Flight quality adaptive evaluation method

Abstract

A flight quality adaptive evaluation method. First, organize the flight actions to be evaluated, then select relevant standard flight action entries from historical flight data to establish a standard action benchmark library; construct a deep variational autoencoder network. Then, calculate the distance metric between the organized flight actions to be evaluated and the corresponding standard action compressed features obtained after being processed by the deep variational autoencoder network using dynamic programming, and then calculate the similarity metric between the flight actions to be evaluated and the standard actions. Then, according to the similarity metric between the flight actions to be evaluated and the standard actions, considering both global information and local information, form a multi-perspective fusion scoring rule, thereby constructing a scoring model, that is, score considering the completion of individual actions and also combining the performance within the complete flight sorties, so as to realize the multi-perspective automatic comprehensive evaluation of flight action quality.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining, and in particular to a method for adaptively evaluating flight quality. Background Art

[0002] The evaluation of flight quality is mainly divided into subjective method and objective method. The subjective method is generally to score the driver by experts observing the flight process or the three-dimensional trajectory reproduction process and the parameter transformation law. It depends on the expert experience of flight instructors and has strong subjective factors. Therefore, the scoring discrimination in actual applications is not high. It not only consumes a lot of manpower and material resources, but also is difficult to objectively and fairly evaluate the training quality of pilots correctly. The objective method is to select parameters in flight data to evaluate the flight skills of pilots. At present, there are few literatures using the objective method for flight quality evaluation. In 2018, Xiao Yanping et al. comprehensively analyzed the lateral-directional flight quality evaluation indexes of aircraft from the perspective of system dynamics models. In 2019, Wang Yuwei et al. constructed a comprehensive evaluation structure system for flight quality with five levels and used the comprehensive weighting method to weight the scores of basic actions. Therefore, for different flight actions, it is necessary to first extract the standard flight parameter indexes of this action and then calculate, which is time-consuming and laborious, and does not consider the influence of other flight parameters on quality evaluation. Affected by various factors such as aircraft performance, pilot operating habits, and flight environment, and at the same time, complex maneuvering actions have strong randomness and ambiguity. Since the recognition features of the above algorithms are not diverse enough, the hard division using precise thresholds, and the knowledge expression and reasoning process do not fully reflect the randomness and ambiguity of complex maneuvering actions, so the promotion value in actual applications is not high. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for adaptively evaluating flight quality to solve the problems in the above background art.

[0004] The technical problem to be solved by the present invention is realized by the following technical solutions:

[0005] The method for adaptively evaluating flight quality is as follows:

[0006] 1. Sort out the flight actions to be evaluated

[0007] Since the applicable scenarios and target tasks of different aircraft are different, and their flight action categories are different, according to the tasks in actual flight sorties, the flight process to be evaluated is divided into 10 categories of actions, including roll, ascending roll, half-roll reversal, lazy eight, takeoff, up turn, low-altitude level flight, down turn, landing, and ground taxiing. Among them, if a flight process does not belong to any category of actions, it belongs to other categories. Therefore, the goal is to divide the flight sorties into 11 flight states;

[0008] 2. Establish a standard action benchmark library

[0009] Combined with the flight training syllabus of the specified aircraft type and the experience and knowledge of flight experts, select relevant standard flight action items from rich historical flight data to establish a standard action benchmark library with typical representativeness. The relevant standards include flight parameter standards and three-dimensional trajectory standards for various flight actions. The flight parameter standards include the value-taking rules of flight parameters such as entry speed, exit speed, take-off pitch angle, time, altitude, and intake pressure. The three-dimensional trajectory standard is used to judge whether the three-dimensional trajectory of the flight action conforms to the action description and whether its smoothness meets the requirements;

[0010] According to the above standards, extract a number of standard data for each type of flight action after sorting from the flight historical data, record the start and end times of the action, and classify them into the standard action benchmark library;

