A method and apparatus for predicting aircraft hard landing based on dynamic window attention mechanism

Abstract

This application relates to the field of aviation safety technology, and in particular to a method and apparatus for predicting aircraft hard landing based on a dynamic window attention mechanism. The method includes: preprocessing raw flight data of the target aircraft's real-time position; assigning weights to multiple physical quantities of the real-time position to generate an importance weight matrix of the physical quantities, thereby generating a splicing matrix; obtaining the position code corresponding to the real-time position based on the splicing matrix; obtaining the time window corresponding to the real-time position based on the position code and the splicing matrix to obtain the time unit corresponding to the real-time position; and generating a predicted value of gravity acceleration for the real-time position to generate a hard landing prediction result for the aircraft. This application can understand the current flight status of the target aircraft in real time from a certain altitude to the landing point through the quantile distribution of physical quantities, and predict the specific value of gravity acceleration at each moment during the entire landing phase, thereby achieving more accurate early warning for hard landing scenarios.

Description

Technical Field

[0001] This application relates to the field of aviation safety technology, and in particular to a method and apparatus for predicting aircraft hard landings based on a dynamic window attention mechanism. Background Technology

[0002] Flight safety is the lifeline of the civil aviation industry. The risk of hard landings can cause significant damage to aircraft and easily lead to casualties. Research on flight risk identification and early warning technologies has received widespread attention, especially the analysis of flight risks during hard landings.

[0003] In related technologies, prediction and modeling for aircraft hard landings are mostly constructed using one-hot codes, that is, they mostly use classification to divide the gravitational acceleration value at landing into several risk values, and determine the risk of aircraft hard landing based on the risk values.

[0004] However, in related technologies, the data processing methods for predicting and modeling aircraft hard landings mostly follow a fixed number of seconds before and after the touchdown point. This does not take into account the problem that the landing time cannot be accurately controlled in the actual landing scenario. It is only suitable for the overall flight assessment after the flight ends, which is not conducive to predicting whether a hard landing will occur. Moreover, its use of data is relatively superficial, and it does not pay attention to the deep correlation of flight data, which seriously affects the accuracy of aircraft hard landing predictions and urgently needs to be addressed. Summary of the Invention

[0005] This application provides a method and apparatus for predicting aircraft hard landings based on a dynamic window attention mechanism. This addresses the problems in related technologies where the prediction and modeling of aircraft hard landings mostly involve processing data by taking a fixed number of seconds before and after the landing point. This approach fails to consider the inability to accurately control the landing time in actual aircraft landing scenarios and is only suitable for overall flight assessment after the flight ends. It is not conducive to predicting whether a hard landing will occur during landing. Furthermore, its use of data is relatively superficial, neglecting the deep correlation of flight data, which seriously affects the accuracy of aircraft hard landing predictions.

[0006] The first aspect of this application provides a method for predicting aircraft hard landing based on group-individual contrast feature enhancement, comprising the following steps: acquiring raw flight data of the target aircraft's real-time position; aligning the granularity of the raw flight data according to the data sampling frequency of the raw flight data to obtain real-time flight data of the real-time position; determining the actual quantiles of multiple physical quantities of the target aircraft at the real-time position based on the real-time flight data; generating an importance weight matrix of the multiple physical quantities based on the quantile interval of the actual quantiles; and concatenating the importance weight matrix and the numerical matrix corresponding to the real-time flight data to generate a concatenated matrix for each of the multiple physical quantities; acquiring the position code corresponding to the real-time position based on the concatenated matrix; acquiring the time window corresponding to the real-time position based on the position code and the concatenated matrix; acquiring the time unit corresponding to the real-time position based on the time window; and generating a predicted value of gravity acceleration for the real-time position based on the time unit corresponding to the real-time position to generate a hard landing prediction result for the target aircraft.

[0007] A second aspect of this application provides an aircraft hard landing prediction device based on group-individual comparative feature enhancement, comprising: a preprocessing module for acquiring raw flight data of the target aircraft's real-time position, aligning the granularity of the raw flight data according to the data sampling frequency of the raw flight data to obtain real-time flight data of the real-time position; a weighting module for determining the actual quantiles of multiple physical quantities of the target aircraft at the real-time position based on the real-time flight data, generating an importance weight matrix of the multiple physical quantities according to the quantile interval of the actual quantiles, and concatenating the importance weight matrix and the numerical matrix corresponding to the real-time flight data to generate a concatenated matrix for each of the multiple physical quantities; and a prediction module for acquiring a position code corresponding to the real-time position based on the concatenated matrix, acquiring a time window corresponding to the real-time position based on the position code and the concatenated matrix, acquiring a time unit corresponding to the real-time position based on the time window, and generating a predicted value of gravity acceleration for the real-time position based on the time unit to generate a hard landing prediction result for the target aircraft.

[0008] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aircraft hard landing prediction method based on group-individual contrast feature enhancement as described in the above embodiments.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described aircraft hard landing prediction method based on group-individual contrast feature enhancement.

[0010] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, is used to implement the above-described method for predicting aircraft hard landings based on enhanced group-individual contrast features.

[0011] This application embodiment can generate a splicing matrix by assigning corresponding weights to the quantiles of multiple physical quantities in the target aircraft's flight data, and combine a dynamic window attention mechanism to focus on calculating the predicted values ​​of gravitational acceleration before and after the landing point of the target aircraft, thereby obtaining the hard landing prediction result of the target aircraft. Thus, it achieves the construction of a splicing matrix that helps reduce prediction errors by focusing on the quantile distribution of individual physical quantity data in the group physical quantity data, i.e., querying the comprehensive distribution of a single data sample in the overall normal landing flight dataset; and enhances the model's sensitivity to the dynamic window data near the landing point through the dynamic window attention mechanism, obtaining the predicted value of gravitational acceleration at each position in the landing segment, effectively improving the prediction accuracy of the specific value of gravitational acceleration at the landing point. This solves the problem that most related technologies for predicting and modeling aircraft hard landings process data by taking a fixed number of seconds before and after the touchdown point, without considering the problem of not being able to accurately control the landing time in the actual landing scenario. It is only suitable for the overall flight assessment after the flight ends, which is not conducive to predicting whether a hard landing will occur. Moreover, its use of data is relatively superficial and does not focus on the deep correlation of flight data, which seriously affects the accuracy of aircraft hard landing prediction.

[0012] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0013] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0014] Figure 1 This is a flowchart of an aircraft hard landing prediction method based on group-individual contrast feature enhancement provided in an embodiment of this application;

[0015] Figure 2 This is a schematic diagram of the quantile intervals of altitude in a normal landing sample according to an embodiment of this application;

[0016] Figure 3This is a schematic diagram of the quantile intervals of pitch angle in a normal landing sample according to an embodiment of this application;

[0017] Figure 4 This is a schematic diagram of the quantile intervals of vertical velocity in a normal landing sample according to an embodiment of this application;

[0018] Figure 5 This is a schematic diagram of the quantile distribution of altitude in a hard landing sample according to an embodiment of this application;

[0019] Figure 6 This is a schematic diagram of the quantile distribution of pitch angles in a hard landing sample according to an embodiment of this application;

[0020] Figure 7 This is a schematic diagram of the quantile distribution of vertical velocity in a hard landing sample according to an embodiment of this application;

[0021] Figure 8 This is a schematic diagram of the structure of an aircraft hard landing prediction model according to an embodiment of this application;

[0022] Figure 9 This is a schematic diagram of the structure of a gating unit based on a dynamic landing window attention mechanism according to an embodiment of this application;

[0023] Figure 10 This is a flowchart illustrating aircraft hard landing prediction according to one embodiment of this application;

[0024] Figure 11 Weights as intermediate variables in the attention mechanism of one embodiment of this application The query mapping at different time units of the global weight Q is reflected in the weight after slicing. A diagram illustrating the magnitude of attention weights;

[0025] Figure 12 This is a schematic diagram illustrating the prediction results of a normal landing sample according to an embodiment of this application;

[0026] Figure 13 This is a schematic diagram illustrating the prediction results of a re-landing sample according to an embodiment of this application;

[0027] Figure 14 This is a schematic diagram of the confusion matrix of the prediction results according to an embodiment of this application;

[0028] Figure 15 This is a schematic diagram of the structure of an aircraft hard landing prediction device based on a dynamic window attention mechanism according to an embodiment of this application;

[0029] Figure 16 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application.

