A method and system for detecting aircraft flight anomalies

Abstract

The application provides an aircraft flight anomaly detection method, comprising: obtaining historical flight monitoring data; based on a predefined highest risk stage of aircraft flight and a key monitoring parameter, extracting corresponding original data from the historical flight monitoring data and performing data preprocessing; using a principal component analysis method, performing dimension reduction processing on the preprocessed original data to obtain flight samples; constructing an OC-DNN anomaly detection model and training, and further using the flight samples to test the trained OC-DNN anomaly detection model to obtain an anomaly detection result of each flight sample. The application also provides an aircraft flight anomaly detection system. Based on the anomaly detection technology, the application can identify abnormal flight from conventional flight data without predefined safety standards, and can effectively improve the accuracy of aircraft flight anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of computer data detection technology, and in particular to a method and system for detecting aircraft flight anomalies. Background Technology

[0002] Currently, airlines proactively identify abnormal flights by analyzing the Flight Data Recorder (FDR) data in the Flight Operation Quality Assurance (FOQA) program. The FOQA program mainly uses out-of-limit detection methods to identify abnormal flight data.

[0003] Exceedance detection methods detect predefined abnormal aircraft operation events by monitoring whether specific flight parameters exceed predefined limits under specific conditions. While they perform well on known safety issues, they ignore safety critical conditions that do not belong to the predefined set of abnormal events when capturing flight data records.

[0004] In recent years, with the increasing availability of data and computing resources, the application of anomaly detection technology in the aerospace field has seen tremendous growth. Given the advantage of anomaly detection technology in systematically utilizing available information across the entire dataset and effectively detecting various anomalies without any pre-defined restrictions, there is an urgent need to apply it to aircraft anomaly flight identification. This would enable the identification of anomalous flights within regular flight data without predefined safety standards, effectively improving the accuracy of aircraft flight anomaly identification. Summary of the Invention

[0005] The technical problem to be solved by the embodiments of the present invention is to provide an aircraft flight anomaly detection method and system. Based on anomaly detection technology, it can identify abnormal flight from conventional flight data without predefined safety standards, and can effectively improve the accuracy of aircraft flight anomaly identification.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for detecting aircraft flight anomalies, the method comprising the following steps:

[0007] Acquire historical flight monitoring data;

[0008] Based on the predefined highest risk phase of aircraft flight and key monitoring parameters, the corresponding raw data is extracted from the historical flight monitoring data and preprocessed.

[0009] Principal component analysis was used to reduce the dimensionality of the preprocessed raw data to obtain flight samples.

[0010] An OC-DNN anomaly detection model is constructed and trained, and then the trained OC-DNN anomaly detection model is tested using the flight samples to obtain the anomaly detection results for each flight sample.

[0011] The historical data was collected through the aircraft's flight data recorder.

[0012] The highest-risk phases of aircraft flight include the takeoff phase and the approach / landing phase; the key monitoring parameters include altitude, true airspeed, vertical speed, pitch angle, roll angle, and power.

[0013] The preprocessing steps for the raw data include denoising, interpolation, and normalization.

[0014] The specific steps for using principal component analysis to reduce the dimensionality of the preprocessed raw data to obtain flight samples include:

[0015] Step 1: Let the feature vector of each flight sample be obtained by concatenating each parameter data into a single vector. in, It is the original data of the i-th parameter at the j-th time point among the key monitoring parameters corresponding to a certain stage of the highest risk phase of aircraft flight after preprocessing;

[0016] Step 2: Normalize according to the z-score formula The data are standardized to the same range; where n is the number of flights collected, j is the length of all flight distances from the start of the collection time point, i is the parameter number, and σ represents the standard deviation of the observations. The average value of a specific parameter i for j unit time lengths at the start time points of all flight distances;

[0017] Step 3: Calculate the correlation coefficient matrix R = (r jq ) k×k ,in, Let r be the correlation coefficient between parameters j and q at time j, and r ii =1, r iq =r qi ;

