CN116244837B - Flap fault sensing method and system - Google Patents

Abstract

A flap fault perception method and system of the present invention belong to the field of automatic detection; the method steps include: monitoring flap overrun events, constructing a measurement data set of parameters to be monitored; preprocessing the measurement data, and obtaining threshold samples for judging faults Data; based on the preprocessed measured value data and threshold sample data, construct the GRU neural network trailing edge flap performance evaluation model; save the trained GRU neural network trailing edge flap performance evaluation model, and monitor the parameters during the data prediction stage Predict the time series data and get the fault prediction result. The invention reduces the error in the calculation process of the fault data, and makes accurate fault perception through learning and training of the fault data, and solves the problem that the fault response cannot be made in time in the prior art.

Description

A flap fault sensing method and system

Technical Field

The invention belongs to the field of automatic detection, and in particular relates to a flap fault sensing method and system.

Background Art

The high lift of modern large aircraft is provided by the flaps located at the trailing edge of the wing. During low-speed stages such as takeoff and landing, the trailing edge flaps are extended outward and bent downward to increase the wing area, change the configuration and provide lift to ensure a reasonable taxiing distance and safe takeoff speed, while improving the aircraft's climb rate, approach speed and approach attitude.

Common flap failures currently include:

(1) Actuator disengagement and wing tilt. One of the actuators of a single wing or the hinge connected to the body is stuck, and it tilts due to external force, or one of the actuators itself is jammed or freewheeling, while at the same time another actuator of the wing is still driving the wing to continue moving.

(2) Airfoil asymmetry: A single airfoil does not move synchronously with other airfoils. This may be a secondary fault of airfoil tilt.

(3) Non-command of the wing surface, that is, the actual position of the wing surface is inconsistent with the command of the flap and slat handle.

If one or more of these failures occur during takeoff or landing, the aircraft structure may be seriously damaged or even crash. The current aircraft operating system will only issue a fault alarm when a fault occurs, which is used to prompt the pilot to make the next decision, but there is no advance prediction of the occurrence of the fault, and it cannot fundamentally solve the problems of flap jamming, imbalance on both sides, and angle inaccuracy. If the pilot makes an operational error, there are still major safety hazards.

Therefore, there is a need in the art for an improved method and device for flap fault sensing.

Summary of the invention

Technical issues to be solved:

In order to avoid the shortcomings of the prior art, the present invention provides a flap fault perception method and system, which reduces the dimension of the collected data through the principal component analysis method to reduce the influence of variables on the prediction results; at the same time, when the aircraft is flying, the GRU neural network intelligent algorithm is used to automatically detect the flap status, and an early warning is given before the fault occurs. When a fault occurs, the pilot is alarmed, and the backup plan is activated to operate the flap to ensure flight safety. The present invention reduces the error in the fault data calculation process, and makes accurate fault perception by learning and training the fault data, solving the problem that the prior art cannot respond to faults in a timely manner.

The technical solution of the present invention is: a flap fault sensing method, the specific steps are as follows:

Step 1: flap over-limit event monitoring, constructing a measurement data set of the parameters to be monitored;

Step 2: pre-process the measurement data obtained in step 1 to obtain threshold sample data for fault judgment;

S2.1, standardize the measurement data;

S2.2, calculate the covariance matrix based on the standardized data;

S2.3, calculate the eigenvalues and eigenvectors of the matrix;

S2.4, determine the principal components according to the component contribution rate for all features, that is, complete the dimensionality reduction processing of the measurement data;

S2.5, determine the principal component space and the residual space, and then determine the sample data of the threshold;

Step 3: Based on the preprocessed measurement data and threshold sample data, a GRU neural network trailing edge flap performance evaluation model is constructed;

Step 4: Save the trained GRU neural network trailing edge flap performance evaluation model, predict the time series data of the monitoring parameters in the prediction data stage, and obtain the fault prediction results.

A further technical solution of the present invention is: the flap over-limit event monitoring method in step 1 is:

S1.1: Determine the out-of-limit events and corresponding monitoring parameters, and confirm that the sensors function normally and the data is correct;

S1.2: Determine the threshold for limit-exceeding events;

S1.3: Compile messages and decode data for over-limit events;

S1.4: Message data processing: Import QAR data for decoding. When decoding is completed, each message will generate a corresponding TXT file.

A further technical solution of the present invention is: when constructing the measurement data set of the parameter to be monitored in step 1, in order to eliminate the accidental influence of the flight load on the opening time, the trailing edge flap opening time is preprocessed by a sliding window method; assuming that within a certain period of time, the trailing edge flap opening time series is:

in, Corresponding to the 1st, 2nd, 3rd and jth trailing edge flap opening time in the trailing edge flap opening time series;

Take a sliding window with a width of j for the open time series, where j represents the number of data in a sliding window, and then take the average of j consecutive data in the window: the average width of the sliding window is X _k , k=1…j; It refers to the kth trailing edge flap opening time in a sliding window;

In the next iteration, the starting position of the sliding window is shifted by a set step size, that is, the width j, and the mean of the data in the window is calculated; the sliding continues until the end position of the window is the last data; by setting a reasonable sliding window forward step size, the influence of random factors on the parameters is eliminated, and the performance of the trailing edge flap is more accurately reflected.

A further technical solution of the present invention is: the parameters to be monitored include flap handle gear information FLAPCONFIGURATION, left trailing edge flap position sensor angle TE FLAP POSN LT, right trailing edge flap position sensor angle TE FLAP POSN RT; using the measured values after the flaps are opened at three gears: 0, 5, and 30 for analysis;

Extract characteristic values from the measured data of the parameter to be monitored, subtract the baseline value from the data on the fault side to obtain the difference CZ; extract the mean, variance, standard deviation, maximum value, maximum ten-digit average, minimum value, minimum ten-digit average, skewness, kurtosis, lower bound value of outliers, upper bound value of outliers, lower bound value of extreme values, upper bound value of extreme values, mean square value, and margin for the difference CZ, a total of 15 characteristic values; obtain a vector consisting of 15 characteristic values for the 0, 5, and 30 gears of each flight.

