Rapid fault diagnosis method for airborne vibration sensor

Abstract

The invention belongs to the technical field of aircraft airborne testing, and particularly relates to a rapid fault diagnosis method for an airborne vibration sensor. Comprising the following steps that 1, flight test sensor record data are obtained, data of a vibration sensor are separated according to parameter names, and the data of the sensor comprise a fault sample data sequence and a normal sample data sequence; 2, respectively forming a training set, a verification set and a test set according to a random selection mode; 3, respectively constructing normal sampling training subsets for the training set, and constructing an unbalanced sampling training subset by diluting normal samples; 4, wavelet transformation is carried out on the normal sampling training subset data sequence and the unbalanced sampling training subset data sequence respectively, and a normal sampling training subset input matrix and an unbalanced sampling training subset input matrix based on wavelet transformation modulus maxima are constructed respectively; 5, inputting the normal sampling training subset input matrix and the unbalanced sampling training subset input matrix into the bilateral branch network model for training, and obtaining prediction output through different weight accumulation learning of two sample sets; step 6, using a cross entropy loss function to realize convergence of the network model, and completing network model training; and 7, performing sensor fault judgment on output of the sensor and prediction output of the network model by adopting a majority consistent voting principle. The problem of sensor abnormal phenomenon detection or abnormal missing in a flight test is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft airborne testing, and in particular relates to a method for rapid fault diagnosis of an airborne vibration sensor. Background Art

[0002] During flight tests, a large number of sensors need to be installed on the aircraft to obtain test flight data. To ensure the reliability and effectiveness of the test data, it is necessary to effectively manage the status of the sensors before installation and during the test flight. Traditional periodic calibration can easily lead to the lack of detection or warning of sensor abnormalities during the calibration cycle, resulting in invalid data and increasing the difficulty of troubleshooting the test system. Summary of the invention

[0003] The purpose of the present invention is to provide a method for rapid fault diagnosis of airborne vibration sensors, so as to solve the problem of detecting abnormal phenomena or missing abnormalities of sensors in flight tests. The method can realize real-time monitoring of abnormal states of sensors in flight tests and provide real-time feedback on abnormal conditions, thereby eliminating invalid test data and improving the reliability of the test system.

[0004] The technical solution of the present invention:

[0005] Technical solution of the present invention: According to a first aspect of the present invention, a method for rapid fault diagnosis of an airborne vibration sensor is proposed, comprising the following steps:

[0006] Step 1: Obtain the flight test sensor recording data and separate the vibration sensor data according to the parameter name. The sensor data includes a fault sample data sequence and a normal sample data sequence;

[0007] Step 2: Form training set, validation set and test set respectively by random selection;

[0008] Step 3: construct normal sampling training subsets for the training set, and construct unbalanced sampling training subsets by diluting the normal samples;

[0009] Step 4: Perform wavelet transform on the normal sampling training subset and the unbalanced sampling training subset data sequences respectively, and construct the normal sampling training subset input matrix and the unbalanced sampling training subset input matrix based on the wavelet transform modulus maximum respectively;

[0010] Step 5: Input the normal sampling training subset input matrix and the imbalanced sampling training subset input matrix into the bilateral branch network model for training, and obtain the prediction output by accumulating learning with different weights for the two sample sets;

[0011] Step 6: The cross entropy loss function is used to achieve the convergence of the network model and complete the network model training;

[0012] Step 7: Use the majority consensus voting principle to determine sensor faults based on the sensor output and the network model prediction output.

[0013] In a possible embodiment, in step 2, the sensor data sequence is divided into three parts: a training set, a validation set, and a test set in a ratio of 7:1:2.

[0014] In a possible embodiment, in step 3, the specific process of constructing an unbalanced sampling training subset by diluting normal samples includes: slicing the flight test ultra-long time series data set according to time, cutting each flight sortie data into K pieces, sampling K complete fault samples, and sampling K / i normal samples.

[0015] In a possible embodiment, in step 4, the specific process of constructing the input matrix based on the wavelet transform modulus maximum value includes:

[0016] The daubechies wavelet basis function is selected, and the wavelet type is selected as db1~db10. The sensor data is decomposed by 6 layers of wavelets to obtain wavelet coefficients d1~d6. The modulus processing |d1|~|d6| is performed, that is, D1~D6; the normal flight test data and fault data samples are constructed from the modulus maximum value based on the two dimensions of wavelet decomposition layer number and wavelet type to construct the training set 1 input matrix X i1 , training set 2 input matrix X i2 .