[0018] Step 4: Calculate the eigenvalues ​​of the correlation coefficient matrix R: λ1≥λ2≥…≥λ k ≥0, and the corresponding eigenvectors u1, u2, ..., u k , where u i =(u 1i ,u 2i ,…,u ji );

[0019] Step 5: Based on the results of Step 4, we can obtain P new variables composed of the feature vectors. Where m′ P It is the Pth principal component;

[0020] Step Six: Obtain each flight sample and denote it as f. n ′=[m1 (n)′ m2 (n)′ ,…,m p (n)′ ].

[0021] The OC-DNN anomaly detection model is described as follows:

[0022] Where w is the scalar output obtained from the hidden to the output layer; V is the weight matrix from the input to the hidden unit; r is the hyperplane bias; ν is a trade-off parameter used to control the maximum distance between the hyperplane and the origin and the number of data points allowed to cross the hyperplane; and g(·) is the sigmoid activation function.

[0023] The specific training steps for the OC-DNN anomaly detection model are as follows:

[0024] Step 1: Update parameters w and V; where, given r, use the objective function. To update parameters w and V, where,

[0025] Step 2: Update parameter r; where, based on w and V obtained in Step 1, further update parameter r according to the objective function. We obtain that the parameter r is The νth quartile;

[0026] Step 3: Obtain the trained OC-DNN anomaly detection model.

[0027] The specific steps for testing the trained OC-DNN anomaly detection model using the flight samples to obtain the anomaly detection results for each flight sample include:

[0028] Step 1: Input flight sample f n ′, initialize r to r (0) ;

[0029] Step 2: Update parameters w, V, and r until convergence is achieved;

[0030] Step 3: Based on the parameter update results, obtain the predicted value for each input flight sample.

[0031] Step 4: Calculate the decision score parameters for each sample.

[0032] Step 5: Perform anomaly detection based on sample decision scores. When a flight sample score S... n When ≥0, the flight sample f n ′ represents a normal sample; when a certain flight sample scores S n When <0, the flight sample f n ′ represents an abnormal sample.

[0033] The method further includes:

[0034] Based on the preprocessed raw data, a reference profile is generated using statistical methods. Combined with the anomaly detection results of each flight sample, each abnormal flight sample is compared with the reference nominal profile to achieve visualization processing of abnormal flights.

[0035] This invention also provides an aircraft flight anomaly detection system, comprising:

[0036] The data acquisition unit is used to acquire historical flight monitoring data;

[0037] The data extraction unit is used to extract the corresponding raw data from the historical flight monitoring data and perform data preprocessing based on the predefined highest risk phase of aircraft flight and key monitoring parameters.

[0038] The data analysis unit is used to perform dimensionality reduction on the preprocessed raw data using principal component analysis to obtain flight samples.

[0039] The data anomaly detection unit is used to construct and train an OC-DNN anomaly detection model, and further use the flight samples to test the trained OC-DNN anomaly detection model to obtain the anomaly detection result for each flight sample.

[0040] Implementing the embodiments of the present invention has the following beneficial effects:

[0041] This invention is based on anomaly detection technology. It collects historical flight monitoring data from the aircraft flight data recorder and extracts key monitoring parameters based on the highest risk phase of the aircraft flight. Then, it obtains flight samples through principal component analysis and further uses a pre-trained OC-DNN anomaly detection model for testing to obtain the anomaly detection results for each flight sample. Thus, it can be driven by data from the flight data recorder without the need for predefined standards, which effectively improves the accuracy of aircraft flight anomaly identification and has important theoretical and application value for the continuous airworthiness management of aircraft. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0043] Figure 1 A flowchart of an aircraft flight anomaly detection method provided in an embodiment of the present invention;

[0044] Figure 2 A statistical chart showing the percentage of fatal accidents and deaths at Boeing from 2009 to 2018 in an application scenario of an aircraft flight anomaly detection method provided in an embodiment of the present invention.