A further technical solution of the present invention is: the specific steps in step 2 are as follows:

S2.1, standardize the measurement data;

Assume x ₁ , x ₂ , …, x _N are the data extracted from the features, each of which is 15-dimensional data. The data is standardized to eliminate the influence of dimension and the influence of the variable's own variation and value. The standardization method is to subtract the average value of the corresponding dimension from the data of each dimension. , and then divided by the variance of the corresponding dimension , get the standardized data , where N represents the number of data;

(2.3)

Wherein, X _pq is the data of the qth dimension in the pth data;

S2.2, calculate the covariance matrix based on the standardized data;

Based on the data obtained by standardization , the calculated covariance matrix is:

Where, the matrix elements r _uv are the correlation coefficients of the original variables x _u and x _v , u, v = 1, 2, …, m;

S2.3, calculate the eigenvalues and eigenvectors of the matrix;

The characteristic equation of the matrix in step S2.2 is: |λI-R|=0, where I represents an m-order identity matrix with 1 on the diagonal and 0 on the rest;

First, the solution λ _i of the eigenvalue is obtained according to the conversion rule; then, the eigenvalues are arranged in order of λ ₁ ≥λ ₂ ≥…≥λ _m ≥0; then, the eigenvector e _i corresponding to each eigenvalue λ _i is calculated, i = 1, 2, …, m, and ||e _i || = 1, that is, , where e _ij represents the j-th component of the eigenvector e _i ;

S2.4, determine the principal component according to the component contribution rate for all features, that is, complete the dimensionality reduction processing of the measured data, and determine the first r eigenvalues as the principal component;

The contribution of the ingredients is: ;

S2.5, based on the principal components determined in S2.4, determine the principal component space and the residual space, and then determine the sample data of the threshold;

Arrange the eigenvectors corresponding to the r eigenvalues in rows from top to bottom and get Matrix; by calculation and The product of will get the principal component subspace projection matrix C; further calculation will get the residual subspace projection matrix After principal component analysis, the original data is expressed by the following formula: The input fault sample data is divided into two parts:

in, The dimensional space composed is a subspace of the principal component. The composed dimensional space is a subspace of the residual; yes Projection on the PCS principal component subspace; yes Projection on the RS residual subspace, C and is the corresponding projection matrix;

The fault sample data is divided into the principal component indicating the normal value and the residual component indicating the measurement fault by the principal component analysis method; under normal conditions, It is mainly used to measure noise, so in fault mode, The value will be significantly larger;

The calculation formula of the variable monitoring SPE statistic acting on the residual subspace RS is:

When the monitored object is in normal working condition, The projection in the residual subspace is within a normal range. If a fault occurs, the projection of the monitored variable in the residual subspace will change, causing the SPE statistic to exceed the threshold. :

in, is a standard normal distribution Confidence.

A further technical solution of the present invention is: in step 3, a GRU neural network trailing edge flap performance evaluation model is constructed The specific steps are:

in, represents the update gate, Represents the reset gate, represents the input of the current neuron, represents the output of the current neuron, represents the output of the previous neuron, Represents the pending output value of the current neuron, represents the weight of the update gate, represents the weight of the reset gate, represents the output weight, represents the weight of the pending output, Represents the sigmoid function.

The further technical solution of the present invention is: S3.1, the GRU neural network merges the forget gate and the exit gate in the LSTM into an update gate, and optimizes the combination with the output gate into a cell structure of two "gates"; when the update gate The smaller the value, the less information the current neuron needs to retain, and the more information the previous neuron needs to retain. According to the process of each "gate" in the GRU neural network, each output information is affected by each neuron, so each neuron is interdependent. Under normal circumstances, it is more active in the reset gate for short-distance learning and in the update gate for long-distance learning.

A further technical solution of the present invention is: the specific steps of the verification method of the GRU neural network trailing edge flap performance evaluation model are as follows:

First, a GRU neural network model with 6 hidden units is established, that is, the number of hidden neurons is 6;

Then, through the formula derivation, it is concluded that the weight parameter to be learned is , where the first three weights are concatenated, so they must be separated during learning, i.e.

in, Reset the weights between the gate and the input of the current neuron for the current neuron, Reset the weights between the gate and the output of the previous neuron for the current neuron, Update the weight between the gate and the input of the neuron in power for the current neuron update gate, Update the weight between the gate of the current neuron and the previous neuron, is the weight between the current neuron output and input value, The current neuron output and the previous neuron output;

Input to the output layer , the output is ;

The loss function in neural network training is defined as:

In the formula, represents the loss value, Represents the output value of each neuron, Represents the output of the output layer, that is, the real original data;

Finally, the output value of the neuron Subtract the true value , then square it and average it to get the variance; the variance obtained is the loss value of the loss function; the loss function is a function used to evaluate the gap between the neural network target and the actual output. The smaller the function value, the smaller the difference between the actual data and the target output, which means that the weight is more appropriate.

A further technical solution of the present invention is: using the gradient descent method to optimize the objective function First, define the learning rate of the gradient descent method and optimize the descent gradient of the training model each time; then use the optimizer in Tensor Flow to optimize the loss function. The optimizer used is Adam Optimizer.

A flap fault sensing method implementation system includes a data acquisition module, a host computer, a main drive module, an auxiliary drive module, and an alarm module;

The data acquisition module is a rotary variable differential transformer installed at the trailing edge flap position, which is a linear sensor capable of collecting continuous position signals;

The host computer collects data from the data acquisition module in real time, performs fault prediction through the aircraft flap fault perception method, and generates control instructions;

After receiving the action instruction from the host computer, the main driving module drives the flap to make attitude adjustment;

The auxiliary drive module is controlled by the host computer. When monitoring determines that a fault occurs, the alarm module is triggered, and the host computer starts the auxiliary drive module to push the flaps outward to stabilize the fuselage attitude.