[0030] Figure label:

[0031] 10-Aircraft hard landing prediction device based on dynamic window attention mechanism; 100-Preprocessing module, 200-Weighting module and 300-Prediction module; 1601-Memory, 1602-Processor and 1603-Communication interface. Detailed Implementation

[0032] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0033] The following describes, with reference to the accompanying drawings, an aircraft hard landing prediction method and apparatus based on a dynamic window attention mechanism according to embodiments of this application. Addressing the problems mentioned in the background art, this application provides an aircraft hard landing prediction method based on a dynamic window attention mechanism. In this method, corresponding weights are assigned to multiple physical quantities in the target aircraft flight data to generate a splicing matrix. The dynamic window attention mechanism is then used to calculate the predicted values ​​of gravitational acceleration before and after the landing point of the target aircraft, thus obtaining the hard landing prediction result. This achieves the construction of a splicing matrix that helps reduce prediction errors by focusing on the quantile distribution of individual physical quantity data in the group physical quantity data—that is, querying the comprehensive distribution of a single data sample in the overall normal landing flight dataset. Furthermore, the dynamic window attention mechanism enhances the model's sensitivity to dynamic window data near the landing point, obtaining the predicted value of gravitational acceleration at each position in the landing segment, effectively improving the prediction accuracy of the specific value of gravitational acceleration at the landing point. This addresses the problems in related technologies, such as the prediction and modeling of aircraft hard landings, where data processing methods mostly rely on taking fixed seconds before and after the touchdown point. This approach fails to consider the inability to accurately control landing time in actual landing scenarios, and is only suitable for overall flight assessment after the flight is completed. It is not conducive to predicting whether a hard landing will occur during landing. Furthermore, its use of data is relatively superficial, neglecting the deep correlation of flight data, which seriously affects the accuracy of aircraft hard landing predictions.

[0034] Specifically, Figure 1 This is a flowchart illustrating an aircraft hard landing prediction method based on a dynamic window attention mechanism, provided as an embodiment of this application.

[0035] like Figure 1 As shown, the aircraft hard landing prediction method based on the dynamic window attention mechanism includes the following steps:

[0036] In step S101, the raw flight data of the target aircraft's real-time position is obtained, and the granularity of the raw flight data is aligned according to the data sampling frequency of the raw flight data to obtain the real-time flight data of the real-time position.

[0037] Those skilled in the art will understand that most of the data processing methods for predicting and modeling aircraft hard landings in related technologies are based on taking a fixed number of seconds before and after the touchdown point. This method of taking a fixed number of seconds does not take into account the problem that the landing time cannot be accurately controlled in the actual landing scenario of the aircraft. It is only suitable for the overall flight assessment after the end of the flight and is not conducive to predicting and judging whether a hard landing will occur when the aircraft lands.

[0038] In some embodiments, this application may, but is not limited to, use the real-time position of the target aircraft as the data reference, because the aircraft's altitude is known, and the landing point, which is at an altitude of 0, is also known. Therefore, this application can accurately control the landing time of the target aircraft in the actual landing scenario by using the real-time position and its corresponding time. Here, the target aircraft can be understood, but is not limited to, as referring to the specific aircraft object for hard landing prediction.

[0039] Furthermore, in this application embodiment, the acquisition of the target aircraft's raw flight data may, but is not limited to, starting from 150ft from the target aircraft's landing point. The specific data acquisition start position may, but is not limited to, be determined by those skilled in the art based on the actual situation. This application embodiment is only for illustrative purposes and does not impose any specific limitations.

[0040] In this application embodiment, the flight data may be, but is not limited to, the aircraft's QAR flight data (Quick Access Reorder, a type of flight data stored in a data storage device that stores the aircraft's own attitude and external information during flight).

[0041] Furthermore, in this embodiment of the application, after obtaining the original flight data of the target aircraft, it is necessary to perform certain model processing on the data. Therefore, this embodiment of the application may, but is not limited to, align the granularity of the original flight data according to the data sampling frequency of the original flight data to obtain usable real-time flight data corresponding to the real-time position of the aircraft.

[0042] In other words, raw flight data contains various types of aircraft data. The data acquisition devices (such as sensors for altitude, speed, acceleration, and heading angle) for these different types of data have different sampling frequencies. This results in different granularities of the raw flight data (the smallest unit interval of data in the time dimension, i.e., the time difference between two adjacent data points). For example, altitude data sampled at 10Hz has a granularity of 0.1 seconds (one data point every 0.1 seconds), while speed data sampled at 1Hz has a granularity of 1 second (one data point every 1 second).

[0043] If these data are used directly for analysis, such as calculating the correspondence between altitude and speed at a certain moment, time misalignment may occur: 0.1 seconds of altitude data cannot find corresponding speed data, and 1.0 second of speed data corresponds to 10 altitude data points.

[0044] Therefore, the embodiments of this application can perform granular alignment processing on the raw flight data to ensure that the time intervals (granularity) of all data remain consistent, ensuring that each time unit has complete multi-dimensional data. In the embodiments of this application, the data can be interpolated or downsampled according to the target granularity (usually the granularity corresponding to the highest sampling frequency, or the granularity required by the business) to complete the granularity alignment processing of the raw flight data, thereby obtaining usable real-time flight data corresponding to the real-time position of the target aircraft.

[0045] The embodiments of this application can align the granularity of the original flight data according to the data sampling frequency of the target aircraft's original flight data, unify the time interval of this multi-source flight data, eliminate the time misalignment caused by different collection frequencies, and provide a time-synchronized data foundation for subsequent data processing.

[0046] Step S102: Determine the actual quantiles of multiple physical quantities of the target aircraft at the real-time position based on the real-time flight data, generate an importance weight matrix of multiple physical quantities based on the quantile interval of the actual quantiles, and splice the importance weight matrix and the numerical matrix corresponding to the real-time flight data to generate a spliced ​​matrix of each physical quantity among the multiple physical quantities.

[0047] Understandably, physical quantities here can be understood as physical parameters that are collected in real time by aircraft sensors and recorded in QAR during flight, reflecting the aircraft's own state, power system performance, and flight environment conditions. These include, but are not limited to, aircraft motion parameters, such as altitude, speed, heading, and attitude angles; power system parameters, such as engine speed, thrust, fuel flow, and lubricating oil pressure; and environmental and control parameters, such as ambient temperature, atmospheric pressure, landing gear position, and control surface deflection angle.

[0048] In other embodiments, this application can obtain the actual quantiles of multiple physical quantities of the target aircraft at the actual position based on real-time flight data, and assign corresponding weights to multiple physical quantities according to the quantile interval in which the actual quantiles are located, thereby generating an importance weight value matrix of multiple physical quantities.

[0049] Each physical quantity has a different impact on the landing gravitational acceleration of the aircraft when the target aircraft is in different positions. For example, the external temperature at a certain position has a greater impact on the aircraft landing, and the atmospheric pressure at a certain position has a greater impact on the aircraft landing. Therefore, the embodiments of this application can assign corresponding weights to multiple physical quantities at the real-time position based on the quantile interval of the actual quantile of multiple physical quantities at the real-time position, and generate an importance weight value matrix of each physical quantity among multiple physical quantities.

[0050] After obtaining the importance weight value matrix, this embodiment of the application can further splice the importance weight value matrix and the numerical matrix corresponding to the real-time flight data to generate a spliced ​​matrix, so as to use the spliced ​​matrix to predict the subsequent hard landing of the target aircraft.

