Training method and training system of flight safety detection model in approaching landing stage of aircraft, flight safety detection method and detection system, computer readable storage medium and computer program product

Abstract

The invention relates to a training method and training system of a flight safety detection model in an aircraft approaching landing stage, a flight safety detection method and detection system, a computer readable storage medium and a computer program product. The training method comprises the following steps: a historical data processing step: performing data cleaning and data standardization on historical flight data related to an aircraft approaching landing stage in the historical data processing step to obtain processed historical data; a clustering step: clustering the processed historical data through an unsupervised clustering algorithm in the clustering step to obtain a clustering point set and an outlier set; a marking step: marking the outlier set by adopting an abnormal label in the marking step so as to obtain abnormal data and unmarked normal data corresponding to the clustering point set; and a training step: inputting the abnormal data and the normal data into a training module of the deep neural network in the training step for supervised training, and obtaining a flight safety detection model after the training is completed.

Description

Technical Field

[0001] The present invention relates to a method for training a flight safety detection model for the aircraft approach and landing phase. The flight safety detection model is trained using a deep neural network, allowing predictions and warnings during the aircraft approach and landing phase. The present invention also relates to a training system for the flight safety detection model, a flight safety detection method, and a flight safety detection system for the aircraft approach and landing phase. This invention belongs to the field of flight safety detection technology. Background Art

[0002] The aircraft approach and landing process is generally divided into three phases: initial approach, intermediate approach, and approach and landing. The approach and landing phase begins with the final approach fix (FAF), where the aircraft aligns with the runway along the glide path (such as an ILS guidance system) and maintains a steady speed and descent rate until touchdown. The approach and landing phase is the most dangerous and has the highest accident rate during the entire flight process. The flight data involved is multidimensional and complex, characterized by complex parameter dimensions (such as multi-source heterogeneous parameters such as attitude, control, and environment), dynamically changing sequence lengths (up to tens of thousands of sampling points for a single flight), and nonlinear redundant correlations between parameters. Therefore, how to conduct real-time safety monitoring to reduce risks during this phase is a key challenge.

[0003] Currently, in the field of real-time flight safety detection during the aircraft landing phase, existing technologies mostly rely on supervised learning and physical model driving.

[0004] One solution discloses a multivariate time series prediction model based on StemGNN, which processes data through sliding windows and introduces graph convolutional networks to predict parameters such as pitch angle and vertical acceleration, and combines flight state transition stability analysis to achieve anomaly detection.

[0005] However, this method relies on parameter out-of-limit judgment and manually defined flight procedure rules, which requires a large workload for data labeling and high prior knowledge.

[0006] In addition, in existing technologies, such supervised learning methods rely on manual labeling of abnormal data at this stage, but abnormal events are relatively sparse in actual scenarios and the labeling cost is high.

[0007] Another solution discloses a graph neural network model based on embedded physical knowledge. By constraining network training with a six-degree-of-freedom longitudinal motion model, it predicts the mean and variance of the glide path boundary to detect anomalies. This method primarily generates simple anomaly alerts based on whether the current data point exceeds the boundary threshold.

[0008] However, this method requires the integration of complex flight dynamics equations, and model training relies on physical mechanisms and expert experience, which lacks flexibility.

[0009] In addition, in the existing technology, such physical mechanism-based modeling methods can only roughly simulate the flight process of the entire approach and landing phase, and it is difficult to accurately construct the flight dynamics equation. Therefore, such modeling is difficult to adapt to the parameter coupling relationship in complex environments. Summary of the Invention

[0010] In order to achieve one or more of the above-mentioned objectives of the present invention, the inventors provide a flight safety detection model for the aircraft approach and landing phase, which can realize unsupervised training detection of anomaly detection models and detect flight anomalies in the final approach and landing phase of civil aircraft in real time.

[0011] A training method for a flight safety detection model for the aircraft approach and landing phase, in a first example, the training method includes: a historical data processing step, in which historical flight data related to the aircraft approach and landing phase are cleaned and standardized to obtain processed historical data; a clustering step, in which the processed historical data are clustered using an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set; a labeling step, in which the outlier point set is labeled with an abnormal label to obtain abnormal data and unlabeled normal data corresponding to the cluster point set; and a training step, in which the abnormal data and normal data are input into a training module of a deep neural network for supervised training, and a flight safety detection model is obtained after the training is completed.

[0012] According to the above structure, anomaly labels are automatically generated for abnormal data through an unsupervised algorithm without presetting physical rules.

