Low-altitude aircraft chip security evaluation method based on electromagnetic channel monitoring

Abstract

The application relates to the technical field of low-altitude aircraft chip safety evaluation, and discloses a low-altitude aircraft chip safety evaluation method based on electromagnetic channel monitoring. The method first collects real-time channel data output by an electromagnetic channel monitoring device during chip operation, and channel data from multiple groups of historical safety evaluation records to form an evaluation data set; then, based on the chip abnormal response characteristics corresponding to each group of channel data in the evaluation data set, the historical channel data is divided into multiple types; next, a safety evaluation model is trained using the evaluation data set, core feature dimensions are screened by analyzing data classification deviation, the importance distribution is calculated according to the feature dispersion of data on the core feature dimensions, and the model is iteratively calibrated; finally, the real-time channel data is input into the trained model to obtain a safety evaluation conclusion and determine the chip safety state. The method can evaluate the safety state of the low-altitude aircraft chip in real time and accurately, and improve the evaluation reliability and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude aircraft chip security evaluation, in particular to a low-altitude aircraft chip security evaluation method based on electromagnetic channel monitoring. BACKGROUND

[0002] With the wide application of low-altitude aircraft in civil, industrial, military and other fields, the security performance of its core chip has become a key factor to ensure the stable operation of the aircraft and the successful execution of the task. In the actual operation process of low-altitude aircraft, the chip may be affected by various factors such as electromagnetic interference, hardware failure, software vulnerability and malicious attack, which may cause abnormal behavior of the chip and affect the performance and safety of the entire aircraft.

[0003] The methods for chip security evaluation mainly focus on the security verification in the design stage of the chip and the performance evaluation of the chip in a specific test environment. However, these methods have obvious limitations. The verification in the design stage mainly targets the theoretical security of the chip, which is difficult to completely simulate the complex electromagnetic channel conditions that the chip may face in the actual low-altitude flight environment. There are various electromagnetic signal sources in the low-altitude environment, such as radar signals, communication signals, natural electromagnetic interference, etc. These signals will affect the operation of the chip through the electromagnetic channel, while the traditional design verification method often fails to fully consider these actual environmental factors.

[0004] The performance evaluation in the existing test environment is usually carried out under fixed and controllable conditions, which cannot truly reflect the actual security state of the chip in the dynamically changing low-altitude flight scene. In the actual operation process of the chip, its working state will change with the flight attitude, flight height and surrounding electromagnetic environment of the aircraft, and the traditional evaluation method is difficult to monitor and evaluate the security performance of the chip in such dynamic environment in real time.

[0005] In addition, most of the existing chip security evaluation methods lack in-depth analysis and effective use of the abnormal response characteristics of the chip. When the chip appears abnormal, its electromagnetic channel data will show specific characteristic changes, but the traditional method often cannot accurately capture and analyze these characteristics, resulting in the inability to timely and accurately judge the security state of the chip, which may cause serious safety accidents.

[0006] The existing evaluation model often has problems such as unreasonable feature dimension selection and inaccurate data importance distribution calculation when processing complex electromagnetic channel data, resulting in low accuracy and reliability of the evaluation model. It cannot meet the high precision and real-time requirements of low-altitude aircraft for chip security evaluation.

[0007] Therefore, there is an urgent need for a method that can monitor low-altitude aircraft chips in real time, accurately and comprehensively based on actual electromagnetic channel monitoring data to ensure the safe operation of low-altitude aircraft. SUMMARY

[0008] The purpose of the present application is to provide a low-altitude aircraft chip safety evaluation method based on electromagnetic channel monitoring to solve the problems raised in the background art.

[0009] To achieve the above-mentioned purpose, the present application provides the following technical solutions: a low-altitude aircraft chip safety evaluation method based on electromagnetic channel monitoring, the method comprising the following steps:

[0010] Collecting real-time channel data output by electromagnetic channel monitoring equipment during the operation of low-altitude aircraft chips and an evaluation data set composed of multiple sets of historical channel data recorded during the safety evaluation process;

[0011] Based on the chip abnormal response characteristics corresponding to each set of channel data in the evaluation data set, the historical channel data is divided into multiple types; based on the evaluation data set, the safety evaluation model is trained, the core feature dimension is selected by analyzing the data classification bias, the data importance distribution is calculated according to the feature dispersion of different data in the core feature dimension; based on the importance distribution, the safety evaluation model is iteratively calibrated to obtain a trained safety evaluation model;

[0012] Inputting the real-time channel data output by the electromagnetic channel monitoring equipment during the operation of the low-altitude aircraft chip into the trained safety evaluation model to obtain a safety evaluation conclusion, and performing a chip safety state judgment operation based on the safety evaluation conclusion.

[0013] Preferably, the historical channel data is divided into multiple types based on the chip abnormal response characteristics corresponding to each set of channel data in the evaluation data set, comprising:

[0014] For the kth set of historical channel data:

[0015] record each training phase of the kth group of historical channel data from first misjudged as normal to first determined as abnormal as an analysis phase of the kth group of historical channel data; arrange abnormal response change values corresponding to the kth group of historical channel data in all analysis phases in chronological order to obtain a feature evolution sequence of the kth group of historical channel data; take a position index of each value in the feature evolution sequence as a horizontal axis coordinate, take a value corresponding to each position index as a vertical axis coordinate, form no less than two coordinate points, take all coordinate points as an input of feature distribution analysis, and obtain each two-dimensional coordinate vector and a corresponding mapping value; take a two-dimensional coordinate vector corresponding to a maximum mapping value as a dominant direction vector, and take an inverse sine value of a ratio of a vertical axis component to a horizontal axis component in the dominant direction vector as a feature distribution trend value;

[0016] based on the training phase corresponding to first misjudgment as normal of the kth group of historical channel data, a stage interval between first misjudgment as normal and first determination as abnormal, and the feature distribution trend value, calculate a classification measure value of the kth group of historical channel data;

[0017] based on the classification measure value, determine a type to which the kth group of historical channel data belongs.

[0018] Preferably, based on the training phase corresponding to first misjudgment as normal of the kth group of historical channel data, the stage interval between first misjudgment as normal and first determination as abnormal, and the feature distribution trend value, calculating the classification measure value of the kth group of historical channel data comprises:

[0019] record a difference between the feature distribution trend value and a preset reference trend value as a second evaluation parameter;

[0020] calculate a standardized result of the training phase corresponding to first misjudgment as normal of the kth group of historical channel data, and determine a product of the stage interval between first misjudgment as normal and first determination as abnormal, the standardized result, and the second evaluation parameter as the classification measure value of the kth group of historical channel data.