[0045] Figure 3 Histogram of decision scores for abnormal and normal flight samples in an application scenario of an aircraft flight anomaly detection method provided in this embodiment of the invention;

[0046] Figure 4 This is a reference nominal cross-sectional view of takeoff altitude in an application scenario of an aircraft flight anomaly detection method provided in an embodiment of the present invention.

[0047] Figure 5 This is a visualization of abnormal flight data samples in an application scenario of an aircraft flight anomaly detection method provided in an embodiment of the present invention.

[0048] Figure 6 This is a schematic diagram of an aircraft flight anomaly detection structure provided in an embodiment of the present invention. Detailed Implementation

[0049] 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.

[0050] like Figure 1 As shown in the figure, an aircraft flight anomaly detection method is provided in an embodiment of the present invention. The method includes the following steps:

[0051] Step S1: Obtain historical flight monitoring data;

[0052] Step S2: Based on the predefined highest risk phase of aircraft flight and key monitoring parameters, extract the corresponding raw data from the historical flight monitoring data and perform data preprocessing.

[0053] Step S3: Use principal component analysis to reduce the dimensionality of the preprocessed raw data to obtain flight samples;

[0054] Step S4: Construct and train an OC-DNN anomaly detection model, and further use the flight samples to test the trained OC-DNN anomaly detection model to obtain the anomaly detection results for each flight sample.

[0055] The specific process is as follows: Before step S1, high-risk flight phases, i.e., the highest-risk phases of aircraft flight, are predefined by analyzing publicly available accident statistics reports, including the takeoff phase and the approach / landing phase. By analyzing safety information documents available in the public domain, key monitoring parameters are selected for the takeoff and approach / landing phases. These key monitoring parameters include altitude, true airspeed (TAS), vertical speed, pitch angle, roll angle, and power (N1).

[0056] In step S1, historical flight monitoring data is collected using the aircraft flight data recorder.

[0057] In step S2, based on the highest risk phase of the aircraft flight and key monitoring parameters, the corresponding raw data is extracted from the historical flight monitoring data and preprocessed. The preprocessing steps for the raw data include denoising, interpolation, and normalization.

[0058] In step S3, the dimensionality reduction of the preprocessed raw data is performed using principal component analysis to obtain the flight samples. The specific steps include:

[0059] Step 1: Let the feature vector of each flight sample be obtained by concatenating each parameter data into a single vector. in, It is the original data of the i-th parameter at the j-th time point among the key monitoring parameters corresponding to a certain stage of the highest risk phase of aircraft flight after preprocessing;

[0060] Step 2: Normalize according to the z-score formula The data are standardized to the same range; where n is the number of flights collected, j is the length of all flight distances from the start of the collection time point, i is the parameter number, and σ represents the standard deviation of the observations. The average value of a specific parameter i for j unit time lengths at the start time points of all flight distances;

[0061] Step 3: Calculate the correlation coefficient matrix R = (r jq ) k×k ,in, Let r be the correlation coefficient between parameters j and q at time j, and r ii =1, r iq =rqi ;

[0062] Step 4: Calculate the eigenvalues ​​of the correlation coefficient matrix R: λ1≥λ2≥…≥λ k ≥0, and the corresponding eigenvectors u1, u2, ..., u k , where u i =(u 1i ,u 2i ,…,u ji );

[0063] Step 5: Based on the results of Step 4, we can obtain P new variables composed of the feature vectors. Where m′ P It is the Pth principal component;

[0064] Step Six: Obtain each flight sample and record it as...

[0065] In step S4, an OC-DNN (one-class deep learning neural network) anomaly detection model is first constructed. This OC-DNN anomaly detection model is described as follows:

[0066] Where w is the scalar output obtained from the hidden to the output layer; V is the weight matrix from the input to the hidden unit; r is the hyperplane bias; ν is a trade-off parameter used to control the maximum distance between the hyperplane and the origin and the number of data points allowed to cross the hyperplane; and g(·) is the sigmoid activation function.