[0013] In the second example of the training method, optionally including the first example, the data cleaning in the historical data processing step includes data alignment of the sliding landing point time window of each flight data, including: first, manually selecting a flight data of a standard approach landing, and using its landing segment data as the reference window land_std for data alignment; second, performing a sliding window interception of a length L on the data of each other flight to obtain any number of current windows land_recent of a length L; then, performing Euclidean distance calculation between each current window land_recent of a certain flight and the reference window land_std through an error loss function, wherein the error loss function is: Loss = sum(k*

[0014] Euler(land_recent-land_std)), where Loss is the loss value; k is the Euclidean distance coefficient, k = 2; land_recent is the current window; sum is the summation function, thus obtaining the loss value of each current window land_recent; then, the current window land_recent corresponding to the minimum loss value is selected as the alignment window for the current flight, and the alignment window is used as the landing point time window for alignment and data interception; the remaining flight data is processed to align all flight data with the reference window.

[0015] According to the above-described configuration, data distortion, information loss, and model deviation can be minimized.

[0016] In a third example of the training method, optionally including one or more of the above examples, data standardization in the historical data processing step includes one or more of the following: coordinate transformation of longitude and latitude parameters to obtain standardized coordinate parameters, coordinate transformation of altitude parameters to obtain standardized altitude parameters.

[0017] According to the above configuration, the influence of different longitude and latitude data in different regions and different approach and landing directions of different airports can be eliminated, and the influence of the altitude of airports in different regions can be eliminated.

[0018] In a fourth example of the training method, which optionally includes one or more of the above examples, the historical data processing step includes: using the Pearson correlation coefficient to calculate the correlation between various flight data parameters in the data related to the aircraft approach and landing phase to form a correlation coefficient matrix, and selecting n types of weakly correlated flight data parameters for data cleaning and data standardization. The selected n types of weakly correlated flight data parameters include multiple items of the following: radio altitude, indicated airspeed, descent rate, pitch angle, roll angle, engine speed, fuel quantity, rudder deflection, elevator deflection, normal acceleration, lateral acceleration, longitudinal acceleration, aileron, angle of attack, yaw rate, and true airspeed.

[0019] In the fifth example of the training method, which optionally includes one or more of the above examples, the clustering step includes: clustering the selected n weakly correlated flight data parameters in pairs in turn, and obtaining outlier point sets and cluster point sets under a total of n*(n-1) / 2 flight data parameter combinations through every two flight data parameter combinations.

[0020] According to the above structure, it is possible to intuitively see which parameter combinations are clustered as anomalies, and it is possible to indirectly infer which parameter has a problem based on the results of the outlier set (ie, anomaly).

[0021] In a sixth example of the training method, which optionally includes one or more of the above examples, the marking step includes: counting the number of outliers appearing in each pair of flight data parameter combinations in the data of different aircraft flights; if a flight data parameter of a certain aircraft flight is marked as an outlier in the clustering results of greater than or equal to q flight data parameter combinations, where q is a preset positive integer threshold, then the flight data parameter data of the flight flight is considered to be abnormal and is marked with an abnormal label; the above process is repeated n times to obtain all n-dimensional flight data parameter data points and corresponding n-dimensional abnormal labels.

[0022] According to the above structure, abnormal labels are automatically generated to avoid the difficulties of manual operation on large amounts of data, improve processing efficiency, reduce or even avoid the prior knowledge requirements for manually defining abnormal data, reduce the subjective bias introduced by manual labeling, and provide a traceable decision-making basis for subsequent tracing of abnormal situations.

[0023] In the seventh example of the training method, one or more of the above examples may be optionally included, and the training steps include: using a deep neural network based on a self-attention mechanism, and arranging n-dimensional data points and corresponding n-dimensional abnormal labels according to aircraft flight numbers and sampling time series, intercepting p data points each time to form a sequence, forming M data sequence tensors with a shape of n*p and a label sequence tensor with a shape of n*p, where p is a preset positive integer and M is the total number of intercepted data tensors, and inputting them into the training module of the deep neural network for supervised training.

[0024] Based on the above structure, the sequence modeling and parallel computing capabilities of the self-attention mechanism architecture can better capture long-range dependencies in the sequence. Therefore, it is possible to observe which parts of the sequence the model focuses on when making predictions, which helps explain why a certain sequence data or data point is judged to be anomalous.

[0025] In the eighth example of the training method, one or more of the above examples are optionally included, and the clustering algorithm selected in the clustering step includes one or more of the following: K-Means algorithm, K-Medoids algorithm, DBSCAN algorithm, Gaussian mixture model GMM algorithm, OPTICS algorithm, BIRCH algorithm, and CURE algorithm.

[0026] In order to achieve one or more of the above-mentioned objectives of the present invention, the present invention provides a training system for a flight safety detection model for the aircraft approach and landing phase. The training system includes the following modules: a historical data processing module, which cleans and standardizes historical flight data related to the aircraft approach and landing phase to obtain processed historical data; a clustering module, which clusters the processed historical data using an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set; a labeling module, which labels the outlier point set with an abnormal label, thereby obtaining abnormal data and unlabeled normal data corresponding to the cluster point set; and a training module, which inputs the abnormal data and normal data into a deep neural network training module for supervised training. After the training is completed, a flight safety detection model is obtained.