Beneficial Effects

The beneficial effects of the present invention are as follows: The present invention proposes a flap fault perception method and system, which utilizes a GRU neural network to perform fault diagnosis, evaluate the health status of the flap, and ensure the flight safety of the aircraft by turning on the auxiliary flap drive device when a fault occurs. This solves the problem that the traditional aircraft flaps only alarm when a fault occurs and cannot track the health status of the flaps throughout their life cycle, allowing the crew and maintenance personnel to grasp the flap status in a timely manner, thereby ensuring the flight safety of the aircraft to the greatest extent.

Furthermore, the flap data was used as the basis for the flight QAR data and the flap over-limit event criteria were developed by consulting the AMM manual.

Furthermore, principal component analysis is used to reduce the dimensionality of various parameters of the flap sensor and find the key characteristic parameters to prepare for the GRU neural network training.

Furthermore, by training with the GRU neural network, the predicted data of flap operation can be obtained, which has a high accuracy compared with the real data.

Furthermore, a backup auxiliary drive module is provided for the flap. When the main drive module is stuck and cannot drive the bearing to operate, the backup auxiliary drive device can be started to complete the deflection of the flap.

The GRU neural network model designed by the present invention is simple, and the training parameters are fewer than those of the LSTM network, so it is more suitable for networks with large data volumes. The gradient can be effectively suppressed from disappearing or exploding during training, the problem of local optimality in training is solved, and good diagnostic accuracy can be obtained. By using the BUS communication bus owned by the flight control computer, the communication speed can reach 100kbps, the sensor data is stripped, and the data is passed into the trained neural network for calculation, and the response time is about 200ms, which is fast.

From the above content, it can be seen that the present invention has established a flap fault perception method and system, which uses neural networks for parameter learning and can sense the occurrence of flap failures. At the same time, a backup flap drive device is equipped to ensure the flight safety of the aircraft, with a simple structure and high economic benefits.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of principal component analysis fault monitoring;

FIG2 is a schematic diagram of a flap rail device;

Figure 3 is a schematic diagram of the SPE statistic of the residual space projection of the test sample (5 levels);

Explanation of reference numerals: 1 is the inner slide rail of the wing, 2 is the connecting rod, 3 is the side slide rail of the flap, and 4 is the flap.

DETAILED DESCRIPTION

The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.

This embodiment provides a flap fault perception method and system, which reduces the dimension of the collected data through the principal component analysis method to reduce the influence of variables on the prediction results; at the same time, when the aircraft is flying, the GRU neural network intelligent algorithm is used to automatically detect the flap status, and an early warning is given before the fault occurs. When a fault occurs, the pilot is alarmed and the backup plan is activated to operate the flap to ensure flight safety. The present invention reduces the error in the fault data calculation process, and makes accurate fault perception through learning and training of fault data, solving the problem of the inability to respond to faults in a timely manner in the prior art. The specific steps of the method are as follows:

1. Flap Overlimit Event Monitoring

1.1 Determine the out-of-limit events and corresponding monitoring parameters, and confirm that the sensor functions normally and the data is correct. Combine the AMM manual, FIM manual and actual project needs to determine the out-of-limit events that need to be monitored. Apply the sensor fault diagnosis method to analyze the sensor performance to ensure that the sensor performance is normal and the data is accurate.

1.2 Determine the threshold of the over-limit event. There are three main methods: (a) Query the AMM manual or other relevant manuals to determine the event threshold. (b) Determine the event threshold by training a large amount of historical data. (c) Determine the event threshold by analogy with similar parameters of other models. In this embodiment, the threshold is determined by the AMM manual.

1.3 Compile messages and decode data for over-limit events. Use Airfase software to compile messages and realize automatic data extraction. Compiling messages has the following two main contents: (a) Determine the message content: including the parameters contained in the message and the time interval for parameter extraction. (b) Determine the message trigger logic: Determine the message trigger time by superimposing the restrictions of various parameter conditions.

1.4 Message data processing Import QAR data for decoding. When decoding is completed, each message will generate a corresponding TXT file.

The trailing edge flap position sensor is located at the torque tube of the left and right trailing edge flaps. Its function is to record the position information of the trailing edge flaps and feed the information back to the FSEU flap aileron electronic assembly. The trailing edge flap position sensor is a rotary variable differential transformer, a linear sensor that can transmit continuous position signals. During the takeoff or landing process of the aircraft, in order to improve the flight performance of the aircraft, the trailing edge flaps are retracted and extended. The retraction and extension of the flaps drives the sensor input rod to move, and the spoiler/iron core connected to the input rod moves accordingly. The induced voltage/inductance in the coil changes, generating a voltage/current signal proportional to the rotation angle. The trailing edge flap angle is obtained by solving the voltage/current signal. When the aircraft is flying, the flight computer collects QAR (Quick Access Recorder) data in real time. After analysis, the following parameters can be used for trailing edge flap position sensor fault analysis: flap handle gear information (FLAP CONFIGURATION), left trailing edge flap position sensor angle (TEFLAP POSN LT), right trailing edge flap position sensor angle (TE FLAP POSN RT), see Table 1.1.

Table 1.1 Parameter extraction results

In a flight cycle, the trailing edge flaps are operated as follows: during takeoff, the trailing edge flaps are opened and maintained at position 5; during flight, the trailing edge flaps are retracted and maintained at position 0; during landing, the trailing edge flaps are opened and maintained at 30 (under special conditions, they need to be opened and maintained at 40). The trailing edge flaps are in positions 0, 5, and 30 for a long time in a flight cycle, and the data samples are richer, so this embodiment extracts the three parameters again according to the gear position.

When the flaps are changed from 0, 5, 30 to other positions, the angle value is continuously changing. When performing data analysis, the non-steady-state data should be eliminated, and the values after the flaps are opened in the three positions should be used for analysis. Taking the data of position 0 as an example, the data finally used for analysis is shown in Table 1.2.

