A method and system for evaluating the quality of aircraft probe data based on machine learning

Abstract

The application discloses an aircraft detection data quality evaluation method and system based on machine learning, which comprises the following steps: acquiring original detection data of an aircraft-mounted device; checking the original detection data based on multi-source consistency to obtain preliminary screening data; calculating comprehensive confidence according to the flight path data, taking the preliminary screening data with the comprehensive confidence higher than a threshold as credible data, and taking the preliminary screening data with the comprehensive confidence lower than the threshold as uncertain data; performing scene scarcity analysis on the uncertain data based on a memory-enhanced self-encoder of meteorological characteristics and space-time distribution characteristics to obtain scarce data; calculating original scores of the credible data and the scarce data by using an asymmetric penalty mechanism; evaluating the scarce data based on space-time consistency and prediction value to obtain a rare feature vector; constructing a time-series cumulative deviation quality evaluation model based on a basic score function according to the original scores and the rare feature vector; and outputting an evaluation result according to to-be-evaluated data.

Description

Technical Field

[0001] This invention relates to the field of aircraft exploration data quality assessment, and in particular to a machine learning-based method and system for aircraft exploration data quality assessment. Background Technology

[0002] Aircraft meteorological observation is a crucial method for upper-air atmospheric sounding. Commercial aircraft, equipped with sensors for temperature, pressure, and wind speed, continuously collect meteorological data along their flight paths. This data offers advantages such as high sampling frequency, wide coverage, and dense vertical data layers, making it one of the most important sources of upper-air observation data in global numerical weather prediction assimilation systems. However, the quality of aircraft meteorological observation data varies significantly due to factors such as changes in flight attitude, sensor drift, extreme weather interference, and equipment start-up and shutdown. Accurately evaluating and selecting high-quality observation data is a key technical challenge in ensuring the effectiveness of numerical weather prediction assimilation.

[0003] Existing quality assessment methods mainly rely on fixed threshold judgments or simple statistical filtering. Although they can identify anomalous sampling points that clearly violate physical laws, they have significant limitations in practical applications: they do not fully consider the differences in dynamic characteristics of different flight phases such as takeoff, cruise, and landing; the use of a uniform standard for trajectory smoothness evaluation is prone to introducing systematic misjudgments during maneuvering flight phases; they lack a mechanism to distinguish between anomalous and scarce data. When aircraft pass through extreme weather systems such as typhoons and fronts or enter sparse airspace from historical observations, outlier data with important assimilation value are often simply regarded as anomalous and removed; the scoring mechanism is symmetrical and fails to implement asymmetric penalty correction for scarce scenarios; and they lack cumulative deviation analysis in the time-series dimension, making it difficult to reflect the dynamic evolution characteristics of data quality.

[0004] Therefore, there is an urgent need for a new method for assessing the quality of aircraft detection data, which can accurately separate reliable data from uncertain data, avoid misjudging scarce data, balance physical consistency and scarcity value, and improve the accuracy of quality assessment. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for evaluating the quality of aircraft detection data based on machine learning.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: The raw detection data from the aircraft's onboard equipment is acquired, including flight path data, meteorological parameters, and cloud microphysical data. The raw detection data is then verified based on multi-source consistency to obtain preliminary screening data. The overall confidence level of the initial screening data is calculated based on the flight track data. The initial screening data with an overall confidence level higher than the threshold is regarded as reliable data, and the rest is regarded as uncertain data. The memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data. The reliable data and the scarce data are used as probe data, and an asymmetric penalty mechanism is used to calculate the original score of the probe data; the scarce data is evaluated based on spatiotemporal consistency and prediction value to obtain a scarce feature vector; Based on the original scores and the rare feature vectors, a time-series cumulative deviation quality assessment model based on the basic scoring function is constructed; the data to be assessed is input into the time-series cumulative deviation quality assessment model, and the assessment results are output.

[0007] Furthermore, the method for verifying the original detection data based on multi-source consistency includes: Sampling points that do not meet the physical consistency constraints in the original detection data will be removed. Physical consistency constraints include the relationship between temperature and altitude violating the standard atmospheric decrease law, the relationship between cloud microphysical phase and temperature violating the basic thermodynamic law, and the altitude at which cloud microphysical characteristics appear does not match the flight track altitude. The verified sampling points are arranged in chronological order to generate a preliminary screening data sequence.

[0008] Furthermore, the method for calculating the overall confidence level of the initial screening data based on the trajectory data includes: The proportion of the number of default fields in the initial screening data to the total number of fields is used as the basic default rate; the distribution position of the default fields in the data sequence is obtained, and the default patterns are divided into head default, middle default, tail default, and random default based on the distribution position; the data completeness is calculated based on the basic default rate and the default patterns. The residuals of the trajectory data are calculated using filtering and fitting. Based on the rate of change of altitude and rate of change of speed in the trajectory data, it is determined whether the current moment is the takeoff phase, cruise phase, or landing phase. The first adjustment coefficient is used for the takeoff and landing phases, and the second adjustment coefficient is used for the cruise phase. The first adjustment coefficient is greater than the second adjustment coefficient. The trajectory smoothness is calculated based on the residuals and the adjustment coefficients. The data completeness and track smoothness are weighted and fused to obtain the overall confidence level. The initial screening data with an overall confidence level greater than the preset confidence level threshold are regarded as reliable data, and the rest are regarded as uncertain data. The preset confidence level threshold ranges from [0.6, 0.8].