[0027] In order to achieve one or more of the above-mentioned objectives of the present invention, the present invention provides a flight safety detection method during the approach and landing phase of an aircraft.

[0028] In the first example of the flight safety detection method, the flight safety detection method includes: a real-time data processing step, in which the time-series flight data is acquired, stored and updated in real time, and data cleaning and data standardization are performed to obtain processed real-time data; a flight safety detection step, in which the processed real-time data is input into the flight safety detection model of any one of the above examples in real time to obtain a flight safety detection result; and an alarm step, in which an alarm of a corresponding level is issued based on the flight safety detection result; and the real-time data processing step is executed again until the aircraft completes landing or the corresponding system is shut down.

[0029] In a second example of the flight safety detection method, optionally including the first example, the flight safety detection step includes: setting n types of weakly correlated flight data parameters into three importance levels, wherein the flight data parameters directly related to the aircraft's approach and landing maneuvers are set as first-level parameters among the n types of weakly correlated flight data parameters; and dividing the remaining flight data parameters of the n types of weakly correlated flight data parameters into two parts according to the degree of correlation with the first-level parameters, wherein the parts with high correlation are set as second-level parameters, and the parts with low correlation are set as third-level parameters; and the flight safety detection result includes: the number of detected anomalies in each of the first-level parameters, the second-level parameters, and the third-level parameters. According to the above structure, graded alarms can be designed based on parameter correlation.

[0030] In a third example of the flight safety detection method, which optionally includes one or more of the above examples, the warning step includes:

[0031] When the flight safety detection result shows that n weakly correlated flight data parameters are all within the normal range, there is no alarm; when the flight safety detection result shows that at least one third-level parameter is detected to be abnormal, and both the second-level and first-level parameters are normal, a reminder is issued; when the flight safety detection result shows that at least one second-level parameter is detected to be abnormal, and there is no abnormality in the first-level parameters, a warning is issued; when the flight safety detection result shows that at least one first-level parameter is detected to be abnormal, a serious warning is issued.

[0032] To achieve one or more of the aforementioned objectives of the present invention, the present invention provides a flight safety monitoring system for an aircraft during its approach and landing phase. The flight safety monitoring system comprises the following modules: a real-time data processing module that acquires, stores, and updates time-series flight data in real time, performs data cleaning and standardization, and generates processed real-time data; a flight safety monitoring module that inputs the processed real-time data into a flight safety monitoring model based on any of the above examples in real time to generate flight safety monitoring results; and an alarm module that issues an alarm of a corresponding level based on the flight safety monitoring results.

[0033] In order to achieve one or more of the above-mentioned purposes of the present invention, the present invention provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the training method in the above example and / or the flight safety detection method in the above example.

[0034] In order to achieve one or more of the above-mentioned objectives of the present invention, the present invention provides a computer program product, including a computer program, which is executed by a processor to implement the training method in the above example and / or the flight safety detection method in the above example.

[0035] This application uses unsupervised flight data learning and real-time safety detection strategies to unsupervisedly train anomaly detection models and detect flight anomalies in the final approach and landing phase of civil aircraft in real time.

[0036] This application focuses on historical data (non-pilot operations). By analyzing a large amount of historical data, it automatically learns the abnormal responses of the aircraft throughout the entire process through an unsupervised algorithm, and provides cluster point sets and outlier point sets (abnormal points) during the approach and landing process, without the need to pre-define risk stages and key monitoring parameters.

[0037] This application uses deep neural networks for prediction and warning, combined with a self-attention mechanism to deeply mine temporal dependencies. The trained flight safety detection model allows real-time input of time-series flight data fragments to determine and detect flight safety during the final approach and landing phase of civil aircraft.

[0038] This application can be activated during the final approach phase of the aircraft to monitor landing safety in real time. It can be applied to scenarios such as civil aircraft test flights and route operations, significantly reducing the risks of accidents such as hard landings, runway deviations, and low energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to describe the implementation of the above and other features of the present invention, a more particular description of the invention briefly described above will be presented with reference to exemplary embodiments of the invention shown in the accompanying drawings. It will be understood that these drawings depict only exemplary embodiments of the invention and are not to be considered limiting of its scope, and the invention will be described and explained with additional specificity and detail through the use of the accompanying drawings. In the drawings:

[0040] Figure 1 is a flow chart of a method for training a flight safety detection model for an aircraft approach and landing phase according to an embodiment of the present invention;

[0041] Figure 2 is a block diagram of a training system for a flight safety detection model for an aircraft approach and landing phase according to an embodiment of the present invention;

[0042] Figure 3 is a flow chart of a flight safety detection method for an aircraft during approach and landing according to an embodiment of the present invention;

[0043] Figure 4 4 is a block diagram of a flight safety detection system for an aircraft during approach and landing according to an embodiment of the present invention.

