into two categories, namely those provided by the inertial navigation system (φ tm ,θ tm ,ψ tm ) and (V tm ,α tm ,β tm The Euler angle measurement value provided by the inertial navigation system is the result of integrating information from various sensors such as gyroscopes, accelerometers, and magnetic sensors. In this invention, it is assumed that the inertial navigation system that provides such measurement values ​​has no faults. Then there is

[0057] y 1t =H1x t +w 1t

[0058] Among them, H1 is

[0059]

[0060] According to the concept of redundant design, for each type of atmospheric data, the redundant ADS system is provided by several sensors of the same type. Therefore, assuming that the number of airspeed sensors, angle of attack sensors, and sideslip angle sensors are K1, K2, and K3 respectively, the corresponding measurement equation can be written as

[0061]

[0062] For convenience, define The above formula can be written as

[0063] y 2t =H2x t +w 2t

[0064] Among them, y 2t ∈R K (K=K1+K2+K3) contains all sensor measurements in the redundant ADS system, and H2 is based on y 2t The measurement matrix is ​​formed by the sources of different components.

[0065] The unified measurement equation of the aircraft is obtained, namely

[0066] y t =Hx t +w t

[0067] in, Contains all measurements of the aircraft's dynamic system, is the measurement matrix composed of H1 and H2, w t is the noise vector consisting of all sensor noises, usually assumed to be Gaussian distributed

[0068] In order to establish the mathematical model of the above-mentioned sensor fault adaptive detection-removal mechanism, the present invention proposes to introduce a binary indicator variable to the sensor to mark its attribute (i.e., whether a fault occurs), and use the corresponding probability distribution model to model the binary indicator variable to realize the adaptive mechanism. Specifically, the sensor observation value y t The i-th component introduces a binary indicator variable z it To characterize its properties. If y it is a valid observation value, then z it =1, otherwise, if y it is an outlier, then z it =0, that is

[0069]

[0070] Among them, h i is the i-th row of the measurement matrix H. When yit is the effective observation value, i.e. z it =1, y it The likelihood function can be written as σ ii,t is the diagonal matrix R t The i-th main diagonal element of ; and when y it is an outlier value, i.e. z it =0, y it The likelihood function value of should be a constant. Combining the above two points, y t The likelihood function can be unified as

[0071]

[0072] It should be noted that the measured value y t It also includes the Euler angle provided by the navigation system. In the present invention, it is assumed that the navigation system has no faults, so for the measured value y t The Euler angle measurement value in , let the corresponding z it =1.

[0073] In the unified form of the likelihood probability density function given by it Taking different values ​​to implement the sensor fault detection-removal strategy. However, the indicator variable z in the model it It still needs to be set manually. In order to realize the adaptive and autonomous execution of the sensor fault detection-removal strategy, the indicator variable z is further it Modeled as a random variable following a Beta-Bernoulli distribution:

[0074]

[0075] Among them, Beta(·,·) represents the beta function, e it and f it is the parameter of the Beta-Bernoulli distribution. By adjusting e it and f it The correct adjustment can be achieved.

[0076] Combining the above sensor fault adaptive detection-removal model with the SSM model for estimating atmospheric data, we can obtain an atmospheric data estimation model based on sensor fault adaptive detection-removal. and π t ={π it}, theoretically the unknown quantity can be obtained through the following maximum a posteriori probability estimation problem:

[0077]

[0078] However, the maximum a posteriori probability density is almost impossible to solve, so the variational Bayesian inference method (VBI) is used to approximate the solution. Specifically, the following mean field method is used to approximate the maximum a posteriori probability density, that is,

[0079] p(x t ,z t ,π t |y 1:t )≈q(x t ,z t ,π t )=q(x t )q(z t )q(π t )

[0080] According to the VBI theory, the distribution of different factors can be solved through the following sequential iterative process:

[0081]

[0082] Where k is the iteration index of the VBI inference process, is the total probability density function. k (x t ) can be solved with the help of Kalman smoothing theory solutions such as UKS and CKS, q k (z t ) and q k (π t ) Due to the relationship between the conjugate prior distribution, there is a closed solution. After a certain iterative process, the VBI process will converge. At this time, q k (x t ) is usually approximated to a Gaussian distribution, so its expectation can be used as an estimate of the state, that is,

[0083]

[0084] According to the definition of system state, x t Contains atmospheric data and Euler angles, and obtains the estimated value of atmospheric data.

[0085] Taking the longitudinal short-period linear system of a certain aircraft as an example, the feasibility of the proposed solution to provide angle of attack estimation when the redundant angle of attack sensor fails is verified. The discrete dynamic equation of the longitudinal short-period of a certain aircraft is:

[0086]

[0087] Among them, x t =[α t ,q t] is the state of the system, including the angle of attack α t and the pitch rate q t ;u t =δ t is the input of the elevator. Considering the measurement values ​​provided by one pitch rate sensor and two angle of attack sensors of the aircraft, the measurement equation is

[0088]

[0089] Assume that the pitch rate sensor works normally, but the angle of attack sensor may fail. To this end, introduce a binary random variable z for each of the two angle of attack sensors. 1t and z 2t As an indicator of whether a fault has occurred, the estimated value of the angle of attack is calculated according to the solution proposed by the present invention.