[0051] The following section provides a further explanation of how the embodiments of this application determine the weights to be assigned when the quantiles of multiple physical quantities at multiple real-time locations are in different quantile intervals.

[0052] Optionally, in one embodiment of this application, before generating the importance weight matrix of multiple physical quantities based on the quantile interval where the actual quantile is located, the method further includes: constructing a comprehensive distribution of multiple physical quantities in multiple normal landing samples based on the values ​​of multiple physical quantities in multiple time units of the target landing segment; dividing each time unit in the target landing segment into multiple quantile intervals based on a pre-set quantile boundary value; and determining the weights corresponding to the multiple quantile intervals by combining the comprehensive distribution and the values ​​of the multiple time units, so as to determine the weight value of the quantile interval where the actual quantile is located based on the weights corresponding to the multiple quantile intervals.

[0053] In some embodiments, this application may, but is not limited to, determine the weight values ​​that should be assigned to the quantile intervals of multiple physical quantities at multiple real-time locations based on multiple normal landing samples.

[0054] First, in the embodiments of this application, the comprehensive distribution of multiple physical quantities in multiple normal landing samples can be constructed based on the values ​​of multiple physical quantities in multiple time units during the target landing segment.

[0055] In this context, the comprehensive distribution of multiple physical quantities can be understood as the combined data of the quantiles of each physical quantity in each time unit during the landing segment, across multiple normal landing flight data samples. Therefore, when concatenating the original numerical matrix and the importance weight feature matrix corresponding to multiple physical quantities, in addition to storing the individual information of the physical quantities at the current time step, the distribution relationship of a certain physical quantity in the comprehensive distribution at the current time step can also be stored. This can be understood as a data representation of the group-individual characteristic relationship of aircraft physical quantities.

[0056] Specifically, in order to maintain data richness during the construction process, this application embodiment can simulate and construct a certain number of normal landing samples. To simulate the actual landing process of an aircraft, this application embodiment does not extract data from the dataset according to a fixed number of time units before and after the touchdown point, as is the conventional method for analyzing hard landings. Instead, it uses an altitude that better reflects the aircraft landing situation as a benchmark, setting the starting point for data collection at an altitude of 150 feet above the ground. Multiple normal landing samples are extracted from this time unit, making the samples more consistent with the dynamic process of an actual aircraft landing.

[0057] Then, embodiments of this application may collect, but are not limited to, 16 physical quantities that are highly correlated with the aircraft's state at landing. The sampling frequencies of these 16 key physical quantities vary, ranging from 8Hz to 1Hz. Table 1 shows the physical quantities and their sampling frequencies in one embodiment of this application, and may be represented, but is not limited to, as follows:

[0058]

[0059] Similar to the data processing of the original flight data, considering that the data in the 10-20s range from 150 feet to the landing point is generally quite scattered, and that the sampling frequencies of 1Hz and 8Hz are different, the number of data points for different physical quantities within the same time period will vary greatly. To align the various physical quantities in terms of time granularity, embodiments of this application may, but are not limited to, use interpolation to supplement physical quantity data that are less than 8Hz to 8Hz, while taking a fixed 200 data points (corresponding to 200 flight positions) to ensure that the data points and the landing points among the data points must exist in the data sample.

[0060] Then, embodiments of this application may, but are not limited to, determine the comprehensive distribution corresponding to each physical quantity based on the multiple quantiles of multiple physical quantities in the normal landing sample for each time unit (i.e., time units corresponding to 200 flight positions) corresponding to 200 data points in the landing segment.

[0061] For example, this application can perform statistics on a massive number of normal landing samples (assuming tens of thousands of normal landings) at each point of the 200 time units after alignment in the previous step. The value of each physical quantity in each time unit of the landing segment in the tens of thousands of normal landings constitutes the comprehensive distribution of each physical quantity in the normal landing samples.

[0062] Then, in this embodiment, each time unit in the target landing segment can be divided into multiple quantile intervals according to the pre-set quantile boundary values. The weights corresponding to these quantile intervals are determined by combining the comprehensive distribution of multiple physical quantities in multiple normal landing samples and the values ​​of multiple time units. Here, the target landing segment can be understood as the period (time) from 150 feet to a certain position after the landing point, such as from 150 feet to 20 feet after landing. The specific time range can be determined by those skilled in the art based on actual conditions. This embodiment is merely illustrative and does not impose specific limitations.

[0063] In this embodiment of the application, the pre-defined quantile boundary values ​​include, but are not limited to, six values: 2.5%, 10%, 90%, and 97.5%. The 2.5%, 10%, 90%, and 97.5% quantile boundary values ​​respectively represent how many data points in the sample are smaller than the physical quantity value corresponding to the current quantile. For example, the 2.5% quantile boundary value means that only 2.5% of all data points are smaller than the data value corresponding to the 2.5% quantile boundary value, and the 97.5% quantile boundary value means that only 2.5% of all data points are larger than the data value corresponding to the 97.5% quantile boundary value.

[0064] For example, taking the descent rate of the 100th time unit as an example: In this embodiment of the application, the descent rate value of each sample in the 100th time unit is extracted from tens of thousands of normal landing samples and sorted. The descent rate value in the 100th time unit among these tens of thousands of normal landing samples is the comprehensive distribution, which belongs to the group (characteristic), and the descent rate value of a single sample in the 100th time unit is the individual (characteristic).

[0065] Then, based on the rate of decline value at the 100th time unit, we can obtain how much data falls within which interval. For example, 2.5% of the values ​​are below value A, and 5% of the values ​​are below value B. This application can obtain the physical quantity value corresponding to each physical quantity at the pre-set 2.5%, 5%, 10%, 90%, 95%, and 97.5% quantile boundary values.

[0066] Then, by processing the data of each physical quantity in each of the 200 time units in the same way, we can obtain the physical quantity values ​​corresponding to the pre-set 2.5%, 5%, 10%, 90%, 95%, and 97.5% quantile boundary values ​​of each physical quantity in the 200 time units. By connecting the 2.5%, 5%, 10%, 90%, 95%, and 97.5% quantile boundary values ​​of each physical quantity in the 200 time units, we can obtain four curves that change with time (200 time units): the 2.5% quantile line, the 5% quantile line, the 10% quantile line, the 90% quantile line, the 95% quantile line, and the 97.5% quantile line. The six lines for each physical quantity divide each time unit in the target landing phase into seven quantile intervals: 0%-2.5%, 2.5%-5%, 5%-10%, 10%-90%, 90%-95%, 95%-97.5%, and 97.5%-100%.

[0067] The above process is repeated for each of the 16 physical quantities, and finally, for each physical quantity, the quantile intervals shown above are obtained.

[0068] Figure 2 This is a schematic diagram of the quantile intervals of altitude in a normal landing sample according to an embodiment of this application; Figure 3 This is a schematic diagram of the quantile intervals of pitch angle in a normal landing sample according to an embodiment of this application; Figure 4 This is a schematic diagram illustrating the quantile intervals of vertical velocity in a normal landing sample according to an embodiment of this application. Figure 2 , Figure 3 and Figure 4 As shown, the dark black line in the three figures represents the actual quantile values ​​of the physical quantities altitude, pitch angle, and vertical velocity over 200 time units in a normal landing flight data set. Excluding this dark black line, from top to bottom are the 97.5% quantile line, 95% quantile line, 90% quantile line, 10% quantile line, 5% quantile line, and 2.5% quantile line, along with their respective intervals.

[0069] as well as, Figure 5 This is a schematic diagram of the quantile distribution of altitude in a hard landing sample according to an embodiment of this application; Figure 6 This is a schematic diagram of the quantile distribution of pitch angles in a hard landing sample according to an embodiment of this application; Figure 7 This is a schematic diagram of the quantile distribution of vertical velocity in a heavy landing sample according to an embodiment of this application. By observing the overall distribution diagrams of multiple physical quantities in normal landing and heavy landing, it can be clearly observed that the distribution of physical quantities in the heavy landing sample is more dispersed than that in the normal landing sample in certain time units.