[0044] List of reference numerals:

[0045] 11 Historical data processing module

[0046] 13 Clustering Module

[0047] 15 Marking Module

[0048] 17 Training Modules

[0049] 21 Real-time data processing module

[0050] 23 Flight safety detection module

[0051] 25 Alarm Module DETAILED DESCRIPTION

[0052] The term "window," as used in this article, is an important method for analyzing time series data. It refers to a fixed or variable-length subsequence of continuous time series data. By defining a window, you can perform local analysis, feature extraction, or anomaly detection on the data, thereby exploring temporal dependencies and local patterns in the data.

[0053] The term "clustering" used in this article refers to an unsupervised learning method used to group a set of data objects into clusters. Data objects within the same cluster have high similarity, while data objects in different clusters have low similarity. The goal of clustering is to exploit the inherent structure of data, achieving "like attracts like," and helping people discover hidden patterns or regularities in the data.

[0054] The term "self-attention mechanism" used in this article is a core architecture for modeling internal dependencies in sequence data in deep learning. It calculates the dependencies between computational elements within the same sequence, dynamically assigns weights, and captures long-distance dependencies in parallel and efficiently.

[0055] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0057] Figure 1 This is a flow chart of a method for training a flight safety detection model for an aircraft approach and landing phase according to an embodiment of the present invention. The flight safety detection model obtained according to this embodiment can be applied to real-time flight safety detection during an aircraft approach and landing phase.

[0058] The training method of the present invention generally includes: a historical data processing step 101 , a clustering step 103 , a labeling step 105 and a training step 107 .

[0059] In the historical data processing step 101 , data cleaning and data standardization are performed on the historical flight data related to the aircraft approach and landing phase to obtain processed historical data.

[0060] Preferably, the data cleaning in the historical data processing step 101 may include collating multiple flight data and / or noise filtering.

[0061] Preferably, the data cleaning in the historical data processing step 101 may include aligning the data of each flight's sliding landing point time window. Aligning the data of the sliding landing point time window includes defining a reference window for the landing point time window, land_std = [height1, ..., height50], comparing the current glide curve of each flight with the reference glide curve using the sliding landing point time window, calculating the error loss, and then selecting the time window with the smallest error value as the alignment window.

[0062] Specifically, first, flight data of a certain standard approach and landing is manually selected, and its landing segment data is used as the reference window land_std for data alignment.

[0063] Secondly, the flight data of each other flight is intercepted with a sliding window of length L to obtain any number of current windows land_recent of length L.

[0064] Subsequently, the Euclidean distance between each current window land_recent of a certain flight and the reference window land_std is calculated using the error loss function, where the error loss function is: Loss = sum(k*Euler(land_recent-land_std)), where Loss is the loss value; k is the Euclidean distance coefficient, k = 2; land_recent is the current window; sum is the summation function, thereby obtaining the loss value of each current window land_recent.

[0065] Then, the current window land_recent corresponding to the minimum loss value is selected as the alignment window of the current flight, and the alignment window is used as the landing point time window for alignment and data capture.

[0066] Repeat the above process to align all flight data with the reference window. By selecting the time window with the least loss for data alignment and data truncation, data distortion, information loss, and model bias can be minimized.

[0067] Preferably, the data standardization in the historical data processing step 101 may include performing coordinate conversion on the longitude and latitude parameters to obtain standardized coordinate parameters. To eliminate the effects of different longitude and latitude data in different regions and different approach and landing directions at different airports, a standardized coordinate system (right-handed system) is defined: the landing point of each flight is used as the origin, and the x, y, and z coordinate axes are parallel to the landing direction, perpendicular to the landing direction, and in the standardized altitude direction.

[0068] Preferably, the data standardization in the historical data processing step 101 may include performing coordinate conversion on the altitude parameter to obtain a standardized altitude parameter. To eliminate the influence of the altitude of airports in different regions, the corresponding altitude of each aligned flight landing point is selected as the reference altitude, and the altitude data of each flight is corrected.

[0069] Preferably, the historical data processing step 101 includes calculating the correlation between various flight data parameters in the data related to the aircraft's approach and landing phase using the Pearson correlation coefficient to form a correlation coefficient matrix, and selecting n weakly correlated flight data parameters as targets for data cleaning and data standardization. Here, weak correlation means that a significant change in one parameter does not lead to a significant change in another parameter.