[0009] Furthermore, a method for performing scene scarcity analysis on the uncertain data based on a memory-enhanced autoencoder with meteorological and spatiotemporal distribution characteristics includes: Meteorological data such as temperature, air pressure, humidity, wind speed, and vertical velocity are acquired to form a meteorological feature vector; longitude, latitude, altitude, and timestamp of flight track data are acquired to form a spatiotemporal distribution feature vector. The meteorological feature vector and spatiotemporal distribution feature vector of the uncertain data are input into the memory-enhanced autoencoder trained with historical flight detection data; the memory module of the memory-enhanced autoencoder stores multiple feature prototypes under normal meteorological conditions; The multidimensional feature vector of uncertain data is input into the trained encoder to obtain the query vector; the cosine similarity between the query vector and K feature prototypes in the memory module is calculated to obtain the matching distribution {s1,s2,…,sK}; the peak probability of the query vector and the matching distribution is calculated. ; Based on cloud penetration duration, temperature-climatological deviation, regional scarcity, and date distribution, the single-factor scarcity is calculated using the threshold method or historical frequency statistics, and the factors and combination types that exceed the threshold are counted. Based on the factor combination type, identify the scenario category to which the uncertain data belongs; if both cloud penetration duration and temperature-climatological deviation exceed the threshold, it is determined to be an extreme weather crossing scenario; if both regional scarcity and date distribution scarcity exceed the threshold, it is determined to be a route scarcity scenario; calculate the scarcity score of the uncertain data based on the scenario type: in, The duration of cloud penetration; This refers to temperature climatological deviation; In terms of regional scarcity; The scarcity of date distribution; For scene weighting factors, 1.1≤ ≤1.8; λ is the matching distribution suppression coefficient, 0.2≤λ≤1; If the scarcity score is higher than the preset scarcity threshold, uncertain data will be treated as scarce data; the preset scarcity threshold ranges from [0.5, 1]. Furthermore, the method for calculating the raw score of the probe data using an asymmetric penalty mechanism includes: The measured position, measured velocity, and measured acceleration of the flight path data are obtained. Numerical differentiation is used to calculate the difference between the rate of change of position and the actual velocity, and the difference between the rate of change of velocity and the measured acceleration. These are then summed to obtain the dynamic residual. ; The measured temperature, measured air pressure, and measured altitude of meteorological data are obtained. The difference between the standard atmospheric temperature calculated from the measured altitude and the measured temperature is calculated using the ISA temperature deviation, which is used as the third deviation. The difference between the standard atmospheric pressure corresponding to the measured altitude and the measured air pressure is calculated using the ISA pressure deviation, which is used as the fourth deviation. The third and fourth deviations are weighted and fused to obtain the thermodynamic residual. ; Obtain the cloud microphysical phase state from the cloud microphysical data, and calculate the cloud microphysical residual based on the measured temperature T and the cloud microphysical phase. ;like And the phase is ice crystal, or And its phase is liquid water. ;otherwise ; The raw score of the probe data is calculated using a penalty coefficient; the physical residuals of the reliable data are directly mapped to the raw score. ,in Let j be the original score for the j-th dimension, where j∈{dyn,ther,cld}; p is the penalty coefficient. cld >p ther p cld >p dyn ; For physical residuals; calculate the original score of scarce data based on the scarcity score: , where R is the scarcity score; The asymmetric adjustment coefficient is 0.5 ≤ ≤2.

[0010] Furthermore, a method for evaluating the scarce data based on spatiotemporal consistency and forecast value includes: The background field of the numerical weather prediction model is obtained at the location and time of the scarce data. The difference between the observed value and the background field of the numerical weather prediction model is calculated to obtain the observation background difference. The cosine similarity is used to calculate the matching degree of meteorological features in the historical rare event database. The observation background difference and the matching degree are weighted and fused to obtain the forecast value index. Calculate the spatiotemporal consistency index based on meteorological data: in, Let i be the temperature value at the i-th sampling point. Let be the temperature value of the i-th adjacent sampling point; N is the preset number of adjacent sampling points; This is the measured temperature value; Estimate temperature values ​​for space; For preset weighting coefficients, ; The scarcity value feature vector is obtained by fusing the forecast value index and the spatiotemporal consistency index.

[0011] Furthermore, the construction of the time-series cumulative deviation quality assessment model based on the basic scoring function includes: Construct a basic scoring function based on the original scores and scarce feature vectors: in This is the base score for the i-th sampling point. ∈[0,100]; The original score for the i-th sampling point in the j-th physical dimension; Let be the weight coefficient for the j-th dimension, where j∈{dyn,ther,cld}. v(i) is the scarce feature vector of the i-th sampling point; The scarcity intensity amplification factor is 0.2 ≤ ≤0.8; N rare N represents the number of scarce data points. total This represents the total number of data points. For the normalized Euclidean norm; A trusted dataset; This is a scarce dataset; Calculate single-point deviation based on the base score. The cumulative deviation of the time series is calculated using the exponentially weighted moving average formula: ,in The cumulative temporal deviation for the first i sampling points, where i ≥ 2, and the first sampling point = ; The smoothing coefficient is 0.05 ≤ ≤0.2; An adaptive parameter optimization model is constructed based on the cumulative deviation over time, and the expression for the loss function is: in, Score the cumulative temporal deviation of the first i sampling points; The regularization coefficient is 10. -4 ≤ ≤10 -1 ; The initial value is ; The online gradient descent method is used to adjust the weight coefficients. Perform iterative updates, repeating the above iterations until the loss function converges or the preset maximum number of iterations is reached, to obtain the time-series cumulative deviation sequence { }; Calculate the final quality score ; Final quality score Detection data exceeding the preset acceptable threshold is marked as acceptable; otherwise, it is marked as unacceptable. The preset acceptable threshold ranges from [60, 80].