[0021] Preferably, based on the classification measure value, determining the type to which the kth group of historical channel data belongs comprises:

[0022] if the classification measure value is greater than a preset type division threshold value, determine that the kth group of historical channel data belongs to a high-risk chip category; otherwise, determine that the kth group of historical channel data belongs to a low-risk chip category.

[0023] Preferably, the core feature dimension is screened by analyzing data classification bias conditions, comprising:

[0024] for the qth type:

[0025] constructing a third feature matrix based on all the channel data determined to be correct in the qth type, wherein each row in the third feature matrix is a group of channel data determined to be correct; processing the third feature matrix by using a feature screening method to obtain a third feature screening result, wherein each column of data in the third feature screening result constitutes a third candidate feature dimension;

[0026] constructing a fourth feature matrix based on the channel data determined to be incorrect in the qth type, wherein each row in the fourth feature matrix is a group of channel data determined to be incorrect; processing the fourth feature matrix by using a feature screening method to obtain a fourth feature screening result, wherein each column of data in the fourth feature screening result constitutes a fourth candidate feature dimension;

[0027] performing correlation analysis on the third candidate feature dimension and the fourth candidate feature dimension, and screening a core feature dimension based on the analysis result.

[0028] Preferably, the correlation analysis on the third candidate feature dimension and the fourth candidate feature dimension, and the screening of the core feature dimension based on the analysis result, comprise:

[0029] matching the third candidate feature dimension and the fourth candidate feature dimension by using a feature coupling method to obtain a plurality of feature coupling pairs; calculating the coupling degree of the two dimensions in each feature coupling pair, and taking a feature coupling pair with a coupling degree greater than a preset coupling threshold as a target coupling pair;

[0030] taking the intermediate dimension of the two dimensions in each target coupling pair as a core feature dimension.

[0031] Preferably, the calculation of the data importance distribution according to the feature dispersion of different data on the core feature dimension comprises:

[0032] for the rth group of channel data determined to be correct:

[0033] taking the channel data determined to be incorrect in the type in which the rth group of channel data determined to be correct is located as reference data of the rth group of channel data determined to be correct; calculating the mean value of the feature value of each core feature dimension of all the reference data of the rth group of channel data determined to be correct, respectively; and taking the absolute deviation value of the feature value of each core feature dimension of the rth group of channel data determined to be correct from the corresponding mean value as the dispersion index of each core feature dimension of the rth group of channel data determined to be correct;

[0034] taking the sum of the standardization results of the dispersion indexes of all the core feature dimensions of the rth group of channel data determined to be correct as the importance coefficient of the rth group of channel data determined to be correct; and calculating the target importance of the rth group of channel data determined to be correct in each training stage based on the importance coefficient.

[0035] Preferably, the acquisition of the chip abnormal response feature corresponding to each group of channel data in the evaluation data set comprises:

[0036] The chip abnormal response trajectory corresponding to each group of channel data in the evaluation data set is statistically analyzed to obtain a corresponding trajectory distribution atlas, and the atlas is divided into not less than two division intervals by using a region division method.

[0037] The average value of the abnormal response change rate of all channel data in each division interval is calculated respectively, and is taken as the chip abnormal response feature of each group of channel data in the corresponding division interval.

[0038] Preferably, the calculation of the target importance of the rth group of correct determination data in each training stage based on the importance coefficient comprises:

[0039] For any one group of channel data with correct determination: the initial importance of the group of data in each training stage is added to the corresponding importance coefficient to obtain the target importance of the group of data in each training stage.

[0040] Preferably, the iterative calibration of the security evaluation model based on the importance distribution comprises:

[0041] The target importance of all correct determination data in each training stage is weighted and summed to obtain a comprehensive calibration parameter of the training stage; the comprehensive calibration parameter is difference calculated with the current parameter of the model, and the internal weight of the model is adjusted based on the calculation result; the parameter adjustment operation is repeatedly executed until the comprehensive calibration parameter is less than a preset calibration threshold, and the iterative calibration of the security evaluation model is completed.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] The method provides rich and comprehensive data support for chip security evaluation by collecting real-time channel data in the running process of the low-altitude flying vehicle chip and a plurality of groups of historical channel data recorded in the security evaluation process. The historical channel data is divided into a plurality of types based on the chip abnormal response feature corresponding to each group of channel data in the evaluation data set, and the classification method can fully consider the electromagnetic channel feature difference of the chip under different abnormal conditions, so that the subsequent security evaluation is more accurate.

[0044] When training the security evaluation model, the core feature dimension is screened by analyzing the data classification bias, which can effectively remove redundant features, select the most critical feature dimension for chip security evaluation, and improve the training efficiency and evaluation accuracy of the model. According to the feature dispersion of different data on the core feature dimension, the data importance distribution is calculated, so that the model can pay more attention to the data that has a greater impact on the evaluation result, further improving the reliability of the evaluation. Based on the importance distribution, the security evaluation model is iteratively calibrated to continuously optimize the model parameters, making the model better adapt to the actual electromagnetic channel data and chip security evaluation requirements.

[0045] The real-time channel data is input into the trained security evaluation model to obtain the security evaluation conclusion, and based on this, the chip security state judgment operation is performed, realizing the real-time monitoring and accurate evaluation of the chip security state of the low-altitude flying vehicle. When classifying historical channel data, for each group of historical channel data, each training phase from the first misjudgment as normal to the first judgment as abnormal is taken as the analysis phase, the abnormal response change value is arranged in time sequence to obtain the feature evolution sequence, and then the coordinate points are formed for feature distribution analysis, the dominant direction vector and feature distribution trend value are obtained, and the classification measure value is calculated combined with the training phase corresponding to the first misjudgment as normal, the phase interval and the feature distribution trend value, so as to judge the type. This detailed classification process can deeply mine the time evolution characteristics and distribution trend of chip abnormal response, improving the accuracy and scientificity of historical channel data classification.

[0046] When calculating the classification measure value, the difference between the feature distribution trend value and the preset reference trend value is taken as the second evaluation parameter, combined with the standardized result of the training phase and the phase interval, so that the classification measure value can comprehensively reflect the influence of multiple key factors on chip abnormalities, further ensuring the accuracy of classification. By comparing the classification measure value with the preset type division threshold, the chip category can be determined, which can clearly distinguish high-risk and low-risk chips, providing a strong basis for subsequent safety management and decision-making.