Table 1.2 Trailing edge flap sensor 0 gear data

2. Data Preprocessing

In order to eliminate the accidental influence of flight load on the opening time, the sliding window method is used to preprocess the opening time of the trailing edge flap. Assume that within a certain period of time, the opening time series of the trailing edge flap is:

in, Corresponding to the 1st, 2nd, 3rd and jth trailing edge flap opening time in the trailing edge flap opening time series;

Take a sliding window with a width of j for the open time series, where j represents the number of data in a sliding window, and then take the average of j consecutive data in the window: the average width of the sliding window is X _k , k=1…j; It refers to the kth trailing edge flap opening time in a sliding window;

In the next iteration, the starting position of the sliding window is moved forward by one step, the sliding window size remains unchanged, the mean of the data in the window is calculated, and the sliding continues until the end position of the window is exactly the last data. By selecting a reasonable sliding window width j (i.e., step size), the influence of some random factors on the parameters can be eliminated, and the performance of the trailing edge flap can be more accurately reflected.

For the sensor data of the 0, 5, and 30 gears of each flight, 200 seconds of data are selected for the 0 gear, 100 seconds of data for the 5 gear, and 60 seconds of data for the 30 gear. The baseline value is subtracted from the data on the fault side to obtain the difference CZ. For the difference CZ, the mean, variance, standard deviation, maximum value, maximum ten-digit average, minimum value, minimum ten-digit average, skewness, kurtosis, lower bound of abnormal value, upper bound of abnormal value, lower bound of extreme value, upper bound of extreme value, mean square value, and margin are extracted, a total of 15 eigenvalues. The extraction formula is shown in Table 2.1. A vector consisting of 15 eigenvalues can be obtained for the 0, 5, and 30 gears of each flight.

Table 2.1 Eigenvalues

The principal component analysis method can be used to reduce the dimension of the flap sensor data eigenvalues and find the most effective parameters for fault diagnosis.

Principal Component Analysis (PCA) is a statistical method. Through orthogonal transformation, a set of a-dimensional variables that may be correlated is converted into a set of linearly uncorrelated b-dimensional variables (a < b). In the research and application of many fields, it is often necessary to observe multiple variables that reflect things and collect a large amount of data in order to analyze and find patterns. On the other hand, there may be correlations between many variables, which increases the complexity of problem analysis. Therefore, it is necessary to find a reasonable method to reduce the indicators that need to be analyzed while minimizing the loss of information contained in the original indicators. The principal component analysis method belongs to this type of dimensionality reduction method.

Step 1: Data Standardization

Assume x ₁ , x ₂ , …, x _N are the data extracted from the features, each of which is 15-dimensional data. The data is standardized to eliminate the influence of dimension and the influence of the variable's own variation and value. The standardization method is to subtract the average value of the corresponding dimension from the data of each dimension. , and then divided by the variance of the corresponding dimension , get the standardized data , where N represents the number of data;

(2.3)

Wherein, X _pq is the data of the qth dimension in the pth data;

Step 2: Calculate the covariance matrix of the data.

Based on the data obtained by standardization , the calculated covariance matrix is:

The matrix elements r _uv (u, v = 1, 2, …, m) in the above formula are the correlation coefficients of the original variables _xu and _xv .

Step 3: Calculate the matrix eigenvalues and eigenvectors.

|λI-R|=0 is the characteristic equation of the above matrix, and I represents an m-order unit matrix with 1 on the diagonal and 0 on the rest.

According to the conversion rule, the solution of the eigenvalue λ _i is obtained, and then the order of the eigenvalue is λ ₁ ≥λ ₂ ≥…≥λ _m ≥0. For each eigenvalue λ _i mentioned above, the corresponding eigenvector e _i ( i =1, 2, …, m) is calculated, and ||e _i || =1, that is, , where e _ij represents the j-th component of the eigenvector e _i .

Step 4: Determine the principal component according to the component contribution rate for all features, that is, complete the dimensionality reduction processing of the measured data and determine the first r eigenvalues as the principal component;

The contribution of the components is ;

There are m features in total (that is, the 15 features used in the initial feature extraction). After listing each eigenvalue λ ₁ ≥λ ₂ ≥…≥λ _m ≥0, find the first r features out of the m features and ensure that the sum of the first r eigenvalues accounts for 85% of all features. These r features are considered to be the main components.

(λ ₁₊ λ ₂ +…+λ _r )/(λ ₁₊ λ _2+…+ λ _m )≥0.85

Step 5: Determine the principal component space and residual space.

Arrange the eigenvectors corresponding to the r eigenvalues in rows from top to bottom and get Matrix; by calculation and The product of will get the principal component subspace projection matrix C; further calculation will get the residual subspace projection matrix After principal component analysis, the original data is expressed by the following formula: The input fault sample data is divided into two parts:

in, The dimensional space composed is a subspace of the principal component. The composed dimensional space is a subspace of the residual; yes Projection on the PCS principal component subspace; yes Projection on the RS residual subspace, C and is the corresponding projection matrix;

The fault sample data is divided into the principal component indicating the normal value and the residual component indicating the measurement fault by the principal component analysis method; under normal conditions, It is mainly used to measure noise, so in fault mode, The value will increase significantly.

The principal component analysis algorithm mainly monitors and diagnoses faults through statistical means. There are two crucial statistics in the statistical process. The first is the T2 statistic used for variable testing, which mainly acts on the principal component subspace; the second is the SPE statistic used for variable monitoring ^, which mainly acts on the residual subspace RS .

The SPE statistic calculation formula is:

When the monitored object is in normal working condition, The projection in the residual subspace is within a normal range. If a fault occurs, the projection of the monitored variable in the residual subspace will change, causing the SPE statistic to exceed the threshold. :

in, is a standard normal distribution Confidence.

Taking the Boeing 737-NG trailing edge flap as an example, the trailing edge flap position indication system consists of a flap position indicator, left and right flap position sensors, and FSEU; the flap position sensor (FPS) and position indicator (FPI) form a linkage mechanism, and defects in any component will affect the accuracy of the flap position indication.