[0070] The selected n weakly correlated flight data parameters include multiple items of the following: radio altitude, indicated airspeed, descent rate, pitch angle, roll angle, engine speed, fuel quantity, rudder deflection, elevator deflection, normal acceleration, lateral acceleration, longitudinal acceleration, aileron, angle of attack, yaw rate, and true airspeed.

[0071] In the clustering step 103 , the processed historical data is clustered using an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set.

[0072] The clustering algorithm selected in the clustering step may include one or more of the following: K-Means algorithm, K-Medoids algorithm, DBSCAN algorithm, Gaussian mixture model GMM algorithm, OPTICS algorithm, BIRCH algorithm, CURE algorithm. Preferably, density-based DBSCAN algorithm may be used.

[0073] Preferably, the clustering step includes sequentially clustering each pair of n weakly correlated flight data parameters, obtaining outlier point sets and cluster point sets for a total of n*(n-1) / 2 flight data parameter combinations. Conventional clustering methods typically cluster all parameters together, but high-dimensional parameter coupling clustering easily obscures local correlation characteristics between parameters, making it difficult to distinguish the contribution of different parameters to the clustering results. Pairwise parameter clustering, on the other hand, allows intuitive identification of which parameter combinations are clustered as anomalies. The results of the outlier point sets (i.e., anomalies) can indirectly infer which parameter has a problem.

[0074] In the labeling step 105, the outlier set is labeled with anomaly labels to obtain anomaly data. Automatically generating anomaly labels can avoid the difficulty of manually processing large amounts of data, improve processing efficiency, reduce or even eliminate the prior knowledge required to manually define anomaly data, reduce the subjective bias introduced by manual labeling, and provide a traceable decision-making basis for subsequent anomaly tracing.

[0075] Preferably, the marking step 105 may include counting the number of outliers that appear in each pair of flight data parameter combinations for data from different aircraft flights. If a flight data parameter for a particular aircraft flight is marked as an outlier in the clustering results of q or more flight data parameter combinations, the flight data parameter for that flight flight is considered an outlier and is labeled with an outlier label. Specifically, a single flight parameter may participate in multiple pairwise parameter combination clusterings. In different combination clustering results, some may be normal while others may be abnormal. Therefore, statistics are required. If q outliers occur, the parameter is considered an outlier.

[0076] Here, q is a preset positive integer threshold that can be selected based on the specific aircraft model and airport. Its value should be determined based on scenario sensitivity requirements and historical data distribution characteristics to balance anomaly detection coverage and misjudgment risk. Repeat the above process n times to obtain all n-dimensional flight data parameter data points and their corresponding n-dimensional anomaly labels.

[0077] In the training step 107, the abnormal data corresponding to the outlier set and the processed historical data corresponding to the outlier set are input into the training module of the deep neural network for supervised training. After the training is completed, the flight safety detection model is obtained.

[0078] Preferably, training step 107 may include employing a deep neural network based on a self-attention mechanism. The self-attention mechanism architecture, due to its superior sequence modeling capabilities and parallel computing advantages, excels in time series analysis, better capturing long-range dependencies within sequences. The self-attention mechanism allows for observing which parts of the sequence the flight safety detection model focuses on when making predictions, helping to explain why certain sequence data or data points are identified as anomalous.

[0079] Preferably, the training step 107 may include: arranging n-dimensional data points and corresponding n-dimensional abnormal labels according to the aircraft flight number and sampling time sequence, intercepting p data points each time to form a sequence, forming M data sequence tensors with a shape of n*p and a label sequence tensor with a shape of n*p, where p is a preset positive integer and M is the total number of intercepted data tensors. The value needs to comprehensively consider the data sampling frequency, the time series dependency span, and the model calculation efficiency. Optionally, each time B data tensors and corresponding B label tensors are input into the training module of the deep neural network for supervised training, where B is the batch size and is a preset positive integer. The value of B can be dynamically adjusted according to the hardware memory capacity and the complexity of the model, and takes into account the gradient update stability and parallel computing efficiency. Usually, a power of 2 is selected to adapt to the matrix operation optimization characteristics of the GPU tensor core. The above method can be used to train a flight safety detection model.

[0080] The deep neural network training process of this embodiment can be performed offline. The flight safety detection model obtained after training can be deployed in an onboard display system or a ground monitoring system, and receive real-time data through real-time onboard data or telemetry networks for real-time flight safety detection during the aircraft's approach and landing phases.

[0081] Figure 2 4 is a block diagram of a training system for a flight safety detection model for an aircraft approach and landing phase according to an embodiment of the present invention.

[0082] The training system of the present invention generally includes: a historical data processing module 11, a clustering module 13, a labeling module 15, and a training module 17.

[0083] The historical data processing module 11 performs data cleaning and data standardization on the historical flight data related to the aircraft approach and landing phase to obtain processed historical data.

