阅读说明：本技术 一种航空发动机传感器故障检测方法及装置 (Method and device for detecting faults of aero-engine sensor ) 是由 王雷 裴紫焱 鲁统超 马乐乐 孙震宇 于 2021-08-02 设计创作，主要内容包括：一种航空发动机传感器故障检测方法及装置。方法包括采集快速访问记录器数据；利用高斯混合模型划分数据；训练CNN-LSTM混合模型；利用上CNN-LSTM混合模型进行特征提取与故障识别；显示识别结果和故障信息等步骤。本发明融合高斯混合模型、CNN模型与LSTM模型,可实现对航空发动机传感器故障的精准检测。通过高斯混合模型聚类划分飞行阶段后,模型训练时的收敛速度与最终的检测精度都有明显提升。将CNN-LSTM混合模型与传统的单一CNN模型、LSTM模型进行对比,本发明的CNN-LSTM混合模型对QAR数据具有较强的特征提取能力,极大地方便了飞机驾驶员和机务维修人员对航空发动机传感器的故障诊断。(A method and a device for detecting faults of an aircraft engine sensor. The method comprises collecting quick access recorder data; dividing data by using a Gaussian mixture model; training a CNN-LSTM mixed model; performing feature extraction and fault identification by using the upper CNN-LSTM mixed model; and displaying the identification result and the fault information. The method disclosed by the invention integrates the Gaussian mixture model, the CNN model and the LSTM model, and can realize accurate detection of the faults of the sensor of the aero-engine. After the flight phases are divided by Gaussian mixture model clustering, the convergence speed and the final detection precision during model training are obviously improved. Compared with the traditional single CNN model and the traditional LSTM model, the CNN-LSTM mixed model has stronger characteristic extraction capability on QAR data, and greatly facilitates the fault diagnosis of aircraft drivers and crew maintenance personnel on the aeroengine sensor.)

1. A method for detecting faults of an aircraft engine sensor is characterized by comprising the following steps: the method for detecting the faults of the aero-engine sensor comprises the following steps in sequence:

1) the method comprises the steps that 8 types of quick access recorder data in the flight process are collected in advance according to the time sequence, wherein the 8 types of quick access recorder data are respectively the throttle lever angle PLA, the high-pressure rotor rotating speed N1, the deflation valve opening VBV, the compressor adjustable stationary blade angle CVV, the atmospheric total temperature TAT, the fuel flow FF, the aircraft height ALT and the flight Mach number MACH, the first 6 types of data are detection data of 6 types of sensors at different positions, and then sliding average filtering processing is carried out on the detection data of each sensor respectively to obtain data after extreme value average filtering; forming a data set by all the average filtered data of the depolarization value;

2) using the aircraft height ALT and the flight Mach MACH as clustering data, inputting the data in the data set into a Gaussian mixture model GMM, setting clustering numbers according to the number of the flight stages divided in the flight process, training the Gaussian mixture model GMM, continuously adjusting model parameters, and finally obtaining the trained Gaussian mixture model GMM, thereby dividing the average filtered data of the depolarization values of 6 sensors in the flight process into a plurality of groups of data of different flight stages;

3) respectively adding a sensor dead jamming fault, a constant gain fault and a constant deviation fault to the data of each group of different flight stages to obtain a plurality of groups of sensor fault data of different flight stages, then giving a label to each type of sensor fault data, and then randomly dividing the sensor fault data of each group of different flight stages with the labels into a training set and a testing set according to the proportion of 8: 2; then, carrying out Z-Score standardization processing on the training set of each flight stage, and then respectively inputting the training set into a CNN-LSTM mixed model for training to obtain a plurality of trained CNN-LSTM mixed models;

4) performing Z-Score standardization processing on the test set of each flight stage, then respectively inputting the test set into the trained CNN-LSTM mixed model of the corresponding flight stage for identification, performing gradient descent training on the trained CNN-LSTM mixed model by using an error between an identification result and sensor fault data with labels in the corresponding training set, and ending the training until the identification accuracy reaches a set value to obtain the trained CNN-LSTM mixed model;

5) after receiving a quick access recorder data set of an aeroengine, a system platform carries out sliding depolarization value average filtering processing and clustering on detection data of 6 sensors including a throttle lever angle PLA, a high-pressure rotor rotating speed N1, a blow-off valve opening VBV, a compressor adjustable stator blade angle CVV, an atmospheric total temperature TAT and a fuel flow FF according to the methods of the step 1) and the step 2), and then carries out feature extraction and fault identification on the clustered data by using the trained CNN-LSTM mixed model so as to determine whether the sensors have faults or not;

6) and if the fault information exists in the identification result, displaying the identification result and the fault information on the system platform to remind an aircraft pilot or a crew member to pay attention.