[0067] Secondly, the OC-DNN anomaly detection model is trained, and the specific steps are as follows:

[0068] Step 1: Update parameters w and V; where, given r, use the objective function. To update parameters w and V, where,

[0069] Step 2: Update parameter r; where, based on w and V obtained in Step 1, further update parameter r according to the objective function. We obtain that the parameter r is The νth quartile;

[0070] Step 3: Obtain the trained OC-DNN anomaly detection model.

[0071] Finally, the trained OC-DNN anomaly detection model is tested using flight samples to obtain anomaly detection results for each flight sample. The specific steps include:

[0072] Step 1: Input flight sample f n ′, initialize r to r (0) ;

[0073] Step 2: Update parameters w, V, and r until convergence is achieved;

[0074] Step 3: Based on the parameter update results, obtain the predicted value for each input flight sample.

[0075] Step 4: Calculate the decision score parameters for each sample.

[0076] Step 5: Perform anomaly detection based on sample decision scores. When a flight sample score S... n When ≥0, the flight sample f n ′ represents a normal sample; when a certain flight sample scores S n When <0, the flight sample f n ′ represents an abnormal sample.

[0077] In this embodiment of the invention, abnormal flight visualization processing is achieved. The method specifically includes: generating a reference profile based on the preprocessed raw data using statistical methods, and comparing each abnormal flight sample with the reference nominal profile in combination with the abnormal detection results of each flight sample, thereby achieving abnormal flight visualization processing.

[0078] like Figures 2 to 5 As shown, the application scenario of an aircraft flight anomaly detection method in an embodiment of the present invention will be further explained as follows:

[0079] Step 1: Identifying High-Risk Flight Phases. A chart showing Boeing's fatal accidents and death percentages from 2009 to 2018 is shown. Figure 2 As shown, although the takeoff and approach / landing phases account for only 6% of the total flight time in a 1.5-hour flight, they account for 52% of all fatalities and 64% of all fatal accidents. The takeoff and approach / landing phases can be considered the riskiest phases of the entire flight.

[0080] The second step: Selection of key monitoring parameters. By analyzing available safety information documents in the public domain, including FAA Advisory Circular AC 120-82, FAA-established stable approach standards, and parameters recorded by flight data recorders, six key monitoring parameters were ultimately determined: altitude, true airspeed (TAS), vertical speed, pitch angle, roll angle, and power (N1).

[0081] The third step: Historical flight monitoring data collection. Flight monitoring data was collected from 1,034 flights of four A320 aircraft operated by the airline throughout 2018. Each flight monitoring data point included the entire flight process from takeoff to approach / landing, involving a total of 2,667 parameters.

[0082] Step 4: Data Preprocessing. Based on the key monitoring parameters selected in Step 2, relevant flight data within 120 seconds after the aircraft's takeoff power is applied are extracted, ultimately generating a dataset of 1034 raw data points, each 120×7 in size, with the first column representing the time parameter.

[0083] First, a simple moving average filter is used to eliminate any noise that may be present in the original data. Then, "TIME" is selected as the time axis, and spline interpolation is performed on the collected aircraft takeoff data in 1-second increments. Finally, a normalization method is used to reduce the impact of differences in the absolute values ​​of different measurements.

[0084] Step 5: Principal Component Analysis. Based on the data preprocessing results in Step 4, principal component analysis is used to reduce the dimensionality of the data. The size of the flight sample after principal component analysis is 120×1.

[0085] Step 6: Construction and training of the OC-DNN anomaly detection model.

[0086] Step 7: Testing the OC-DNN anomaly detection model. Two abnormal flight simulation datasets and one normal flight simulation dataset are generated. These simulation datasets, along with collected historical monitoring data, are used as test data and input into the trained OC-DNN model to obtain the decision score for each flight. Histograms of the decision scores for abnormal and normal flight samples are also generated, as shown below. Figure 3 As shown.