[0084] The clustering module 13 clusters the processed historical data using an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set.

[0085] The marking module 15 marks the outlier point set with an abnormal label, so as to obtain abnormal data and unlabeled normal data corresponding to the cluster point set.

[0086] The training module 17 inputs the abnormal data and normal data into the training module of the deep neural network for supervised training, and obtains the flight safety detection model after the training is completed.

[0087] Figure 3 FIG2 is a flow chart of a flight safety detection method for an aircraft during approach and landing according to an embodiment of the present invention. The flight safety detection model described above can be applied to real-time flight safety detection during the aircraft during approach and landing.

[0088] The flight safety monitoring method of the present invention generally comprises a real-time data processing step 201, a flight safety monitoring step 203, and an alerting step 205. Onboard flight data records are acquired in real time, stored, updated, and preprocessed. Inference is performed in parallel with flight data storage and updating to ensure the real-time nature of each inference. Each time, the latest processed data sequence tensor is output and input into a neural network in real time. Inference identifies safety and anomalies, and based on the results, generates possible alerts until the aircraft successfully lands or the system shuts down.

[0089] In the real-time data processing step 201, the time-series flight data is acquired, stored, and updated in real time, and data cleaning and standardization are performed to obtain processed real-time data. Preferably, the time-series flight data is cleaned and standardized similarly to the data cleaning and standardization described above for the historical flight data, ultimately obtaining the processed real-time data as the data sequence tensor to be evaluated.

[0090] In the flight safety check step 203, the processed real-time data is input into the trained flight safety check model described above in real time to obtain flight safety check results. Preferably, the deep neural network inference is performed in parallel with the storage and preprocessing of the real-time flight data to ensure the real-time nature of the inference results.

[0091] Preferably, the flight safety detection step 203 includes: outputting an anomaly confidence rating based on the input data, setting each parameter's warning confidence threshold based on the parameter's importance, and determining parameter anomalies based on the anomaly confidence rating and threshold. Here, the n types of weakly correlated flight data parameters described above are set to three importance levels. Among the n types of weakly correlated flight data parameters, those directly related to the aircraft's approach and landing maneuvers are set as first-level parameters. The remaining flight data parameters of the n types of weakly correlated flight data parameters are divided into two parts based on their correlation with the first-level parameters, with the highly correlated parts being set as second-level parameters and the less correlated parts being set as third-level parameters. The flight safety detection results may include: how many anomalies are detected in each of the first-level parameters, the second-level parameters, and the third-level parameters.

[0092] In the warning step 205, a warning of a corresponding level is issued according to the flight safety detection result.

[0093] Preferably, the alarm step 205 may include four alarm levels: safe, low priority, medium priority and high priority.

[0094] When the flight safety detection result is that n weakly correlated flight data parameters are all within the normal range, that is, there are no abnormal conditions, it is safe and there is no alarm.

[0095] If the flight safety check detects at least one abnormality in the Level 3 parameter, and no abnormalities in the Level 2 or Level 1 parameters, a low-priority warning is issued. This warning generally does not affect flight safety and serves only to alert the pilot.

[0096] If the flight safety check detects at least one abnormal Level 2 parameter, with a high probability of multiple Level 3 parameter abnormalities, and no Level 1 parameter abnormalities, a medium-priority alert is issued, requiring the pilot's attention.

[0097] If the flight safety check detects at least one abnormality in a Level 1 parameter, with a high probability of multiple abnormalities in Level 2 and Level 3 parameters, a high-priority alert is issued, posing a serious threat to flight safety and requiring immediate action by the pilot.

[0098] Real-time flight safety monitoring equipment during the final approach and landing phase can be deployed as needed in an onboard display system (such as the Electronic Flight Instrument System (EFIS)), delivering varying degrees of warning information to the pilot via the display system or voice prompt system. For special scenarios, it can also be deployed in a ground monitoring system for flight test engineers to monitor flight safety.

[0099] Preferably, the real-time data processing step 201 is performed again and repeated until the aircraft completes landing or the corresponding system is shut down.

[0100] Figure 4 4 is a block diagram of a flight safety detection system for an aircraft during approach and landing according to an embodiment of the present invention.

[0101] The flight safety detection system of the present invention generally comprises: a real-time data processing module 21, a flight safety detection module 23, and an alarm module 25. Optionally, the flight safety detection system is activated by determining whether the aircraft has entered the final approach and landing phase based on the radio altitude.

[0102] The real-time data processing module 21 acquires, stores and updates the time-series flight data in real time, performs data cleaning and data standardization, and obtains processed real-time data.

[0103] The flight safety detection module 23 inputs the processed real-time data into the flight safety detection model described above in real time to obtain a flight safety detection result.

