A method for monitoring accelerometer failure in an aircraft flight control system

Abstract

The present invention discloses an accelerometer fault monitoring method for an aircraft flight control system, and belongs to the field of accelerometer fault monitoring. The present invention constructs a mathematical model for the accelerometer fault, selects a residual sequence as a monitoring signal through model analysis, and detects whether the fault occurs by comparing the residual sequence. The present invention can specifically monitor three types of accelerometer faults through a machine learning method of a support vector machine, and can determine whether the fault belongs to a constant gain, a constant deviation or a stuck fault at the first time when the fault signal appears. The present invention uses a support vector machine to train and learn the collected training set data, thereby improving the processing capability of a large number of samples of accelerometer fault monitoring and improving the accuracy of accelerometer fault monitoring.

Description

Technical Field

[0001] The invention belongs to the field of accelerometer fault monitoring, in particular to an accelerometer fault monitoring method for an aircraft flight control system. Background Art

[0002] At present, drones are getting closer and closer to people's lives, which puts higher demands on their safety and reliability. Since drones are highly integrated automated equipment, the causes and types of their failures are varied. Among them, sensor failure is one of the most common failures. The sensor is responsible for maintaining the flight balance and navigation of the drone. If the sensor fails, the drone may be forced to land at the least, or crash at the worst, causing serious losses. Therefore, fault monitoring of drone sensors is very important.

[0003] Accelerometers are widely used in inertial navigation systems, such as civil aircraft, logistics and transportation drones, and automotive navigation. Accelerometers are used to measure the acceleration parameters of moving bodies in inertial navigation systems, and their working conditions directly affect the performance of inertial navigation systems. Accelerometers will have errors that change slowly over time during operation. When the errors accumulate to a certain extent, they are prone to failure. Once a failure occurs, it will cause navigation and positioning errors, and may even lead to huge catastrophic events. The errors generated by accelerometers include deterministic errors and random errors. Deterministic errors include: installation errors, scale factors, etc., which are caused by process errors or installation imbalances. They can be solved by improving the process level or external calibration. Random errors are generated during the working process and have no definite rules, so it is difficult to eliminate them from the outside. Since the output signal of the accelerometer is an irregular nonlinear signal containing a variety of high and low frequency information, it is difficult to mathematically model it. The feature extraction of the accelerometer output signal is the key to fault monitoring. Whether its feature extraction is sufficient is related to the accuracy of the fault diagnosis results. Summary of the invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art and provide an accelerometer fault monitoring method for an aircraft flight control system.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for monitoring aircraft accelerometer faults based on a machine learning method comprises the following steps:

[0007] Step 1: Establish an output model of an accelerometer of an aircraft flight control system;

[0008] Step 2: construct a fault output model based on the three faults of accelerometer sensor: constant deviation, constant gain and stuck;

[0009] Step 3: Use the EKF algorithm to estimate the accelerometer output, and then make a difference between the estimated accelerometer output value and the actual output to obtain the residual sequence of the system;

[0010] Step 4: Use the residual sequence of the system under the normal state and three fault states of the accelerometer to train the SVM decision tree. The trained SVM decision tree can determine whether the system is faulty and the fault type.

[0011] Step 5: Use the SVM decision tree to classify the residual sequence of the unknown category of the aircraft system, and monitor the three types of accelerometer sensor faults based on the residual sequence classification results.

[0012] Furthermore, step 1 is specifically as follows:

[0013] The established accelerometer output model a is:

[0014]

[0015]

[0016] in, Represents the acceleration of the aircraft in the navigation coordinate system, g 0 represents the gravitational acceleration in the inertial system, The attitude quaternion q=(q 0 ,q 1 ,q 2 ,q 3 ) T represents the attitude transfer matrix, η represents the Gaussian white noise process;

[0017] when When , the approximate expression of formula (1) is:

[0018]

[0019] Normalize a in (3) to get z, and discretize (3) to get

[0020] z k+1 =g(q k+1 )+δz k+1 (4)

[0021]

[0022] Among them, z k+1 represents the normalized, discrete a; δz k+1 represents the discretized Gaussian white noise η;

[0023] The quaternion differential equation is:

[0024]

[0025] Where w=(w x ,w y ,w z ) is the output of the three-axis gyroscope, q=((q 0 ,q 1 ,q 2 ,q 3 ) T is the attitude quaternion, Represents quaternion multiplication;

[0026] After further solving and discretization, we get:

[0027] q k+1 =φ k+1 / k q k +Γ k+1 / k δθ k (7)

[0028] Taking equation (4) as the observation equation and equation (7) as the state equation, a nonlinear attitude estimation system with observation equation is formed as equation (8):

[0029]

[0030] Among them, q k Represents discrete attitude quaternion; z k+1 is the monitoring quantity, i.e. the normalized output of the accelerometer; φ k+1 / k and Γ k+1 / k are the relative values ​​of the two measurements of the accelerometer before and after; δ is θ k The normalization coefficient of .