[0087] Step 8: First, generate the reference profile. Based on the data preprocessing results in Step 4, the generated reference nominal profile of the aircraft's flight altitude during takeoff is as follows: Figure 4As shown. The black dashed line represents the average of all flight samples, and the shaded area shows the distribution of flight data records. Dark gray represents the 50th percentile of the flight data records, and light gray represents the 90th percentile of the flight data records.

[0088] Finally, the abnormal flight visualization process compares each abnormal flight sample with a reference nominal profile to achieve visualization. The visualization results for each key monitored parameter are as follows: Figure 5 As shown.

[0089] like Figure 6 As shown in the figure, an aircraft flight anomaly detection system is provided in an embodiment of the present invention, comprising:

[0090] Data acquisition unit 110 is used to acquire historical flight monitoring data;

[0091] The data extraction unit 120 is used to extract the corresponding raw data from the historical flight monitoring data and perform data preprocessing based on the predefined highest risk phase of aircraft flight and key monitoring parameters.

[0092] The data analysis unit 130 is used to perform dimensionality reduction on the preprocessed raw data using principal component analysis to obtain flight samples.

[0093] The data anomaly detection unit 140 is used to construct and train an OC-DNN anomaly detection model, and further use the flight samples to test the trained OC-DNN anomaly detection model to obtain the anomaly detection result for each flight sample.

[0094] Implementing the embodiments of the present invention has the following beneficial effects:

[0095] This invention is based on anomaly detection technology. It collects historical flight monitoring data from the aircraft flight data recorder and extracts key monitoring parameters based on the highest risk phase of the aircraft flight. Then, it obtains flight samples through principal component analysis and further uses a pre-trained OC-DNN anomaly detection model for testing to obtain the anomaly detection results for each flight sample. Thus, it can be driven by data from the flight data recorder without the need for predefined standards, which effectively improves the accuracy of aircraft flight anomaly identification and has important theoretical and application value for the continuous airworthiness management of aircraft.

[0096] It is worth noting that the various units included in the above system embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0097] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as ROM / RAM, disk, optical disk, etc.

[0098] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for detecting aircraft flight anomalies, characterized in that, The method includes the following steps: Acquire historical flight monitoring data; Based on the predefined highest risk phase of aircraft flight and key monitoring parameters, the corresponding raw data is extracted from the historical flight monitoring data and preprocessed. Principal component analysis was used to reduce the dimensionality of the preprocessed raw data to obtain flight samples. An OC-DNN anomaly detection model is constructed and trained, and the trained OC-DNN anomaly detection model is further tested using the flight samples to obtain the anomaly detection results for each flight sample. The OC-DNN anomaly detection model is described as follows: ; in, w It is a scalar output obtained from the hidden layer to the output layer; V It is the weight matrix from the input to the hidden unit; r It is the deviation of the hyperplane; ν It is a trade-off parameter used to control the maximum distance between the hyperplane and the origin and the number of data points allowed to cross the hyperplane; It is the sigmoid activation function; The specific steps for testing the trained OC-DNN anomaly detection model using the flight samples to obtain the anomaly detection results for each flight sample include: Step 1: Input flight sample ,initialization r for r (0) ; Step 2: Update parameters w , V and r until convergence is satisfied; Step 3: Based on the parameter update results, obtain the predicted value for each input flight sample. ; Step 4: Calculate the decision score parameters for each sample. ; Step 5: Perform anomaly detection based on sample decision scores. When a flight sample scores... At that time, the flight sample This is a normal sample; when a certain flight sample scores... At that time, the flight sample This is an abnormal sample.

2. The aircraft flight anomaly detection method as described in claim 1, characterized in that, The historical flight monitoring data was collected through the aircraft flight data recorder.