[0012] Secondly, a machine learning-based aircraft detection data quality assessment system includes: Data acquisition and verification module: used to acquire raw detection data from aircraft-mounted equipment, including flight path data, meteorological parameters, and cloud microphysical data; verify the raw detection data based on multi-source consistency to obtain preliminary screening data; Data filtering module: used to calculate the overall confidence level of the initial screening data based on the flight track data, and to regard the initial screening data with an overall confidence level higher than a threshold as reliable data, and the rest as uncertain data; a memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data; The original scoring module is used to calculate the original score of the probe data by using the reliable data and the scarce data as probe data and employing an asymmetric penalty mechanism; and to evaluate the scarce data based on spatiotemporal consistency and prediction value to obtain a scarce feature vector. Model building module: used to construct a time-series cumulative deviation quality assessment model based on the basic scoring function according to the original score and the rare feature vector; input the data to be evaluated into the time-series cumulative deviation quality assessment model and output the evaluation results.

[0013] The beneficial effects of this invention are: This invention is a machine learning-based method and system for evaluating the quality of aircraft reconnaissance data. Compared with existing technologies, this invention has the following technical advantages: This invention achieves accurate separation of reliable and uncertain data through multi-source consistency verification and comprehensive confidence level classification; it performs scene scarcity analysis based on memory-enhanced autoencoders to effectively distinguish between abnormal and scarce data, avoiding the misjudgment and exclusion of high-value data such as extreme weather crossings; it employs an asymmetric penalty mechanism to implement differentiated scoring correction for scarce data, taking into account both physical consistency and scarcity value; and it constructs a time-series cumulative deviation quality assessment model, using exponentially weighted moving averages to accumulate single-point deviations along the time dimension, making the assessment results have both single-point accuracy and trend sensitivity, thus improving the accuracy of aircraft detection data quality assessment. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a machine learning-based aircraft detection data quality assessment method according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The present invention provides a machine learning-based method and system for evaluating the quality of aircraft reconnaissance data, comprising the following steps: like Figure 1As shown, this embodiment includes the following steps: The raw detection data from the aircraft's onboard equipment is acquired, including flight path data, meteorological parameters, and cloud microphysical data. The raw detection data is then verified based on multi-source consistency to obtain preliminary screening data. In this embodiment, taking a commercial aircraft meteorological observation data for numerical weather prediction data assimilation as an example, the aircraft was equipped with a GPS positioning system, airborne meteorological sensors, and airborne cloud microphysics detection equipment; the flight period was from 10:00 to 12:30 Beijing time, the sampling frequency was 1Hz, the flight altitude range was 1500m to 6200m, and a total of about 9000 sampling points were recorded; the raw detection data contained 16 key fields, including 7 track data items: longitude, latitude, altitude, ground speed, vertical speed, heading angle, and timestamp; 5 meteorological parameters items: temperature, air pressure, humidity, horizontal wind speed, and vertical wind speed; and 4 cloud microphysics data items: phase state, liquid water content, number concentration, and effective particle radius; The verification rules are set based on the fundamental laws of atmospheric physics: temperature and altitude should follow the standard atmospheric decreasing law; cloud microphysical phases and temperature should satisfy thermodynamic constraints; and the detection altitude of cloud microphysical characteristics should be consistent with the flight track altitude. If the measured temperature deviates from the standard atmospheric temperature by more than 15℃, it is rejected. While weather phenomena such as temperature inversions and fronts do exist in the actual atmosphere, a deviation exceeding 15℃ is extremely unlikely under special weather conditions, usually indicating a sensor malfunction. For example, sampling point #623 recorded an altitude of 6500m and a measured temperature of 12℃, while the standard atmospheric temperature at 6500m is approximately -9℃, a deviation of 21.5℃, thus it is rejected. Similarly, sampling point #1842 recorded a temperature of -2℃ but was marked as liquid water, thus it is rejected. And sampling point #3715 had a liquid water content of 0.5g / m³. 3 The local ground elevation was 450m, but the flight