[0047] When screening the core feature dimension, the channel data feature matrix of correct and incorrect judgment is constructed for each type respectively, and after processing by the feature screening method, the correlation analysis is carried out, the target coupling pair is matched by the feature coupling method, and the intermediate dimension is taken as the core feature dimension. This method can fully consider the feature difference between correct and incorrect judgment data, screen out the core feature dimension that has the greatest impact on the classification result, and improve the feature selection ability and evaluation performance of the model.

[0048] In the calculation of the data importance distribution, the data with incorrect determination in the type of correct data is taken as the reference data to calculate the discrete index on each core feature dimension, and then the importance coefficient and the target importance are obtained. This calculation method can determine the importance of data according to the discrete degree of the data on the core feature dimension, so that the model can more reasonably allocate weights in the training process, and improve the attention of the model to key data and the accuracy of evaluation.

[0049] In the acquisition of the chip abnormal response feature, the chip abnormal response trajectory corresponding to each group of channel data in the evaluation data set is counted to obtain a trajectory distribution atlas, and the average value of the abnormal response change rate of each interval is calculated as a feature after the region division method is used for division. This method can comprehensively and accurately capture the feature distribution and change rule of the chip abnormal response, and provides a reliable feature basis for the classification of historical channel data and the training of the safety evaluation model.

[0050] In the calculation of the target importance based on the importance coefficient, the initial importance in each training stage is added to the importance coefficient, so that the target importance can dynamically reflect the importance change of data in different training stages, and the training process of the model is further optimized. The target importance of all correct determination data in each training stage is weighted and summed to obtain a comprehensive calibration parameter, and the internal weight is adjusted after the difference between the comprehensive calibration parameter and the current parameter of the model is calculated, until the comprehensive calibration parameter is less than a preset threshold, the iterative optimization of the safety evaluation model is realized, the evaluation accuracy and reliability of the model are continuously improved, and the model can better adapt to the actual needs of the chip safety evaluation of low-altitude aircraft, and provides strong technical support for ensuring the safe operation of low-altitude aircraft. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The working principle diagram of the low-altitude aircraft chip safety evaluation method based on electromagnetic channel monitoring is described in the present application.

[0052] Figure 2 The flowchart for the classification of historical channel data types is shown.

[0053] Figure 3 The flowchart for the calculation of the classification measure value is shown.

[0054] Figure 4 The flowchart for the calculation of the data importance distribution is shown.

[0055] Figure 5 The flowchart for the acquisition of the chip abnormal response feature is shown. DETAILED DESCRIPTION

[0056] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0057] Please refer to Figures 1-5 The application relates to a low-altitude aircraft chip safety evaluation method based on electromagnetic channel monitoring, and the implementation steps are as follows:

[0058] Real-time channel data output by an electromagnetic channel monitoring device during the running of a low-altitude aircraft chip and an evaluation data set composed of a plurality of groups of channel data recorded in a historical safety evaluation process are collected. The electromagnetic channel monitoring device can be a monitoring device based on an ultra-wideband antenna, which collects electromagnetic radiation signals during the operation of the chip in real time and converts the signals into digital signals to form real-time channel data. The evaluation data set is composed of at least 1000 groups of channel data recorded in a chip safety evaluation process, and each group of data contains electromagnetic channel characteristic parameters during the operation of the chip and a corresponding safety state identifier.

[0059] Based on the chip abnormal response characteristics corresponding to each group of channel data in the evaluation data set, the historical channel data is divided into a plurality of types. Specifically, for the kth group of historical channel data, each training stage from the first misjudgment as normal to the first judgment as abnormal is determined as an analysis stage, the abnormal response change values corresponding to all analysis stages are arranged in time sequence to form a feature evolution sequence. The position index of each value in the sequence is taken as the horizontal axis coordinate, and the corresponding value is taken as the vertical axis coordinate. No less than two coordinate points are generated, which are taken as the input of feature distribution analysis, each two-dimensional coordinate vector and the corresponding mapping value are obtained, the two-dimensional coordinate vector corresponding to the maximum mapping value is selected as the dominant direction vector, the inverse sine value of the ratio of the vertical axis component to the horizontal axis component in the dominant direction vector is calculated, and the feature distribution trend value is obtained.

[0060] The safety evaluation model is trained based on the evaluation data set, and the core feature dimension is selected by analyzing the data classification deviation. First, for each type, the channel data correctly judged in the type is used to construct a third feature matrix, each row corresponds to a group of data, and the recursive feature elimination method is used to process the third feature matrix to obtain a third feature screening result, and each column of data constitutes a third candidate feature dimension. The channel data incorrectly judged is used to construct a fourth feature matrix, and the recursive feature elimination method is also used to process the fourth feature matrix to obtain a fourth candidate feature dimension.

[0061] The importance distribution of data is calculated according to the feature dispersion of different data in the core feature dimension. For the rth group of correctly judged channel data, the misjudged data in the type in which the rth group of data is located is taken as reference data, the mean value of the feature values of the reference data in each core feature dimension is calculated, the absolute deviation value of the feature values of the rth group of data in each core feature dimension from the corresponding mean value is taken as a dispersion index, and the importance coefficient of the rth group of data is obtained by summing the standardized dispersion indexes. Then, the target importance of the rth group of data in each training stage is calculated.

[0062] The importance distribution is used to iteratively calibrate the security evaluation model to obtain a trained security evaluation model. The target importance of all correctly judged data in each training stage is weighted and summed to obtain a comprehensive calibration parameter, and the comprehensive calibration parameter is subtracted from the current parameter of the model. The internal weight of the model is adjusted according to the difference, and the operation is repeated until the comprehensive calibration parameter is less than a preset calibration threshold.

[0063] Real-time channel data output by an electromagnetic channel monitoring device during the operation of a low-altitude aircraft chip is input into the trained security evaluation model to obtain a security evaluation conclusion, and a chip security state judgment operation is performed based on the conclusion. The security evaluation model can use a support vector machine model. After inputting the feature parameters of real-time channel data, the model outputs corresponding security risk levels such as high risk, medium risk, and low risk, and the current security state of the chip is determined according to the output result.

[0064] Embodiment 1:

[0065] When classifying the historical channel data, the operation is specifically performed on the kth group of historical channel data. The analysis stage of the data needs to be determined. The analysis stage is defined as the training stage in which the kth group of historical channel data is first misjudged as normal to the training stage in which the kth group of historical channel data is first judged as abnormal. Each training stage during this period is recorded as the analysis stage of the kth group of historical channel data. For example, if the kth group of data is first misjudged as normal in the third training, and is first correctly judged as abnormal in the seventh training, then the third to seventh training stages constitute the analysis stage of the data, and the abnormal response change value corresponding to each training stage needs to be collected in this process.