There is a synchronizer in both the flap position sensor and the indicator. The synchronizer coil is powered by the same 28VAC power supply to ensure that the phase of the changing magnetic field generated by the coil is consistent. The three groups of coils arranged at 120 degrees generate voltage in the changing magnetic field. The three groups of coils of the sensor and the three groups of coils of the indicator are connected separately. When the sensor coil deflects an angle under the flap drive, the angle of the coil cutting the magnetic field changes, causing the voltage to change. When the angle of the coil in the indicator is inconsistent with the angle of the sensor coil, the voltage on the coil is unbalanced and a current is generated. The indicator is equivalent to a motor rotating. When the indicator angle rotates to the same angle as the sensor coil, the voltage generated and the voltage generated by the sensor reach a balanced state, the current disappears, and the indicator stops rotating. When the flap angle changes, the angle of the sensor and indicator coil cutting the magnetic field also changes, and the generated induced voltage must change. The FSEU detects the voltage on the connection line between the sensor and the indicator to determine the flap angle. Therefore, the FSEU does not participate in the calculation of the flap position indication, and transmits the detected signal to other systems for asymmetric protection. FSEU compares the detected flap position signals on both sides. When the difference exceeds 9 degrees and exceeds 0.5 seconds, the flap bypass protection is triggered, and the flap cannot be operated by the handle. A 15-degree scissor difference appears on the cockpit flap indicator. When one side of the flap sensor itself fails, the synchronizer coil output voltage is abnormal, and there will be a difference between the balance voltage required by the corresponding synchronizer coil in the indicator and the actual flap position. The crew will see the flap pointer in the cockpit deviate from the pointer angle on the normal side. Similarly, if the indicator is blocked for some reason, or the synchronizer coil fails, the force generated by the current in the coil in the magnetic field cannot drive the synchronizer coil to deflect the corresponding angle, and the effect seen in the cockpit is that the two flap pointers have a scissor difference.

3. Constructing the GRU neural network trailing edge flap performance evaluation model

S3.1 Neural Network Gated Recurrent Unit (GRU) is an improved model based on Long Short-Term Memory (LSTM) neural network. GRU network inherits the ability of LSTM network to process and predict time series, and optimizes the network structure. GRU neural network merges the forget gate and the exit gate in LSTM into an update gate, and optimizes the combination with the output gate into a cell structure of two "gates".

From Figure 3, we can see how the GRU model propagates forward. The calculation formula is as follows:

in, represents the update gate, Represents the reset gate, represents the input of the current neuron, represents the output of the current neuron, represents the output of the previous neuron, Represents the pending output value of the current neuron, represents the weight of the update gate, represents the weight of the reset gate, represents the output weight, represents the weight of the pending output, Represents the sigmoid function.

When updating the gate The smaller the value, the less information the current neuron needs to retain, and the more information the previous neuron needs to retain. According to the process of each "gate" in the above GRU neural network, it can be seen that the information of each output is affected by each neuron, so each neuron is interdependent. Under normal circumstances, it is more active in the reset gate for short-distance learning and more active in the update gate for long-distance learning.

S3.2 GRU training of the GRU neural network trailing edge flap performance evaluation model

First, a GRU neural network model with 6 hidden units is established, that is, the number of hidden neurons is 6.

Then, through the formula derivation, it is concluded that the weight parameter to be learned is , where the first three weights are concatenated, so they must be separated during learning, i.e.

in, Reset the weights between the gate and the input of the current neuron for the current neuron, Reset the weights between the gate and the output of the previous neuron for the current neuron, Update the weight between the gate and the input of the neuron in power for the current neuron. Update the weight between the gate of the current neuron and the previous neuron, is the weight between the current neuron output and input value, The current neuron output and the previous neuron output;

Input to the output layer , the output is ;

The loss function in neural network training is defined as:

In the formula, represents the loss value, Represents the output value of each neuron, Represents the output of the output layer, that is, the real original data; the output value of the neuron Subtract the true value , then square it, and finally average it to get the variance. The obtained variance is the loss value of the loss function. The loss function is a function used to evaluate the gap between the neural network target and the actual output. The smaller the function value, the smaller the difference between the actual data and the target output, which means that the weight is more appropriate.

Finally, the objective function is optimized using the gradient descent method , then it is necessary to define the learning rate of the gradient descent method, and perform the descent gradient of each optimization training model. The learning rate selected in this article is 0.004. This article directly uses the optimizer in Tensor Flow to optimize the loss function, and the optimizer used is Adam Optimizer. Common optimization algorithms include: Stochastic Gradient Descent (SGD) algorithm, Adagrad algorithm, and Adam algorithm. The stochastic gradient descent algorithm calculates the gradient of each batch for each iteration and then updates the parameters. It is also the most commonly used optimization algorithm in the optimizer. SGD is simple in form and has been widely used.

4. Save the model trained by the GRU neural network. In the subsequent data prediction stage, the trained model needs to be called to predict the time series data and obtain the prediction results.

This embodiment uses the GRU neural network to perform fault diagnosis and evaluate the health status of the flaps. In the event of a fault, the backup flap drive device is activated to ensure the flight safety of the aircraft. This solves the problem that the traditional aircraft flaps only alarm when a fault occurs and cannot track the health status of the flaps throughout their life cycle. It allows the crew and maintenance personnel to understand the flap status in a timely manner, thus ensuring the flight safety of the aircraft to the greatest extent.

Example:

Referring to Figure 1, take the trailing edge flap position sensor failure of B5631 on March 13, 2016 as an example.

(1) Training sample selection

During the flight, the left trailing edge flap lagged behind the rear flap in gear position 5. The existing sample data spans from January 1, 2016 to November 15, 2016, with a total of 1,194 flights. One hundred flights were selected as training samples. In each flight, 0 bits selected 200 seconds of sensor data on both sides, 5 bits selected 100 seconds of data, and 30 bits selected 60 seconds of data. 13 feature values such as the mean and maximum value of the CZ data of each gear position in each flight were extracted to form the training samples. Taking some feature values of the first 6 flights in gear position 5 as an example, the training samples are shown in Table 1.4.

Table 1.4 B3561 principal component analysis training samples (5 levels)

(2) Data standardization

The B5631 training sample data is standardized, and the standardized data is shown in Table 1.5, taking some characteristic values of the first 6 flights in 5 gears as an example.