3. The aircraft flight anomaly detection method as described in claim 2, characterized in that, The highest-risk phases of aircraft flight include the takeoff phase and the approach / landing phase; the key monitoring parameters include altitude, true airspeed, vertical speed, pitch angle, roll angle, and power.

4. The aircraft flight anomaly detection method as described in claim 3, characterized in that, The preprocessing steps for the raw data include denoising, interpolation, and normalization.

5. The aircraft flight anomaly detection method as described in claim 1, characterized in that, The specific steps for using principal component analysis to reduce the dimensionality of the preprocessed raw data to obtain flight samples include: Step 1: Let the feature vector of each flight sample be obtained by concatenating the data of each parameter into a single vector. ,in, It is the first of the key monitoring parameters corresponding to a certain stage of the highest risk phase of aircraft flight after preprocessing. j The first time point i The original data for each parameter; Step 2: Normalize according to the z-score formula Standardize the data to the same range; among which, n The number of flights collected, j The length of the time point at which all flight distances began to be collected. i As parameter number, The standard deviation of the observed values Start time point for collecting data for all flight distances j Specific parameters for each unit of time length i The average value; Step 3: Calculate the correlation coefficient matrix ,in, For time points j Time, parameters j and parameters q The correlation coefficient between the two parameters, and , ; Step 4: Calculate the correlation coefficient matrix R eigenvalues and the corresponding feature vectors ,in ; Step 5: Based on the results of Step 4, we can obtain the eigenvectors. P A new variable is ,in For the first P principal component; Step Six: Obtain each flight sample and record it as... .

6. The aircraft flight anomaly detection method as described in claim 1, characterized in that, The specific training steps for the OC-DNN anomaly detection model are as follows: Step 1: Update parameters w and V ;wherein, in the given r Under the condition of using the objective function To update parameters w and V ,in, , , ; Step 2: Update parameters r Among them, based on the results obtained in step one w and V Further update parameters r According to the objective function Get the parameters r yes The ν One quartile; Step 3: Obtain the trained OC-DNN anomaly detection model.

7. The aircraft flight anomaly detection method as described in claim 1, characterized in that, The method further includes: Based on the preprocessed raw data, a reference profile is generated using statistical methods. Combined with the anomaly detection results of each flight sample, each abnormal flight sample is compared with the reference nominal profile to achieve visualized processing of abnormal flights.

8. An aircraft flight anomaly detection system, characterized in that, include: The data acquisition unit is used to acquire historical flight monitoring data; The data extraction unit is used to extract the corresponding raw data from the historical flight monitoring data and perform data preprocessing based on the predefined highest risk phase of aircraft flight and key monitoring parameters. The data analysis unit is used to perform dimensionality reduction on the preprocessed raw data using principal component analysis to obtain flight samples. The data anomaly detection unit is used to construct and train an OC-DNN anomaly detection model, and further use the flight samples to test the trained OC-DNN anomaly detection model to obtain the anomaly detection result for each flight sample. The OC-DNN anomaly detection model is described as follows: ; in, w It is a scalar output obtained from the hidden layer to the output layer; V It is the weight matrix from the input to the hidden unit; r It is the deviation of the hyperplane; ν It is a trade-off parameter used to control the maximum distance between the hyperplane and the origin and the number of data points allowed to cross the hyperplane; It is the sigmoid activation function; The specific steps for testing the trained OC-DNN anomaly detection model using the flight samples to obtain the anomaly detection results for each flight sample include: Step 1: Input flight sample ,initialization r for r (0) ; Step 2: Update parameters w , V and r until convergence is satisfied; Step 3: Based on the parameter update results, obtain the predicted value for each input flight sample. ; Step 4: Calculate the decision score parameters for each sample. ; Step 5: Perform anomaly detection based on sample decision scores. When a flight sample scores... At that time, the flight sample This is a normal sample; when a certain flight sample scores... At that time, the flight sample This is an abnormal sample.

Images