[0011] 3. Construct a variational autoencoder feature compression network

[0012] Construct a deep variational autoencoder network to constrain the distribution of hidden layer features, remove redundant information between features, standardize features in each dimension, so as to obtain compressed features of standard actions, make the distance metric calculation more reasonable, and improve the scoring accuracy. The variational autoencoder feature compression network is divided into two stages: model training and feature compression. During model training, use convolutional layers to form the encoder of the model, accept the input of multivariate time signal x, and encode the multivariate time signal x to obtain the latent variable z (z is constrained by the KL divergence). The latent variable z contains feature information of flight actions and is used for subsequent quality evaluation. Subsequently, use the decoder composed of transposed convolution to reconstruct the signal using the latent variable z, and then use the mean square error between the signal reconstructed by the latent variable z and the original signal as the loss function. Train the network parameters of the encoder by minimizing the reconstruction error to obtain the encoder with fixed weights;

[0013] After the network training is completed, input the flight parameters into the encoder with fixed weights to obtain hidden variables presenting a normal distribution, so as to realize feature standardization and redundancy removal;

[0014] 4. Adaptive dynamic time warping

[0015] First, process the standard actions in the standard action benchmark library in step 2 through the variational autoencoder feature compression network, and then measure the distance between the flight actions to be evaluated sorted in step 1 and the corresponding standard action compressed features obtained after being processed by the variational autoencoder feature compression network. Use the method of dynamic programming to calculate the distance metric to find the best matching relationship between sequences, allow sequence points to self-replicate and then perform misaligned matching to measure non-equal-length sequences, and be robust to noise. Specifically as follows:

[0016] Suppose two flight time series X = <x1, x2,..., xm >, Y = <y1, y2, …, y n >, and the action durations are m and n respectively. Then the path after dynamic time warping is W = <w1, w2, …, w k , …, w K >. The calculation process for computing the similarity between X and Y is as follows:

[0017] 1) Construct a matrix D of size n × m. The element d ij in the i-th row and j-th column is i = dist(x j ), y

[0018]

[0019] ), where dist is a distance calculation function, usually the Euclidean distance, as shown in Equation (1): 11 to d nm in matrix D. The search process needs to satisfy the following constraints:

[0020] A. Boundary conditions: w1 = (1, 1), w K = (m, n). The speed of any flight action may change, but the start and end times of the actions match each other. The selected path must start from the lower left corner and end at the upper right corner;

[0021] B. Continuity: If w k-1 = (a k-1 , b k-1 ), then the next point w k = (a k , b k ) of the path needs to satisfy (a k - a k-1 ) <= 1 and (b k - b k-1 ) <= 1, that is, it is impossible to skip a point for matching and only adjacent points can be aligned. This can ensure that each coordinate in X and Y appears in W;

[0022] C. Monotonicity: If w k-1 = (a k-1 , b k-1 ), then the next point w k = (a k , b k ) of the path needs to satisfy (a k - a k-1 ) >= 0 and (b k - b k-1 ) >= 0. This restricts the W path to progress monotonically with time to ensure that the matching points do not intersect;

[0023] Combined with the monotonicity and continuity constraints, there are only three direction choices for the path of any point. If the path W has passed through the point (i, j), the next point can only be one of the following three cases: (i + 1, j), (i, j + 1), or (i + 1, j + 1);

[0024] 3) Search for the shortest path from d 11 to d nm in the matrix D as the similarity between the X and Y sequences. Therefore, the total dynamic time warping distance is shown in Equation (2):

[0025]

[0026] where i = 1, 2, …, m; j = 1, 2, …, n, D dtw (X, Y) represents the dynamic time warping distance between the time series X and Y, x i and y j represent the points in the sequences X and Y respectively, d i,j (x i , y j ) represents the Euclidean distance between x i and y j at two points;

[0027] To achieve fast and accurate flight action evaluation, a fixed window is selected to limit the maximum distance of sequence offset. In addition, to suppress the phenomenon that the dynamic time warping distance is too low due to abnormal differences at both ends of the sequence, an offset is selected, which represents the maximum offset allowing abnormal points at both ends of the sequence to be ignored. Finally, for the flight action sequence to be evaluated, the dynamic time warping distance is calculated respectively with each standard flight action in the corresponding category in the standard action benchmark library;

[0028] 5. Global - local information multi - perspective fusion scoring rule

[0029] Based on step 4, calculate the similarity measure between the flight action to be evaluated and the standard action. Then, according to the similarity measure between the flight action to be evaluated and the standard action, considering both global information and local information, form a set of multi - perspective fusion scoring rules based on global - local information, thereby constructing a scoring model, and setting an upper bound and a lower bound of the quality score threshold in the scoring model;

[0030] 6. Apply the scoring model

[0031] Input the flight action to be evaluated into the scoring model, and comprehensively evaluate the flight action to be evaluated through the scoring model.