Table 1.5 B3561 training sample standardization results (5 levels)

(3) Learning and training

After standardizing the training sample consisting of 100 flights, the covariance matrix of the sample is calculated , the covariance matrix is a 13×13 matrix, see the table below.

Table 1.6 B3561 training sample covariance matrix

1 0.62 0.45 0.89 -0.17 -0.98 0.19 0.59 0.69 0.51 0.24 0.39 0.05 0.63 1 0.16 0.52 -0.46 -0.57 0.41 0.22 0.93 0.03 0.48 -0.06 0.37 0.45 0.16 1 0.31 0.61 -0.52 -0.54 0.83 0.18 0.39 -0.11 0.36 -0.19 0.89 0.52 0.31 1 -0.28 -0.87 0.33 0.42 0.57 0.39 0.25 0.29 -0.19 -0.18 -0.46 0.62 -0.28 1 0.04 -0.98 0.62 -0.49 0.38 -0.62 0.45 -0.61 -0.99 -0.57 -0.52 -0.87 0.04 1 -0.06 -0.68 -0.62 -0.60 -0.12 -0.49 0.06 0.19 0.41 -0.54 0.33 -0.98 -0.06 1 -0.61 0.44 -0.38 0.63 -0.45 0.61 0.59 0.22 0.84 0.42 0.62 -0.68 -0.61 1 0.24 0.53 -0.15 0.48 -0.26 0.69 0.93 0.18 0.57 -0.49 -0.62 0.44 0.24 1 0.03 0.53 -0.07 0.41 0.51 0.03 0.39 0.39 0.38 -0.61 -0.38 0.53 0.03 1 -0.69 0.99 -0.81 0.23 0.48 -0.11 -0.11 -0.62 -0.12 0.63 -0.15 0.53 -0.69 1 -0.78 0.98 0.39 -0.06 0.36 0.36 0.45 -0.49 0.45 0.48 -0.07 0.99 -0.78 1 -0.88 0.05 0.37 -0.19 -0.19 -0.61 0.06 0.61 -0.26 0.41 -0.81 0.98 -0.88 1

At the same time, calculate The eigenvalue d and eigenvector v of , and the eigenvalue d are arranged in descending order as follows: 5.3604, 4.9017, 1.6887, 0.6312, 0.2082, 0.1204, 0.0594, 0.0159, 0.0130, 0.0007, 3.6e-05, 1.3e-15, and 5.36e-16.

The principal component space is determined with a contribution rate of 0.99, and the eigenvectors corresponding to the eigenvalues constitute the projection matrix C of the principal component space. Correspondingly, the eigenvectors corresponding to the eigenvalues with smaller contributions constitute the projection matrix of the residual space The corresponding eigenvalues in the principal component space are 5.3604, 4.9017, 1.6887, 0.6312, 0.2082, and 0.1204, and the principal component space projection matrix C composed of the corresponding vectors is shown in Table 1.7.

Table 1.7 Principal component space projection matrix of B3561 training samples

-0.1554 -0.4109 0.0537 -0.1964 0.1391 -0.2444 0.05655 -0.3797 -0.0192 0.6194 -0.0175 0.3007 -0.2832 -0.1342 -0.4516 -0.0319 -0.8018 -0.1436 -0.0998 -0.3834 0.1337 -0.4519 -0.0938 0.7241 -0.3249 0.2061 -0.3616 0.0042 0.0823 0.1196 0.2089 0.3843 -0.0336 0.1891 -0.1889 0.2311 0.3184 -0.2073 0.3417 -0.157 0.3559 -0.131 -0.3269 -0.1736 -0.3778 -0.0321 -0.3647 0.0684 0.0592 -0.4012 -0.0221 0.5006 0.0011 -0.0528 -0.3819 -0.0817 0.3185 0.0341 -0.0007 -0.2771 0.3159 -0.2448 -0.2811 -0.1787 0.1251 -0.281 -0.3883 -0.0226 0.3272 0.0638 -0.0248 -0.183 0.3533 -0.1743 -0.3087 -0.1519 0.1002 -0.1517

The corresponding eigenvalues of the residual space are 0.0594, 0.0159, 0.0130, 0.0007, 3.6e-05, 1.3e-15, and 5.36e-16, and the corresponding vectors form the residual space projection matrix See Table 1.8.

Table 1.8 B3561 training sample residual space projection matrix

-0.0743 -0.0498 -0.3741 -0.0074 -0.7337 -4.70E-12 -1.50E-12 -0.6131 0.0105 -0.0072 0.0406 -0.0054 -5.80E-14 -1.70E-14 -0.0201 0.0098 -0.027 -0.1771 0.0234 2.20E-13 7.50E-14 0.1183 -0.1708 0.2094 -0.0261 6.80E-04 1.30E-14 5.30E-15 -0.0366 -0.2478 0.0033 0.7934 -0.0877 -8.70E-13 -2.90E-13 0.0920 0.1782 0.4369 -0.0617 -0.6691 -4.30E-12 -1.3E-12 -0.0205 0.5034 4.90E-04 0.5599 0.0539 1.30E-13 1.70E-14 0.0853 0.7341 0.1655 -0.0709 0.0512 3.50E-13 1.le-13 0.7389 -0.1407 0.0535 0.1136 6.20E-04 -3.20E-14 -1.40E-14 -0.1022 -0.0935 0.4799 0.0163 0.0048 0.0532 0.6459 -0.1392 -0.1899 0.4514 0.0289 0.0115 -0.5922 -0.1745 -0.0607 -0.0434 0.3238 0.0084 0.0019 0.1790 -0.7422 -0.0844 -0.1272 0.2345 0.0188 0.0079 0.7838 0.0386

B5631 currently has flight data of 1194 flights. After removing 100 flights as training samples, the remaining 1094 flights are used as test samples.

For each flight, 13 eigenvalues are extracted to form a vector X, and the vector X is standardized.

After calculation, the SPE threshold is 0.45.

The SPE statistics of the 5-level test data are shown in Figure 3.

