An approach and landing phase anomaly detection method based on flight data

Abstract

The application relates to the technical field of flight state anomaly detection, in particular to an approach and landing phase anomaly detection method based on flight data, which comprises the following steps: extracting flight data to determine direct and indirect influence parameters affecting the approach and landing of an airplane, and performing pretreatment; a multivariate time series prediction model based on StemGNN is constructed to predict and perform over-limit monitoring on the flight data; according to the flight program of the approach and landing phase, a flight state transition stationarity analysis model is established to perform flight state transition stationarity analysis on the monitoring parameters that are not over-limited; and abnormal analysis is performed according to the output result. Through the pretreatment of the flight data, the direct and indirect influence parameters affecting the flight safety in the approach and landing phase are analyzed, multivariate parameter prediction analysis is performed through the StemGNN network, and flight state stationarity analysis is performed through multi-feature fusion, so that anomaly detection and early warning are realized, and then the prediction of unsafe events can be more accurately realized.

Description

Technical Field

[0001] The present invention relates to the technical field of flight status anomaly detection, and in particular to a method for detecting anomalies in the approach and landing phase based on flight data. Background Art

[0002] Improving civil aviation flight safety is a key focus for the industry. Based on the classification of aircraft flight phases, most accidents occur during the approach and landing phases. The approach phase is the most maneuverable phase of the flight, with a high potential for error. Landing flight safety is currently the aviation industry's greatest safety challenge, as hard landings, tail strikes, and runway overruns are common during this phase. Therefore, conducting safety analyses of approach and landing flight conditions can help reduce the incidence of unsafe incidents.

[0003] Most current flight safety analysis methods are based on data-driven safety analysis of Quick Access Record (QAR) data or Flight Data Recorder (FDR) data. QAR records the aircraft's onboard equipment parameters, pilot control parameters, aircraft flight status parameters, and environmental parameters during flight. The data recorded by QAR is characterized by high frequency, diverse types, and large data volumes. Airline operators and aviation regulators conduct in-depth mining and analysis of QAR data to meet the needs of routine maintenance, flight inspections, system performance monitoring, and flight quality monitoring. Leveraging QAR-recorded flight data for in-depth analysis of aircraft safety flight patterns can reveal patterns of flight data changes during safe flight, which can then be used for real-time in-flight flight safety analysis.

[0004] Currently, civil aviation utilizes numerous safety warning and collision avoidance systems to ensure flight safety. Most of these systems predict a specific parameter through computer networks, comparing the predicted value with a pre-set threshold to determine whether it exceeds the limit and provide a flight safety warning. However, these systems fail to consider the indirect impact of other parameters on the predicted flight parameter. Therefore, using both direct and indirect influencing parameters to predict flight unsafe events can more accurately predict unsafe events.

[0005] Furthermore, the smoothness of flight state transition has a great impact on flight safety, and the change of flight mode is also reflected in the change pattern of flight parameters. Existing research rarely considers extracting effective flight state information that represents flight safety mode.

[0006] Based on this, we propose a StemGNN flight approach and landing anomaly detection method based on flight data. Summary of the Invention

[0007] The purpose of the present invention is to overcome the above problems or at least partially solve the above problems, and a method for detecting anomalies in the approach and landing phase based on flight data is proposed. For the approach and landing phase, according to the monitoring items and monitoring parameters in the "Implementation and Management of Flight Quality Assurance (FOQA)", the main events affecting approach and landing safety are analyzed as tail strikes and hard landings. According to the monitoring items of these two events, their monitoring parameters are determined to be pitch angle and vertical acceleration; according to flight dynamics, there are three types of parameters affecting pitch angle and vertical acceleration: control input parameters (throttle lever control amount, elevator deflection, aileron, rudder), aircraft state parameters (roll angle, roll angular velocity, pitch angular velocity, altitude, latitude, longitude, descent rate, corrected airspeed), and external environmental parameters (wind speed). A multivariate time series prediction model based on StemGNN is constructed, and direct parameters and indirect parameters are used as inputs of the model to predict pitch angle and vertical acceleration and perform over-limit judgment. Combined with the flight program, the multivariate time series prediction model is used to extract the flight status of the control and attitude changes in the approach and landing phase, and a smooth flight status transfer model is established to ensure the safe approach and landing of the aircraft.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting anomalies in the approach and landing phase based on flight data, comprising the following steps:

[0009] Step 1: Extract flight data for preprocessing, including the following steps:

[0010] Extract flight data from historical flights from the Quick Access Recorder;

[0011] Determine the monitoring items and parameters for the approach and landing phase in accordance with the "Implementation and Management of Flight Quality Assessment (FOQA)". The monitoring parameters include direct impact parameters and indirect impact parameters.

[0012] The direct and indirect impact parameters are processed with noise and normalized;

[0013] Step 2: Build a prediction model to predict flight data and monitor over-limits, including the following steps:

[0014] Perform sliding window processing on direct and indirect impact parameters and construct sample data as input of the prediction model;

[0015] Build a multivariate time series prediction model based on StemGNN, extracting the dependencies between sequences and the temporal correlation within sequences through graph convolutional networks and frequency domain analysis;

[0016] The model is trained using prediction loss and reconstruction loss, and the classification weighting method in cost-sensitive learning is introduced to design a loss function that pays more attention to specified features.

[0017] Compare the predicted value with the monitoring parameter to determine whether it exceeds the limit. If so, issue a safety warning. If not, proceed to step three.

[0018] Step 3: Based on the flight procedure of the approach and landing phase, a flight state transition stability analysis model is established, and flight state transition stability analysis is performed on the monitoring parameters that are within the limit;

[0019] Step 4: Perform anomaly analysis based on the output results.

[0020] In a preferred embodiment, in step 1, the monitoring items analyzed during the approach and landing phase include tail strikes and hard landing events, and the monitoring parameters of the tail strikes and hard landing events include pitch angle and vertical acceleration, which are directly influencing parameters.

[0021] In a preferred embodiment, in step 1, according to the principles of flight dynamics, the parameters affecting the pitch angle and vertical acceleration include control input parameters (throttle lever control amount, elevator deflection, aileron, rudder), aircraft state parameters (roll angle, roll angular velocity, pitch angular velocity, altitude, latitude, longitude, descent rate, corrected airspeed) and external environmental parameters (wind speed), which are indirect influencing parameters.

[0022] In a preferred embodiment, corresponding direct influencing parameters and indirect influencing parameters are determined in a quick access recorder, [CK_TLA_PCT1, FLAP_POS_FWC, RUDDER, ELEV1] are control input parameters in the aircraft motion equation, [HEIGH_DEST, LAT_AIR, LONG_AIR, ALT_RATE, ROLL_RATE1, ROLL, TAS_REC, PITCH, PITCH_RATE, VRTG] are state parameters, SIDE_WIND is an external meteorological parameter, and the control input parameters and meteorological parameters affect the state parameters of the aircraft.

[0023] In a preferred embodiment, in step 1, the Kalman filter method is used to perform noise processing on the direct and indirect influencing parameters, and the Min-Max method is used to normalize the parameters, that is,

[0024]

[0025] In the formula, x represents the original data, x min Represents the minimum value in the original data, x max Indicates the maximum value in the original data, x new Represents the normalized data.

[0026] In a preferred embodiment, in step 2, a sliding window process is performed on the direct and indirect impact parameters to construct sample data as the input of the prediction model, that is, Where N represents the number of direct and indirect influencing parameters, here N = 15, S represents the sliding window step size, and T is the input sliding window length. The output is H represents the output sliding window length, that is, the prediction time step.