[0017]

[0018] In a possible embodiment, in step 5, the specific process of obtaining the prediction output by cumulatively learning two sample sets with different weights includes:

[0019] Matrix X i1 and X i2 The feature vector f is obtained by convolutional neural network and average pooling respectively. c and f r , according to each round of learning

[0020]

[0021] The sample confidence sets the feature vector weight coefficient, and the calculation process of the weight coefficient α is:

[0022] where y i is the training sample set X i The output weight coefficient after convolution pooling, T is the current training number, T max is the maximum number of training times;

[0023] The weighted feature vectors are sent to the classifiers respectively, and then the outputs are integrated by accumulation. The predicted output value z is:

[0024] In one possible embodiment, in step 6, the cross entropy loss function is defined as follows:

[0025]

[0026] In a possible embodiment, step 7 specifically includes the following steps: if H pieces of data generated by a sensor in a flight record are D1, D2, ...DH, and its predicted abnormal results of the sensor are PD1, PD2, ..., PDH, respectively, if there are greater than or equal to H / 2 predicted values ​​that indicate that the sensor is abnormal, then the sensor is ultimately considered to be abnormal; otherwise, the sensor is considered to be working normally.

[0027] According to a second aspect of the present invention, a computing device is provided, comprising a processor and a memory, wherein the processor is configured to execute instructions stored in the memory to implement the above method.

[0028] According to a third aspect of the present invention, a computer program product comprising instructions is provided. When the instructions are executed by a computing device, the computing device is caused to perform the above method.

[0029] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, comprising computer program instructions, and when the computer program instructions are executed by a computing device, the computing device is caused to execute the above method.

[0030] Advantages of the present invention:

[0031] The beneficial effect of a method for rapid diagnosis of airborne vibration sensor faults provided by the present invention is that it is simple and effective, and solves the problem of detecting abnormal phenomena or missing abnormalities of sensors in flight tests. Real-time monitoring of the abnormal state of the sensor in the flight test is achieved, and real-time feedback on the abnormal situation is provided, invalid test data can be eliminated, and the reliability of the test system can be improved; at the same time, the strategies of undersampling and oversampling are adopted to alleviate the problem of uneven sample distribution of abnormal detection. The key technical difficulties of rapid diagnosis of airborne vibration sensor faults are: the output signal of the vibration sensor is a high-frequency signal, and the characteristics of the fault signal are difficult to identify; in addition, compared with the frequency of standard events, the frequency of abnormal events is lower. Compared with other sensors of the same type, vibration sensors in flight tests are used in large quantities and are widely distributed, generally distributed in multiple parts inside and outside the fuselage. Due to the different external environments borne by each part, the factors causing sensor abnormalities are also different, and the number of abnormal samples is relatively low. The method for rapid diagnosis of airborne vibration sensor faults provided by the present invention can overcome the above difficulties. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention is a flowchart of a method for rapid fault diagnosis of an airborne vibration sensor according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be described in further detail below with reference to the accompanying drawings.

[0034] The process of rapid diagnosis of airborne vibration sensor fault is as follows: Figure 1 As shown in the figure, it is divided into data preprocessing, model classification learning, and result output. The data preprocessing stage mainly includes steps such as vibration parameter separation and normalization, wavelet transform signal processing, data sampling, normal and fault data labels, etc. Model classification learning mainly includes the design of optimal sampling weights, loss function optimization, etc. Data output includes steps such as calling the model and outputting prediction results. Model upgrade refers to the upgrade and optimization of the model by sending the newly accumulated test flight data into the model after a certain period of time.

[0035] The following is an explanation of the workflow of the rapid diagnosis method for airborne vibration sensor faults.

[0036] A method for rapid fault diagnosis of an airborne vibration sensor comprises the following steps:

[0037] Step 1: Obtain M flight test sensor recording data, separate N vibration sensor data according to parameter names, and form an M*N data sequence, where the fault sample is the M*P data sequence and the normal sample is the M*(NP) data sequence;

[0038] Step 2: Divide the sensor data sequence into three parts: training set, validation set, and test set in a ratio of 7:1:2 by random selection. The three parts of data cannot have the same samples, that is, three sample sets are formed: training set R*N, validation set S*N, and test set T*N.