[0066] The abnormal response change values corresponding to the kth group of historical channel data in all analysis stages are arranged in chronological order to obtain a feature evolution sequence of the group of data. Assuming that the values in the feature evolution sequence are 0.2, 0.5, 0.8, 1.2, and 1.5 in turn, the position indexes of each value in the sequence are 1, 2, 3, 4, and 5 in turn. At this time, each position index is taken as a horizontal axis coordinate, and the value corresponding to each position index is taken as a vertical axis coordinate, so that a plurality of coordinate points are formed, such as (1, 0.2), (2, 0.5), (3, 0.8), (4, 1.2), and (5, 1.5), and the number of these coordinate points is not less than two. Then, all these coordinate points are taken as input data for feature distribution analysis, and each two-dimensional coordinate vector and a corresponding mapping value are obtained by a specific analysis method.

[0067] Among all the obtained two-dimensional coordinate vectors and mapping values, the two-dimensional coordinate vector with the largest mapping value is selected and determined as a dominant direction vector. For example, after calculation and analysis, it is found that the mapping value corresponding to a certain two-dimensional coordinate vector is the largest among all vectors, and the vector is identified as the dominant direction vector. Assuming that the dominant direction vector is (5, 1.3), the horizontal axis component is 5, and the vertical axis component is 1.3. Next, the ratio of the vertical axis component to the horizontal axis component in the dominant direction vector is calculated, that is, 1.3 divided by 5 to obtain 0.26, and the inverse sine value of the ratio is taken to obtain the feature distribution trend value. Through mathematical calculation, the value is about 0.263 radians.

[0068] Based on the training stage corresponding to the first misjudgment of the kth group of historical channel data as normal, the interval between the first misjudgment as normal and the first determination as abnormal, and the feature distribution trend value obtained in the foregoing, the classification measure value of the group of data is calculated. First, the feature distribution trend value is compared with a preset reference trend value, and the difference between the two is recorded as a second evaluation parameter. Here, the preset reference trend value can be set according to actual conditions and historical data, such as 0.2 radians. In the foregoing example, the second evaluation parameter is 0.063, which is obtained by subtracting 0.2 from 0.263.

[0069] The training stage corresponding to the first misjudgment of the kth group of historical channel data as normal is standardized to obtain a standardized result. There are various standardization methods, for example, the method of (training stage-1) / (maximum training stage-1) can be used. Assuming that the maximum training stage is 10 in the entire evaluation data set, and the training stage corresponding to the first misjudgment of the kth group of data as normal is the third time, the standardized result is (3-1) / (10-1), which is calculated to be about 0.222.

[0070] The interval between the first misjudgment as normal and the first judgment as abnormal is calculated. In the previous example, the first misjudgment as normal is at the 3rd training, and the first judgment as abnormal is at the 7th training, so the interval is 7 minus 3, which is 4.

[0071] The classification measure value of the kth set of historical channel data is determined by multiplying the interval, the standardized result, and the second evaluation parameter. That is, 4 times 0.222 times 0.063, which is calculated to be about 0.056.

[0072] After obtaining the classification measure value, it is necessary to judge the type to which the kth set of historical channel data belongs based on the value. At this time, the classification measure value is compared with a preset type division threshold. The preset type division threshold is also determined according to a large amount of historical data and actual application requirements, such as being set to 0.05. If the classification measure value is greater than the preset type division threshold, it is determined that the kth set of historical channel data belongs to the high-risk chip category; otherwise, if the classification measure value is less than or equal to the preset type division threshold, it is determined that it belongs to the low-risk chip category. In the previous example, the classification measure value is 0.056, which is greater than the preset 0.05, so it is determined that the kth set of historical channel data belongs to the high-risk chip category.

[0073] Embodiment 2:

[0074] In the selection of core feature dimensions, the qth type is taken as an example to expand the specific operation. All the channel data judged correctly in the qth type need to be extracted, and it is assumed that there are m sets of channel data judged correctly in the type. The m sets of data are constructed into a third feature matrix, in which each row corresponds to a set of channel data judged correctly, each column corresponds to a feature dimension, and each element in the matrix records the feature value of the corresponding set of data in the feature dimension.

[0075] After the third feature matrix is constructed, a feature selection method is used to process it. There are many feature selection methods, such as variance threshold method, recursive feature elimination method, etc. Taking the variance threshold method as an example, a variance threshold is set, such as 0.1. The core idea of this method is to delete those feature dimensions with a variance less than the set threshold, because the data fluctuation in these feature dimensions is small, which may not contribute much to classification. After processing, the third feature selection result is obtained, in which each column of data constitutes a third candidate feature dimension. It is assumed that after screening, n third candidate feature dimensions are obtained, which are considered as potential feature dimensions that play an important role in classification in the data judged correctly.

[0076] The channel data determined to be erroneous in this type is processed. All the channel data determined to be erroneous is also selected from the qth type, assuming that there are p groups in total. A fourth feature matrix is constructed using the p groups of data, which has a structure similar to the third feature matrix, with each row corresponding to a group of channel data determined to be erroneous and each column corresponding to a feature dimension, recording the feature values of each group of data in the feature dimension.

[0077] After the fourth feature matrix is constructed, the same feature screening method is used to process the fourth feature matrix as for the third feature matrix, for example, the variance threshold method is also used, and the same variance threshold of 0.1 is set. After processing, the fourth feature screening result is obtained, in which each column of data constitutes a fourth candidate feature dimension, and it is assumed that o fourth candidate feature dimensions are obtained. These fourth candidate feature dimensions are screened from the data determined to be erroneous and may reflect the key feature dimensions that cause classification errors.

[0078] After the third candidate feature dimensions and the fourth candidate feature dimensions are screened, they need to be associated and analyzed to screen the core feature dimensions. The association and analysis process uses a feature coupling method, for example, the mutual information method. The mutual information method can measure the correlation between two feature dimensions, and by calculating the mutual information value between the third candidate feature dimensions and the fourth candidate feature dimensions, the coupling degree between them can be determined.

[0079] Specifically, each third candidate feature dimension is matched with each fourth candidate feature dimension to form multiple feature coupling pairs. For each feature coupling pair, the mutual information value is calculated, and the greater the mutual information value, the stronger the correlation between the two feature dimensions. A preset coupling threshold is set, such as 0.5, and the feature coupling pairs with a mutual information value greater than the preset coupling threshold are taken as target coupling pairs. Assuming that q target coupling pairs are obtained through calculation, the feature dimensions in these target coupling pairs have a strong correlation.