[0070] like Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, through the overall distribution diagrams of multiple physical quantities during normal landing and re-landing, the reasonable range of each key physical quantity parameter at each instant during normal landing can be accurately quantified. For example, 95% of normal values ​​fall between the 2.5% and 97.5% quantiles. Then, embodiments of this application can construct the weight value that should be assigned to each physical quantity when it is within each quantile interval in each time unit.

[0071] In this embodiment, the weight values ​​can be designed based on, but are not limited to, the criteria for a hard landing: According to Airbus's "Flight Quality Assurance (FOQA) Implementation and Management" definition of a large vertical overload during landing: within 2 seconds before touchdown to 5 seconds after touchdown, a vertical overload >1.60g is considered a slight deviation, and >1.80g is considered a severe deviation. Considering safety, this embodiment uses a vertical acceleration >1.50g during landing as the criterion for a hard landing.

[0072] Based on this, for sample points located in the 10%-90% quantile range (the largest number of data points, representing the most normal range), this embodiment of the application uses a normal acceleration of 1g (9.8m / s²). 2 The weighting is assigned by analogy "1" for sample points located in the 5%-10% and 90%-95% range (a small portion of data that deviates slightly from the normal range); by analogy "1.5" for sample points located in the 2.5%-5% and 95%-97.5% range (a small portion of data that deviates significantly from the normal range); and by analogy "1.75" for sample points located in the 0%-2.5% and 97.5%-100% range (equivalent to a small portion of data that deviates completely from the normal range). By analogy "2" for sample points located in the 0%-2.5% and 97.5%-100% range (equivalent to a small portion of data that deviates completely from the normal range).

[0073] In other words, the greater the deviation of the actual data of a physical quantity from the normal range, the greater its weight, indicating that it is most likely to occur at its maximum acceleration during the landing phase. For example, if the value of a physical quantity in a certain time unit falls within the 95%-97.5% quantile range, it means that this physical quantity of the aircraft in that time unit is larger than that of 95% of flights in history, and is an "outlier." Therefore, the weight assigned to this physical quantity in that time unit should be greater.

[0074] In simple terms, the weight assigned to a physical quantity in this embodiment is a coefficient that indicates the actual quantile of the physical quantity deviates from the main distribution position. The further it deviates, the larger its value. By assigning a value, the distribution of a single physical quantity in a specific time unit can be quantified.

[0075] The expressions used to assign weight values ​​in the embodiments of this application can, but are not limited to, the following:

[0076]

[0077] in, Indicates the first Quantile lines corresponding to each quantile boundary value Indicates the first A physical quantity in Quantiles in time units That is, the first A physical quantity in The quantile weights are calculated from the quantiles within the time unit. Indicates the first A physical quantity in The value of the time unit, Indicates the first The boundary of the quantile, Indicates the first A physical quantity in a time unit The first Each quantile interval.

[0078] By assigning corresponding weights to time units corresponding to multiple physical quantities in a landing sample, this embodiment of the application can generate an importance weight value matrix containing the group-individual characteristics of flight data by splicing together the importance weights of each physical quantity.

[0079] First, the embodiments of this application can construct the first n Weighted eigenvectors of each physical quantity:

[0080]

[0081] in, Represented as the first The weighted eigenvectors of the physical quantities are 200×1 eigenvectors. This is expressed as the physical quantity in a time unit. The importance weights are assigned, such as 1 / 1.5 / 1.75 / 2, etc. Combining the feature vectors of multiple physical quantities, for example, 16 feature vectors, the final result is a 200×16 importance weight matrix. .

[0082] After obtaining the importance weight matrix of multiple physical quantities, the concatenated matrix is ​​obtained by splicing the numerical matrices of multiple physical quantities (all matrices are 200×16 (200 time points × 16 physical quantities)). This splicing process is essentially an interaction between the individual (overall distribution of a single physical quantity) and the group (overall distribution of all physical quantities). By comparing the distribution of the individual in the group, that is, the distribution of each physical quantity (specific numerical information and its quantile interval) and the distribution of all physical quantities (specific numerical information of all physical quantities and their quantile interval), it is clear which physical quantities are abnormal at real-time locations in the target landing segment. It also provides a clearer picture of the target aircraft's flight status at real-time locations (whether there are any physical quantity anomalies) and the differences between normal landing and hard landing (how many physical quantities are abnormal, the degree of the anomaly, etc.). This enriches the group-individual interaction information of the feature matrix, allowing the dataset to retain both specific numerical information and distribution information within the group.

[0083] Use the importance weight matrix For example, in this embodiment of the application, the importance weight value matrix and the numerical matrix corresponding to the historical flight data can be concatenated to obtain the following concatenated matrix:

[0084]

[0085] in, It is a numerical matrix of physical quantities. This is the feature matrix, where the data column dimension is expanded to twice that of the original flight data.

[0086] This application embodiment can, in data processing, integrate the quantifiable relationship between physical quantities in their respective time units (altitude positions) and the distribution of group data, capture the global characteristics of each physical quantity of the target aircraft at its real-time position, and deeply mine the influence of the value of each physical quantity in each time unit and the distribution of group data on the aircraft's gravitational acceleration. This makes up for the shortcomings of most models in related technologies that only focus on the specific physical quantity values ​​or only use the distribution of physical quantities as a visualization means, and effectively improves the accuracy of predicting the specific value of gravitational acceleration at the landing point.

[0087] Optionally, in one embodiment of this application, determining the actual quantiles of multiple physical quantities of the target aircraft at its real-time location based on real-time flight data, and generating an importance weight matrix of multiple physical quantities based on the quantile interval in which the actual quantiles are located, includes: determining the time unit corresponding to the real-time location based on the real-time location; and generating an importance weight matrix of multiple physical quantities based on the time unit corresponding to the real-time location and the quantile interval in which the actual quantiles are located.

[0088] Based on the relevant descriptions in other embodiments, it can be understood that when constructing the importance weight value matrix based on landing samples, this application uses 200 time units in the target landing segment as a basis, and subsequent processing also needs to be based on time units.

[0089] Therefore, in actual execution, for ease of processing, this application can determine the time unit corresponding to the real-time position based on the real-time position of the target aircraft, and then generate an importance weight matrix of multiple physical quantities based on the weights corresponding to the time unit corresponding to the real-time position and the weights of the quantile intervals where the actual quantiles of multiple physical quantities corresponding to the real-time position are located.

[0090] It should be noted that in the actual acquisition of the target aircraft's raw flight data, data collection begins at an altitude of 150 feet. The data collected is the value of multiple physical quantities in the first time unit corresponding to the altitude of 150 feet. The weight feature vector of each physical quantity has only one element t1. The feature vectors of the 16 physical quantities form a 1×16 importance weight value matrix. The 16 physical quantities also correspond to a 1×16 numerical matrix. The combination of these features forms a 2×16 concatenation matrix.

[0091] As the aircraft's real-time position gradually decreases, the weight eigenvectors of each physical quantity gradually increase, corresponding to the weight values ​​t1, t2, t3, t4, t5, ... for each real-time position. The size of the importance weight matrix composed of the eigenvectors of the 16 physical quantities is 1×16, 2×16, 3×16, 4×16, 5×16, ... The numerical matrices corresponding to the 16 physical quantities also increase sequentially to 1×16, 2×16, 3×16, 4×16, 5×16, ... The spliced ​​matrix is ​​formed by splicing the importance weight matrix and the numerical matrix corresponding to each real-time position.

[0092] Optionally, in one embodiment of this application, before obtaining the location code corresponding to the real-time location, the method further includes: constructing a target linear function and a target sine function for obtaining the location code based on the time unit in the target landing segment; and designing the location code based on the target linear function and the target sine function.