track was only 200m, which was considered an invalid detection and was removed. After verification, a total of 78 abnormal sampling points were removed, and the remaining 8922 valid sampling points were arranged in chronological order to generate an initial screening data sequence. The overall confidence level of the initial screening data is calculated based on the flight track data. The initial screening data with an overall confidence level higher than the threshold is regarded as reliable data, and the rest is regarded as uncertain data. The memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data. In this embodiment, the default cases in 16 fields of 8922 sampling points are statistically analyzed. Then, the default positions are multiplied by the attenuation coefficient. Defaults occurring at the beginning or end of the flight are often caused by equipment start-up and shutdown, and have a relatively small impact on the core analysis section. Continuous defaults in the middle are more alarming. Therefore, in this embodiment, the default attenuation is set to 0.8 for the beginning, 0.9 for the middle, 0.7 for the end, and 1.0 for random. Taking sampling point #1200 as an example, it is in the cruise phase, at an altitude of 5500m, under clear sky conditions, with no defaults in any of the 12 fields, a basic default rate of 0%, and a calculated data integrity of 1.0. A third-order polynomial filter was applied to the altitude sequences of 30 points before and after the flight, and the root mean square of the residuals was calculated. The tolerance for residuals varied across different flight phases. During takeoff and landing, the flight path itself fluctuated significantly, so a first adjustment factor of 0.9 was set. During cruise, the aircraft should remain stable, so a second adjustment factor of 0.6 was set. Taking sampling point #1200 as an example, the filtered fitting residual was 0.8 m / s. 2 Currently, during the cruise phase, the second adjustment factor is used, and the calculated track smoothness is 0.952. The data completeness and track smoothness are weighted and fused at a ratio of 0.4:0.6 to obtain a comprehensive confidence level of 0.971. With a preset confidence threshold of 0.7, after calculations were performed on all 8922 sampling points, there were 7860 reliable data points and 1062 uncertain data points. A memory-enhanced autoencoder was employed to perform scene scarcity analysis on uncertain data. The autoencoder was pre-trained using 1.2 million historical data points from the past three years, and the memory module stored 10 feature prototypes obtained through K-means clustering. For each uncertain data point, meteorological feature vectors composed of temperature, air pressure, humidity, wind speed, and vertical velocity, and spatiotemporal feature vectors composed of longitude, latitude, altitude, and timestamp were extracted. Taking sampling point #5600 as an example, the meteorological feature vector and spatiotemporal distribution feature vector were input into the encoder. The cosine similarity between the encoder's output query vector and the feature prototypes was {0.82, 0.68, 0.55, 0.47, 0.39, 0.33, 0.28, 0.22, 0.17, 0.11}, and the peak probability p was calculated. max =0.204; The scarcity thresholds for cloud penetration duration and temperature deviation were calculated using the threshold method, with thresholds of 10 minutes and 5℃, respectively. Based on historical frequency statistics, the aircraft detection coverage rate was 30%, and the historical date detection coverage frequency was 50%, resulting in scarcity thresholds of 0.7 and 0.5 for geographical and date distributions, respectively. At sampling point #1200, the cloud penetration duration was 14 minutes, the temperature climatological deviation was 8℃, the geographical scarcity was 0.92, and the date distribution scarcity was 0.65, simultaneously satisfying the conditions for extreme weather crossing scenarios and rare flight path scenarios. The scenario weighting factors were 1.6 and 1.4, respectively, ω sceneTaking the larger value of 1.6 and λ of 0.5, the scarcity score R = 5.57 is calculated; the preset scarcity threshold is 0.7, and the data is marked as scarce. After calculating the scarcity scores of 1062 uncertain data points one by one, 185 of the sampling points exceeded the threshold and were marked as scarce data; the remaining 877 were ordinary uncertain data. The reliable data and the scarce data are used as probe data, and an asymmetric penalty mechanism is used to calculate the original score of the probe data; the scarce data is evaluated based on spatiotemporal consistency and prediction value to obtain a scarce feature vector; In this embodiment, the penalty coefficient is set to p. dyn =4、p ther =3、p cld =20, the asymmetric adjustment coefficient γ is taken as 1.2, and the original score of the detection data is calculated; taking reliable data #1200 as an example, =0.5, =0.2, =0, calculate the original scores for sampling point #1200 as follows: u dyn =98.0, u ther =99.1, u cld =100; Taking scarce data, #5600, as an example. =6.2, =7.5, =1, scarcity score R=5.57, the original scores for sampling point #5600 are calculated as follows: u dyn =96.77, u ther =97.07, u cld =97.40; For each of the 185 scarce