[0032] Beneficial effects:

[0033] 1. The present invention inputs all the collected flight parameters, considering the influence of all flight parameters on flight quality assessment, without the need to separately extract the flight indicators related to different flight actions;

[0034] 2. The present invention performs redundancy removal processing on the context action information in the entire flight scenario, effectively removing the redundant information between flight parameters;

[0035] 3. The present invention uses the adaptive dynamic time warping distance to measure the distance between the flight actions to be evaluated with different lengths and the standard flight actions;

[0036] 4. The present invention constructs a scoring rule that combines global features and local features, that is, considering the completion of individual actions and combining the performance within the entire flight mission to score, so as to realize the multi-perspective automatic comprehensive evaluation of flight action quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the variational autoencoder feature compression network structure in the preferred embodiment of the present invention.

[0038] Figure 2 It is a schematic diagram of the path search direction in the preferred embodiment of the present invention.

[0039] Figures 3 to 8 It is a schematic diagram of the typical action score distribution after dimensionality reduction (feature dimension = 8) in the preferred embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0040] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below with reference to specific drawings.

[0041] The flight quality adaptive evaluation method specifically comprises the following steps:

[0042] 1. Organize the flight actions to be evaluated

[0043] Since the applicable scenarios and target tasks of different aircraft are different, and their flight action categories are different, according to the tasks in the actual flight mission, the flight process to be evaluated is divided into 10 categories of actions, including roll, ascending roll, half roll reversal, lazy eight, takeoff, up turn, low altitude level flight, down turn, landing and ground taxiing. Among them, if a flight process does not belong to any category of actions, it belongs to other categories. Therefore, the goal is to divide the flight mission into 11 flight states, as shown in Table 1:

[0044] Table 1 Flight action category table to be evaluated

[0045] Flight maneuver Other Roll Ascending roll Half roll and invert Lazy eight Takeoff Label 0 1 2 3 4 5 Flight maneuver Up turn Low-level straight flight Down turn Landing Taxi on the ground Label 6 7 8 9 10

[0046] 2. Establish a standard action benchmark library

[0047] To achieve the quality assessment of actions for different flight categories, it is necessary to select the corresponding standard action benchmarks and establish a standard action benchmark library as a reference benchmark for flight action assessment. The establishment process of the standard action benchmark library is as follows: Combining the flight training syllabus of the specified aircraft type and the expert knowledge of flight experts, select relevant standard flight action entries from rich historical flight data to establish a standard action benchmark library with typical representativeness. The relevant standards include flight parameter standards and three-dimensional trajectory standards for various flight actions. The flight parameter standards include the value-taking rules of flight parameters such as entry speed, exit speed, take-off pitch angle, time, altitude, and intake pressure. The three-dimensional trajectory standard is used to judge whether the three-dimensional trajectory of the flight action conforms to the action description and whether its smoothness meets the requirements;

[0048] According to the above standards, extract several pieces of standard data for each type of flight action after sorting from the flight historical data, record the start and end times of the actions, and incorporate them into the standard action benchmark library;

[0049] 3. Construct a variational autoencoder feature compression network

[0050] When the aircraft enters an action, the flight environment is complex. The flight action parameters monitored by on-board sensors include variables such as spatio-temporal coordinates and state parameters. The variable types are diverse and the variation ranges are different, and the differences between the variables are huge; in addition, when the aircraft executes an action, each collected signal shows the same or opposite change trend, and there is highly redundant information between the feature dimensions. Directly using the original signal for flight quality assessment will produce serious inductive bias and cannot be applied to the actual flight environment. To reduce the impact of the above problems on flight quality assessment, construct a deep variational autoencoder network to constrain the distribution of hidden layer features, remove the redundant information between features, standardize the features of each dimension, make the distance metric calculation more reasonable, and improve the scoring accuracy;