[0039] Step 3: Slice the flight test ultra-long time series data set according to time, cut each flight sortie data into K pieces, and perform unbalanced sampling on normal samples and fault samples to address the problem of fewer fault samples and unbalanced samples. K fault samples are completely sampled, and K / i normal samples are sampled. Generate two training sets, training set 1 is normal sampling: the normal sample data sequence is R*P, and the fault sample data sequence is R*(NP), and training set 2 is unbalanced sampling: the normal sample data sequence is K / i*R*P and the fault sample data sequence is K*R*(NP).

[0040] Step 4: Perform wavelet transform on the data sequences of training set 1 and training set 2, and construct an input matrix based on the maximum value of the wavelet transform modulus;

[0041] Specifically, the daubechies wavelet basis function is selected, and the wavelet type is selected as db1~db10. The sensor data is decomposed into 6 layers of wavelets to obtain wavelet coefficients d1~d6, and the modulus processing |d1|~|d6| is performed, that is, D1~D6. The normal flight test data and fault data samples are constructed from the modulus maximum values ​​based on the two dimensions of wavelet decomposition layer number and wavelet type to construct the training set 1 input matrix X i1 , training set 2 input matrix X i2 .

[0042]

[0043] Step 5: Convert the matrix X i1 and X i2 Input the bilateral branch network model and obtain the output by accumulating learning with different weights for two sample sets.

[0044] Specifically, the matrix X i1 and X i2 The feature vector f is obtained by convolutional neural network and average pooling respectively. c and f r , the feature vector weight coefficient is set according to the sample confidence after each round of learning. The calculation process of the weight coefficient α is:

[0045] where y i is the training sample set X i The output weight coefficient after convolution pooling, T is the current training number, T max For maximum

[0046]

[0047] Number of training sessions.

[0048] The weighted feature vectors are sent to the classifiers respectively, and then the outputs are integrated by accumulation. The predicted output value z is:

[0049]

[0050] Step 6: Convergence problem of network training model, use a unified loss function to constrain the optimization direction of the network. The cross entropy loss function used in this method is defined as follows:

[0051]

[0052] Step 7: After the model training is completed, the inference phase begins. The majority voting principle is used to improve the performance of the model. That is, if a sensor in a flight record generates H data points, D1, D2, ... DH, and its predicted abnormal results for the sensor are PD1, PD2, ..., PDH, if there are greater than or equal to H / 2 predicted values ​​that the sensor is abnormal, then the sensor is finally considered abnormal, otherwise it is considered to be working normally.

[0053] Example 1

[0054] (1) When the aircraft is grounded or in flight, the sensor data is recorded and read to start the data preprocessing phase (including the following steps (2)-(5)).

[0055] (2) Read the flight test sensor data, separate the vibration sensor data according to the parameter name, and divide the samples into different training sets according to the principles of different vibration sensors (piezoelectric, piezoresistive). Select the sensor type that needs to be trained, and randomly select the sensor data of this type into three parts: training set, validation set, and test set in a ratio of 7:1:2. To ensure the training effect, the three parts of data cannot have the same samples. The training set completes the preliminary data processing according to the steps of selecting subsets, renaming columns, deleting duplicate values, processing missing values, normalizing, and removing jump points.

[0056] (3) Since the number of fault samples in the data set is very small, far less than the number of normal samples, the samples are extremely unbalanced. In order to solve this problem, a weighted approach is used in training to deal with the imbalance problem. That is, the weight of the category with a large number of samples in the training set (the majority class) is reduced; the weight of the category with a small number of samples in the training set (the minority class) is increased. This strategy can alleviate the problem of unbalanced sample distribution in state detection.

[0057] (4) After initialization, the data is subjected to discrete wavelet analysis. Daubechies is selected according to the characteristics of vibration parameters.

[0058] Wavelet basis function is used to obtain a low-frequency reconstruction signal A1 and high-frequency reconstruction signals D1~D10, and the low-frequency and high-frequency signal features of normal flight test data and fault data are extracted to construct a feature matrix.

[0059] (5) After completing data preprocessing, start model training (including the following steps (6)-(8)).

[0060] (6) The data feature values ​​are input into the deep learning classification network model in the form of a matrix for training. The normal sampling and reverse sampling data are input into the network training model respectively. Different weights are set for the feature values ​​after convolution pooling. The weight coefficient α = T / T max, where T is the current training number, T max The maximum number of training times is used to extract the potential information contained in the sensor data through the network model.

[0061] (7) Convergence problem of network training model, using a unified loss function to constrain the optimization direction of the network. The cross entropy loss function used in this method is defined as follows:

[0062]

[0063] The smaller the loss function, the better the robustness of the model. The loss function guides model learning, and backpropagation is performed based on the loss function to continuously modify model parameters, ultimately improving the performance of the model on the target task.