data points, the spatiotemporal consistency index and forecast value index were calculated. Taking sampling point #5600 as an example, the ECMWF 6-hour forecast background field showed a temperature difference of 5.2℃ and a humidity difference of 22%. After normalization, the observation background difference index was 0.35, and the cosine similarity with the historical rare event database was 0.81. The calculated forecast value index was 0.645. Take N=3 adjacent points before and after the sampling point, with a sampling interval of 1s; the temperature sequence of the adjacent points of sampling point #5600 is [-6.5, -7.2, -7.8] and [-8.5, -9.0, -9.8]; at the same time, select the 3 nearest radiosonde stations to calculate the spatial estimated temperature value, the temperatures of the radiosonde stations are -7.2℃, -9.0℃, and -10.5℃, and the distances from the sampling point are 15km, 22km, and 31km, respectively. The spatial estimated temperature value is calculated as -8.15℃ using the inverse distance weighting method; the measured temperature of sampling point #5600 is -8.0℃, α is 0.7, β is 0.3, and the spatiotemporal consistency index is calculated to be 0.36; the scarcity value vector v(5600)=[0.645,0.36]; Based on the original scores and the rare feature vectors, a time-series cumulative deviation quality assessment model based on the basic scoring function is constructed; the data to be assessed is input into the time-series cumulative deviation quality assessment model, and the assessment results are output. In this embodiment, the initial weights of the three physical dimensions are set to θ=[0.40,0.35,0.25], and σ is set to 0.6. The base score is calculated point by point. Taking the #1200 trusted data as an example, the base score Q is calculated. 1200 =98.99; Taking the scarce data #5600 as an example, v(5600)=[0.645,0.36], calculate the Euclidean norm of the scarce feature vector. =0.738, divided by the vector dimension, the square root normalized value is 0.552; N rare =185, N total =8045, calculate the base score Q 5600 =97.04; Single-point deviation E 1200 =100-Q 1200 =1.01, E 5600 =100-Q 5600 =2.96; D 5599 =1.72, ρ is taken as 0.12, calculate D 5600 =1.869; δ is 0.01, η is 0.001, and the weights are iteratively updated using the online gradient descent method. Each batch consists of 100 sampling points. After about 200 iterations, the weights converge to ω=[0.36,0.38,0.26]. The weight of the thermodynamic dimension increases from 0.35 to 0.38, indicating that the thermodynamic parameters such as temperature and air pressure have a higher distinguishability than expected in the measured data of this flight mission and contribute more to the quality assessment. The weight of the dynamic dimension decreases from 0.40 to 0.36, indicating that the differences in the trajectory data are relatively small. By traversing all 8922 sampling points, a complete time-series cumulative deviation sequence is obtained, and the final cumulative deviation D is calculated. 8922 =2.31, calculate the final quality score Sfinal =97.69; A comparative experiment was conducted using 600 expert-annotated detection data points as the test sample, including 400 normal data points, 100 abnormal data points, and 100 scarce data points. The control group used the traditional threshold method, without scarcity analysis or asymmetric scoring correction, and did not calculate the cumulative deviation over time. The evaluation results were as follows: among the 400 normal data points, 369 were judged as qualified and 31 were misjudged as unqualified; among the 100 abnormal data points, 72 were judged as unqualified and 28 were misjudged as qualified; among the 100 scarce data points, 35 were judged as qualified and 65 were misjudged as unqualified. The correct identification rate for normal data was 92.3%, the correct identification rate for abnormal data was 72.0%, the correct identification rate for scarce data was 35.0%, and the overall accuracy rate was 79.3%. In the method of this invention, the evaluation results are as follows: out of 400 normal data points, 385 were judged as qualified and 15 were misjudged as unqualified; out of 100 abnormal data points, 88 were judged as unqualified and 12 were misjudged as qualified; out of 100 scarce data points, 92 were judged as qualified and 8 were misjudged as unqualified; the correct identification rate for normal data was 96.3%, the correct identification rate for abnormal data was 88.0%, the correct identification rate for scarce data was 92.0%, and the overall judgment accuracy rate was 94.2%. In this embodiment, the method for verifying the original detection data based on multi-source consistency includes: Sampling points that do not meet the physical consistency constraints in the original detection data will be removed. Physical consistency constraints include the relationship between temperature and altitude violating the standard atmospheric decrease law, the relationship between cloud microphysical phase and temperature violating the basic thermodynamic law, and the altitude at which cloud microphysical characteristics appear does not match the flight track altitude. The verified sampling points are arranged in chronological order to generate a preliminary screening data sequence.