[0080] For the two dimensions in each target coupling pair, an intermediate dimension needs to be determined as a core feature dimension. The calculation method of the intermediate dimension can be to consider the feature values of the two dimensions comprehensively, for example, to take the weighted average of the two dimensions, and the weights can be determined according to their mutual information values, the higher the mutual information value, the greater the corresponding weight, so that the influence of the feature dimension with strong correlation can be highlighted. For example, the two dimensions in the target coupling pair are A and B, and their mutual information values are 0.6 and 0.7 respectively, then dimension A can be assigned a weight of 0.4 and dimension B can be assigned a weight of 0.6, and the intermediate dimension is 0.4 times the feature value of dimension A plus 0.6 times the feature value of dimension B. The intermediate dimension obtained in this way can comprehensively reflect the information of the two related feature dimensions.

[0081] Through the above series of operations, the core feature dimensions are screened out from the third candidate feature dimension and the fourth candidate feature dimension. These core feature dimensions not only consider the key features in the correct judgment data, but also combine the features that cause classification deviation in the incorrect judgment data, can more comprehensively reflect the feature distribution of the data, thereby playing a key role in the training of the security evaluation model and improving the classification accuracy and reliability of the model.

[0082] Embodiment 3:

[0083] When calculating the data importance distribution, for the rth group of channel data judged correct, the data judged incorrect in the type where the rth group of data belongs to needs to be determined first, and these data are taken as the reference data of the rth group of data judged correct. Assuming that the rth group of data belongs to type T, and there are s groups of data judged incorrect in type T, then the s groups of data constitute the reference data set of the rth group of data.

[0084] For each core feature dimension, the mean value of the feature values of the reference data on the dimension is calculated. For example, if the core feature dimension is F1, the feature values of the reference data on the F1 dimension are 1.2, 1.5, 1.8, and 2.1 respectively, then when calculating the mean value of these values, all the values are added and divided by the number of data, that is, (1.2+1.5+1.8+2.1)÷4=1.65, and this value is the mean value of the reference data on the F1 dimension.

[0085] The absolute deviation value of the feature value of the rth group of data judged correct on each core feature dimension from the corresponding mean value is calculated, and the deviation value is recorded as the dispersion index of the rth group of data on the core feature dimension. Continuing to take the F1 dimension as an example, if the feature value of the rth group of data on the F1 dimension is 1.9, then the absolute deviation value is |1.9-1.65|=0.25, and this 0.25 is the dispersion index of the rth group of data on the F1 dimension. According to the same logic and method, the dispersion indices of the rth group of data on all core feature dimensions are sequentially calculated.

[0086] Assuming that there are t core feature dimensions, after the calculation of all dimensions is completed, t dispersion indices are obtained. At this time, the dispersion indices need to be standardized. The purpose of standardization is to eliminate the influence of different dimension quantities and numerical ranges on the results, so that the dispersion indices have comparability. There are many standardization methods, and the common one is the minimum-maximum standardization, that is, (dispersion index-minimum value)÷(maximum value-minimum value). Assuming that among the t dispersion indices, the minimum value is 0.1, the maximum value is 0.5, and one of the dispersion indices is 0.3, then the standardization result is (0.3-0.1)÷(0.5-0.1)=0.5.

[0087] The sum of all standardized discrete indicators is the importance coefficient of the rth group of correct data. For example, if the standardized discrete indicators are 0.5, 0.6, 0.7, 0.8, and 0.6 (assuming t = 5), the sum is 0.5 + 0.6 + 0.7 + 0.8 + 0.6 = 3.2, which is the importance coefficient of the rth group of data.

[0088] After obtaining the importance coefficient, the target importance of the rth group of correct data in each training stage needs to be calculated based on the coefficient. For any group of correct channel data, the target importance in each training stage is calculated by adding the initial importance of the data in the training stage to the corresponding importance coefficient. The initial importance is the basic importance value set for each group of data at the beginning of training. The initial importance of different training stages may be the same or different, and the specific value needs to be determined according to the training strategy.

[0089] Suppose the initial importance of the rth group of data is set to 0.8 in a certain training stage, and the importance coefficient is 3.2, then the target importance of the data in the training stage is 0.8 + 3.2 = 4.0. It needs to be noted that in actual operation, the target importance needs to be recalculated in each training stage, because the initial importance of different training stages may be adjusted according to the training progress or data performance, and if the core feature dimension or reference data changes, the importance coefficient may also change.

[0090] The entire calculation process needs to ensure the accuracy of the data and the rigor of the logic. When determining the reference data, the data type must be accurately divided to avoid including incorrect data of different types in the reference range, otherwise it will cause deviation in the mean calculation, and then affect the accuracy of the discrete indicators and the importance coefficient. When calculating the mean of the core feature dimension, the feature values of all reference data must be accurate, and any data entry error or feature extraction deviation may cause distortion of the mean.

[0091] When standardizing the discrete indicators, the minimum and maximum values need to be correctly determined. If the data range is incorrectly calculated, the standardization result will deviate from the actual value, affecting the calculation of the importance coefficient. In addition, the setting of the initial importance needs to be combined with the training target and data characteristics, and a reasonable initial value setting can make the target importance more accurately reflect the importance of the data.

[0092] For example, in actual operation, if the dispersion index of a certain group of correct data in multiple core feature dimensions is large, it indicates that the difference between it and the reference data is significant, and the importance coefficient obtained by standardization and summation will also be large, which means that this group of data has higher importance in model training and needs to be given greater weight in model parameter adjustment to better fit the characteristics of this type of data and improve the classification ability of the model for similar data. Conversely, if the dispersion index is small, the importance coefficient is also small, indicating that the data is similar to the reference data and has relatively low importance in training.

[0093] Through this method of calculating data importance distribution based on core feature dimension dispersion, the importance of each group of correct data in training can be dynamically evaluated, allowing the model to pay more attention to data that is significantly different from incorrect data and has a more significant impact on classification results during iteration, thereby improving the training efficiency and evaluation accuracy of the model. Each step from the determination of the reference data to the calculation of the target importance needs to be strictly executed to ensure that the final importance distribution accurately reflects the feature differences and importance of the data, providing a reliable basis for the iterative calibration of the security evaluation model.