[0027] In a preferred embodiment, in step 2, the dependency between sequences and the temporal correlation within the sequence are extracted by graph convolutional networks and frequency domain analysis. First, the model uses the potential correlation layer to extract the dependency matrix W between the input multivariate time series. Then, the graph structure G = (X, W) consisting of the input and the dependency matrix is ​​used as the input of the StemGNN Block. The StemGNN Block jointly extracts the temporal dependency in the multivariate time series in the spectral domain. Finally, the model is trained using the prediction loss and the reconstruction loss, as shown in formula (2):

[0028]

[0029] The first term represents the prediction loss, the second term represents the reconstruction loss, X t is the true value to be predicted at time t, For X t The predicted value of B, the difference between the two represents the prediction loss; t-i (X) represents the value obtained by the backtracking module reconstruction, where B represents the entire network that generates the backcast output. By self-encoding the output, the representation ability of the sequence is enhanced, Δ θ Represents all parameters in the network.

[0030] In a preferred embodiment, in step 2, the loss function is as shown in formula (3), the first term is the prediction loss, the second term is the reconstruction loss, W C is the weighting matrix:

[0031]

[0032] In a preferred embodiment, in step 2, the potential correlation layer includes a gated recurrent unit (GRU) and a self-attention mechanism. First, the multivariate time series is input into the GRU, the intrinsic temporal correlation of the sequence is extracted, and the final hidden state of the GRU is obtained. Then, the self-attention mechanism is used to calculate the relationship matrix W of the multivariate time series.

[0033] In a preferred embodiment, in the step two, the StemGNN is constructed by stacking multiple StemGNN Blocks using skip connection, the StemGNN Block is designed by embedding a Spectral-Sequential cell (Spe-Seq) unit into a spectral graph convolution module, the StemGNN converts data from space-time domain to frequency domain, captures periodicity and autocorrelation within time series by spectral sequential convolution Spe-Seq, and extracts spatial dependency between time series by spectral graph convolution.

[0034] In a preferred embodiment, in the step two, two StemGNN blocks are used for residual connection, the residual between the output value of the second StemGNN block and the reconstructed value of the first StemGNN block, finally, the outputs of the two blocks are superimposed and sent into GLU and fully connected layer to generate the predicted value.

[0035] In a preferred embodiment, in the step three, the definition of flight state is a feature matrix extracted from flight parameters that can represent the motion state (control input and attitude) of the aircraft, the input of the last fully connected layer of the StemGNN prediction model (the output of the IDFT layer) is extracted as the expression of the flight state, and the state matrix at time t is denoted as St, as shown in equation (4):

[0036]

[0037] where f t nl is the lth value of the nth feature at time t, n is the number of features, and l is the dimension of the feature.

[0038] In a preferred embodiment, in the step three, the approach and landing phase includes two phases: the approach procedure segment from 1500 feet to 800 feet and the landing segment from 800 feet to the ground.

[0039] In a preferred embodiment, in the step three, the flight state transition of the approach and landing phase in the flight process is a series of data that changes regularly according to the approach procedure, the approach procedure of a specific airport is consulted according to the China Navigation Data Compilation, and the change of the aircraft state in the approach process is analyzed according to a large amount of fleet data.

[0040] In a preferred embodiment, in the step three, the distance between the flight state matrices of the current time and the next time is solved as the change of the flight state, and the change of the state matrices at time t and t+1 is denoted as equation (5):

[0041]

[0042] Where ED is the Euclidean distance, as shown in formula (6):

[0043]

[0044] Comparing the change from the current flight state to the next flight state, if it exceeds the state change range, the flight state transition is considered to be non-stable, with large fluctuations, that is, a state mutation occurs, which may cause unsafe flight events.

[0045] In a preferred embodiment, in step 4, in the event of a sudden change in state, the flight parameters at the next moment are further analyzed in detail. If it is not caused by the pilot's manipulation, then there is a malfunction in some part of the aircraft, making it impossible for the aircraft to fly according to flight principles. If it is caused by the pilot's manipulation, the system prompts the pilot to confirm whether the manipulation is wrong.