[0051] According to the above variational autoencoder principle, construct a variational autoencoder feature compression network to obtain the compressed features of the standard action. As Figure 1 shown, the variational autoencoder feature compression network is divided into two stages: model training and feature compression. During model training, use the convolutional layer to form the encoder of the model, accept the input of the multivariate time signal x, and encode the multivariate time signal x to obtain the hidden variable z (z is constrained by the KL divergence). The hidden variable z contains the feature information of the flight action and is used for subsequent quality assessment. Subsequently, the decoder composed of transposed convolution uses the hidden variable z to reconstruct the signal, and then uses the mean square error between the reconstructed signal and the original signal as the loss function. Train the network parameters of the encoder by minimizing the reconstruction error to obtain the encoder with fixed weights. The specific parameter settings of the encoder and decoder are shown in Table 2:

[0052] Table 2 Parameter Settings of Variational Autoencoder Feature Compression Network

[0053]

[0054] After the network training is completed, the flight parameters are input into the encoder with fixed weights to obtain hidden variables that present a normal distribution, so as to realize the standardization and redundancy removal of features, which is beneficial to improving the accuracy of subsequent distance measurement;

[0055] 4. Adaptive Dynamic Time Warping

[0056] First, the standard actions in the standard action benchmark library in Step 2 are processed by the variational autoencoder feature compression network, and then the distance between the to-be-evaluated flight actions sorted out in Step 1 and the corresponding standard action compression features obtained after being processed by the variational autoencoder feature compression network is measured to measure the quality of its flight actions. Common distance measurement methods such as Euclidean distance, Mahalanobis distance, and cosine pair distance can calculate the similarity between two time series. The flight data during the flight is affected by various factors such as aircraft performance, pilot operation, and environment. There are random noises and various complex deformations (translation, stretching, discontinuity, etc.) in the flight time series, and its deformation time and deformation degree cannot be predicted. For the time series of the same flight action data, it has a relatively obvious stretching characteristic in the horizontal time length, while the vertical amplitudes are not much different; due to the difference in the duration of the same flight action, it cannot meet the requirement that the measurement objects of common distance measurement methods have the same scale. In this embodiment, the distance is calculated by the dynamic programming method to find the best matching relationship between sequences, allowing sequence points to be self-replicated and then misaligned for matching to measure non-equal-length sequences and being robust to noises. Specifically as follows:

[0057] Suppose two flight time series X = <x1, x2,..., x m >, Y = <y1, y2,..., y n >, and the action durations are m and n respectively. Then the path after using dynamic time programming is W = <w1, w2,..., w k ,..., w K >. The calculation process of calculating the similarity between X and Y is as follows:

[0058] 1) Construct a matrix D of size n×m, and the element d ij = dist(x i , y j ), where dist is the distance calculation function, usually using the Euclidean distance, as shown in Equation (1):

[0059]

[0060] 2) Use dynamic path planning to search from matrix D for the path from d11 to d nm The shortest path, and the search process needs to satisfy the following constraints:

[0061] A. Boundary conditions: w1 = (1, 1), w K = (m, n). The speed of any flight action may change, but the start and end times of the actions match each other. The selected path must start from the lower left corner and end at the upper right corner;

[0062] B. Continuity: If w k-1 = (a k-1 , b k-1 ), then the next point w k = (a k , b k ) needs to satisfy (a k - a k-1 ) <= 1 and (b k - b k-1 ) <= 1, that is, it is impossible to skip a point for matching, and only adjacent points can be aligned, so as to ensure that each coordinate in X and Y appears in W;

[0063] C. Monotonicity: If w k-1 = (a k-1 , b k-1 ), then the next point w k = (a k , b k ) needs to satisfy (a k - a k-1 ) >= 0 and (b k - b k-1 ) >= 0, which restricts that the W path progresses monotonically with time to ensure that the matching points do not intersect;

[0064] Combining the monotonicity and continuity constraints, there are only three direction choices for the path of any point. If the path W has passed through the point (i, j), then the next point can only be one of the following three cases: (i + 1, j), (i, j + 1) or (i + 1, j + 1), as Figure 2 shown;

