An intelligent detection method for aircraft pneumatic sensor faults

Abstract

The present invention discloses an intelligent detection method for aircraft pneumatic sensor faults, comprising the following steps: S1, obtaining flight data of aircraft pneumatic sensor faults under various flight states of multiple aircraft; S2, using a control numerical experiment to select data directly related to the detection of aircraft pneumatic sensor faults from the flight data; S3, preprocessing the selected data, and stacking the preprocessed data into an "image"; The present invention discloses an intelligent detection method for aircraft pneumatic sensor faults, which can accurately detect the faults of pneumatic sensors of different aircraft or the same aircraft under different flight states, and the relevant detection results have good interpretability and credibility; By using the deep neural network method, intelligent detection of pneumatic sensor faults of different aircraft or the same aircraft under different flight states can be realized without parameter adjustment.

Description

Technical Field

[0001] The present invention belongs to the field of aviation technology, and particularly relates to an intelligent detection method for aircraft pneumatic sensor faults. Background Art

[0002] A large number of pneumatic sensors are installed in civil airliners, military fighter jets, general aircraft, etc. These sensors measure pneumatic information such as the speed and angle of attack of the aircraft, which is crucial for ensuring the safe operation of the aircraft. However, pneumatic sensors are generally assembled on the outer surface of the aircraft and are prone to failure due to rain and icing; these failures seriously affect the operation of the aircraft; therefore, aircraft pneumatic sensor fault detection technology is very necessary.

[0003] The existing aircraft pneumatic sensor fault detection technology in the industrial field is developed based on hardware redundancy, that is, multiple sets of sensors are assembled on the aircraft, redundant measurements are carried out for the same pneumatic information, and each set of sensors is monitored through voting logic to detect faults. However, this technology requires the assembly of multiple sets of sensors, resulting in high costs; flight accidents caused by pneumatic sensor faults in recent years, such as A330 and B737MAX, also indicate that the existing aircraft pneumatic sensor fault detection technology in the industrial field still has deficiencies.

[0004] Different from hardware redundancy, there are currently a large number of aircraft pneumatic sensor fault detection technologies based on software redundancy; however, these methods mostly rely on the dynamic and kinematic models and parameters of the aircraft, and need to adjust parameters separately for different aircraft or different flight states of the same aircraft, resulting in a large amount of work and low engineering generality, and have not been widely applied.

[0005] In the research on fault detection of mechanical equipment, aero-engines, etc., a large number of sensor fault detection methods using deep neural networks have emerged. These methods have good generality for different equipment or different operating states of the same equipment, and the fault detection accuracy is relatively high. However, these methods are widely regarded as "black boxes", and the algorithm framework and internal operation mechanism are not clear, which affects the interpretability mechanism analysis, reduces the engineering credibility, and restricts their application in the engineering field. Summary of the Invention

[0006] The present invention provides an intelligent detection method for aircraft pneumatic sensor faults, aiming to solve the above existing problems.

[0007] The present invention is implemented as follows. An intelligent detection method for aircraft pneumatic sensor faults includes the following steps:

[0008] S1. Obtain flight data of aircraft pneumatic sensor faults in different flight states of various aircraft;

[0009] S2. Use a controlled numerical experiment to select data directly related to the fault detection of aircraft aerodynamic sensors from the flight data;

[0010] S3. Preprocess the selected data, and stack the preprocessed data into an "image"; among them, the flight data includes simulated flight data and real flight data, and the preprocessed flight data is divided into a training set and a test set, and all the preprocessed real flight data is located in the test set;

[0011] S4. Use a convolutional neural network and a long short-term memory network to construct a deep neural network model, and perform an interpretability analysis on the network structure of the deep neural network model;

[0012] S5. Use the samples in the training set to train the constructed deep neural network model, use the Adams algorithm to iteratively update the network weights based on the training data, and set the learning efficiency to decay with the number of training iterations until the loss value trend of the deep neural network model converges;

[0013] S6. Use the trained deep neural network model to segment the samples in the test set to generate a segmentation result map.