[0094] Embodiment 4:

[0095] In obtaining the chip abnormal response characteristics corresponding to each group of channel data in the evaluation data set, the chip abnormal response trajectory corresponding to each group of channel data needs to be statistically analyzed. Taking a certain group of channel data as an example, during the operation of the chip, the electromagnetic channel monitoring equipment continuously collects the electromagnetic radiation signals of the chip, and records the electromagnetic channel characteristic parameters (such as signal strength, frequency distribution, phase change, etc.) at different times in chronological order to form the abnormal response trajectory of this group of data. For example, when the chip is subjected to some kind of interference or attack, its electromagnetic radiation signals will change accordingly, and these changes are recorded point by point to form trajectory data reflecting the abnormal response of the chip.

[0096] The abnormal response trajectory data obtained by statistical analysis is plotted into a trajectory distribution map. The horizontal axis of the map can be set as a time sequence (such as sample point number, running time, etc.), and the vertical axis corresponds to the specific value of the electromagnetic channel characteristic parameter. Taking the horizontal axis as sample points 1 to 100 and the vertical axis as signal strength values as an example, the signal strength values at each sample point are marked in the coordinate system and connected by a curve or scattered points to form an intuitive trajectory distribution map for subsequent analysis.

[0097] The atlas is divided by a region division method to obtain no less than two division intervals. The region division method can be selected in various ways, such as a k-means clustering algorithm based on data distribution characteristics, or a threshold division method set according to domain knowledge. Taking the k-means clustering algorithm as an example, assuming that the atlas is divided into 3 division intervals, the algorithm will automatically classify similar data points into the same interval according to the distribution of the trajectory data points, so that the data points in each interval have a high degree of similarity, and the data points in different intervals have large differences.

[0098] After completing the atlas division, for each division interval, the average value of the abnormal response change rate of all channel data in the interval is calculated, and the average value is taken as the chip abnormal response feature of each group of channel data in the corresponding division interval. The calculation of the abnormal response change rate is based on the feature parameter values of adjacent sampling points, that is, the ratio of (the feature parameter value of the current sampling point - the feature parameter value of the previous sampling point) to the feature parameter value of the previous sampling point. For example, a division interval contains 10 sampling point data, and the signal strength values are 10, 12, 15, 13, 16, 18, 20, 19, 21, and 23 in turn. Calculate the change rate of each adjacent sampling point: the first change rate is (12-10) / 10=0.2, the second is (15-12) / 12=0.25, the third is (13-15) / 15≈-0.13, and so on, 9 change rates corresponding to 10 data points are obtained. Add the 9 change rates and take the average, that is, (0.2+0.25-0.13+…+0.09) / 9, the result is the average value of the abnormal response change rate of the division interval, which will be taken as the chip abnormal response feature of each group of channel data in the interval.

[0099] Taking another group of channel data as an example, assuming that its abnormal response trajectory atlas is divided into 2 intervals by k-means clustering. The first interval contains the first 50 sampling points, and the second interval contains the last 50 sampling points. In the first interval, the electromagnetic channel feature parameter (such as frequency offset) changes relatively smoothly, and the change rate of adjacent sampling points is mostly between 0.05 and 0.1; in the second interval, the change of the feature parameter is more violent, and the change rate is mostly between 0.15 and 0.25, and even individual large fluctuation values appear. By calculating the average value of the change rate of each interval, the abnormal response feature values reflecting the characteristics of the interval can be obtained, such as the average value of the first interval is 0.08, and the average value of the second interval is 0.22.

[0100] In actual operation, the accuracy of region division directly affects the extraction effect of abnormal response features. If the division interval is too many, it may lead to too small data volume in each interval, and the average value calculation lacks representativeness; if the division interval is too small, it may not be able to accurately capture different change patterns of trajectory data. Therefore, the division method and interval number need to be reasonably selected according to the distribution characteristics of trajectory data. For example, when the trajectory data presents obvious stage change, threshold division based on domain knowledge can be used to divide stages with similar change trend into the same interval; when the data distribution is complex, unsupervised learning methods such as k-means clustering can more objectively divide.

[0101] In addition, the calculation of abnormal response change rate needs to pay attention to the continuity and accuracy of data. If the data of a sampling point is abnormal due to noise interference or equipment failure, data preprocessing (such as filtering, interpolation, etc.) is needed to avoid the influence of abnormal values on the calculation of change rate. For example, the signal strength value of a sampling point suddenly drops due to sudden interference, causing the adjacent change rate to appear a large negative value, which needs to be corrected by median filtering method, and then the change rate is calculated to ensure the reliability of the average value.

[0102] For each set of channel data in the evaluation data set, the above operations need to be repeated: first, the abnormal response trajectory is counted, the map is drawn, the interval is divided, and then the average value of the change rate of each interval is calculated as the chip abnormal response feature of the data. For example, the evaluation data set contains 1000 groups of channel data, each group of data corresponds to an independent trajectory map, by dividing the interval and calculating the average value, the abnormal response feature of each group of data is obtained, which will be used for subsequent historical channel data type division and safety evaluation model training.

[0103] When processing large-scale data sets, computer programs can be used to automate the above processes and improve processing efficiency. For example, k-means clustering is used to divide the interval using Python's scikit-learn library, and NumPy library is used for change rate calculation and average value statistics. At the same time, a data verification mechanism needs to be established to check the rationality of the division results and calculation results, such as checking whether the interval division conforms to the data distribution law, whether the average value is within a reasonable range, etc., to ensure that the extracted abnormal response features truly reflect the abnormal response of the chip.

[0104] The chip abnormal response features obtained in this way can quantitatively describe the abnormal change patterns of the chip in different running stages, and provide a key basis for subsequent historical channel data type division based on abnormal response features. For example, the abnormal response features of a group of channel data show that the average value of the change rate in multiple intervals is large, indicating that the chip abnormal response is relatively intense, which may correspond to a high-risk chip category; while the average value of the change rate of another group of data is small, which may correspond to a low-risk category. The extraction of such quantitative features makes the chip safety evaluation process more objective and operable, and lays a foundation for building an accurate safety evaluation model. Each link from trajectory data collection to feature extraction needs to be strictly implemented to ensure the accuracy and effectiveness of the features, thereby providing reliable data support for the safety evaluation of low-altitude aircraft chips.

[0105] Embodiment 5:

[0106] When iterating and calibrating the safety evaluation model, take a specific training stage as an example. Assume that there are v groups of correctly judged channel data in the current training stage, and each group of data has obtained its target importance in the previous calculation, for example, the target importance of the first group of data is 3.5, the second group is 4.2, the third group is 2.8, and so on, until the vth group of data.