[0065] 3) Search for the shortest path from d 11 to d nm in the search matrix D as the similarity of the X and Y sequences. Therefore, the total dynamic time warping distance is shown in Equation (2):

[0066]

[0067] where i = 1, 2,..., m; j = 1, 2,..., n, D dtw(X,Y) represents the dynamic time warping distance between time series X and Y, where x i and y j represent points in sequences X and Y respectively, and d i,j (x i , y j ) represents the Euclidean distance between points x i and y j ; the Euclidean distance between two points;

[0068] However, calculating all paths within the entire matrix will result in extremely high time complexity. To achieve fast and accurate flight action evaluation, a fixed window is selected to limit the maximum distance of sequence offset. In addition, to suppress the phenomenon that the dynamic time warping distance is too low due to abnormal differences at both ends of the sequence, an offset is selected, indicating the maximum offset that allows ignoring abnormal points at both ends of the sequence. Finally, for the flight action sequence to be evaluated, the dynamic time warping distance is calculated separately with each standard flight action in the corresponding category in the standard action benchmark library for subsequent flight action quality evaluation;

[0069] 5. Global - Local Information Multi - Perspective Fusion Scoring Rule

[0070] A reasonable scoring algorithm should take into account both the completion of all indicators and the completion of individual specific indicators during flight action quality evaluation. If all indicators of the flight action are above the standard, a weighted sum is obtained based on the completion quality of each indicator to get the overall comprehensive score. However, if a single specific indicator in the flight action does not meet the requirements, the final score should not exceed a fixed score. In this embodiment, both global information and local information are considered to form a set of multi - perspective fusion scoring rules based on global - local information, thereby constructing a scoring model to achieve multi - perspective comprehensive evaluation:

[0071] 5.1 Rule Description

[0072] Suppose a flight action x ∈ R l×n to be evaluated and k standard actions of the same type where l and l i represent the durations of the flight action to be evaluated and each standard flight action respectively, and n is the total number of flight parameters. The flight action to be evaluated and the standard actions of the same type are respectively input into the variational auto - encoder feature compression network to obtain the standardized hidden - layer features and where z is the dimension of the independent features compressed by the variational auto - encoder feature compression network. Then, using the adaptive dynamic time warping algorithm, the dynamic time warping distances d z ={d 1,z , d 2,z , …, d k,z} are calculated on each dimension between the action to be evaluated and each standard action, where Take the minimum value of the calculated k distances as the similarity measure between the action to be evaluated and the standard action benchmark in this dimension, i.e., D z = min(d 1,z , d 2,z , …, d k,z );

[0073] According to the similarity measure between the flight action to be evaluated and the standard action, perform a multi - perspective comprehensive score of global - local information fusion. Select two threshold vectors a and b, and divide the overall score into three intervals: 100 points, 60 - 100 points, and 60 points. Considering the global information, if the distances between the flight action to be evaluated and the standard action benchmark in each feature dimension are all less than the lower threshold a, then the flight action gets 100 points; considering the local information, if there is a distance measure in a certain feature dimension that is greater than the lower threshold, then the action is less than 100 points. On this basis, if the distances in each feature dimension are all less than the upper threshold b, the comprehensive score is determined by the sum of the margins exceeding the lower threshold in each dimension, and it is assumed that the score has a linear relationship with the total margin exceeding the lower threshold in each dimension. The larger the total margin, the lower the score; the smaller the total margin, the higher the score; when the total margin is 0, the score reaches 100 points; if there is a distance measure in a certain feature dimension that is greater than the upper threshold, then the action gets 60 points. The scoring rules are shown in Equation (3):

[0074]

[0075] After simplification, the final flight action quality score can be expressed as Equation (4):

[0076]

[0077] 5.2 Threshold Selection

[0078] As described in 5.1, the quality score is closely related to the selection of the upper threshold and the lower threshold. If the threshold value is too high, the algorithm tends to output a higher score, resulting in a falsely high score; if the threshold value is too low, the algorithm is too strict, leading to a low overall score for the flight action quality. Therefore, a reasonable selection of the upper threshold b and the lower threshold a is crucial for the scoring result;