[0014] Further, in step S4, it specifically includes: first, use a sufficient number of convolutional kernel numbers and long short-term memory network node numbers to determine the convolutional kernel dimension; then, use the selected convolutional kernel dimension and a sufficient number of long short-term memory network node numbers to determine the convolutional kernel number; finally, use the selected convolutional kernel dimension and number to determine the long short-term memory network node number.

[0015] Further, in step S4, it also includes: through the method of feature visualization, determine the process of the convolutional kernel extracting features from the image data input to the network; then, associate the effective operations of all convolutional kernels inside the network to the input data of the network to determine the input data area concerned by the convolutional operations inside the network.

[0016] Further, the convolutional kernel dimension is 2*2, the number of convolutional kernels is 48, and the number of long short-term memory network nodes is 16.

[0017] Further, in step S2, the selected data are: aircraft speed, angle of attack, sideslip angle, and the angular velocity and acceleration of the three axes within the aircraft body axis system.

[0018] Further, the aircraft speed, angle of attack, and sideslip angle extract features using a convolutional neural network and a long short-term memory network at the same time, and the angular velocity and acceleration of the three axes within the aircraft body axis system extract features using a long short-term memory network.

[0019] Further, in step S3, the preprocessing steps specifically include: using a spatio-temporal stacking method to downsample all selected data, normalizing it according to the state, and then stacking it to form a two-dimensional matrix form.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention discloses an intelligent detection method for aircraft pneumatic sensor faults, which can accurately detect pneumatic sensor faults of different aircraft or the same aircraft under different flight states, and the relevant detection results have good interpretability and credibility; by using the deep neural network method, intelligent detection of pneumatic sensor faults of different aircraft or the same aircraft under different flight states can be realized without parameter adjustment; at the same time, referring to technical methods in fields such as image processing, research on the interpretability analysis mechanism of the deep neural network is carried out, effectively exploring the interpretability analysis mechanism of this method and improving its engineering credibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the network structure of the present invention;

[0022] Figure 2 Schematic diagram of data preprocessing of the present invention;

[0023] Figure 3 Schematic diagram of the attenuation of the network training learning efficiency of the present invention;

[0024] Figure 4 Effect diagram of the convolution kernel dimension setting of the network structure of the present invention;

[0025] Figure 5 Effect diagram of the convolution kernel number setting of the network structure of the present invention;

[0026] Figure 6 Effect diagram of the network node number setting of the network structure of the present invention;

[0027] Figure 7 Schematic diagram of the ablation learning of the network structure of the present invention;

[0028] Figure 8 Schematic diagram of the visualization of convolution kernel feature extraction of the present invention;

[0029] Figure 9 Schematic diagram of the classification activation mapping of the convolution layer of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0031] In the description of the present invention, the meaning of "a plurality" is two or more, unless otherwise specifically defined.

[0032] Embodiment

[0033] As Figure 1 shown, the present invention provides a technical solution: an intelligent detection method for aircraft pneumatic sensor faults, including the following steps:

[0034] S1. Obtain flight data of aircraft pneumatic sensor faults in various flight states of multiple aircraft;

[0035] S2. Adopt a control numerical experiment to select data directly related to the detection of aircraft pneumatic sensor faults from the flight data;

[0036] S3. Preprocess the selected data, and stack the preprocessed data into an "image"; wherein, the flight data includes simulated flight data and real flight data, and the preprocessed flight data is divided into a training set and a test set, and all the preprocessed real flight data is located in the test set;

[0037] S4. Adopt a convolutional neural network and a long short-term memory network to construct a deep neural network model, and perform interpretability analysis on the network structure of the deep neural network model;

[0038] S5. Train the constructed deep neural network model with the samples in the training set, adopt the Adams algorithm to iteratively update the network weights based on the training data, and set the learning efficiency to decay with the number of training iterations until the loss value trend of the deep neural network model converges;

[0039] S6. Use the trained deep neural network model to segment the samples in the test set to generate a segmentation result map.

[0040] In this embodiment, it is necessary to collect and summarize data of multiple types of aircraft in different flight states and various pneumatic sensor fault conditions, and form a relevant data set; to avoid the "overfitting" problem in the training of the deep neural network, the data set needs to be segmented into training data and test data.

[0041] It is also necessary to determine the input state of the deep neural network directly related to the detection of aircraft pneumatic sensor faults; this generally requires a control numerical experiment to select directly related states as network inputs from the measurable flight states of the aircraft.