[0104] The warning module 25 issues a warning of a corresponding level based on the flight safety detection result.

[0105] The present invention has the following beneficial effects:

[0106] By processing flight data of flights in different regions, at different altitudes, and with different approach and landing directions, the present invention can be applied to a certain aircraft model, different flights, and different airports, solving the problem of small amounts of data for a single flight and inability to directly use data from multiple airports.

[0107] Traditional rule-based flight data security detection requires expert knowledge and relies on manually designed judgment rules. It can only make judgments based on conditions and lacks adaptability and generalization capabilities. The present invention uses a data-driven flight data security detection method, which does not require manually designed judgment rules and has better adaptability.

[0108] Unsupervised clustering uses the two-dimensional DBSCAN algorithm, which produces better pairwise parameter clustering diagrams, enhancing the interpretability of the resulting labels while reducing computational costs. Furthermore, the labeled flight data obtained through clustering is used to train the neural network, achieving the goal of unsupervised learning. Furthermore, the neural network can comprehensively learn the connections between the sequence of fitting parameter points, avoiding the loss of correlation information between parameters at a single parameter point and the time series information between multiple parameter points that would be lost if cluster analysis alone were used.

[0109] We designed a neural network based on a self-attention architecture. Due to its superior sequence modeling capabilities and parallel computing advantages, the self-attention architecture excels in time series analysis, better capturing long-range dependencies within sequences. Furthermore, the self-attention architecture's attention mechanism provides a degree of interpretability for the model, as it allows us to observe which parts of the sequence the model focuses on when making predictions, helping to explain why a particular sequence data point is considered an anomaly.

[0110] Traditional time series anomaly detection models require the input of complete time series data sequences and are therefore only suitable for offline judgment. The method proposed in this invention can judge flight safety online and issue warnings, with good real-time performance.

[0111] Design warning devices with different risk levels, taking into account the importance of different parameters and the different risk levels in the final approach and landing phase, identify the abnormal type and transmit warning information, provide operational suggestions, and improve flight safety.

[0112] Above, in order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention are clearly and completely described in conjunction with the specific embodiments of the present invention and the accompanying drawings. Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, a magnetic disk, an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0113] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A training method for a flight safety detection model for an aircraft approach and landing phase, characterized in that: The training method comprises: a historical data processing step, in which historical flight data related to the aircraft's approach and landing phase are cleaned and standardized to obtain processed historical data; A clustering step, in which the processed historical data is clustered by an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set; a labeling step, in which the outlier point set is labeled with an abnormal label, thereby obtaining abnormal data and unlabeled normal data corresponding to the cluster point set; A training step, in which the abnormal data and the normal data are input into a training module of a deep neural network for supervised training, and the flight safety detection model is obtained after the training is completed.

2. The training method according to claim 1, characterized in that The data cleaning in the historical data processing step includes aligning the sliding landing point time window of each flight data, including: First, we manually select flight data of a certain standard approach and landing, and use its landing segment data as the reference window land_std for data alignment. Secondly, perform a sliding window interception of length L on each other flight data to obtain any number of current windows land_recent of length L; Subsequently, the Euclidean distance between each current window land_recent and the reference window land_std of a certain flight is calculated using an error loss function, wherein the error loss function is: Loss = sum(k*Euler(land_recent-land_std)), where Loss is the loss value; k is the Euclidean distance coefficient, k=2; land_recent is the current window; sum is the summation function, thereby obtaining the loss value of each current window land_recent; Then, the current window land_recent corresponding to the minimum loss value is selected as the alignment window of the current flight, and the alignment window is used as the landing point time window to perform data alignment and data interception; The remaining flight data are processed to align all flight data with the reference window.

3. The training method according to claim 1, characterized in that The data standardization in the historical data processing step includes one or more of the following: performing coordinate conversion on longitude and latitude parameters to obtain standardized coordinate parameters, and performing coordinate conversion on altitude parameters to obtain standardized altitude parameters.

4. The training method according to claim 1, characterized in that The historical data processing step includes: using the Pearson correlation coefficient to calculate the correlation between various flight data parameters in the data related to the aircraft approach and landing phase to form a correlation coefficient matrix, and selecting n weakly correlated flight data parameters for data cleaning and data standardization, wherein the selected n weakly correlated flight data parameters include multiple items of the following: radio altitude, indicated airspeed, descent rate, pitch angle, roll angle, engine speed, fuel quantity, rudder deflection, elevator deflection, normal acceleration, lateral acceleration, longitudinal acceleration, aileron, angle of attack, yaw rate, and true airspeed.

5. The training method according to claim 4, characterized in that The clustering step includes: clustering the selected n weakly correlated flight data parameters in pairs in turn, and obtaining outlier point sets and cluster point sets under a total of n*(n-1) / 2 flight data parameter combinations through every two flight data parameter combinations.