[0064] (8) After the model training is completed, the inference phase begins. The majority consensus voting principle is used to improve the performance of the model. That is, if a sensor in a flight record generates H data points, D1, D2, ..., DH, and its predicted abnormal results for the sensor are PD1, PD2, ..., PDH, respectively, if there are greater than or equal to H / 2 predicted values ​​that the sensor is abnormal, then the sensor is ultimately considered abnormal, otherwise it is considered to be working normally.

[0065] (9) After the inference phase, the model upgrade begins. The new input model data is added to the training samples for joint training. As the amount of data increases, the performance of the model will continue to evolve and its generalization will gradually improve.

Claims

1. A method for rapid fault diagnosis of an airborne vibration sensor, characterized in that: The steps include: Step 1: Obtain the flight test sensor recording data and separate the vibration sensor data according to the parameter name. The sensor data includes a fault sample data sequence and a normal sample data sequence; Step 2: Form training set, validation set and test set respectively by random selection; Step 3: construct normal sampling training subsets for the training set, and construct unbalanced sampling training subsets by diluting the normal samples; Step 4: Perform wavelet transform on the normal sampling training subset and the unbalanced sampling training subset data sequences respectively, and construct the normal sampling training subset input matrix and the unbalanced sampling training subset input matrix based on the wavelet transform modulus maximum respectively; Step 5: Input the normal sampling training subset input matrix and the imbalanced sampling training subset input matrix into the bilateral branch network model for training, and obtain the prediction output by accumulating learning with different weights for the two sample sets; Step 6: The cross entropy loss function is used to achieve the convergence of the network model and complete the network model training; Step 7: Use the majority consensus voting principle to determine sensor faults based on the sensor output and the network model prediction output.

2. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 1, characterized in that: In step 2, the sensor data sequence is divided into three parts: a training set, a validation set, and a test set in a ratio of 7:1:

2.

3. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 1, characterized in that: In step 3, the specific process of constructing an unbalanced sampling training subset by diluting normal samples includes: slicing the flight test ultra-long time series data set according to time, cutting each flight sortie data into K pieces, sampling K complete fault samples, and sampling K / i normal samples.

4. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 1, characterized in that: In step 4, the specific process of constructing the input matrix based on the wavelet transform modulus maximum value includes: The daubechies wavelet basis function is selected, and the wavelet type is selected as db1~db10. The sensor data is decomposed by 6 layers of wavelets to obtain wavelet coefficients d1~d6. The modulus processing |d1|~|d6| is performed, that is, D1~D6; the normal flight test data and fault data samples are constructed from the modulus maximum value based on the two dimensions of wavelet decomposition layer number and wavelet type to construct the training set 1 input matrix X i1 , training set 2 input matrix X i2 .

5. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 5, characterized in that: In step 5, the specific process of obtaining the prediction output by cumulatively learning two sample sets with different weights includes: Matrix X i1 and X i2 The feature vector f is obtained by convolutional neural network and average pooling respectively. c and f r ,root The feature vector weight coefficient is set according to the sample confidence after each round of learning. The calculation process of the weight coefficient α is: where y i is the training sample set X i The output weight coefficient after convolution pooling, T is the current training number, T max is the maximum number of training times; The weighted feature vectors are sent to the classifiers respectively, and then the outputs are integrated by accumulation. The predicted output value z is: .

6. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 5, characterized in that: In step 6, the cross entropy loss function is defined as follows:

7. The method for rapid fault diagnosis of an airborne vibration sensor according to claim 6, characterized in that: In the step 7, the following steps are specifically included: if H data points generated by a sensor in a certain flight record are D1, D2, ...DH respectively, and the predicted abnormal results of the sensor are PD1, PD2, ..., PDH respectively, if there are greater than or equal to H / 2 predicted values ​​that the sensor is abnormal, then the sensor is finally considered to be abnormal, otherwise it is considered that the sensor is working normally.

8. A computing device, characterized in that The method comprises a processor and a memory, wherein the processor is used to execute instructions stored in the memory to implement the method according to any one of claims 1 to 7.

9. A computer program product comprising instructions, characterized in that When the instructions are executed by a computing device, the computing device is caused to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The method comprises computer program instructions, which, when executed by a computing device, cause the computing device to perform the method according to any one of claims 1 to 7.

Images