[0031] Furthermore, step 2 constructs a model based on the three faults of constant deviation, constant gain and stuck of the accelerometer sensor, specifically:

[0032] The sensor constant deviation fault is manifested as the accelerometer outputting a constant deviation at a certain moment. The corresponding fault model is:

[0033]

[0034] Among them, α is a constant and α≠0;

[0035] The constant gain fault is manifested as the accelerometer starting to output a multiplier factor at a certain moment. The corresponding fault model is:

[0036]

[0037] Where β is the constant gain proportional coefficient and β≠1;

[0038] The stuck fault is manifested as the accelerometer output maintaining a constant value at a certain moment, and the corresponding fault model is:

[0039]

[0040] Among them, γ is a constant and γ≠0;

[0041] The above various accelerometer failure forms are uniformly expressed in the following form:

[0042]

[0043] Among them, β=1, γ=0 are not true at the same time;

[0044] When the system works normally, the system output residual is:

[0045]

[0046] The corresponding residual expectation is:

[0047]

[0048] When a system failure occurs, the system output residual is:

[0049]

[0050] The corresponding residual expectation is:

[0051]

[0052] However, β=1 and γ=0 do not hold simultaneously.

[0053] Furthermore, it is characterized in that

[0054] When E(e k+1 )=0, the accelerometer has no fault.

[0055] When E(e k+1 )≠0, the accelerometer is faulty.

[0056] Furthermore, in step 3, the state of the aircraft attitude estimation system is estimated by EKF, and then the state is substituted into the observation equation to obtain the estimate of the accelerometer output value, and then the estimate of the accelerometer output value is subtracted from the actual output to obtain the residual sequence of the system.

[0057] Furthermore, in step 3, the state of the aircraft attitude estimation system is estimated by EKF, specifically:

[0058] (301) Initialization

[0059] Assign values ​​to the three-axis gyroscope noise covariance matrix Q and the three-axis accelerometer noise covariance R according to actual conditions; and the state covariance matrix P 0 / 0 The initial value is:

[0060]

[0061]

[0062] (302) Time Update

[0063] q k+1 / k =φ k+1 / k q k

[0064]

[0065]

[0066] (303) Observation Update

[0067]

[0068] q k+1 / k+1 =q k+1 / k +K k (z k+1 -g(q k+1 / k ))P k+1 / k

[0069]

[0070] Furthermore, in step 4, the training of the SVM decision tree is performed from bottom to top.

[0071] Furthermore, the training process is:

[0072] (401), collect the residual sequence training sample set of the aircraft system in the three fault states and normal state of the accelerometer as {X 1 ,X 2 ,X 3 ,X 4}, each training set X i Contains j The training samples are

[0073] (402), calculate the inter-class separation degree {s 12 ,s 13 ,s 14 ,s 23 ,s 24 ,s 34}, select the inter-class separation s ij The smallest two-class training set trains the binary classification SVM;

[0074] (403), merging the inter-class separation degree s ij The smallest two-category training set is one category;

[0075] (404), looping steps (402) and (403) until the binary classification SVM at the root node.

[0076] Furthermore, the classification process of step 5 is:

[0077] When the SVM decision tree is used to classify an aircraft system residual sequence of unknown category, the SVM binary classifier at the root node is used to classify the category of the aircraft system residual sequence, and then the sub-SVM binary classifier of the current node is used to classify the category of the aircraft system residual sequence downward according to the structure of the decision tree until it is classified into a certain category.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] The present invention discloses an accelerometer fault monitoring method for an aircraft flight control system. The method constructs a mathematical model for the accelerometer fault, selects a residual sequence as a monitoring signal through model analysis, and detects whether the fault occurs by comparing the residual sequence. The present invention can specifically monitor three types of accelerometer faults through a machine learning method of a support vector machine, and can judge whether the fault belongs to a constant gain, a constant deviation, or a stuck fault at the first time when the fault signal appears. The present invention uses a support vector machine to train and learn the collected training set data, thereby improving the processing capability of a large number of samples of accelerometer fault monitoring and improving the accuracy of accelerometer fault monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is the overall flow chart of accelerometer failure;