[0079] First, considering the sparsity of the 100-score actions, the lower bound b of the threshold is determined using the standard action benchmark library. The similarity measures between the standard actions in the standard action benchmark library are calculated using the adaptive dynamic time warping algorithm, and the minimum value among them is selected as the lower bound of the threshold. Since the flight parameter indicators between the standard actions are very close, using the minimum value as the lower bound of the threshold not only ensures that the 100-score actions meet the requirements of the standard actions but also has a high degree of sparsity. Second, the upper bound a of the threshold divides the boundaries of the two interval segments of 60 - 100 points and 60 points. Assuming that various flight actions follow a Gaussian distribution, where the standard flight actions are distributed near the mean of the Gaussian distribution, the maximum value of the similarity measures between the standard actions in the standard action benchmark library is taken as the variance σ of the Gaussian distribution. According to the 3σ principle, the probability of the existence of flight actions beyond 3σ from the mean is almost 0. Therefore, three times the maximum value of the similarity measures between the standard actions is selected as the upper bound a of the threshold to ensure the sparsity of the 60-score flight actions. In summary, the values of the lower bound b and the upper bound a of the threshold are as shown in Equation (5) and are respectively:

[0080]

[0081] Combined with the above content, the multi-perspective fusion evaluation process based on global-local information is shown in Table 3 as follows:

[0082] Table 3 Evaluation process table

[0083] Table 1 Overall process of scoring algorithm

[0084]

[0085] 5.3. Test results

[0086] Since the deep convolutional autoencoder imposes a prior of independent and identically distributed on the hidden layer features, the features after dimensionality reduction by the deep convolutional autoencoder often approximately follow an independent and identically distributed standard normal distribution, which can eliminate the scale differences between dimensions and is beneficial for calculating the distance between features;

[0087] When different dimensionality reduction dimensions are selected, it will affect the training effect of the network. When the dimension is too low, feature information is missing and the original signal cannot be effectively reconstructed; when the dimension is too high, redundancy inevitably appears between dimensions, breaking the independence between different features and affecting the performance of the scoring model. Therefore, the overall distribution of scores obtained under different feature dimensions was tested to achieve hyperparameter optimization. By comparing the score distribution results under 5 groups of feature dimensions [5, 6, 8, 10, 12], when the feature dimension is greater than or equal to 10, the scores of some actions are significantly inclined to the low score segment, indicating that the variational autoencoder feature compression network will produce some invalid dimensions at this time; when the feature dimension is less than or equal to 6, the scores of some actions are significantly inclined to the high score segment, indicating that the variational autoencoder feature compression network is difficult to converge at this time and the discrimination of features is insufficient. Therefore, the final scoring model selects a feature dimension of 8, as Figure 3 shown, and then a scoring model that meets the standards is obtained;

[0088] 6. Application of the scoring model

[0089] Input the flight action to be evaluated into the scoring model, and comprehensively evaluate the flight action to be evaluated through the scoring model.

Claims

1. Adaptive evaluation method for flight quality, characterized in that, The specific steps are as follows:

1. Organize the flight actions to be evaluated According to the tasks in the actual flight sorties, divide the flight process to be evaluated into 11 flight states; 2. Establish a standard action benchmark library Combined with the flight training syllabus of the specified aircraft type and the experience and knowledge of flight experts, select relevant standard flight action entries from historical flight data to establish a standard action benchmark library with typical representativeness; 3. Construct a variational autoencoder feature compression network Construct a deep variational autoencoder network to constrain the distribution of hidden layer features in order to obtain the compressed features of standard actions; The variational autoencoder feature compression network is divided into two stages: model training and feature compression. During model training, use convolutional layers to form the encoder of the model, accept the input of the multivariate time signal x, and encode the multivariate time signal x to obtain the latent variable z. The latent variable z contains the feature information of the flight action. Then, use the decoder composed of transposed convolution to reconstruct the signal using the latent variable z, and use the mean square error between the signal reconstructed by the latent variable z and the original signal as the loss function. Train the network parameters of the encoder by minimizing the reconstruction error to obtain the encoder with fixed weights; After the network training is completed, input the flight parameters into the encoder with fixed weights to obtain the hidden variables presenting a normal distribution; 4. Adaptive dynamic time warping First, process the standard actions in the standard action benchmark library in step 2 through the variational autoencoder feature compression network, and then calculate the distance metric between the flight actions to be evaluated sorted out in step 1 and the corresponding standard action compressed features obtained after being processed by the variational autoencoder feature compression network using the dynamic programming method; 5. Global-local information multi-perspective fusion scoring rule On the basis of step 4, calculate the similarity metric between the flight actions to be evaluated and the standard actions, and then, according to the similarity metric between the flight actions to be evaluated and the standard actions, consider both global information and local information to form a set of multi-perspective fusion scoring rules based on global-local information, thereby constructing a scoring model, and set an upper threshold and a lower threshold for quality scoring in the scoring model; 6. Apply the scoring model Input the flight actions to be evaluated into the scoring model, and comprehensively evaluate the flight actions to be evaluated through the scoring model.