[0093] As one possible approach, embodiments of this application can focus the prediction of aircraft hard landing on a certain time unit before and after the landing point in the target landing segment. Therefore, embodiments of this application can, but are not limited to, dynamically capture the time units before and after the landing point in the target landing segment to focus on the prediction of gravitational acceleration in this part of the time unit.

[0094] In actual implementation, this application may, but is not limited to, obtain the dynamic window corresponding to the real-time position based on the position code corresponding to the real-time position of the target aircraft.

[0095] Before being applied to the target aircraft, this application first constructs a target linear function and a target sine function for obtaining the position code based on the time unit in the target landing segment of the landing sample, and then designs the position code according to the target linear function and the target sine function.

[0096] Here, the target linear function and the target sine function can be understood as linear and sine functions that reflect the mathematical positional relationship between the aircraft's real-time position and the landing point.

[0097] For ease of application, embodiments of this application may, but are not limited to, design the location encoding and subsequent gravity acceleration prediction as a certain model, and the basic architecture of the model may, but is not limited to, adopt the Transformer architecture. Figure 8 This is a schematic diagram of the structure of an aircraft hard landing prediction model according to an embodiment of this application, as shown below. Figure 8 As shown in the embodiments of this application, the aircraft hard landing prediction model includes, but is not limited to, an input part, a position encoding part, an encoder part, a decoder part, and an output part. The input part is a concatenation matrix, and the position encoding part encodes the real-time position of the target aircraft. Since the Transformer model itself does not have the ability to directly identify sequential position information, the position encoding module in this embodiment uses the sum of the position encoding and the concatenation matrix to allow the aircraft hard landing prediction model to clearly determine which time unit the physical quantity is currently in during the landing phase. Simultaneously, this facilitates the determination of the dynamic window center point of each landing sample, i.e., the position encoding of the landing point, preparing for subsequent dynamic window design.

[0098] First, embodiments of this application can construct linear time coding units and periodic time coding units for each time unit, and their expressions can be, but are not limited to, as follows:

[0099]

[0100] in, Let be a learnable parameter matrix. If we consider the change of gravitational acceleration over time as a waveform, then ... This is a frequency parameter matrix, whose values ​​reflect the period of a local waveform within the overall waveform. If... A smaller period indicates a shorter period, which can capture high-frequency changes in gravitational acceleration over time; conversely, a larger period indicates a longer period, which can capture low-frequency changes in gravitational acceleration over time. The phase offset parameter matrix determines the start and end positions of the waveform; each time unit They will all go through a process based on learnable parameters (Frequency parameter matrix and phase offset parameter matrix) The Middle The above consists of (number of parameters). The mapping yields the time code corresponding to that position, where... It can be understood as a mapping function between the frequency parameter matrix and the phase offset parameter matrix obtained through mapping training and the time unit, based on the known frequency parameter matrix and phase offset parameter matrix; This is the input to the model. It is a linear time encoding unit used to represent the length of a period; For periodic time encoding units, a sine function is used to capture periodic time information; , express The first position, , They represent From the second position to the last position, , , , These represent the dimensions corresponding to each parameter.

[0101] The linear time coding unit and the periodic time coding unit can be used to obtain the linear time code and periodic time code corresponding to the real-time position of the target aircraft, respectively.

[0102] Combining the two yields:

[0103]

[0104] in, This indicates a splicing operation. This represents the dimension after position encoding. This refers to the position encoding (time encoding) part in the aircraft hard landing prediction model, which is equivalent to the final position encoding obtained by integrating the linear time encoding and periodic time encoding corresponding to the real-time position of the target aircraft.

[0105]

[0106] in, The input to the encoder is actually a fused position code. With batch splicing matrix X The information that follows.

[0107] Optionally, in one embodiment of this application, before generating the predicted value of gravity acceleration for the real-time position based on the time unit corresponding to the real-time position, the method further includes: combining the splicing matrix of multiple physical quantities of at least one landing sample with the position encoding to obtain a splicing result; inputting the splicing result into a pre-built encoder to output the key parameter matrix, value parameter matrix, and query parameter matrix of the pre-built encoder; and constructing a dynamic window centered on the landing point and a soft gating unit based on the target Gaussian function based on the position information in the key parameter matrix, value parameter matrix, and query parameter matrix; and determining the predicted value of gravity acceleration based on the time deviation between each flight position and the landing point in the target landing segment of at least one landing sample. The embedding matrix of the dynamic window in the target landing segment is determined. Based on the dynamic window, causal logic slices of the key parameter matrix and value parameter matrix are obtained. Based on the causal logic slices of the key parameter matrix and value parameter matrix and the pre-built soft gating unit based on the target Gaussian function for capturing the landing window, a decoder is constructed to decode the causal logic slices of the key parameter matrix and value parameter matrix and the causal logic slices of the query parameter matrix. Based on the decoder, the causal logic slices of the key parameter matrix and value parameter matrix and the causal logic slices of the query parameter matrix are decoded to obtain the gravity acceleration prediction value for each flight position. The gravity acceleration prediction value for the real-time position is determined based on the gravity acceleration prediction value for each flight position.

[0108] Based on the descriptions of other embodiments, it is understood that the embodiments of this application can focus the attention mechanism on the time units before and after the landing point, in order to focus on calculating the predicted value of gravitational acceleration in this part. In the embodiments of this application, this part can be implemented based on the gating unit of the dynamic landing window attention mechanism.

[0109] Figure 9 This is a schematic diagram of the structure of a gating unit based on a dynamic landing window attention mechanism according to an embodiment of this application. Figure 9 As shown, the gating unit based on the dynamic landing window attention mechanism in this embodiment mainly includes, but is not limited to, encoder and decoder modules. The encoder is primarily used to generate the key parameter matrix, value parameter matrix, and query parameter matrix of the encoder based on the splicing matrix of multiple physical quantities and the splicing result of the position encoding. In this system, there is a one-to-one correspondence between time units and physical quantities, with each time unit having a corresponding physical quantity.

[0110] Specifically, Figure 9 The frame to the right of the bottommost physical quantity represents the time unit (different time window) corresponding to the entire landing segment. The circle under each time unit represents 16 physical quantity values ​​and their 16 weight values ​​enhanced based on the comparative features of the group and individuals.

[0111] The slicing unit corresponds to the K and V parameter slicing operation at the acquired time window, the Gaussian soft coding layer corresponds to the operation of the soft gating mechanism of the attention part designed based on the Gaussian function, and the causal masking layer corresponds to the operation related to the causal logic mechanism. These parts belong to the decoder part.

[0112] In actual implementation, the gating unit based on the dynamic landing window attention mechanism of this application is mainly implemented through an encoder-decoder module. The encoder module in the embodiments of this application is mainly implemented using a pre-built encoder. Here, the pre-built encoder can be understood as adopting a conventional encoder architecture, which is used to construct an encoder with three global parameters, Q (query), K (key value), and V (value), based on the position encoding corresponding to the real-time position.

[0113] The dynamic window attention mechanism dynamically captures landing points in the training dataset and focuses attention on the intervals near the landing points. Based on this dynamic window attention mechanism, the embodiments of this application can achieve enhanced attention to the landing interval, that is, reduce the weight attention of the K and V parts in the Transformer architecture from the entire landing segment to only focus on the landing interval, that is, focus more on the positional information before and after landing, such as 50 feet before landing and 20 feet after landing.

[0114] In the decoder module, this application mainly designs a window encoding mechanism for K and a soft gating mask based on Gaussian function. The main implementation results are as follows: under the condition of ensuring causal order, the K and V parameters are sliced ​​according to the dynamic window part, and the slicing process is the extraction process of dynamic window data; the sliced ​​parameter V and the global parameter Q are calculated by an attention mechanism to obtain the intermediate variable A; the intermediate variable A is multiplied element-wise with the soft attention mask based on Gaussian function to realize the model adaptively adjusts the attention intensity of the landing point at different time steps; subsequently, the decoder output result is mapped by a linear layer to obtain the real-time gravity acceleration prediction value of the current time unit in the prediction module.