[0017] In this embodiment, the method for calculating the overall confidence level of the initial screening data based on the trajectory data includes: The proportion of the number of default fields in the initial screening data to the total number of fields is used as the basic default rate; the distribution position of the default fields in the data sequence is obtained, and the default patterns are divided into head default, middle default, tail default, and random default based on the distribution position; the data completeness is calculated based on the basic default rate and the default patterns. The residuals of the trajectory data are calculated using filtering and fitting. Based on the rate of change of altitude and rate of change of speed in the trajectory data, it is determined whether the current moment is the takeoff phase, cruise phase, or landing phase. The first adjustment coefficient is used for the takeoff and landing phases, and the second adjustment coefficient is used for the cruise phase. The first adjustment coefficient is greater than the second adjustment coefficient. The trajectory smoothness is calculated based on the residuals and the adjustment coefficients. The data completeness and track smoothness are weighted and fused to obtain the overall confidence level. The initial screening data with an overall confidence level greater than the preset confidence level threshold are regarded as reliable data, and the rest are regarded as uncertain data. The preset confidence level threshold ranges from [0.6, 0.8].

[0018] In this embodiment, the method for performing scene scarcity analysis on the uncertain data based on a memory-enhanced autoencoder with meteorological and spatiotemporal distribution characteristics includes: Meteorological data such as temperature, air pressure, humidity, wind speed, and vertical velocity are acquired to form a meteorological feature vector; longitude, latitude, altitude, and timestamp of flight track data are acquired to form a spatiotemporal distribution feature vector. The meteorological feature vector and spatiotemporal distribution feature vector of the uncertain data are input into the memory-enhanced autoencoder trained with historical flight detection data; the memory module of the memory-enhanced autoencoder stores multiple feature prototypes under normal meteorological conditions; The multidimensional feature vector of uncertain data is input into the trained encoder to obtain the query vector; the cosine similarity between the query vector and K feature prototypes in the memory module is calculated to obtain the matching distribution {s1,s2,…,sK}; the peak probability of the query vector and the matching distribution is calculated. ; Based on cloud penetration duration, temperature-climatological deviation, regional scarcity, and date distribution, the single-factor scarcity is calculated using the threshold method or historical frequency statistics, and the factors and combination types that exceed the threshold are counted. Based on the factor combination type, identify the scenario category to which the uncertain data belongs; if both cloud penetration duration and temperature-climatological deviation exceed the threshold, it is determined to be an extreme weather crossing scenario; if both regional scarcity and date distribution scarcity exceed the threshold, it is determined to be a route scarcity scenario; calculate the scarcity score of the uncertain data based on the scenario type: in, The duration of cloud penetration; This refers to temperature climatological deviation; In terms of regional scarcity; The scarcity of date distribution; For scene weighting factors, 1.1≤ ≤1.8; λ is the matching distribution suppression coefficient, 0.2≤λ≤1; If the scarcity score is higher than the preset scarcity threshold, uncertain data will be treated as scarce data; the preset scarcity threshold ranges from [0.5, 1]. In this embodiment, the method for calculating the original score of the detection data using an asymmetric penalty mechanism includes: The measured position, measured velocity, and measured acceleration of the flight path data are obtained. Numerical differentiation is used to calculate the difference between the rate of change of position and the actual velocity, and the difference between the rate of change of velocity and the measured acceleration. These are then summed to obtain the dynamic residual. ; The measured temperature, measured air pressure, and measured altitude of meteorological data are obtained. The difference between the standard atmospheric temperature calculated from the measured altitude and the measured temperature is calculated using the ISA temperature deviation, which is used as the third deviation. The difference between the standard atmospheric pressure corresponding to the measured altitude and the measured air pressure is calculated using the ISA pressure deviation, which is used as the fourth deviation. The third and fourth deviations are weighted and fused to obtain the thermodynamic residual. ; Obtain the cloud microphysical phase state from the cloud microphysical data, and calculate the cloud microphysical residual based on the measured temperature T and the cloud microphysical phase. ;like And the phase is ice crystal, or And its phase is liquid water. ;otherwise ; The raw score of the probe data is calculated using a penalty coefficient; the physical residuals of the reliable data are directly mapped to the raw score. ,in Let j be the original score for the j-th dimension, where j∈{dyn,ther,cld}; p is the penalty coefficient. cld >p ther p cld >p dyn ; For physical residuals; calculate the original score of scarce data based on the scarcity score: , where R is the scarcity score; The asymmetric adjustment coefficient is 0.5 ≤ ≤2.

[0019] In this embodiment, the method for evaluating the scarce data based on spatiotemporal consistency and forecast value includes: The background field of the numerical weather prediction model is obtained at the location and time of the scarce data. The difference between the observed value and the background field of the numerical weather prediction model is calculated to obtain the observation background difference. The cosine similarity is used to calculate the matching degree of meteorological features in the historical rare event database. The observation background difference and the matching degree are weighted and fused to obtain the forecast value index. Calculate the spatiotemporal consistency index based on meteorological data: in, Let i be the temperature value at the i-th sampling point. Let be the temperature value of the i-th adjacent sampling point; N is the preset number of adjacent sampling points; This is the measured temperature value; Estimate temperature values ​​for space; For preset weighting coefficients, ; The scarcity value feature vector is obtained by fusing the forecast value index and the spatiotemporal consistency index.

[0020] In this embodiment, constructing a time-series cumulative deviation quality assessment model based on a basic scoring function includes: Construct a basic scoring function based on the original scores and scarce feature vectors: in This is the base score for the i-th sampling point. ∈[0,100]; The original score for the i-th sampling point in the j-th physical dimension; Let be the weight coefficient for the j-th dimension, where j∈{dyn,ther,cld}. v(i) is the scarce feature vector of the i-th sampling point; The scarcity intensity amplification factor is 0.2 ≤ ≤0.8; N rare N represents the number of scarce data points. total This represents the total number of data points. For the normalized Euclidean norm; A trusted dataset; This is a scarce dataset; Calculate single-point deviation based on the base score. The cumulative deviation of the time series is calculated using the exponentially weighted moving average formula: ,in The cumulative temporal deviation for the first i sampling points, where i ≥ 2, and the first sampling point = ; The smoothing coefficient is 0.05 ≤ ≤0.2; An adaptive parameter optimization model is constructed based on the cumulative deviation over time, and the expression for the loss function is: in, Score the cumulative temporal deviation of the first i sampling points; The regularization coefficient is 10. -4 ≤ ≤10 -1 ; The initial value is ; The online gradient descent method is used to adjust the weight coefficients. Perform iterative updates, repeating the above iterations until the loss function converges or the preset maximum number of iterations is reached, to obtain the time-series cumulative deviation sequence { }; Calculate the final quality score ; Final quality score Detection data exceeding the preset acceptable threshold is marked as acceptable; otherwise, it is marked as unacceptable. The preset acceptable threshold ranges from [60, 80].