[0090] Three failure scenarios of the angle of attack sensors are considered. The first scenario is that one of the two angle of attack sensors fails while the other works normally. The measured value is as follows: Figure 1 (a) As shown in the second case, the two angle of attack sensors fail alternately, but at the same time one sensor can work normally, and the measured value is as follows: Figure 2 (a) As shown; the third situation is that the two angle of attack sensors fail alternately, and both sensors fail at the same time for a period of time. The measured values ​​are as follows Figure 3 (a) shows. Three schemes are used to obtain the estimated value of the angle of attack. First, the measurement value of a single sensor is directly used; second, the measurement value of the two angle of attack sensors is processed without introducing the adaptive fault detection and removal process; third, the measurement value of the two angle of attack sensors is processed by the adaptive sensor fault detection and removal process. Under the three different situations, the estimation results are shown as follows: Figure 1 (b) Figure 2 (b) and Figure 3 (b) These results show that the estimated angle of attack is close to the true value only when the sensor is fault-free or when the sensor fault adaptive detection and removal process is used. This preliminarily demonstrates the effectiveness and feasibility of the proposed robust atmospheric data estimation algorithm for redundant ADS systems based on sensor fault adaptive detection and removal.

[0091] In summary, this paper develops a robust atmospheric data estimation method for redundant aircraft ADS systems, resistant to sensor failures. By introducing a mechanism that automatically detects and removes sensor failures without human intervention, a robust atmospheric data estimation algorithm is designed. Simulation results demonstrate the feasibility of the proposed robust atmospheric data estimation scheme in a real-world flight environment.

Claims

1. An atmospheric data fusion method based on a fault adaptive detection and removal strategy, characterized in that: include: The fault-resistant atmospheric data estimation model is established as: y t =Hx t +w t in, Contains all measurements of the aircraft's dynamic system, is the measurement matrix composed of H1 and H2, w t is the noise vector composed of all sensor noises, y 1t It is obtained from the measurement value of the aircraft inertial navigation system: y 1t =H1x t +w 1t where x t is the aircraft system status, including atmospheric data and Euler angles, w 1t is the inertial navigation system sensor noise; y 2t It is obtained by measuring the value of the aircraft's atmospheric data sensor. The atmospheric data sensor is defined as K1 airspeed sensor, K2 angle of attack sensor, and K3 sideslip angle sensor. The measurement equation is: in, is the k1th airspeed sensor measurement value, H as is the airspeed sensor measurement matrix, The k1th airspeed sensor measurement noise, is the measurement value of the k1th angle of attack sensor, H aoa is the angle of attack sensor measurement matrix, is the k1 angle of attack sensor measurement noise, is the measured value of the k1th sideslip angle sensor, H ssa is the sideslip angle sensor measurement matrix, For the k1th sideslip angle sensor measurement noise, define We can get: y 2t =H2x t +w 2t Among them, y 2t ∈R K Contains all sensor measurements in the redundant ADS system, K = K1 + K2 + K3, H2 is the measurement matrix formed according to the sources of different components; Based on the established estimation model, the estimated value of the atmospheric data is obtained through the fault adaptive detection and removal strategy, specifically: The sensor observation value y t The i-th component introduces a binary indicator variable z it To characterize its properties, if y it is a valid observation value, then z it =1, otherwise y it is an outlier, then z it =0, that is: Among them, h i is the i-th row of the measurement matrix H, when y it is the effective observation value, i.e. z it =1, y it The likelihood function is σ ii,t is the diagonal matrix R t The i-th main diagonal element of ; and when z it =0, y it The likelihood function value is a constant, so y t The likelihood function is unified as: Where m is the index number of the sensor; The indicator variable z it Modeled as a random variable following a Beta-Bernoulli distribution: Among them, Beta(·,·) represents the beta function, e it and f it is the parameter of the Beta-Bernoulli distribution; definition and π t ={π it }, the unknown quantity to be determined is obtained through the following maximum a posteriori probability estimation problem: The variational Bayesian inference method is used to approximate the solution: p(x t ,z t ,π t |y 1:t )≈q(x t ,z t ,π t )=q(x t )q(z t )q(π t ) According to the VBI theory, the distribution of different factors is solved through the following sequential iterative process: Where k is the iteration index of the VBI inference process, is the total probability density function, q k (x t ) is solved by the solution of Kalman smoothing theory, q k (z t ) and q k (π t ) all have closed solutions due to the conjugate prior distribution relationship; After a certain iterative process, the VBI process converges, and q k (x t ) is approximately Gaussian, and its expectation is used as the estimated value of the state, that is, According to the definition of system state, x t Contains atmospheric data and Euler angles to obtain estimated values ​​of atmospheric data.

Images