2. The aircraft engine sensor fault detection method of claim 1, wherein: in step 1), the specific method of the sliding depolarization average filtering process is as follows:

and taking 30 detection data of a certain sensor at a certain time point, the first 15 detection data and the last 14 detection data of the detection data, removing the maximum value and the minimum value from the detection data, then calculating the average value of the rest 28 detection data, obtaining the average filtered data of the depolarization value of the sensor at the time point, and sequentially sliding backwards to perform the average filtering of the depolarization value at the later time point.

3. The aircraft engine sensor fault detection method of claim 1, wherein: in the step 2), the flight process is divided into 12 flight stages, namely a takeoff stage, a stair climbing stage, a cruise stage, a descent stage, a near approach stage and a landing stage, and the transition stages of the flight stages are respectively; the number of clusters is 12, and 12 sets of data of different flight phases are obtained.

4. The aircraft engine sensor fault detection method of claim 1, wherein: in the step 3), the number of the CNN-LSTM mixed models and the number of the trained CNN-LSTM mixed models are both 12;

the CNN-LSTM mixed model is formed by connecting a CNN model and an LSTM model in series and comprises 4 CNN layers, 2 LSTM layers and 3 full-connection layers; the operation process of each CNN layer is divided into four steps: two-dimensional convolution operation, batch standardization operation, activation and maximum pooling operation;

each LSTM layer consists of a forgetting gate, an input gate and an output gate; the forgetting gate is used for determining the degree of discarding the previous memory information, the input gate is used for determining the degree of storing the information into the cell memory, and the output gate is used for determining which information in the cell is output.

5. The aircraft engine sensor fault detection method of claim 1, wherein: in step 4), the recognition accuracy is set to be greater than 95%.

6. The aircraft engine sensor fault detection method of claim 1, wherein: in step 6), the identification result and the fault information comprise the associated flight phase, fault sensor type and fault type.

7. The utility model provides an aeroengine sensor fault detection device which characterized in that: the aircraft engine sensor fault detection device includes:

the data preprocessing module is used for carrying out sliding extremum removing average filtering processing on detection data of the sensor in the quick access recorder data set and the quick access recorder data which are collected in advance and received by the system platform so as to eliminate the influence of noise and inhibit accidental pulse interference;

the identification module is used for inputting detection data of a sensor into a trained Gaussian mixture model GMM after a system platform receives a quick access recorder data set of the aircraft engine so as to judge a flight stage to which the detection data belongs, then sending the detection data into a trained CNN-SLTM mixture model corresponding to the flight stage for feature extraction and fault identification, and sending an identification result into the display module;

the display module is used for displaying the identification result and the fault information in a window form, wherein the identification result and the fault information comprise the flight stage, the fault sensor type and the fault type;

the training module is used for extracting the characteristics of each group of data in the pre-collected quick access recorder data and training 12 CNN-LSTM hybrid models in 12 flight phases to obtain model parameters for the identification module to use;

and the verification module is used for carrying out feature extraction and fault detection on the data of the pre-collected QAR data set test set, comparing the detection result with the pre-identified label and calculating the identification accuracy so as to verify the accuracy of the trained CNN-LSTM hybrid model.

Technical Field

The invention belongs to the technical field of civil aviation, and particularly relates to a method and a device for detecting faults of an aircraft engine sensor.

Background

At present, research aiming at the fault detection of an aircraft engine sensor is rapidly developed, but a great deal of literature is consulted to find that the previous research work is only to establish a fault detection model aiming at a certain local flight stage or directly aiming at the whole flight process, and the great difference existing between the data characteristics of the sensor of an aircraft in different flight stages is not considered, so that the use scene of the fault detection model is directly limited. In addition, the inherent time series characteristics of the data are ignored, and the traditional single fault detection model is adopted, so that the potential information of the data cannot be fully utilized.

Disclosure of Invention

In order to solve the problems, the invention aims to provide a method and a device for detecting faults of an aircraft engine sensor.