[0115] First, in this embodiment, the stitching result can be obtained by combining the stitching matrix of multiple physical quantities of at least one landing sample with the position encoding. The stitching result is then input into a pre-built encoder to output the key parameters, value parameters, and query parameters of the pre-built encoder. Based on the position information (which comes from the position encoding) in the key parameters, value parameters, and query parameters, a dynamic window centered on the landing point and a soft gating unit based on the target Gaussian function for capturing the landing window are constructed.

[0116] Specifically, the stitching result of the stitching matrix of multiple physical quantities of at least one landing sample and the position encoding After inputting the encoder, the embodiments of this application can... and That is, the learnable parameter matrices corresponding to the encoder parameters such as key parameters, value parameters, and query parameters are multiplied together to obtain the encoder's key parameter matrix, value parameter matrix, and query parameter matrix. :

[0117]

[0118] Then, in this embodiment of the application, the embedding matrix of the dynamic window in the target landing segment can be determined based on the time deviation between each flight position and the landing point in at least one landing sample target landing segment.

[0119] Specifically, for the decoder of dynamic windows, an embedding matrix is ​​constructed to describe the dynamic window. as follows:

[0120]

[0121] in, , These represent the initial point and the second point of the dynamic window, respectively. Indicates the landing point. The window length is 48 in this embodiment, including 24 time units before landing and 24 time units after landing.

[0122] Here, since the focus is on the dynamic window, the position encoding in this embodiment is the position encoding of the current position relative to the landing point. That is, the position encoding corresponds to the physical quantity being at a certain position before the landing point (e.g., 24 time units before the landing point) and a certain position after the landing point (e.g., 24 time units after the landing point).

[0123] In the above formula This is represented as the absolute position information within the window relative to the landing point, i.e., the time deviation between each flight position and the landing point during the target landing segment:

[0124]

[0125] in, The time-series label is used to represent information at the position of the dynamic window in each batch, i.e., the first time unit is 0, the second time unit is 1, and so on; , They represent the first time. , The time index corresponding to each location.

[0126] Then, in this embodiment of the application, causal logic slices of the key parameter matrix and value parameter matrix can be obtained according to the dynamic window. Based on the causal logic slices of the key parameter matrix and value parameter matrix and the pre-built soft gating unit based on the target Gaussian function for capturing the landing window, causal logic slices for decoding the key parameter matrix and value parameter matrix and causal logic slices for querying the parameter matrix can be constructed (using...). Figure 9 The decoder (of the slice unit in the code) is used to decode the causal logic slices of the key parameter matrix and value parameter matrix, and the causal logic slices of the query parameter matrix, to obtain the predicted value of gravity acceleration for each flight position.

[0127] Specifically, embodiments of this application may further define the start node (12th time unit before landing point) and end node (12th time unit after landing point) of the dynamic window:

[0128]

[0129] in, s , e These represent the start and end nodes of the dynamic window, respectively. After slicing K and V, they are added to the aforementioned landing point position matrix, embedding the time information of the dynamic window relative to the landing point. In this process, the key parameters and value parameters of the decoder were obtained: ,in,

[0130]

[0131] in, This is the embedding matrix for the time unit corresponding to the window position of the landing point. , These represent the K and V parameters after slicing, respectively.

[0132] Meanwhile, this application's embodiments design a causal logic mechanism for the attention mechanism (corresponding to...). Figure 9 (the causal mask layer in the middle), that is, ensuring that in the first... Only one time step can be used Information about the time step:

[0133]

[0134] in, In response to The mask mechanism during querying ensures that the attention mechanism will not focus on a step when the position is negative infinity, thus guaranteeing the rigor of the sequential logic. For intermediate variables in the intermediate attention mechanism that are not activated by softmax, To make the hyperparameters easier to learn after softmax activation, To ensure causality, a masking function is used such that subsequent information is invisible at any given time unit, preventing information leakage. The initial attention weights are then obtained after activation by the softmax function. .

[0135] Then, in this embodiment of the application, a soft gating mechanism for the attention part is designed based on the target Gaussian function (corresponding to...). Figure 9 The Gaussian soft-coding layer in the model enables the model to adaptively adjust the intensity of attention at different time steps, achieving more refined temporal modeling during the critical landing phase. Here, the target Gaussian function can be understood as a mathematical tool used in the attention mechanism to dynamically adjust the attention intensity at different time steps. In this embodiment, the target Gaussian function can be, but is not limited to, expressed as follows:

[0136]

[0137] in, This indicates that the time unit The Gaussian correction parameter value, here Indicates the dynamic window length. The time deviation of each time point in the entire time series relative to the landing point can be expressed as, but is not limited to, as:

[0138]

[0139] in, Indicates the first The time index corresponding to each location. Statement No. Each location corresponds to a time index of the landing point.

[0140] Then, the designed Gaussian function (to integrate all time units) (functions) and Calculate using the following formula:

[0141]

[0142] in, This represents the attention weight matrix after soft gating using a Gaussian function. Indicates the initial attention weights. Represents element-wise product. Represents a learnable proportional parameter. The Gaussian function is designed.

[0143] Therefore, the aforementioned soft gating compensates for the lack of attention to the landing point in non-dynamic windows, for The overall distribution of the function shows that the value decreases as the distance from the landing point increases and the value increases as the distance from the landing point increases. This gating unit enables the differentiated representation of different positions at the entire landing segment level.

[0144] At the same time, in order to ensure that the parameters of the attention mechanism are always 1 in the summation, and to maintain The relative proportions of all parameters in this application can be determined in the following ways in the embodiments. Perform secondary normalization by row:

[0145]

[0146] Finally, the module outputs the following:

[0147]

[0148] in, In the attention mechanism, and The value produced by multiplication for The corresponding learnable parameter matrix. Finally, after projection through the linear layer, the predicted value of gravitational acceleration for each time unit can be obtained. .

[0149] Specifically, the loss function of the aircraft hard landing prediction model in the embodiments of this application can be, but is not limited to, expressed as follows:

[0150]

[0151] in, This represents the global mean square error. It is the mean square error over the 24 time units before and after the landing interval. , It is a weighting coefficient, which can be, but is not limited to, 0.4 or 0.6 in the model of this application embodiment.

[0152] During the training process of the aircraft's gravity acceleration prediction model, the parameters of the learnable parameter matrix in each module of the model are changed by learning, such as... This allows for the extraction of deep relationships between various physical quantities and gravitational acceleration within time units corresponding to different flight positions. The value of the loss function can then continuously decrease, and after training, the learnable parameter matrices ( Once the parameters (such as gravity acceleration, gravitational acceleration, etc.) are determined, a gravity acceleration prediction function can be obtained, which uses a concatenated matrix of multiple physical quantities as variables and gravity acceleration as the function value. This function can be directly applied. In practical applications, only the flight physical variables for the current time unit need to be input, and the model can calculate the predicted gravity acceleration value for the current time unit based on the existing parameters.

[0153] In summary, Figure 10 This is a flowchart illustrating aircraft hard landing prediction according to one embodiment of this application. Figure 10 As shown in the figure, N represents the number of layers of the dynamic window attention gating unit used in the application embodiment. In this application embodiment, the value is 3, that is, there are three such units. In actual application, the number of layers of the dynamic window attention gating unit can be determined by those skilled in the art according to actual needs. This application embodiment is only for illustrative purposes and does not impose any specific limitations.

[0154] The specific data stream based on the dynamic landing window attention gating unit has been included. Figure 9 As shown in the diagram, the data preprocessing module corresponds to the operation of aligning the granularity of the original flight data according to the data sampling frequency of the original flight data. The weight calculation module and the interaction module correspond to the operation of enhancing the comparative features of the group and individuals. That is, the weight calculation module is used to assign corresponding weights to each physical quantity, and the interaction module is used to concatenate the weighted physical quantities with the numerical matrix corresponding to the original real-time flight data to obtain the concatenated matrix.