[0021] Secondly, a machine learning-based aircraft detection data quality assessment system includes: Data acquisition and verification module: used to acquire raw detection data from aircraft-mounted equipment, including flight path data, meteorological parameters, and cloud microphysical data; verify the raw detection data based on multi-source consistency to obtain preliminary screening data; Data filtering module: used to calculate the overall confidence level of the initial screening data based on the flight track data, and to regard the initial screening data with an overall confidence level higher than a threshold as reliable data, and the rest as uncertain data; a memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data; The original scoring module is used to calculate the original score of the probe data by using the reliable data and the scarce data as probe data and employing an asymmetric penalty mechanism; and to evaluate the scarce data based on spatiotemporal consistency and prediction value to obtain a scarce feature vector. Model building module: used to construct a time-series cumulative deviation quality assessment model based on the basic scoring function according to the original score and the rare feature vector; input the data to be evaluated into the time-series cumulative deviation quality assessment model and output the evaluation results.

Claims

1. A machine learning-based method for evaluating the quality of aircraft reconnaissance data, characterized in that, Includes the following steps: The raw detection data from the aircraft's onboard equipment is acquired, including flight path data, meteorological parameters, and cloud microphysical data. The raw detection data is then verified based on multi-source consistency to obtain preliminary screening data. The overall confidence level of the initial screening data is calculated based on the flight track data. The initial screening data with an overall confidence level higher than the threshold is regarded as reliable data, and the rest is regarded as uncertain data. The memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data. The reliable data and the scarce data are used as probe data, and an asymmetric penalty mechanism is used to calculate the original score of the probe data; the scarce data is evaluated based on spatiotemporal consistency and prediction value to obtain a scarce feature vector; Based on the original scores and the rare feature vectors, a time-series cumulative deviation quality assessment model based on the basic scoring function is constructed; the data to be assessed is input into the time-series cumulative deviation quality assessment model, and the assessment results are output.

2. The machine learning-based aircraft detection data quality assessment method according to claim 1, characterized in that, A method for verifying the original detection data based on multi-source consistency includes: Sampling points that do not meet the physical consistency constraints in the original detection data will be removed. Physical consistency constraints include the relationship between temperature and altitude violating the standard atmospheric decrease law, the relationship between cloud microphysical phase and temperature violating the basic thermodynamic law, and the altitude at which cloud microphysical characteristics appear does not match the flight track altitude. The verified sampling points are arranged in chronological order to generate a preliminary screening data sequence.

3. The method for evaluating the quality of aircraft detection data based on machine learning according to claim 1, characterized in that, A method for calculating the overall confidence level of the initial screening data based on the trajectory data includes: The proportion of the number of default fields in the initial screening data to the total number of fields is used as the basic default rate; the distribution position of the default fields in the data sequence is obtained, and the default patterns are divided into head default, middle default, tail default, and random default based on the distribution position; the data completeness is calculated based on the basic default rate and the default patterns. The residuals of the trajectory data are calculated using filtering and fitting. Based on the rate of change of altitude and rate of change of speed in the trajectory data, it is determined whether the current moment is the takeoff phase, cruise phase, or landing phase. The first adjustment coefficient is used for the takeoff and landing phases, and the second adjustment coefficient is used for the cruise phase. The first adjustment coefficient is greater than the second adjustment coefficient. The trajectory smoothness is calculated based on the residuals and the adjustment coefficients. The data completeness and track smoothness are weighted and fused to obtain the overall confidence level. The initial screening data with an overall confidence level greater than the preset confidence level threshold are regarded as reliable data, and the rest are regarded as uncertain data. The preset confidence level threshold ranges from [0.6, 0.8].

4. The method for evaluating the quality of aircraft detection data based on machine learning according to claim 1, characterized in that, A method for performing scene scarcity analysis on the uncertain data using a memory-enhanced autoencoder based on meteorological and spatiotemporal distribution characteristics includes: Meteorological data such as temperature, air pressure, humidity, wind speed, and vertical velocity are acquired to form a meteorological feature vector; longitude, latitude, altitude, and timestamp of flight track data are acquired to form a spatiotemporal distribution feature vector. The meteorological feature vector and spatiotemporal distribution feature vector of the uncertain data are input into the memory-enhanced autoencoder trained with historical flight detection data; the memory module of the memory-enhanced autoencoder stores multiple feature prototypes under normal meteorological conditions; The multidimensional feature vector of uncertain data is input into the trained encoder to obtain the query vector; the cosine similarity between the query vector and K feature prototypes in the memory module is calculated to obtain the matching distribution {s1,s2,…,sK}; the peak probability of the query vector and the matching distribution is calculated. ; Based on cloud penetration duration, temperature-climatological deviation, regional scarcity, and date distribution, the single-factor scarcity is calculated using the threshold method or historical frequency statistics, and the factors and combination types that exceed the threshold are counted. Based on the factor combination type, identify the scenario category to which the uncertain data belongs; if both cloud penetration duration and temperature-climatological deviation exceed the threshold, it is determined to be an extreme weather crossing scenario; if both regional scarcity and date distribution scarcity exceed the threshold, it is determined to be a route scarcity scenario; calculate the scarcity score of the uncertain data based on the scenario type: ； in, The duration of cloud penetration; This refers to temperature climatological deviation; In terms of regional scarcity; The scarcity of date distribution; For scene weighting factors, 1.1≤ ≤1.8; λ is the matching distribution suppression coefficient, 0.2≤λ≤1; If the scarcity score is higher than the preset scarcity threshold, uncertain data will be treated as scarce data; the preset scarcity threshold ranges from [0.5, 1].