[0107] The target importance of these correctly judged data needs to be weighted and summed to obtain the comprehensive calibration parameter of this training stage. The weighted sum method can be set according to actual needs, such as using uniform weight, that is, the weight of each group of data is the same, which is 1 / v. In this case, the calculation method of the comprehensive calibration parameter is to add the target importance of all v groups of data and then divide by v. For example, if v=3, the target importance of the three groups of data is 3.5, 4.2, and 2.8 respectively, then the comprehensive calibration parameter is (3.5+4.2+2.8) ÷ 3=10.5 ÷ 3=3.5.

[0108] After obtaining the comprehensive calibration parameter, it is calculated by the difference value with the current parameter of the model. The current parameter of the model can be the set of weight parameters, bias parameters and other types of adjustable parameters of the model before the start of this training stage. Assume that the current key parameter of the model is θ, and its initial value is set to 2.0, and the calculated comprehensive calibration parameter is 3.5, then the difference is 3.5-2.0=1.5.

[0109] Based on the difference calculation result, the internal weights of the model are adjusted. The adjustment method usually uses optimization algorithms such as gradient descent, the core idea of which is to update the model parameters in the direction that reduces the loss function according to the size and direction of the difference. For example, if the difference is positive, it means that the current model parameter is smaller than the comprehensive calibration parameter, and the parameter value needs to be increased; if the difference is negative, the parameter value needs to be decreased. The adjustment amplitude is related to the size of the difference, and the larger the difference, the larger the adjustment amplitude is usually.

[0110] Take a certain weight parameter w in the model as an example, whose initial value is w0, and the adjusted weight w1 = w0 + a x difference, where a is the learning rate, which is a pre-set constant, used to control the step size of each adjustment. Assuming that the learning rate a = 0.1 and the difference is 1.5, the adjusted weight w1 = w0 + 0.1 x 1.5 = w0 + 0.15.

[0111] After completing one parameter adjustment, it is necessary to determine whether to continue to perform iteration operation. The basis for the determination is whether the comprehensive calibration parameter is less than the pre-set calibration threshold. The pre-set calibration threshold is a small value set according to the accuracy requirement of model training and actual application scenario, for example, 0.01. If the current comprehensive calibration parameter is greater than or equal to the threshold, it means that the model parameter still has a large adjustment space, and the next round of training and parameter adjustment needs to be continued; if the comprehensive calibration parameter is less than the threshold, it is considered that the model has converged, and the calibration process is completed.

[0112] Suppose that after the first adjustment, the comprehensive calibration parameter decreases from 3.5 to 2.8, which is still greater than 0.01, so it needs to enter the next round of training. In the new round of training, the target importance of each group of correct data is recalculated (may change due to the change of data importance distribution), and the new comprehensive calibration parameter is obtained by weighted summation again, such as 2.1, which is compared with the current model parameter and adjusted until the comprehensive calibration parameter is less than 0.01.

[0113] Taking a specific iterative process as an example: Assume the initial model parameter θ=2.0. In the first round of training, the comprehensive calibration parameter is 3.5 with a difference of 1.5. After adjustment, θ=2.0+0.1×1.5=2.15. In the second round of training, due to parameter adjustment, the target importance of the data may change. Assume the new target importance is 3.2, 4.0, and 2.5 respectively. The comprehensive calibration parameter is (3.2+4.0+2.5)÷3=3.23 with a difference of 3.23-2.15=1.08. After adjustment, θ=2.15+0.1×1.08=2.258. In the third round of training, the comprehensive calibration parameter is 2.9 with a difference of 0.642. After adjustment, θ=2.3222. This process is repeated iteratively, with the comprehensive calibration parameter gradually decreasing each time, until after a certain training, the comprehensive calibration parameter is 0.008, which is less than the preset threshold of 0.01. At this point, the iteration stops, and the calibration of the safety assessment model is completed.

[0114] In practice, model parameter tuning is not limited to a single parameter, but rather involves adjusting all trainable parameters in the model simultaneously. For example, for a support vector machine model, the parameters that need to be adjusted include kernel function parameters, regularization parameters, etc., and each parameter is updated based on its corresponding difference and learning rate.

[0115] Furthermore, the learning rate setting is crucial to the efficiency and stability of the iterative process. If the learning rate is too large, it may lead to excessive parameter adjustments, causing oscillations or even failure to converge; if the learning rate is too small, the number of iterations will increase, reducing training efficiency. Therefore, in practical applications, a strategy of dynamically adjusting the learning rate may be adopted, such as gradually decreasing the learning rate as the number of iterations increases, to ensure training stability and convergence speed.

[0116] Meanwhile, the target importance of correctly identified data in each training phase may vary due to factors such as changes in the core feature dimensions and updates to the data importance distribution. For example, after a training iteration, if the model improves its classification accuracy for certain data, the target importance of these data may be adjusted accordingly, thus affecting the overall calibration parameters and the direction of parameter adjustments in the next training round.

[0117] Taking a model for assessing the safety of chips in low-altitude aircraft as an example, suppose the model needs to assess the safety status of the chip when subjected to electromagnetic interference. During initial training, the model has a low accuracy rate in classifying high-risk data, which has high target importance. During iterative calibration, the model will continuously adjust parameters to improve its ability to identify high-risk data. As training progresses, the classification accuracy of high-risk data improves, and its target importance may gradually decrease. The model then begins to focus on other data that are more difficult to classify, until the classification error of all data meets the requirements and the overall calibration parameters are less than the threshold.

[0118] The entire iterative calibration process is a continuous optimization cycle, gradually approaching the true value of the security evaluation result of the model on the channel data by continuously adjusting the model parameters. Each iteration is based on the current training data and model state, dynamically adjusting the parameters until the model performance meets the expected requirements. In this process, the comprehensive calibration parameter serves as an indicator of the difference between the current state and the target state of the model, guiding the direction and amplitude of parameter adjustment, ensuring that the model can accurately evaluate the security status of the low-altitude aircraft chip.

[0119] It should be noted that during the iteration process, local optimization may occur, that is, the comprehensive calibration parameter fluctuates within a small range and cannot be further reduced. At this time, it may be necessary to adjust the training strategy, such as introducing a regularization term, replacing the optimization algorithm, etc., to help the model jump out of the local optimum and continue to optimize.

[0120] Through this iterative calibration method based on importance distribution, the security evaluation model can dynamically adjust parameters according to the importance of data, focusing on data that has a greater impact on classification results, improving the evaluation accuracy and robustness of the model. From the calculation of the comprehensive calibration parameter to the adjustment of the model parameter, each link is closely connected, ensuring that the model is continuously optimized during the training process and eventually becomes a reliable low-altitude aircraft chip security evaluation tool.