[0155] Furthermore, the residual connections, layer normalization, and feedforward neural networks are based on the network connection structure in the attention gating unit of the dynamic landing window.

[0156] After obtaining the splicing matrix, this embodiment of the application can use linear and sine functions to write the position of the splicing matrix to obtain the position code of each position; and take 12 time units before and after the landing point as the center as the window length to divide the dynamic window position within the target range of the landing point in the landing segment, so as to further design the K and V parameters of the decoder.

[0157] Then, in this embodiment, the three basic parameters Q, K, and V can be obtained based on the Transformer encoder architecture. The K and V parameters are then sliced. Simultaneously, based on a Gaussian function soft-gated mask, the relative deviation of each time unit in the landing phase relative to the landing time unit is obtained. In terms of the Gaussian function mapping rule, the greater the relative deviation, the lower the value, and the value decreases significantly after a certain absolute deviation. Based on the sliced ​​K and V parameters, as well as the Q parameter and Gaussian function, a decoder for the prediction model is designed. The output of the decoder, after being mapped by a linear layer, is the real-time gravitational acceleration prediction value of the target aircraft in each time unit.

[0158] Step S103: Based on the splicing matrix, obtain the position code corresponding to the real-time position; based on the position code and splicing matrix, obtain the time window corresponding to the real-time position; obtain the time unit corresponding to the real-time position based on the time window; generate the gravity acceleration prediction value of the real-time position based on the time unit corresponding to the real-time position; and generate the heavy landing prediction result of the target aircraft.

[0159] In actual execution, the processing flow after splicing the matrix is ​​the same as in this paper. This application can obtain the position code corresponding to the real-time position of the target aircraft based on the splicing matrix of multiple physical quantities in the real-time flight data. Combine the position code corresponding to the real-time position with the splicing matrix, and input the combined data into the aircraft's gravity acceleration prediction model to obtain the time window corresponding to the real-time position. Then, the time unit corresponding to the real-time position is obtained according to the time window corresponding to the real-time position. That is, the time window corresponding to the real-time position is determined according to the time deviation corresponding to the position deviation of the real-time position before and after the landing point, such as the nth time window before the landing point. This window corresponds to the time unit corresponding to the real-time position, that is, the nth time window before the landing point.

[0160] Then, the embodiments of this application can be based on the aircraft gravity acceleration prediction model and the time unit corresponding to the real-time position. For example, the time window of the landing segment corresponds to the k-23 to kth time unit before the landing point. Based on the attention and calculation method of the gravity acceleration prediction value calculated in the time unit corresponding to the real-time position and the time unit in the landing sample, the gravity acceleration prediction value of the real-time position is calculated, and finally the heavy landing prediction result of the target aircraft is generated.

[0161] For example, if the predicted real-time gravitational acceleration at each location is greater than 1.50g, especially at the real-time location corresponding to the second time unit before the landing point, a risk of a hard landing can be predicted. Based on the predicted gravitational acceleration at the real-time location, the pilot can be assisted in determining the possibility of a hard landing and adjust flight operations accordingly to address this risk.

[0162] Furthermore, to demonstrate the effectiveness of this gating unit, Figure 11 Weights as intermediate variables in the attention mechanism of one embodiment of this application The query mapping at different time units of the global weight Q is reflected in the weight after slicing. A diagram illustrating the magnitude of attention weights. (From...) Figure 11 It can be seen that within the dynamic window near the landing point, the size distribution of the attention weights is larger and exhibits a linear bar pattern, indicating that the dynamic window attention mechanism in the encoder-decoder module is effective.

[0163] Furthermore, Figure 12 This is a schematic diagram illustrating the prediction results of a normal landing sample according to one embodiment of this application. Figure 13 This is a schematic diagram illustrating the prediction results of a re-landing sample according to an embodiment of this application. Figure 14 This is a schematic diagram of the confusion matrix of the prediction results of one embodiment of this application. It can be seen that the focus on the landing time unit is the highest in this embodiment of the application, and the prediction effect of the most critical time unit for re-landing prediction and early warning - the grounding point (landing point) - is significantly enhanced compared with other location points.

[0164] Next, referring to the accompanying drawings, an aircraft hard landing prediction device based on a dynamic window attention mechanism according to an embodiment of this application is described.

[0165] Figure 15 This is a schematic diagram of the structure of the aircraft hard landing prediction device based on the dynamic window attention mechanism according to an embodiment of this application.

[0166] like Figure 15 As shown, the aircraft hard landing prediction device 10 based on dynamic window attention mechanism includes: a preprocessing module 100, a weighting module 200 and a prediction module 300.

[0167] The preprocessing module 100 is used to acquire the raw flight data of the target aircraft's real-time position, align the granularity of the raw flight data according to the data sampling frequency, and obtain the real-time flight data of the real-time position. The weighting module 200 is used to determine the actual quantiles of multiple physical quantities of the target aircraft at the real-time position based on the real-time flight data, generate an importance weight matrix of multiple physical quantities based on the quantile interval of the actual quantiles, and concatenate the importance weight matrix and the numerical matrix corresponding to the real-time flight data to generate a concatenated matrix of each physical quantity. The prediction module 300 is used to obtain the position code corresponding to the real-time position based on the concatenated matrix, obtain the time window corresponding to the real-time position based on the position code and the concatenated matrix, obtain the time unit corresponding to the real-time position based on the time window, and generate the predicted value of gravity acceleration of the real-time position based on the time unit to generate the heavy landing prediction result of the target aircraft.

[0168] Optionally, in one embodiment of this application, it further includes: a first construction module, configured to construct a comprehensive distribution of multiple physical quantities in multiple normal landing samples based on the values ​​of multiple physical quantities in multiple time units of the target landing segment before generating an importance weight matrix of multiple physical quantities based on the quantile intervals in which the actual quantile is located; a division module, configured to divide each time unit in the target landing segment into multiple quantile intervals based on pre-set quantile boundary values; and a first determination module, configured to determine the weights corresponding to the multiple quantile intervals by combining the comprehensive distribution and the values ​​of the multiple time units, so as to determine the weight value of the quantile interval in which the actual quantile is located based on the weights corresponding to the multiple quantile intervals.

[0169] Optionally, in one embodiment of this application, the weighting module 200 includes: a determining unit, configured to determine the time unit corresponding to the real-time position based on the real-time position; and a generating unit, configured to generate an importance weight matrix of multiple physical quantities based on the time unit corresponding to the real-time position and the quantile interval in which the actual quantile is located.

[0170] Optionally, in one embodiment of this application, it further includes: a second construction module, configured to construct a target linear function and a target sine function for obtaining the position code based on the time unit in the target landing segment before obtaining the position code corresponding to the real-time position; and a first design module, configured to design the position code according to the target linear function and the target sine function.

[0171] Optionally, in one embodiment of this application, it further includes: an encoding module, configured to, before generating the predicted value of gravity acceleration for the real-time position based on the time unit corresponding to the real-time position, combine the splicing matrix of multiple physical quantities of at least one landing sample with the position encoding to obtain a splicing result, input the splicing result into a pre-built encoder to output the key parameter matrix, value parameter matrix, and query parameter matrix of the pre-built encoder, and construct a dynamic window centered on the landing point and a soft gating unit based on the position information in the key parameter matrix, value parameter matrix, and query parameter matrix; and a second determination module, configured to, based on the time deviation between each flight position and the landing point in the target landing segment of at least one landing sample, determine the time deviation to determine the landing point. The target landing segment includes an embedding matrix for the dynamic window; a third construction module, which acquires causal logic slices of the key parameter matrix and value parameter matrix based on the dynamic window, and constructs a decoder for decoding the causal logic slices of the key parameter matrix and value parameter matrix and the causal logic slices of the query parameter matrix based on the causal logic slices of the key parameter matrix and value parameter matrix and a pre-built soft gating unit based on the target Gaussian function for capturing the landing window; and a decoding module, which decodes the causal logic slices of the key parameter matrix and value parameter matrix and the causal logic slices of the query parameter matrix based on the decoder to obtain the gravity acceleration prediction value for each flight position, and determines the gravity acceleration prediction value for the real-time position based on the gravity acceleration prediction value for each flight position.