When the flight reached the 72nd flight, the SPE statistic exceeded the threshold, and the SPE value of the next flight was lower than the threshold again. Considering that only a period of data was selected when selecting the data sample of the 72nd flight, and different flight environments will lead to different flight loads, the eigenvalue vector obtained is accidental. The SPE statistic occasionally exceeded the threshold and it cannot be completely determined that it was a sensor failure. When the flight reached the 124th flight, the same situation occurred, and it cannot be determined that the sensor failed. After the flight reached the 161st flight, the SPE statistic began to frequently exceed the threshold, judging that the sensor was about to fail. On the 243rd flight, the sensor failed, and the flap position sensor was replaced after the flight. The SPE value was always less than the threshold.

The above model building and fault monitoring are based on the 5-speed data of B5631. The model is built for the 0 and 30 speeds of B5631 according to the same steps, and the SPE thresholds of the two speeds and the SPE statistics of the test samples are calculated.

In actual flight, the QAR data is read by the onboard computer, and after feature extraction, the characteristic value of the trailing edge flap under the current flight state is obtained. The characteristic value is then brought into the principal component analysis method and compared with the known SPE thresholds of each gear data to obtain the current health status of the flap driver.

2 , a flap fault sensing method implementation system includes a data acquisition module, a host computer, a main drive module, an auxiliary drive module, and an alarm module;

The data acquisition module is a rotary variable differential transformer installed at the trailing edge flap position, which is a linear sensor capable of collecting continuous position signals; the host computer collects data from the data acquisition module in real time, performs fault prediction through the flap fault sensing method, and generates control instructions; after receiving the action instructions from the host computer, the main drive module drives the flap to make attitude adjustment; the auxiliary drive module is controlled by the host computer, and when the monitoring determines that a fault has occurred, the alarm module is triggered, and the host computer starts the auxiliary drive module to push the flap outward to stabilize the fuselage attitude.

The auxiliary drive module includes an internal wing slide rail 1, a connecting rod 2, and a flap side slide rail 3. The internal wing slide rail 1 is vertically arranged in the wing, and the flap side slide rail 3 is horizontally located in the flap. One end of the connecting rod 2 is located in the internal wing slide rail 1, and the other end is located in the flap side slide rail 3. It can slide in the two slide rails through an internal motor.

Before landing, the aircraft needs to extend the flaps to provide lift and resistance to ensure safe landing. If the flap drive is blocked due to mechanical structural reasons and cannot be extended normally, the pilot can lift it by manipulating the inner slide rail 1 of the wing to provide the flap extension angle, and extend the flaps through the motor inside the connecting rod 2 to achieve compensation effect and ensure the flight quality of the aircraft. The above technical effect can be achieved by adding a connecting rod, which has a simple structure and low modification cost.

Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims (9)