[0081] Figure 2 Flowchart for SVM decision tree training;

[0082] Figure 3 Flowchart for SVM decision tree classification. DETAILED DESCRIPTION

[0083] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0084] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0085] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0086] like Figure 1 As shown in the figure, the aircraft accelerometer fault diagnosis method is mainly divided into three stages: residual sequence solution, SVM training and residual sequence classification. When the switch in the figure is turned to "1", the SVM is trained offline, and when the switch is turned to "2", the residual sequence solution and residual sequence classification are performed online to realize the signal processing and monitoring of the accelerometer fault.

[0087] Residual sequence solution: The UAV flight control system has many small sensors that can realize the flight control of the UAV by outputting velocity and acceleration information. The three-axis accelerometers used in UAVs are usually MEMS sensors with poor accuracy. Generally, the attitude of the aircraft is solved by fusing the accelerometer and gyroscope data through a filter algorithm. The accelerometer output model a is as follows:

[0088]

[0089]

[0090] in, Represents the acceleration of the aircraft in the navigation coordinate system, g 0 represents the gravitational acceleration in the inertial system, The attitude quaternion q=(q 0 ,q 1 ,q2 ,q 3 ) T represents the attitude transfer matrix, η represents Gaussian white noise;

[0091] when When , we can get the approximate expression of formula (1)

[0092]

[0093] Normalize a in the above formula to get z, and discretize formula (3) to get

[0094] z k+1 =g(q k+1 )+δz k+1 (4)

[0095]

[0096] Among them, z k+1 represents the normalized, discrete a; δz k+1 represents the discretized Gaussian white noise η.

[0097] The quaternion differential equation is expressed as:

[0098]

[0099] Where w=(w x ,w y ,w z ) is the output of the three-axis gyroscope, q=((q 0 ,q 1 ,q 2 ,q 3 ) T is the attitude quaternion, Represents quaternion multiplication.

[0100] After further solving and discretization, we can get:

[0101] q k+1 =φ k+1 / k q k +Γ k+1 / k δθ k (7)

[0102] Taking equation (4) as the observation equation and equation (7) as the state equation, a nonlinear attitude estimation system with observation equation is formed as follows:

[0103]

[0104] Among them, q k Represents discrete attitude quaternion; the monitoring quantity is the normalized output z of the accelerometer k+1 .

[0105] The model is constructed based on three typical faults of accelerometer sensor: constant deviation, constant gain and stuck:

[0106] Constant deviation fault: It is manifested as the accelerometer starts to output a constant deviation at a certain moment. The sensor fault model is:

[0107]

[0108] Among them, α≠0 is a constant.

[0109] Constant gain fault: It is manifested as the accelerometer starting to output with a multiplier factor at a certain moment. The sensor fault model is:

[0110]

[0111] Among them, β≠1 is a certain constant gain proportional coefficient.

[0112] Stuck fault: The accelerometer starts to output a constant value at a certain moment. The sensor fault model is:

[0113]

[0114] Among them, γ≠0 is a constant.

[0115] The above various accelerometer failure forms can be uniformly expressed in the following form:

[0116]

[0117] Among them, β=1 and γ=0 do not hold simultaneously.

[0118] The residual is obtained by subtracting the actual output value of the accelerometer from the estimated output value of the acceleration. When the system works normally, the system output residual is as follows:

[0119]

[0120] Then the expectation of the residual is:

[0121]

[0122] When a system failure occurs, the output residual is as follows:

[0123]

[0124] Then the expectation of the residual is:

[0125]

[0126] However, β=1 and γ=0 do not hold simultaneously.

[0127] The present invention reflects whether the aircraft accelerometer is faulty by monitoring the residual error of the system, so the residual error e is used. k+1 Perform fault monitoring of aircraft accelerometers:

[0128] E(e k+1 )=0, the accelerometer has no fault.

[0129] E(e k+1 )≠0, the accelerometer is faulty.

[0130] In addition to detecting whether the accelerometer is faulty, it is also necessary to identify the type of fault based on the residual sequence. For different types of faults, the residual sequence will show different change trends. Therefore, the type of fault can be identified by analyzing the change trend of the residual sequence.