[0172] It should be noted that the foregoing explanation of the aircraft hard landing prediction method based on dynamic window attention mechanism also applies to the aircraft hard landing prediction device based on dynamic window attention mechanism in this embodiment, and will not be repeated here.

[0173] Figure 16 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0174] The memory 1601, the processor 1602, and the computer program stored on the memory 1601 and executable on the processor 1602.

[0175] When the processor 1602 executes the program, it implements the aircraft hard landing prediction method based on group-individual contrast feature enhancement provided in the above embodiments.

[0176] Furthermore, electronic devices also include:

[0177] Communication interface 1603 is used for communication between memory 1601 and processor 1602.

[0178] The memory 1601 is used to store computer programs that can run on the processor 1602.

[0179] The memory 1601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0180] If the memory 1601, processor 1602, and communication interface 1603 are implemented independently, then the communication interface 1603, memory 1601, and processor 1602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 16 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0181] Optionally, in a specific implementation, if the memory 1601, processor 1602, and communication interface 1603 are integrated on a single chip, then the memory 1601, processor 1602, and communication interface 1603 can communicate with each other through an internal interface.

[0182] The processor 1602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0183] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described aircraft hard landing prediction method based on a dynamic window attention mechanism.

[0184] This application also provides a computer program product, including a computer program that can run computer instructions. When the computer instructions are executed by a processor, they implement the aircraft hard landing prediction method based on a dynamic window attention mechanism provided in this application.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0188] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0189] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0190] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

Claims

1. An aircraft heavy landing prediction method based on a dynamic window attention mechanism, characterized in that, The method comprises the following steps: obtaining raw flight data of a real-time position of a target aircraft, aligning granularity of the raw flight data according to a data sampling frequency of the raw flight data, and obtaining real-time flight data of the real-time position; determining actual quantile numbers of a plurality of physical quantities of the target aircraft at the real-time position according to the real-time flight data, generating an importance weight value matrix of the plurality of physical quantities according to a quantile interval in which the actual quantile numbers are located, and splicing the importance weight value matrix and a value matrix corresponding to the real-time flight data to generate a splicing matrix of each of the plurality of physical quantities; based on the splicing matrix, obtaining a position code corresponding to the real-time position, obtaining a time window corresponding to the real-time position according to the position code corresponding to the real-time position and the splicing matrix, obtaining a time unit corresponding to the real-time position according to the time window corresponding to the real-time position, generating a gravity acceleration prediction value of the real-time position according to the time unit corresponding to the real-time position, and generating a heavy landing prediction result of the target aircraft; before generating the gravity acceleration prediction value of the real-time position according to the time unit corresponding to the real-time position, the method further comprises the following steps: combining a splicing matrix and a position code of a plurality of physical quantities of at least one landing sample to obtain a splicing result, inputting the splicing result into a pre-constructed encoder to output a key parameter matrix, a value parameter matrix and a query parameter matrix of the pre-constructed encoder, and constructing a dynamic window centered on a landing point and a soft gating unit based on a target Gaussian function according to position information in the key parameter matrix, the value parameter matrix and the query parameter matrix; determining an embedding matrix of the dynamic window in a target landing section according to a time deviation between each flight position in the target landing section and the landing point; obtaining a causal logic slice of the key parameter matrix and the value parameter matrix according to the dynamic window; and constructing a decoder for decoding the causal logic slice of the key parameter matrix and the value parameter matrix and the causal logic slice of the query parameter matrix according to the causal logic slice of the key parameter matrix and the value parameter matrix and the soft gating unit based on the target Gaussian function pre-constructed for capturing a landing window; based on the decoder, decoding the causal logic slice of the key parameter matrix and the value parameter matrix and the causal logic slice of the query parameter matrix to obtain a gravity acceleration prediction value of each flight position, and determining the gravity acceleration prediction value of the real-time position according to the gravity acceleration prediction value of each flight position.

2. The dynamic window attention mechanism based aircraft heavy landing prediction method according to claim 1, wherein, before generating the importance weight value matrix of the plurality of physical quantities according to the quantile interval in which the actual quantile numbers are located, the method further comprises the following steps: constructing a comprehensive distribution of the plurality of physical quantities in a plurality of normal landing samples according to values of the plurality of physical quantities in a plurality of time units in a target landing section in the plurality of normal landing samples; dividing each time unit in the target landing section into a plurality of quantile intervals according to pre-set quantile boundary values; Determine weights corresponding to the plurality of quantile intervals in combination with the comprehensive distribution and the values of the plurality of time units, to determine a weight value of the quantile interval in which the actual quantile is located based on the weights corresponding to the plurality of quantile intervals.

3. The dynamic window attention mechanism based aircraft heavy landing prediction method of claim 2, wherein, The method for determining the actual quantile of the plurality of physical quantities of the target aircraft at the real-time position according to the real-time flight data, to generate the importance weight value matrix of the plurality of physical quantities according to the quantile interval in which the actual quantile is located, comprises: Determine the time unit corresponding to the real-time position according to the real-time position; Generate the importance weight value matrix of the plurality of physical quantities according to the time unit corresponding to the real-time position and the quantile interval in which the actual quantile is located.

4. The method of claim 2, wherein the dynamic window attention mechanism is based on a dynamic window attention mechanism. Before acquiring the position code corresponding to the real-time position, further comprising: Construct a target linear function and a target sine function for acquiring the position code based on the time unit in the target landing segment; Design the position code according to the target linear function and the target sine function.

5. An apparatus for predicting a heavy landing of an aircraft based on a dynamic window attention mechanism, employing the method for predicting a heavy landing of an aircraft based on a dynamic window attention mechanism according to any one of claims 1 to 4, characterized in that Comprise: The preprocessing module is used for acquiring the original flight data of the real-time position of the target aircraft, aligning the granularity of the original flight data according to the data sampling frequency of the original flight data, and obtaining the real-time flight data of the real-time position; The weighting module is used for determining the actual quantile of the plurality of physical quantities of the target aircraft at the real-time position according to the real-time flight data, generating the importance weight value matrix of the plurality of physical quantities according to the quantile interval in which the actual quantile is located, and splicing the importance weight value matrix and the value matrix corresponding to the real-time flight data to generate the splicing matrix of each physical quantity in the plurality of physical quantities; The prediction module acquires the position code corresponding to the real-time position based on the splicing matrix, acquires the time window corresponding to the real-time position according to the position code corresponding to the real-time position and the splicing matrix, acquires the time unit corresponding to the real-time position according to the time window corresponding to the real-time position, generates the gravity acceleration prediction value of the real-time position according to the time unit corresponding to the real-time position, and generates the heavy landing prediction result of the target aircraft.

6. The dynamic window attention mechanism based aircraft heavy-landing prediction apparatus according to claim 5, wherein, Further comprise: The construction module is used for constructing the comprehensive distribution of the plurality of physical quantities in the plurality of normal landing samples according to the values of the plurality of physical quantities in the plurality of time units in the target landing segment before generating the importance weight value matrix of the plurality of physical quantities according to the quantile interval in which the actual quantile is located; The division module is used for dividing each time unit in the target landing segment into a plurality of quantile intervals according to a pre-set quantile boundary value; The determination module is used for determining weights corresponding to the plurality of quantile intervals in combination with the comprehensive distribution and the values of the plurality of time units, to determine a weight value of the quantile interval in which the actual quantile is located based on the weights corresponding to the plurality of quantile intervals.

7. An electronic device, comprising: Comprise: A memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the method of claim 1-4.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor for implementing the method of claim 1-4.

9. A computer program product comprising a computer program, characterized in that, The computer program is executed for implementing the method of claim 1-4.

Images