In order to achieve the above object, the present invention provides a method for detecting a failure of an aircraft engine sensor, comprising the following steps performed in sequence:

1) the method comprises the steps that 8 types of quick access recorder data in the flight process are collected in advance according to the time sequence, wherein the 8 types of quick access recorder data are respectively the throttle lever angle PLA, the high-pressure rotor rotating speed N1, the deflation valve opening VBV, the compressor adjustable stationary blade angle CVV, the atmospheric total temperature TAT, the fuel flow FF, the aircraft height ALT and the flight Mach number MACH, the first 6 types of data are detection data of 6 types of sensors at different positions, and then sliding average filtering processing is carried out on the detection data of each sensor respectively to obtain data after extreme value average filtering; forming a data set by all the average filtered data of the depolarization value;

2) using the aircraft height ALT and the flight Mach MACH as clustering data, inputting the data in the data set into a Gaussian mixture model GMM, setting clustering numbers according to the number of the flight stages divided in the flight process, training the Gaussian mixture model GMM, continuously adjusting model parameters, and finally obtaining the trained Gaussian mixture model GMM, thereby dividing the average filtered data of the depolarization values of 6 sensors in the flight process into a plurality of groups of data of different flight stages;

3) respectively adding a sensor dead jamming fault, a constant gain fault and a constant deviation fault to the data of each group of different flight stages to obtain a plurality of groups of sensor fault data of different flight stages, then giving a label to each type of sensor fault data, and then randomly dividing the sensor fault data of each group of different flight stages with the labels into a training set and a testing set according to the proportion of 8: 2; then, carrying out Z-Score standardization processing on the training set of each flight stage, and then respectively inputting the training set into a CNN-LSTM mixed model for training to obtain a plurality of trained CNN-LSTM mixed models;

4) performing Z-Score standardization processing on the test set of each flight stage, then respectively inputting the test set into the trained CNN-LSTM mixed model of the corresponding flight stage for identification, performing gradient descent training on the trained CNN-LSTM mixed model by using an error between an identification result and sensor fault data with labels in the corresponding training set, and ending the training until the identification accuracy reaches a set value to obtain the trained CNN-LSTM mixed model;

5) after receiving a quick access recorder data set of an aeroengine, a system platform carries out sliding depolarization value average filtering processing and clustering on detection data of 6 sensors including a throttle lever angle PLA, a high-pressure rotor rotating speed N1, a blow-off valve opening VBV, a compressor adjustable stator blade angle CVV, an atmospheric total temperature TAT and a fuel flow FF according to the methods of the step 1) and the step 2), and then carries out feature extraction and fault identification on the clustered data by using the trained CNN-LSTM mixed model so as to determine whether the sensors have faults or not;

6) and if the fault information exists in the identification result, displaying the identification result and the fault information on the system platform to remind an aircraft pilot or a crew member to pay attention.

In step 1), the specific method of the sliding depolarization average filtering process is as follows:

and taking 30 detection data of a certain sensor at a certain time point, the first 15 detection data and the last 14 detection data of the detection data, removing the maximum value and the minimum value from the detection data, then calculating the average value of the rest 28 detection data, obtaining the average filtered data of the depolarization value of the sensor at the time point, and sequentially sliding backwards to perform the average filtering of the depolarization value at the later time point.

In the step 2), the flight process is divided into 12 flight stages, namely a takeoff stage, a stair climbing stage, a cruise stage, a descent stage, a near approach stage and a landing stage, and the transition stages of the flight stages are respectively; the number of clusters is 12, and 12 sets of data of different flight phases are obtained.

In the step 3), the number of the CNN-LSTM mixed models and the number of the trained CNN-LSTM mixed models are both 12;

the CNN-LSTM mixed model is formed by connecting a CNN model and an LSTM model in series and comprises 4 CNN layers, 2 LSTM layers and 3 full-connection layers; the operation process of each CNN layer is divided into four steps: two-dimensional convolution operation, batch standardization operation, activation and maximum pooling operation;

each LSTM layer consists of a forgetting gate, an input gate and an output gate; the forgetting gate is used for determining the degree of discarding the previous memory information, the input gate is used for determining the degree of storing the information into the cell memory, and the output gate is used for determining which information in the cell is output.

In step 4), the recognition accuracy is set to be greater than 95%.

In step 6), the identification result and the fault information comprise the associated flight phase, fault sensor type and fault type.