[0042] Among them, the data input of the convolutional neural network is generally in the form of an image; therefore, it is also necessary to preprocess the selected data and stack it into an "image".

[0043] It is necessary to expand the structural design and training of the deep neural network; the network is intended to use both convolutional neural network and long short-term memory network simultaneously, and it is necessary to determine the network structure with the highest accuracy and the simplest possible structure through comparative numerical experiments.

[0044] Through the above steps, it is possible to realize technical points such as data acquisition, data preprocessing, network training and testing in the research and development of aircraft aerodynamic sensor fault detection, and a deep neural network for aircraft aerodynamic sensor fault detection with good accuracy can be obtained.

[0045] The network interpretability analysis in the present invention is intended to be carried out from two aspects of "large" and "small" structures.

[0046] The "large" structure refers to the network structure parameters such as the convolutional kernel dimension, the number of convolutional kernels, and the number of long short-term memory network nodes of the fault detection deep neural network, which can be optimized by objective methods. For the unified training and test data, taking the fault detection accuracy of the test data as the criterion, the "large" structure of the network is determined through comparative numerical experiments.

[0047] In addition, considering the correlation between each "large" structure parameter and the network detection result, the idea of the control variable method is adopted, that is, first use a sufficient number of convolutional kernels and the number of long short-term memory network nodes to determine the convolutional kernel dimension; then use the selected convolutional kernel dimension and a sufficient number of long short-term memory network nodes to determine the number of convolutional kernels; finally, use the selected convolutional kernel dimension and number to determine the number of long short-term memory network nodes.

[0048] The "small" structure mainly refers to the key operations of the network such as convolutional kernels. Referring to the methods in the fields of image processing, etc., first, through the method of feature visualization, the process of the convolutional kernel extracting features from the image data input to the network is determined; then, referring to the method of class activation mapping diagram commonly used in the field of image processing, the effective operations of all convolutional kernels inside the network are associated with the input data of the network; thus, the area of the input data concerned by the convolutional operation inside the network can be determined. If this area coincides with the actual position where the fault occurs, it can effectively illustrate that the convolutional kernel operation of the network is reasonable, that is, the internal operation of the network has good credibility.

[0049] Experimental example

[0050] The present invention has collected data of 5 types of different aircraft under 6 flight states and 5 types of aerodynamic sensor fault types in total, as shown in Table 1 in detail. Among them, B1 and B2 are large airliners, F is a fighter plane, Y is a transport plane, and D is a general-purpose plane. The B1 cruise and Y manual operation are simulation data, and the B2 takeoff and landing, F manual operation, D cruise and B1 manual operation are real flight data. Referring to relevant literature materials, the 5 types of aerodynamic sensor faults respectively include the failure of the airspeed tube, angle of attack, and sideslip angle sensors, as well as the abnormal noise of the angle of attack and sideslip angle sensors. These faults have all caused serious flight accidents.

[0051] To avoid the problem of "overfitting" in the network, the collected data is divided into training data and test data. To focus on testing the performance of the network for real flight data, all real flight data is placed in the test data and does not participate in network training. The division of network training and test data is shown in Table 1.

[0052]

[0053] Table 1

[0054] "-" indicates that the altitude data was not recorded during real flight; fault category "0" indicates no fault; fault categories "1 - 5" respectively represent the failure of speed, angle of attack, sideslip angle sensors and the abnormal noise faults of angle of attack and sideslip angle sensors.

[0055] The output of the network corresponds to the types of pneumatic sensor faults in Table 1, a total of 6 categories (1 category of flight without fault and 5 categories of flight with faults). To effectively establish the neural network structure, it is necessary to select the flight state of the aircraft that is directly related to the network output. Considering that the aircraft is equipped with both pneumatic and inertial measurement sensors, the evolution equations of the measurement states (speed V, angle of attack α, sideslip angle β) of the pneumatic sensors are as follows:

[0056]

[0057]

[0058]

[0059] Among them, g is the acceleration due to gravity, S * , C * respectively represent sine and cosine operations, {w x , w y , w z}, {A x , A y , A z} respectively represent the angular velocities and accelerations of the three axes in the body axis system of the aircraft, which can be directly measured by inertial measurement sensors.