6. The training method according to claim 5, characterized in that The labeling step includes: counting the number of outliers appearing in each pair of flight data parameter combinations in the data of different flights of aircraft; if a flight data parameter of a certain flight of aircraft is marked as an outlier in the clustering results of greater than or equal to q flight data parameter combinations, where q is a preset positive integer threshold, then the flight data parameter of the flight is considered to be abnormal and is marked with the abnormal label; the above process is repeated n times to obtain all n-dimensional flight data parameter data points and corresponding n-dimensional abnormal labels.

7. The training method according to claim 6, characterized in that The training step includes: using the deep neural network based on the self-attention mechanism, and Arrange n-dimensional data points and corresponding n-dimensional anomaly labels according to aircraft flight numbers and sampling time sequence, intercept p data points each time to form a sequence, and form M data sequence tensors with a shape of n*p and a label sequence tensor with a shape of n*p, where p is a preset positive integer and M is the total number of intercepted data tensors, which are input into the training module of the deep neural network for supervised training.

8. The training method according to claim 1, characterized in that The clustering algorithm selected in the clustering step includes one or more of the following: K-Means algorithm, K-Medoids algorithm, DBSCAN algorithm, Gaussian mixture model GMM algorithm, OPTICS algorithm, BIRCH algorithm, CURE algorithm.

9. A training system for a flight safety detection model during the aircraft approach and landing phase, characterized in that: The training system includes the following modules: a historical data processing module, the historical data processing module performing data cleaning and data standardization on historical flight data related to the aircraft approach and landing phase to obtain processed historical data; A clustering module, wherein the clustering module clusters the processed historical data using an unsupervised clustering algorithm to obtain a cluster point set and an outlier point set; a labeling module, wherein the labeling module labels the outlier point set with an abnormal label, thereby obtaining abnormal data and unlabeled normal data corresponding to the cluster point set; A training module inputs the abnormal data and the normal data into a deep neural network training module for supervised training, and obtains the flight safety detection model after the training is completed.

10. A flight safety detection method for an aircraft during approach and landing, characterized in that: The flight safety detection method comprises: a real-time data processing step, in which the time-series flight data is acquired, stored, and updated in real time, and data cleaning and data standardization are performed to obtain processed real-time data; a flight safety detection step, in which the processed real-time data is input in real time into a flight safety detection model according to any one of claims 1 to 9 to obtain the flight safety detection result; and an alarm step, in which an alarm of a corresponding level is issued according to the flight safety detection result; and The real-time data processing step is performed again until the aircraft completes landing or the corresponding system is shut down.

11. The flight safety detection method according to claim 10, characterized in that: In the case of adopting the flight safety detection model according to any one of claims 4 to 7, the flight safety detection step includes: Setting the n types of weakly correlated flight data parameters into three importance levels, wherein the flight data parameters directly related to the aircraft approach and landing maneuver among the n types of weakly correlated flight data parameters are set as first-level parameters, and the remaining flight data parameters among the n types of weakly correlated flight data parameters are divided into two parts according to the degree of correlation with the first-level parameters, wherein the part with high correlation is set as second-level parameters, and the part with low correlation is set as third-level parameters; and The flight safety detection result includes: several abnormalities detected in each of the first-level parameters, the second-level parameters, and the third-level parameters.

12. The flight safety detection method according to claim 11, characterized in that: The alarm step includes: When the flight safety detection result is that all of the n weakly correlated flight data parameters are detected to be within a normal range, no alarm is issued; When the flight safety detection result indicates that at least one of the third-level parameters is abnormal, and both the second-level parameters and the first-level parameters are normal, a reminder is issued; When the flight safety detection result indicates that at least one of the secondary parameters is abnormal and the primary parameters are normal, issuing a warning; When the flight safety detection result shows that at least one of the first-level parameters is abnormal, a serious warning is issued.

13. A flight safety detection system for an aircraft during approach and landing, characterized in that: The flight safety detection system includes the following modules: A real-time data processing module, which acquires, stores, and updates time-series flight data in real time, performs data cleaning and data standardization, and obtains processed real-time data; a flight safety detection module, which inputs the processed real-time data into a flight safety detection model according to any one of claims 1 to 9 in real time to obtain the flight safety detection result; as well as An alarm module is configured to generate an alarm of a corresponding level based on the flight safety detection result.

14. A computer-readable storage medium storing a computer program, characterized in that: The computer program is executed by a processor to implement the training method according to any one of claims 1 to 8 and / or the flight safety detection method according to any one of claims 10 to 12.

15. A computer program product comprising a computer program, characterized in that The computer program is executed by a processor to implement the training method according to any one of claims 1 to 8 and / or the flight safety detection method according to any one of claims 10 to 12.

Images