The invention provides a failure detection device for an aircraft engine sensor, which comprises:

the data preprocessing module is used for carrying out sliding depolarization average filtering processing on the QAR data set of the aircraft engine received by the system platform and the detection data of the sensor in the QAR data collected in advance so as to eliminate the influence of noise and inhibit accidental pulse interference;

the identification module is used for inputting detection data of a sensor into a trained Gaussian mixture model GMM after a system platform receives a QAR data set of the aircraft engine so as to judge a flight stage to which the detection data belongs, then sending the detection data into a trained CNN-SLTM mixture model corresponding to the flight stage for feature extraction and fault identification, and sending an identification result to the display module;

the display module is used for displaying the identification result and the fault information in a window form, wherein the identification result and the fault information comprise the flight stage, the fault sensor type and the fault type;

the training module is used for extracting the characteristics of each group of data in the pre-collected QAR data and training 12 CNN-LSTM hybrid models in 12 flight stages to obtain model parameters for the identification module to use;

and the verification module is used for carrying out feature extraction and fault detection on the data of the pre-collected QAR data set test set, comparing the detection result with the pre-identified label and calculating the identification accuracy so as to verify the accuracy of the trained CNN-LSTM hybrid model.

The method and the device for detecting the faults of the sensor of the aircraft engine have the following beneficial effects: compared with the prior art, the method disclosed by the invention integrates the Gaussian mixture model, the CNN model and the LSTM model, and can realize accurate detection of the faults of the sensor of the aero-engine. After the flight phases are divided by Gaussian mixture model clustering, the convergence speed and the final detection precision during model training are obviously improved. Compared with the traditional single CNN model and the traditional LSTM model, the CNN-LSTM mixed model has stronger characteristic extraction capability on QAR data, and greatly facilitates the fault diagnosis of aircraft drivers and crew maintenance personnel on the aeroengine sensor.

Drawings

FIG. 1 is a flow chart of a method for detecting aircraft engine sensor faults provided by the present invention;

FIG. 2 is a schematic diagram of a sampling process in the method for detecting a failure of an aircraft engine sensor according to the present invention;

FIG. 3 is a diagram illustrating a CNN model architecture according to the present invention;

FIG. 4 is a schematic diagram of the cell structure of the LSTM model of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.

As shown in fig. 1, the method for detecting a failure of an aircraft engine sensor according to the present invention comprises the following steps performed in sequence:

1) the method comprises the steps that 8 QARs (quick access recorders) data in the flight process are collected in advance according to the time sequence and respectively comprise an accelerator lever angle PLA, a high-pressure rotor rotating speed N1, a deflation valve opening VBV, a compressor adjustable stationary blade angle CVV, an atmospheric total temperature TAT, a fuel flow FF, an aircraft height ALT and a flight Mach number MACH, wherein the first 6 data are detection data of 6 sensors at different positions, and then sliding depolarization value average filtering processing is carried out on the detection data of each sensor respectively to eliminate the influence of noise and inhibit accidental pulse interference to obtain data after extremum average filtering; forming a data set by all the average filtered data of the depolarization value;

the specific method of the sliding depolarization average filtering process is as follows:

and taking 30 detection data of a certain sensor at a certain time point, the first 15 detection data and the last 14 detection data of the detection data, removing the maximum value and the minimum value from the detection data, then calculating the average value of the rest 28 detection data, obtaining the average filtered data of the depolarization value of the sensor at the time point, and sequentially sliding backwards to perform the average filtering of the depolarization value at the later time point.

The process of acquiring QAR data during flight is shown in FIG. 2. In the present invention, the number of QAR data is 8, and the data are arranged in time sequence, the time sequence length of the data selection window is 30, and data N are sequentially selected from the start position by the step length x of 1 until the end of the data.

2) Using the aircraft height ALT and the flight Mach MACH as clustering data, inputting the data in the data set into a Gaussian mixture model GMM, setting clustering numbers according to the number of the flight stages divided in the flight process, training the Gaussian mixture model GMM, continuously adjusting model parameters, and finally obtaining the trained Gaussian mixture model GMM, thereby dividing the average filtered data of the depolarization values of 6 sensors in the flight process into a plurality of groups of data of different flight stages;

the flight process is divided into 12 flight stages, namely takeoff, stair climbing, cruising, descending, approach and landing and transition stages of the flight stages; the clustering number is 12, and 12 groups of data in different flight phases are obtained;