[0046] Compared with the existing technology, the present invention has the following beneficial effects: the present invention preprocesses the fast access recorder data, combines the "Flight Quality Assurance (FOQA) Implementation and Management" and flight dynamics, analyzes the direct and indirect influencing parameters affecting flight safety in the approach and landing phase, performs multivariate parameter prediction analysis through the StemGNN network, and performs flight status stability analysis through multi-feature fusion, realizes anomaly detection and early warning, and thus can more accurately predict unsafe events. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is an overall block diagram of the present invention;

[0048] Figure 2 This is the StemGNN multivariate time series data prediction model diagram of the present invention;

[0049] Figure 3 This is a flow chart of the stability analysis method for flight state transition of the present invention;

[0050] Figure 4 This is a diagram of the sample data construction method based on the sliding window of the present invention;

[0051] Figure 5 This is a state change diagram of a flight during the approach phase at Nanjing Lukou International Airport of the present invention;

[0052] Figure 6 This is a flight parameter change diagram of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] Example 1

[0055] This embodiment is based on the QAR historical data of 104 Airbus A320 flights, and takes the abnormality during the approach and landing phase at Nanjing Lukou Airport as an example to illustrate its specific implementation.

[0056] The QAR data of 104 A320 flights were collected and analyzed based on the flight phases and the division of flight altitudes during the approach and landing phases in the flight procedures. Data with a flight altitude of 3000ft to 0ft were selected for safety warning analysis.

[0057] Step 1: Based on the monitoring parameters and flight quality monitoring parameters, select 15 corresponding flight data from the QAR data, and perform noise filtering and normalization processing;

[0058] Step 2: Build a multivariate time series prediction model based on StemGNN to perform parameter prediction and over-limit analysis;

[0059] 1) The normalized data were arranged in chronological order to form a two-dimensional matrix. A sliding window was used to construct the training set, with a sliding window step size of S = 1, an input sliding window length of T = 20, and an output sliding window length of H = 10. This method ultimately yielded 41,449 samples, with a training set, validation set, and test set ratio of 7:2:1.

[0060] 2) In the StemGNN network structure, the channel size of the graph convolution layer was set to 64, and the kernel size of the 1D convolution was set to 3. The RMSprop optimizer was used, and the number of iterations was set to 30. The initial value of the learning rate was set to 0.0001 and was decayed by 0.7 after every 5 iterations. A dropout layer was added with a coefficient of 0.5.

[0061] 3) The weighted matrix of the loss function of the network in formula (3) is set as shown in formula (7).

[0062]

[0063] 4) Root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to evaluate the prediction performance of different model methods.

[0064]

[0065]

[0066]

[0067] 5) A performance comparison analysis was conducted on StemGNN (N=15, direct and indirect parameters, weighted loss function), StemGNN (N=15, direct and indirect parameters), StemGNN (N=2, direct parameters), LSTM, and CNN-LSTM. Table 1 shows the performance indicators of each model method.

[0068] Table 1 Performance indicators of each model method

[0069]

[0070] Table 1 shows that the StemGNN-based model significantly outperforms the other two models on three metrics of vertical acceleration (VRTG). The StemGNN model based on cost-sensitive learning achieves the best overall performance for pitch angle (PITCH) and vertical acceleration. The StemGNN-based model also outperforms the LSTM model in predicting PITCH, similar to the CNN-LSTM model. Compared with the LSTM and CNN-LSTM, the StemGNN is more effective in extracting latent features from time series data, contributing to improved prediction accuracy. Compared with the StemGNN model with direct and indirect influencing parameters, the StemGNN models with both direct and indirect influencing parameters further improve. This indicates that indirect influencing parameters influence the prediction accuracy of VRTG and PITCH. Among the StemGNN models with both direct and indirect influencing parameters, the StemGNN model based on cost-sensitive learning achieves the greatest improvement in pitch angle prediction accuracy, indicating that adding an appropriate weight to the prediction loss of the target parameter in the loss function can improve the accuracy of target parameter prediction.