[0060] The coupling relationship between the three Euler attitude angles {ψ, θ, φ} of the aircraft and the angular velocities of the three axes of the aircraft is:

[0061]

[0062]

[0063]

[0064] In the present invention, the final network structure selects three aerodynamic quantities, namely aircraft speed, angle of attack, and sideslip angle, and six measurement values of the inertial sensors as the network inputs.

[0065] As Figure 2 shown, for the convenience of convolutional kernel scanning and feature extraction, it is also necessary to encapsulate the network inputs in an image format. For this purpose, a spatio-temporal stacking method is adopted to downsample all the states of the network inputs (three aerodynamic quantities of speed, angle of attack, and sideslip angle, and six measurement values of the inertial sensors), and after normalizing according to the states, stack them into a two-dimensional matrix form. Thus, the preprocessing of the network input data can be completed.

[0066] As Figure 3 shown, during the training process of the network, the training data in Table 1 is directly called, and the Adams training algorithm is adopted, setting the learning efficiency (learning rate) to decay with the number of training iterations. During the network training, on a GTX3090 graphics card, using Keras (TensorFlow), the training of more than 1000 epochs can be completed within 2 hours, achieving good results.

[0067] In the network test, the test data in Table 1 is directly called, and the network performance is judged according to the detection accuracy of different fault types. The network after training can obtain an aerodynamic sensor fault detection accuracy of not less than 90% for all 4 types of test aircraft (Table 2). The fault intelligent detection method proposed by the present invention has excellent effects.

[0068]

[0069] Table 2

[0070] Table 2 gives the test confusion matrices corresponding to 4 flight states; each matrix horizontally corresponds to 6 fault states (1 fault-free state and 5 fault states), and vertically corresponds to the network detection results; then the diagonal (which has been bolded) is the detection accuracy of the network for each fault state.

[0071] The interpretability analysis of the network "large" structure is coupled with the network structure design and training process. On the one hand, it is used to determine the network "large" structure parameters such as the optimal convolutional kernel dimension, the number of convolutional kernels, and the number of convolutional layers of the network; at the same time, this process also effectively reflects the interpretability mechanism of the parameter setting related to the network "large" structure. The interpretability analysis of the network "large" structure is shown in Figures 4 - 6 .

[0072] As Figure 4 shown, it is expanded for the convolutional kernel dimension; among them, the number of convolutional kernels and the number of convolutional layers are both based on the previous research results and adopt a sufficient quantity. From Figure 4It can be seen that after training, the 2×2 convolutional kernel dimension has the optimal test performance; therefore, the 2×2 convolutional kernel is adopted in the final network structure.

[0073] As Figure 5 shown, for the expansion of the number of convolutional kernels, the convolutional kernel dimension directly adopts the 2×2 determined in Figure 4 , and the number of convolutional layers also adopts a sufficient number. It can be seen from Figure 5 that after training, the number of 48 convolutional kernels has the optimal test performance. Therefore, in the final network structure, the number of convolutional kernels in the convolutional layer is 48.

[0074] As Figure 6 shown, for the expansion of the number of nodes in the long short-term memory network, the convolutional kernel dimension directly adopts the 2×2 determined in Figure 4 , and the number of convolutional kernels adopts the 48 determined in Figure 5 . It can be seen from Figure 6 that the number of 16 nodes in the long short-term memory network has the optimal test performance. Therefore, in the final network structure, the number of nodes in the long short-term memory network is 16.

[0075] For the final formed network structure, see Figure 1 . For the inputs of speed, angle of attack, and sideslip angle, both convolutional neural network (CNN) and long short-term memory network (LSTM) are used to extract features; for the inputs of angular velocity and acceleration of the inertial measurement unit, only the long short-term memory network is used to extract features. To test the effectiveness of the "large" structure of this network, the ablation learning method is adopted, that is, each branch of the optimal network structure (DNN-opt) is deleted and then trained.

[0076] It can be seen from Figure 7 that deleting any branch will result in a decrease in the network detection accuracy, which proves the rationality of the DNN-opt network structure from the opposite side.