3) respectively adding a sensor dead jamming fault, a constant gain fault and a constant deviation fault to the data of each group of different flight stages to obtain a plurality of groups of sensor fault data of different flight stages, then giving a label to each type of sensor fault data, and then randomly dividing the sensor fault data of each group of different flight stages with the labels into a training set and a testing set according to the proportion of 8: 2; then, carrying out Z-Score standardization processing on the training set of each flight stage to reduce the numerical range, and respectively inputting the training set into a CNN-LSTM mixed model for training to obtain a plurality of trained CNN-LSTM mixed models;

the number of the CNN-LSTM mixed models and the number of the trained CNN-LSTM mixed models are both 12.

The CNN-LSTM mixed model is formed by connecting a CNN model and an LSTM model in series and comprises 4 CNN layers, 2 LSTM layers and 3 full-connection layers; as shown in fig. 3, the operation process of each CNN layer is divided into four steps: two-dimensional convolution operation (Conv2D), batch normalization operation (BN), activation (Relu), and max-pooling operation (max-pooling). When the CNN layers extract features, the convolutional layers and the pooling layers are alternately arranged into convolutional groups, the features are extracted layer by layer, and the extracted features are gradually thinned from the whole to the local along with the increase of the layer number; the CNN layer extracts the characteristics of sensor fault data in different flight stages, and the extracted characteristics are used as the input of the LSTM layer, and the LSTM layer performs fault identification.

As shown in fig. 4, each LSTM layer consists of a forgetting gate, an input gate, and an output gate; the forgetting gate is used for determining the degree of discarding the previous memory information, the input gate is used for determining the degree of storing the information into the cell memory, and the output gate is used for determining which information in the cell is output. The LSTM layer has the memory capacity of the early-stage data, so that the features extracted by the CNN layer can be more effectively processed, and the recognition result is finally output.

4) Performing Z-Score standardization processing on the test set of each flight stage, then respectively inputting the test set into the trained CNN-LSTM mixed model of the corresponding flight stage for identification, performing gradient descent training on the trained CNN-LSTM mixed model by using an error between an identification result and sensor fault data with labels in the corresponding training set, and ending the training until the identification accuracy reaches a set value to obtain the trained CNN-LSTM mixed model;

the recognition accuracy in the present invention is set to be greater than 95%.

5) After receiving a QAR data set of an aeroengine, a system platform carries out sliding depolarization value average filtering processing and clustering on detection data of 6 sensors including a throttle lever angle PLA, a high-pressure rotor rotating speed N1, a deflation valve opening VBV, a compressor adjustable stator blade angle CVV, an atmospheric total temperature TAT and a fuel flow FF according to the methods of the step 1) and the step 2), and then carries out feature extraction and fault identification on the clustered data by utilizing the trained CNN-LSTM mixed model so as to determine whether the sensors have faults or not;

6) and if the fault information exists in the identification result, displaying the identification result and the fault information on the system platform to remind an aircraft pilot or a crew member to pay attention.

The identification result and the fault information comprise the flight stage, the fault sensor type and the fault type;

the invention provides a failure detection device for an aircraft engine sensor, which comprises:

the data preprocessing module is used for carrying out sliding depolarization average filtering processing on the QAR data set of the aircraft engine received by the system platform and the detection data of the sensor in the QAR data collected in advance so as to eliminate the influence of noise and inhibit accidental pulse interference;

the identification module is used for inputting detection data of a sensor into a trained Gaussian mixture model GMM after a system platform receives a QAR data set of the aircraft engine so as to judge a flight stage to which the detection data belongs, then sending the detection data into a trained CNN-SLTM mixture model corresponding to the flight stage for feature extraction and fault identification, and sending an identification result to the display module;

the display module is used for displaying the identification result and the fault information in a window form, wherein the identification result and the fault information comprise the flight stage, the fault sensor type and the fault type; for example, in the flight phase 1 takeoff phase, the atmospheric total temperature TAT sensor has a jamming fault.

The training module is used for extracting the characteristics of each group of data in the pre-collected QAR data and training 12 CNN-LSTM hybrid models in 12 flight stages to obtain model parameters for the identification module to use;

and the verification module is used for carrying out feature extraction and fault detection on the data of the pre-collected QAR data set test set, comparing the detection result with the pre-identified label and calculating the identification accuracy so as to verify the accuracy of the trained CNN-LSTM hybrid model.