1. A flap fault sensing method, characterized by the following specific steps: Step 1: flap over-limit event monitoring, constructing a measurement data set of the parameters to be monitored; Step 2: Preprocess the measurement data obtained in step 1 to obtain threshold sample data for fault judgment; the specific steps are as follows: S2.1, standardize the measurement data; Assume x ₁ , x ₂ , …, x _N are the data extracted from the features, each of which is 15-dimensional data. The data is standardized to eliminate the influence of dimension and the influence of the variable's own variation and value. The standardization method is to subtract the mean value u _q of the corresponding dimension from the data of each dimension, and then divide it by the variance σ _q of the corresponding dimension to obtain the standardized data. Where N represents the number of data; Wherein, X _pq is the data of the qth dimension in the pth data; S2.2, calculate the covariance matrix based on the standardized data; Based on the data obtained by standardization The calculated covariance matrix is: Wherein, the matrix elements r _uv are the correlation coefficients of the original variables x _u and x _v , u, v = 1, 2, ..., m; S2.3, calculate the eigenvalues and eigenvectors of the matrix; The characteristic equation of the matrix in step S2.2 is: |λI-R|=0, where I represents an m-order identity matrix with 1 on the diagonal and 0 on the rest; First, the solution λ _i of the eigenvalue is obtained according to the conversion rule; then, the eigenvalues are arranged in the order of λ ₁ ≥λ ₂ ≥…≥λ _m ≥0; then, the eigenvector e _i corresponding to each eigenvalue λ _i is calculated, i = 1, 2, …, m, and ||e _i || = 1, that is, Where e _ij represents the j-th component of the eigenvector e _i ; S2.4, determine the principal component according to the component contribution rate for all features, that is, complete the dimensionality reduction processing of the measured data, and determine the first r eigenvalues as the principal component; The contribution of the ingredients is: S2.5, based on the principal components determined in S2.4, determine the principal component space and the residual space, and then determine the sample data of the threshold; Arrange the eigenvectors corresponding to the r eigenvalues in rows from top to bottom and get Matrix; by calculation and The product of will get the principal component subspace projection matrix C; further calculation will get the residual subspace projection matrix After principal component analysis, the original data is expressed by the following formula: The input fault sample data is divided into two parts: in, The dimensional space composed is a subspace of the principal component. The composed dimensional space is a subspace of the residual; is the projection of x on the PCS principal component subspace; is the projection of x on the RS residual subspace, C and is the corresponding projection matrix; The fault sample data is divided into the principal component indicating the normal value and the residual component indicating the measurement fault by the principal component analysis method; under normal conditions, It is mainly used to measure noise, so in fault mode, The value will be significantly larger; The calculation formula of the variable monitoring SPE statistic acting on the residual subspace RS is: When the monitored object is in normal working condition, The projection in the residual subspace is within a normal range. If a fault occurs, the projection of the monitored variable in the residual subspace will change, causing the SPE statistic to exceed the threshold cl(α): in, c _α is the (1-α) confidence level of the standard normal distribution; Step 3: Based on the preprocessed measurement data and threshold sample data, a GRU neural network trailing edge flap performance evaluation model is constructed; Step 4: Save the trained GRU neural network trailing edge flap performance evaluation model, predict the time series data of the monitoring parameters in the prediction data stage, and obtain the fault prediction results. 2. A flap fault sensing method according to claim 1, characterized in that: the flap over-limit event monitoring method in step 1 is: S1.1: Determine the out-of-limit events and corresponding monitoring parameters, and confirm that the sensors function normally and the data is correct; S1.2: Determine the threshold for limit-exceeding events; S1.3: Compile messages and decode data for over-limit events; S1.4: Message data processing: Import QAR data for decoding. When decoding is completed, each message will generate a corresponding TXT file. 3. A flap fault sensing method according to claim 1, characterized in that: when constructing the measurement data set of the parameter to be monitored in step 1, in order to eliminate the accidental influence of the flight load on the opening time, the trailing edge flap opening time is preprocessed by a sliding window method; assuming that within a certain period of time, the trailing edge flap opening time series is: ε＝[ε ₁ ε ₂ ε ₃ …ε _j …] (2.1) Among them, ε ₁ , ε ₂ , ε ₃ , ε _j correspond to the first, second, third and jth trailing edge flap opening times in the trailing edge flap opening time series; Take a sliding window with a width of j for the opening time series, where j represents the number of data in a sliding window, and then take the average of j consecutive data in the window: the average width of the sliding window is X _k , k = 1…j; ε _k refers to the kth trailing edge flap opening time in a sliding window; In the next iteration, the starting position of the sliding window is shifted by a set step size, that is, the width j, and the mean of the data in the window is calculated; the sliding continues until the end position of the window is the last data; by setting a reasonable sliding window forward step size, the influence of random factors on the parameters is eliminated, and the performance of the trailing edge flap is more accurately reflected. 4. A flap fault sensing method according to claim 3, characterized in that: the parameters to be monitored include flap handle gear information FLAP CONFIGURATION, left trailing edge flap position sensor angle TE FLAP POSN LT, right trailing edge flap position sensor angle TE FLAP POSN RT; and the measured values after the flap is opened at three gears of 0, 5 and 30 are used for analysis; Extract characteristic values from the measured data of the parameter to be monitored, subtract the baseline value from the data on the fault side to obtain the difference CZ; extract the mean, variance, standard deviation, maximum value, maximum ten-digit average, minimum value, minimum ten-digit average, skewness, kurtosis, lower bound value of outliers, upper bound value of outliers, lower bound value of extreme values, upper bound value of extreme values, mean square value, and margin for the difference CZ, a total of 15 characteristic values; a vector consisting of 15 characteristic values is obtained for the 0, 5, and 30 gears of each flight. 5. A flap fault perception method according to claim 4, characterized in that: the specific steps of constructing the GRU neural network trailing edge flap performance evaluation model _yt in step 3 are: r _t =σ(W _r ×[h _t-1 ,x _t ]) (3.1) z _t =σ(W _z ×[h _t-1 ,x _t ]) (3.2) y _t = sig(W _o ·h _t ) (3.5) Among them, _Zt represents the update gate, _rt represents the reset gate, _xt represents the input of the current neuron, _ht represents the output of the current neuron, ht _-1 represents the output of the previous neuron, represents the pending output value in the current neuron, _Wz represents the weight of the update gate, _Wr represents the weight of the reset gate, and _Wo represents the output weight. Represents the weight of the pending output, and sig represents the sigmoid function. 6. A flap fault perception method according to claim 5, characterized in that: S3.1, the GRU neural network merges the forget gate and the exit gate in the LSTM into an update gate, and optimizes the combination with the output gate into a cell structure of two "gates"; when the value of the update gate z _t is smaller, it means that the current neuron needs to retain less information, and the previous neuron needs to retain more information; according to the process of each "gate" in the GRU neural network, it can be seen that each output information is affected by each neuron, so each neuron is interdependent; under normal circumstances, it is more active in the reset gate for short-distance learning and more active in the update gate for long-distance learning. 7. A flap fault perception method according to claim 6, characterized in that: the verification method of the GRU neural network trailing edge flap performance evaluation model has the following specific steps: First, a GRU neural network model with 6 hidden units is established, that is, the number of hidden neurons is 6; Then, through the formula derivation, it is concluded that the weight parameter to be learned is The first three weights are concatenated, so they need to be separated during learning, i.e. W _r =W _rx +W _rh (3.6) W _z =W _zx +W _zh (3.7) Among them, W _rx is the weight between the current neuron reset gate and the current neuron input, W _rh is the weight between the current neuron reset gate and the output of the previous neuron, W _zx is the weight between the current neuron update gate and the input of the current neuron, and W _zh is the weight between the current neuron update gate and the previous neuron. is the weight between the current neuron output and input value, The current neuron output and the previous neuron output; Input to the output layer The output is The loss function in neural network training is defined as: In the formula, _Et represents the loss value, _yd represents the output value of each neuron, Represents the output of the output layer, that is, the real original data; Finally, subtract the true value from the neuron’s output value _yd Then square it and average it to get the variance; the obtained variance is the loss value of the loss function; the loss function is used to evaluate the gap between the neural network target and the actual output. The smaller the function value, the smaller the difference between the actual data and the target output, which means that the weight is more appropriate. 8. A flap fault perception method according to claim 7, characterized in that: the objective function y _t is optimized using the gradient descent method, firstly the learning rate of the gradient descent method is defined, and the descent gradient of each optimization training model is performed; then the optimizer in Tensor Flow is used to optimize the loss function, and the optimizer used is Adam Optimizer. 9. An implementation system of the flap fault sensing method according to any one of claims 1 to 8, characterized in that it comprises a data acquisition module, a host computer, a main drive module, an auxiliary drive module, and an alarm module; The data acquisition module is a rotary variable differential transformer installed at the trailing edge flap position, which is a linear sensor capable of collecting continuous position signals; The host computer collects data from the data acquisition module in real time, performs fault prediction through the flap fault sensing method, and generates control instructions; After receiving the action instruction from the host computer, the main driving module drives the flap to make attitude adjustment; The auxiliary drive module is controlled by the host computer. When monitoring determines that a fault occurs, the alarm module is triggered, and the host computer starts the auxiliary drive module to push the flaps outward to stabilize the fuselage attitude.