[0077] See Figure 8 . The area where the sensor failure occurs (the sideslip angle sensor fails) is marked within the red frame; Figure 8 shows the features extracted by the first convolutional layer of DNN-opt. It can be seen from Figure 8 that within the area corresponding to the failure, the convolutional kernel can extract relatively obvious features. Figure 8 The convolutional kernel features in

[0078] See Figure 9, to display the features extracted by the convolution kernels of all convolutional layers of the DNN-opt network, the Class Activation Mapping (CAM) method can be used. In the figure, the features extracted by each convolutional layer are mapped to the network input, and a heatmap is drawn; the areas of focus in the heatmap coincide with the areas where various faults occur, which also confirms that the convolution kernels in DNNopt can effectively extract fault features, that is, the structure of DNN-opt is reasonable.

[0079] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent detection method for aircraft pneumatic sensor faults, characterized in that, it includes the following steps: S1. Obtain flight data of aircraft pneumatic sensor faults under various flight states of the aircraft; S2. Adopt a comparative numerical experiment to select data directly related to the detection of aircraft pneumatic sensor faults from the flight data; S3. Preprocess the selected data, and the preprocessed data is stacked and formed into an "image"; wherein, the flight data includes simulated flight data and real flight data, and the preprocessed flight data is divided into a training set and a test set, and all the preprocessed real flight data is located in the test set; S4. Use a convolutional neural network and a long short-term memory network to construct a deep neural network model, and perform interpretability analysis on the network structure of the deep neural network model; S5. Use the samples in the training set to train the constructed deep neural network model, adopt the Adams algorithm to iteratively update the network weights based on the training data, and set the learning efficiency to decay with the number of training iterations until the loss value trend of the deep neural network model converges; S6. Use the trained deep neural network model to segment the samples in the test set to generate a segmentation result map; The evolution equations of the pneumatic sensor measurement states, namely the velocity V, angle of attack α, and sideslip angle β, are: where g is the acceleration due to gravity, S * , C * respectively represent sine and cosine operations, {w x , w y , w z}, {A x , A y , A z} respectively represent the angular velocities and accelerations of the three axes in the aircraft body axis system, which are directly measured by inertial measurement sensors; The coupling relationship between the three Euler attitude angles {ψ, θ, φ} of the aircraft and the three-axis angular velocity of the aircraft is:

2. The intelligent detection method for aircraft pneumatic sensor faults according to claim 1, characterized in that, in step S4, it specifically includes: first, use a sufficient number of convolutional kernel numbers and long short-term memory network nodes to determine the convolutional kernel dimension; then, adopt the selected convolutional kernel dimension and a sufficient number of long short-term memory network nodes to determine the convolutional kernel number; finally, adopt the selected convolutional kernel dimension and number to determine the long short-term memory network nodes.

3. The intelligent detection method for aircraft pneumatic sensor faults according to claim 2, characterized in that, in step S4, it further includes: by means of feature visualization, determine the process of convolutional kernels extracting features from the image data input to the network; then, associate the effective operations of all convolutional kernels inside the network to the input data of the network to determine the input data regions concerned by the convolutional operations inside the network.

4. The intelligent detection method for aircraft pneumatic sensor faults according to claim 3, characterized in that: The convolutional kernel dimension is 2*2, the number of convolutional kernels is 48, and the number of long short-term memory network nodes is 16.

5. The intelligent detection method for aircraft pneumatic sensor faults according to claim 1, characterized in that, in step S2, the selected data are: aircraft speed, angle of attack, sideslip angle, and the three-axis angular velocity and acceleration within the aircraft body axis system.

6. The intelligent detection method for aircraft pneumatic sensor faults according to claim 5, characterized in that, The aircraft speed, angle of attack, and sideslip angle simultaneously extract features using a convolutional neural network and a long short-term memory network, and the three-axis angular velocity and acceleration within the aircraft body axis system extract features using a long short-term memory network.

7. An intelligent detection method for aircraft pneumatic sensor faults according to claim 1, characterized in that, in step S3, the preprocessing steps specifically include: adopting a spatio-temporal stacking method, downsampling all selected data, normalizing according to the state, and stacking to form a two-dimensional matrix form.

Images