[0071] Step 3: Based on the flight procedure of the approach and landing phase, a flight state transition stability analysis model is established to analyze the stability of the flight state transition process;

[0072] 1) For the case that the parameters in step two are not out of limits, the output features of the IDFT layer in StemGNN model are extracted as the representation matrix St of the flight state of the aircraft, see equation (4), which has a dimension of (15, 100);

[0073] 2) According to equations (5) and (6), the change of the two states is calculated to determine whether a large fluctuation or state mutation has occurred. Assuming that the change of the flight state matrix is greater than 0.02 in the minimum and maximum values of the change within 4s, as shown in equation (10), it means that a large fluctuation has occurred in a short time.

[0074]

[0075] 3) Combined with the flight approach procedure of Nanjing Lukou Airport, the state change of a certain flight aircraft during the approach phase is analyzed, from Figure 5 It can be seen that the aircraft will have a very obvious state change at 150s-200s, about 1950ft. The real flight height of 1950ft is close to the 1926ft of the last approach positioning point of Lukou International Airport, and the flight height change graph is roughly consistent with the approach flight height profile of the 06, 07 runway of Lukou International Airport. Therefore, it can be proved that the large fluctuation of the flight state occurs at the final approach positioning point FAF of the flight procedure.

[0076] In step four, according to the output results, abnormal analysis is performed.

[0077] Combined with the analysis of the change of the flight parameters in Figure 6 It can be seen that the pilot performs some large flight control on the aircraft at this point, that is, the large mutation of the aircraft state is caused by the control of the pilot.

[0078] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting anomalies during approach and landing phases based on flight data, characterized in that: The following steps are involved: Step 1: Extract flight data for preprocessing, including the following steps: Extract flight data from historical flights from the Quick Access Recorder; Determine the monitoring items and monitoring parameters for the approach and landing phase. The monitoring items include tail strikes and hard landing events. The monitoring parameters for tail strikes and hard landing events include direct influencing parameters and indirect influencing parameters. Direct influencing parameters include pitch angle and vertical acceleration. Parameters that affect pitch angle and vertical acceleration include control input parameters, aircraft state parameters, and external environment parameters, which are indirect influencing parameters. The direct and indirect impact parameters are processed with noise and normalized; Step 2: Build a prediction model to predict flight data and monitor over-limits, including the following steps: Perform sliding window processing on direct and indirect impact parameters and construct sample data as input of the prediction model; Build a multivariate time series prediction model based on StemGNN, extracting the dependencies between sequences and the temporal correlation within sequences through graph convolutional networks and frequency domain analysis; The model is trained using prediction loss and reconstruction loss, and the classification weighting method in cost-sensitive learning is introduced to design a loss function that pays more attention to specified features. Compare the predicted value with the monitoring parameter to determine whether it exceeds the limit. If so, issue a safety warning. If not, proceed to step three. Step 3: Based on the flight procedures for the approach and landing phase, a flight state transition stability analysis model is established, and a flight state transition stability analysis is performed on the monitored parameters that are within the limits. The approach and landing phase consists of two stages: the approach procedure segment from 1500 feet to 800 feet and the landing segment from 800 feet to the ground. Step 4: Perform an abnormality analysis based on the output results. In the event of a sudden change in state, further analyze the flight state parameters at the next moment in detail. If it is not caused by the pilot's operation, then there is a malfunction in some part of the aircraft, making the aircraft unable to fly according to the flight principle. If it is caused by the pilot's operation, the system prompts the pilot to confirm whether the operation is wrong.