5. The method for evaluating the quality of aircraft reconnaissance data based on machine learning according to claim 1, characterized in that, A method for calculating the raw score of the probe data using an asymmetric penalty mechanism includes: The measured position, measured velocity, and measured acceleration of the flight path data are obtained. Numerical differentiation is used to calculate the difference between the rate of change of position and the actual velocity, and the difference between the rate of change of velocity and the measured acceleration. These are then summed to obtain the dynamic residual. ; The measured temperature, measured air pressure, and measured altitude of meteorological data are obtained. The difference between the standard atmospheric temperature calculated from the measured altitude and the measured temperature is calculated using the ISA temperature deviation, which is used as the third deviation. The difference between the standard atmospheric pressure corresponding to the measured altitude and the measured air pressure is calculated using the ISA pressure deviation, which is used as the fourth deviation. The third and fourth deviations are weighted and fused to obtain the thermodynamic residual. ; Obtain the cloud microphysical phase state from the cloud microphysical data, and calculate the cloud microphysical residual based on the measured temperature T and the cloud microphysical phase. ;like And the phase is ice crystal, or And its phase is liquid water. ;otherwise ; The raw score of the probe data is calculated using a penalty coefficient; the physical residuals of the reliable data are directly mapped to the raw score. ,in Let j be the original score for the j-th dimension, where j∈{dyn,ther,cld}; p is the penalty coefficient. cld >p ther p cld >p dyn ; For physical residuals; calculate the original score of scarce data based on the scarcity score: , where R is the scarcity score; The asymmetric adjustment coefficient is 0.5 ≤ ≤2.

6. The method for evaluating the quality of aircraft detection data based on machine learning according to claim 1, characterized in that, Methods for evaluating scarce data based on spatiotemporal consistency and forecast value include: The background field of the numerical weather prediction model is obtained at the location and time of the scarce data. The difference between the observed value and the background field of the numerical weather prediction model is calculated to obtain the observation background difference. The cosine similarity is used to calculate the matching degree of meteorological features in the historical rare event database. The observation background difference and the matching degree are weighted and fused to obtain the forecast value index. Calculate the spatiotemporal consistency index based on meteorological data: ； in, Let i be the temperature value at the i-th sampling point. Let be the temperature value of the i-th adjacent sampling point; N is the preset number of adjacent sampling points; This is the measured temperature value; Estimate temperature values ​​for space; For preset weighting coefficients, ; The scarcity value feature vector is obtained by fusing the forecast value index and the spatiotemporal consistency index.

7. The machine learning-based aircraft detection data quality assessment method according to claim 1, characterized in that, The construction of the time-series cumulative deviation quality assessment model based on the basic scoring function includes: Construct a basic scoring function based on the original scores and scarce feature vectors: ； in This is the base score for the i-th sampling point. ∈[0,100]; The original score for the i-th sampling point in the j-th physical dimension; Let be the weight coefficient for the j-th dimension, where j∈{dyn,ther,cld}. v(i) is the scarce feature vector of the i-th sampling point; The scarcity intensity amplification factor is 0.2 ≤ ≤0.8; N rare N represents the number of scarce data points. total This represents the total number of data points. For the normalized Euclidean norm; A trusted dataset; This is a scarce dataset; Calculate single-point deviation based on the base score. The cumulative deviation of the time series is calculated using the exponentially weighted moving average formula: ,in The cumulative temporal deviation for the first i sampling points, where i ≥ 2, and the first sampling point = ; The smoothing coefficient is 0.05 ≤ ≤0.2; An adaptive parameter optimization model is constructed based on the cumulative deviation over time, and the expression for the loss function is: ； in, Score the cumulative temporal deviation of the first i sampling points; The regularization coefficient is 10. -4 ≤ ≤10 -1 ; The initial value is ; The online gradient descent method is used to adjust the weight coefficients. Perform iterative updates, repeating the above iterations until the loss function converges or the preset maximum number of iterations is reached, to obtain the time-series cumulative deviation sequence { }; Calculate the final quality score ; Final quality score Detection data exceeding the preset acceptable threshold is marked as acceptable; otherwise, it is marked as unacceptable. The preset acceptable threshold ranges from [60, 80].

8. A machine learning-based aircraft reconnaissance data quality assessment system, used to perform the method according to any one of claims 1-7, characterized in that, include: Data acquisition and verification module: used to acquire raw detection data from aircraft-mounted equipment, including flight path data, meteorological parameters, and cloud microphysical data; The original detection data is verified based on multi-source consistency to obtain preliminary screening data. Data filtering module: used to calculate the overall confidence level of the initial screening data based on the flight track data, and to regard the initial screening data with an overall confidence level higher than a threshold as reliable data, and the rest as uncertain data; a memory-enhanced autoencoder based on meteorological features and spatiotemporal distribution features performs scene scarcity analysis on the uncertain data to obtain scarce data; The original scoring module is used to calculate the original score of the probe data by using the reliable data and the scarce data as probe data and employing an asymmetric penalty mechanism; and to evaluate the scarce data based on spatiotemporal consistency and prediction value to obtain a scarce feature vector. Model building module: used to construct a time-series cumulative deviation quality assessment model based on the basic scoring function according to the original score and the rare feature vector; input the data to be evaluated into the time-series cumulative deviation quality assessment model and output the evaluation results.

Images