[0131] The invention estimates the state of the aircraft attitude estimation system by designing a filter with the best performance of the system residual sequence, namely the extended Kalman filter (EKF), and then substitutes the state into the observation equation to obtain the estimate of the accelerometer output value, and then subtracts the estimate of the accelerometer output value from the actual output to obtain the residual sequence of the system.

[0132] The present invention uses the EKF algorithm to estimate the accelerometer output. Since the state equation is a linear equation, only the observation equation needs to be linearized.

[0133] (1) Initialization

[0134] According to the actual situation, assign values ​​to the three-axis gyroscope noise covariance matrix Q and the three-axis accelerometer noise covariance R; and the state covariance matrix P 0 / 0 Assign the initial value to

[0135]

[0136]

[0137] (2) Time update

[0138] q k+1 / k =φ k+1 / k q k

[0139]

[0140]

[0141] (3) Observation update

[0142]

[0143] q k+1 / k+1 =q k+1 / k +K k (z k+1 -g(q k+1 / k ))P k+1 / k

[0144]

[0145] The invention uses EKF, which is essentially a generalization of Kalman filter in nonlinear systems, ignoring the influence of Taylor second-order and above expansions, achieving first-order estimation accuracy for nonlinear systems, thereby obtaining residual sequence data of the accelerometer.

[0146] The present invention introduces a support vector machine (SVM) to classify the system residuals, and performs accelerometer fault diagnosis by classifying the system residual sequence. The fault diagnosis method based on machine learning needs to be trained with the residual sequence of the system under one normal state and three fault states of the accelerometer, so that it can distinguish whether the system is faulty and the type of fault. The trained machine learning algorithm can then be applied to fault monitoring.

[0147] SVM training: SVM decision tree training adopts a bottom-up approach. The SVM decision tree classifies the accelerometer into four states: normal state, constant deviation fault, constant gain fault, and stuck fault. The residual sequence training sample set of the aircraft system under the four states of the accelerometer is {X 1 ,X 2 ,X 3 ,X 4}, each training set X i Contains j The training samples are Calculate the inter-class separation between each two states of the accelerometer {s 12 ,s 13 ,s 14 ,s 23 ,s 24 ,s 34}, from the inter-class separation s ij The worst two categories X i and X j First, train a binary SVM for X i and X j classify; then X i and X j Form a class X ij , and then take out the ijThe 4-1 is the two categories with the worst inter-class separation among the three categories. The binary SVM is trained to classify them. And so on until the binary SVM at the root node. The SVM decision tree divides the training set into 4 categories. The training process is represented as a flowchart as follows Figure 2 This training method effectively reduces the impact of cumulative errors.

[0148] Residual sequence classification: The classification process of the SVM decision tree is the opposite of the training process. When the SVM decision tree is used to classify the residual sequence of the aircraft system of unknown category, the SVM binary classifier at the root node is first used to classify its category, and then the sub-SVM binary classifier of the current node is used to classify its category downward according to the structure of the decision tree until the sample is classified into a certain category. The classification process is represented by the flowchart 3.

[0149] The present invention analyzes the faults of aircraft accelerometer sensors, builds a suitable monitoring model, selects monitoring signals, and classifies accelerometer faults using a machine learning method of a support vector machine. This allows effective monitoring of accelerometer faults, and allows timely and rapid determination of the type of fault when it occurs, so that remedial measures can be quickly taken to reduce the probability of safety accidents occurring in aircraft, thereby improving the safety and reliability of aircraft.