2. The method for detecting anomalies during approach and landing phases based on flight data according to claim 1, characterized in that: In the step 1, the Kalman filter method is used to perform noise processing on the direct and indirect influencing parameters, and the Min-Max method is used to normalize the parameters, that is, （1） Where, Represents the original data, Represents the minimum value in the original data, Represents the maximum value in the original data, Represents the normalized data.

3. The method for detecting anomalies during approach and landing phases based on flight data according to claim 2, characterized in that: In the second step, the direct and indirect impact parameters are processed by sliding window, and the sample data is constructed as the input of the prediction model, that is, ,in Indicates the number of direct and indirect influencing parameters, where , represents the sliding window step size, is the input sliding window length, and the output is , Represents the output sliding window length, that is, the prediction time step.

4. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 3, characterized in that: In the second step, the dependency between sequences and the temporal correlation within sequences are extracted by graph convolutional networks and frequency domain analysis. First, the model uses the potential correlation layer to extract the dependency matrix between the input multivariate time series. , and then the graph structure consisting of input and dependency matrix As the input of StemGNN Block, StemGNNBlock jointly extracts the temporal dependencies in multivariate time series in the spectral domain. Finally, the model is trained using prediction loss and reconstruction loss, as shown in Equation (2): （2） The first term represents the prediction loss, and the second term represents the reconstruction loss. for The true value to be predicted at time, for The predicted value of , the difference between the two represents the prediction loss; Represents the value obtained by reconstructing the backtracking module, where Represents the entire network that generates the back-broadcast output, and enhances the representation ability of the sequence by self-encoding the output. Represents all parameters in the network.

5. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 4, characterized in that: In the second step, the loss function is as shown in formula (3), the first term is the prediction loss, the second term is the reconstruction loss, is the weighting matrix: （3）。 6. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 5, characterized in that: In step 2, the potential correlation layer includes a gated recurrent unit (GRU) and a self-attention mechanism. First, the multivariate time series is input into the GRU to extract the intrinsic temporal correlation of the sequence and obtain the final hidden state of the GRU. Then, the self-attention mechanism is used to calculate the relationship matrix of the multivariate time series. .

7. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 5, characterized in that: In the second step, StemGNN is constructed by stacking multiple StemGNN Blocks using skip connections. StemGNN Block is designed by embedding the spectral-sequential cell (Spe-Seq) unit into the spectral graph convolution module. StemGNN converts data from the spatiotemporal domain to the frequency domain, captures the periodicity and autocorrelation within the time series through spectral sequence convolution Spe-Seq, and uses spectral graph convolution to extract the spatial dependency between time series.

8. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 5, characterized in that: In the second step, two StemGNN blocks are used for residual connection, and the residual between the output value of the second StemGNN block and the reconstructed value of the first StemGNN block is obtained. Finally, the outputs of the two blocks are superimposed and fed into the GLU and fully connected layers to generate the predicted value.

9. The method for detecting anomalies during approach and landing phases based on flight data according to any one of claims 1 to 8, characterized in that: In step 3, the flight state is defined as a feature matrix extracted from the flight parameters that can characterize the aircraft's motion status (control input and attitude). The input of the last fully connected layer of the StemGNN prediction model (the output of the IDFT layer) is extracted as the expression of the flight state. Let The state matrix at the moment is , as shown in formula (4): (4) in, for The moment The first feature values, is the number of features, is the dimension of the feature.

10. The method for detecting anomalies in the approach and landing phase based on flight data according to claim 9, characterized in that: In step 3, the distance between the flight state matrix at the current moment and the next moment is solved as the change in flight state. The change in the state matrix at time t and time t+1 is expressed as formula (5): （5） Where ED is the Euclidean distance, as shown in formula (6): （6） Comparing the change from the current flight state to the next flight state, if it exceeds the state change range, the flight state transition is considered to be non-stable, with large fluctuations, that is, a state mutation occurs, which may cause unsafe flight events.

Images