2. The adaptive evaluation method for flight quality according to claim 1, characterized in that, In step 1, the 11 flight states include roll, ascending roll, half roll reverse, lazy eight, takeoff, up turn, low altitude level flight, down turn, landing, ground taxiing, and other categories not suitable for any type of action.

3. The adaptive evaluation method for flight quality according to claim 1, characterized in that, In step 2, the relevant standards include the flight parameter standards and three-dimensional trajectory standards for various flight actions.

4. The adaptive evaluation method for flight quality according to claim 3, characterized in that, The flight parameter standards include the value-taking rules of flight parameters. The three-dimensional trajectory standards are used to judge whether the three-dimensional trajectory of the flight action conforms to the action description and whether its smoothness meets the requirements.

5. The adaptive evaluation method for flight quality according to claim 1, characterized in that, In step 4, the calculation of the distance metric is specifically as follows: Assume the flight action to be evaluated and two flight time series X = <x1,x2,…,x m >,Y= <y1,y2,…,y n >, the action durations are m and n respectively, then the path after dynamic time planning is W = <w1,w2,…,w k ,…,w K >, the calculation process of calculating the similarity between X and Y is as follows: 1) Construct a matrix D of size n×m, where the element d in the i-th row and j-th column ij = dist(x i , y j ), where dist is a distance calculation function that uses the Euclidean distance, as shown in Equation (1): 2) Search for the shortest path from d to d in matrix D using dynamic path planning. The search process must satisfy the following constraints: 11 to d nm The search process must satisfy the following constraints: A. Boundary conditions: w1 = (1, 1), w K = (m, n). The speed of any flight action may vary, but the start and end times of the actions match each other. The selected path must start from the lower left corner and end at the upper right corner; B. Continuity: If w k-1 =(a k-1 , b k-1 ), then the next point w k =(a k , b k ) needs to satisfy (a k -a k-1 ) <= 1 and (b k -b k-1 ) <= 1, that is, it is impossible to skip a point for matching and only adjacent points can be aligned, which can ensure that each coordinate in X and Y appears in W; C. Monotonicity: If w k-1 =(a k-1 , b k-1 ), then the next point w k =(a k , b k ) needs to satisfy (a k -a k-1 ) >= 0 and (b k -b k-1 ) >= 0, which restricts the W path to progress monotonically over time to ensure that the matching points do not intersect; Combined with monotonicity and continuity constraints, there are only three direction choices for the path of any point. If the path W has passed through the point (i, j), the next point can only be one of the following three cases: (i + 1, j), (i, j + 1) or (i + 1, j + 1); 3) Search for the shortest path in matrix D from d 11 to d nm as the similarity between X and Y sequences. Therefore, the total dynamic time warping distance is shown in Equation (2): where \(i = 1, 2, \ldots, m\); \(j = 1, 2, \ldots, n\), \(D\) dtw \((X, Y)\) represents the dynamic time warping distance between time series \(X\) and \(Y\), \(x\) i and \(y\) j represent points in sequences \(X\) and \(Y\) respectively, \(d\) i,j \((x\) i , y\) j ) represents the Euclidean distance between \(x\) i and \(y\) j at two points.

Images