[0150] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for monitoring aircraft accelerometer faults based on machine learning, characterized in that: The following steps are involved: Step 1: Establish an output model of an accelerometer of an aircraft flight control system; Step 2: construct an accelerometer output model corresponding to the three faults of the accelerometer sensor, namely, constant deviation, constant gain and stuck; Step 3: Use the EKF algorithm to estimate the output of the accelerometer output model, and then make a difference between the estimated accelerometer output value and the actual output to obtain the residual sequence of the system; Step 4: Use the residual sequence of the system under the normal state and three fault states of the accelerometer to train the SVM decision tree; Step 5: Use the trained SVM decision tree to classify the residual sequence of the unknown category of the aircraft system, and monitor the three types of faults of the accelerometer sensor according to the classification results of the residual sequence; In step 3, the state of the aircraft attitude estimation system is estimated by EKF, and then the state is substituted into the observation equation to obtain the estimate of the accelerometer output value, and then the estimate of the accelerometer output value is subtracted from the actual output to obtain the residual sequence of the system; In step 3, the state of the aircraft attitude estimation system is estimated by EKF, specifically: (301) Initialization Assign values ​​to the three-axis gyroscope noise covariance matrix Q and the three-axis accelerometer noise covariance R according to actual conditions; and the state covariance matrix P 0 / 0 The initial value is: (302) Time Update what k+1 / k =φ k+1 / k what k (303) Observation Update q k+1 / k+1 =q k+1 / k +K k (z k+1 -g(q k+1 / k ))P k+1 / k The classification process of step 5 is: When the SVM decision tree is used to classify an aircraft system residual sequence of unknown category, the SVM binary classifier at the root node is used to classify the category of the aircraft system residual sequence, and then the sub-SVM binary classifier of the current node is used to classify the category of the aircraft system residual sequence downward according to the structure of the decision tree until it is classified into a certain category.

2. The method for monitoring aircraft accelerometer faults based on machine learning method according to claim 1, characterized in that: Step 1 is as follows: The established accelerometer output model a is: in, represents the acceleration of the aircraft in the navigation coordinate system, g0 represents the gravity acceleration in the inertial system, g0=(0,0,|g0|), The attitude quaternion q = (q0, q1, q2, q3) T represents the attitude transfer matrix, η represents Gaussian white noise; when When , the approximate expression of formula (1) is: Normalize a in (3) to get z, and discretize (3) to get z k+1 =g(q k+1 )+δz k+1 (4) Among them, z k+1 represents the normalized, discrete a; δz k+1 represents the discretized Gaussian white noise η; The differential equation for the quaternion q is: Where w=(w x ,w y ,w z ) is the output of the three-axis gyroscope, q=((q0,q1,q2,q3) T is the attitude quaternion, Represents quaternion multiplication; After further solving and discretization, we get: q k+1 =φ k+1 / k q k +C k+1 / k dth k (7) Taking equation (4) as the observation equation and equation (7) as the state equation, a nonlinear attitude estimation system with observation equation is formed as equation (8): Among them, q k Represents discrete attitude quaternion; z k+1 is the monitoring quantity, i.e. the normalized output of the accelerometer; φ k+1 / k and Γ k+1 / k are the relative values ​​of the two measurements of the accelerometer before and after; δ is θ k The normalization coefficient of .

3. The method for monitoring aircraft accelerometer faults based on machine learning method according to claim 2, characterized in that: Step 2 builds a model based on the three faults of accelerometer sensor: constant deviation, constant gain and stuck. Specifically: The sensor constant deviation fault is manifested as the accelerometer outputting a constant deviation at a certain moment. The corresponding fault model is: Where a is a constant and a≠0; The constant gain fault is manifested as the accelerometer starting to output a multiplier factor at a certain moment. The corresponding fault model is: Where β is the constant gain proportional coefficient and β≠1; The stuck fault is manifested as the accelerometer output maintaining a constant value at a certain moment, and the corresponding fault model is: Among them, γ is a constant and γ≠0; The above various accelerometer failure forms are uniformly expressed in the following form: Among them, β=1, γ=0 are not true at the same time; When the system works normally, the system output residual is: The corresponding residual expectation is: When a system failure occurs, the system output residual is: The corresponding residual expectation is: However, β=1 and γ=0 do not hold simultaneously.

4. The method for monitoring aircraft accelerometer faults based on machine learning method according to claim 3, characterized in that: When E(e k+1 )=0, the accelerometer has no fault. When E(e k+1 )≠0, the accelerometer is faulty.

5. The method for monitoring aircraft accelerometer faults based on machine learning method according to claim 1, characterized in that: In step 4, the training of the SVM decision tree is carried out from bottom to top.

6. The method for monitoring aircraft accelerometer faults based on machine learning method according to claim 5, characterized in that: The training process is: (401) The residual sequence training sample set of the aircraft system under the three fault states and normal state of the accelerometer is {X1, X2, X3, X4}, and each training set X i Contains j The training samples are (402), calculate the inter-class separation degree {s 12 ,s 13 ,s 14 ,s 23 ,s 24 ,s 34 }, select the inter-class separation s ij The smallest two-class training set trains the binary classification SVM; (403), merging the inter-class separation degree s ij The smallest two-category training set is one category; (404), looping steps (402) and (403) until the binary classification SVM at the root node.

Images