[0121] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "includes" "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0122] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring, characterized in that, The method includes the following steps: The system collects real-time channel data output by electromagnetic channel monitoring equipment during the operation of low-altitude aircraft chips, as well as an assessment dataset consisting of channel data recorded during multiple historical safety assessments. Based on the chip anomaly response characteristics corresponding to each group of channel data in the evaluation dataset, historical channel data are divided into multiple types; the security evaluation model is trained based on the evaluation dataset, and core feature dimensions are selected by analyzing data classification bias; the data importance distribution is calculated based on the feature dispersion of different data on the core feature dimensions; the security evaluation model is iteratively calibrated based on the importance distribution to obtain the trained security evaluation model; The real-time channel data output by the electromagnetic channel monitoring device during the operation of the low-altitude aircraft chip is input into the trained security assessment model to obtain the security assessment conclusion, and the chip security status determination operation is performed based on the security assessment conclusion. The historical channel data is divided into multiple types based on the chip anomaly response characteristics corresponding to each group of channel data in the evaluation dataset, including: For the k-th group of historical channel data: Each training stage of the k-th historical channel data from its first misclassification as normal to its first classification as abnormal is denoted as the analysis stage of the k-th historical channel data. The abnormal response change values ​​of the k-th historical channel data in all analysis stages are arranged chronologically to obtain the feature evolution sequence of the k-th historical channel data. The position index of each value in the feature evolution sequence is used as the horizontal axis coordinate, and the value corresponding to each position index is used as the vertical axis coordinate, forming at least two coordinate points. All coordinate points are used as inputs for feature distribution analysis to obtain each two-dimensional coordinate vector and its corresponding mapping value. The two-dimensional coordinate vector corresponding to the maximum mapping value is used as the dominant direction vector, and the arcsine of the ratio of the vertical axis component to the horizontal axis component in the dominant direction vector is used as the feature distribution trend value. Based on the training phase corresponding to the first misclassification of the k-th group of historical channel data as normal, the phase interval between the first misclassification as normal and the first misclassification as abnormal, and the feature distribution trend value, calculate the classification measure value of the k-th group of historical channel data. The type of the k-th group of historical channel data is determined based on the classification metric.

2. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The classification metric for the k-th group of historical channel data is calculated based on the training phase corresponding to the first misclassification as normal, the phase interval between the first misclassification as normal and the first misclassification as abnormal, and the feature distribution trend value, including: The difference between the characteristic distribution trend value and the preset benchmark trend value is recorded as the second evaluation parameter; Calculate the standardized result of the training phase corresponding to the first misclassification of the k-th group of historical channel data as normal, and determine the classification measure value of the k-th group of historical channel data by multiplying the phase interval between the first misclassification as normal and the first misclassification as abnormal, the standardized result, and the second evaluation parameter.

3. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The determination of the type of the k-th group of historical channel data based on the classification metric includes: If the classification metric is greater than the preset type classification threshold, then the k-th group of historical channel data is determined to belong to the high-risk chip category; otherwise, it is determined to belong to the low-risk chip category.

4. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The process of selecting core feature dimensions by analyzing data classification bias includes: For the q-th type: A third feature matrix is ​​constructed based on all correctly determined channel data in the q-th type, where each row of the third feature matrix is ​​a set of correctly determined channel data; the third feature matrix is ​​processed using a feature filtering method to obtain the third feature filtering result, where each column of data in the third feature filtering result constitutes a third candidate feature dimension; A fourth feature matrix is ​​constructed based on the channel data that was incorrectly identified in the q-th type, where each row of the fourth feature matrix is ​​a set of channel data that was incorrectly identified; the fourth feature matrix is ​​processed using a feature filtering method to obtain the fourth feature filtering result, where each column of data in the fourth feature filtering result constitutes a fourth candidate feature dimension; The third and fourth candidate feature dimensions are correlated and analyzed, and the core feature dimensions are selected based on the analysis results.

5. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 4, characterized in that, The process of performing correlation analysis on the third and fourth candidate feature dimensions, and selecting core feature dimensions based on the analysis results, includes: The feature coupling method is used to match the third candidate feature dimension with the fourth candidate feature dimension to obtain multiple feature coupling pairs; the coupling degree of the two dimensions in each feature coupling pair is calculated, and the feature coupling pairs with a coupling degree greater than a preset coupling threshold are taken as target coupling pairs; The middle dimension of each target coupling pair is taken as a core feature dimension.

6. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The calculation of data importance distribution based on the feature dispersion of different data in the core feature dimension includes: For the r-th group of channel data that is correctly identified: The incorrect data in the type of the channel data that is correctly judged in the r-th group are denoted as the reference data of the correctly judged data in the r-th group; calculate the mean of the feature values ​​of all reference data of the correctly judged data in the r-th group in each core feature dimension; and denot the absolute deviation between the feature value of the correctly judged data in the r-th group in each core feature dimension and the corresponding mean as the discrete index of the correctly judged data in the r-th group in each core feature dimension. The sum of the standardized results of the discrete indices of the r-th group of correctly judged data on all core feature dimensions is used as the importance coefficient of the r-th group of correctly judged data; the target importance of the r-th group of correctly judged data in each training phase is calculated based on the importance coefficient.

7. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The evaluation dataset includes the acquisition of chip anomaly response characteristics corresponding to each group of channel data, including: The abnormal response trajectories of the chips corresponding to each group of channel data in the evaluation dataset are statistically analyzed to obtain the corresponding trajectory distribution map. The map is divided using a region division method to obtain no less than two division intervals. The average value of the abnormal response change rate of all channel data in each divided interval is calculated and used as the chip abnormal response feature of each group of channel data in the corresponding divided interval.

8. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 6, characterized in that, The calculation of the target importance of the r-th group of correctly judged data in each training phase based on the importance coefficient includes: For any set of channel data that is correctly identified: add the initial importance of the data set in each training phase to the corresponding importance coefficient, and use the sum as the target importance of the data set in each training phase.

9. The method for security assessment of low-altitude aircraft chips based on electromagnetic channel monitoring according to claim 1, characterized in that, The iterative calibration of the security assessment model based on the importance distribution includes: The target importance of all correctly judged data in each training phase is weighted and summed to obtain the comprehensive calibration parameter for that training phase; the difference between the comprehensive calibration parameter and the current parameters of the model is calculated, and the internal weights of the model are adjusted based on the calculation result; the parameter adjustment operation is repeated until the comprehensive calibration parameter is less than the preset calibration threshold, thus completing the iterative calibration of the safety assessment model.

Images