DTW and K-means clustering-based flight training quality evaluation method

Abstract

The invention discloses a flight training quality evaluation method based on DTW and K-means clustering, and belongs to the technical field of pilot training, and the method comprises the following steps: S1, data collection and processing; s2, data preprocessing and standard template generation; s3, DTW similarity measurement and mahalanobis distance weighting are carried out; s4, performing similarity measurement and clustering analysis; s5, generating a personalized training scheme; therefore, the invention provides a scientific, objective and efficient flight trainee quality evaluation method, the landing process of the pilot is subjected to comprehensive quantitative analysis in a data-driven manner, a more accurate and personalized flight training scheme is provided, the flight level of the flight trainee is finally improved, and the risk of flight accidents is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of pilot training, and in particular to a flight training quality evaluation method based on DTW and K-means clustering. Background Art

[0002] In recent years, with the continuous expansion of airlines and the increase in the number of flights, the demand for pilots, especially high-quality pilots, is showing an increasing trend. The flight student period is a critical period for the formation of flight technical capabilities. The training received at this stage directly affects the stability and safety of their subsequent flight operations. Therefore, in the new environment, a scientific and objective flight student quality assessment method has been developed to more accurately measure the students' flight technical level. At the same time, providing personalized flight technical ability improvement suggestions for different students can effectively help students overcome weak links in the flight training process and ensure that they have stronger flying quality. Traditional flight training evaluation methods mainly rely on the experience and subjective judgment of flight instructors. The evaluation standards are relatively simple and cannot fully reflect the students' actual flying capabilities.

[0003] With the gradual improvement of flight training equipment, the flight quality assessment of flight trainees has become an important part of improving flight training effectiveness and ensuring flight safety. However, traditional flight quality assessment methods often rely on subjective judgment, which makes it difficult to accurately reflect the pilot's actual flight performance. Especially in complex flight data analysis, the evaluation criteria are relatively vague and lack objectivity. In addition, existing assessment methods often only focus on a single dimension of flight parameters and cannot comprehensively evaluate the pilot's comprehensive flight ability. Flight data is essentially a multidimensional data with strong time series, involving multiple parameters such as altitude, speed, pitch angle, vertical speed, etc. These parameters are not only strongly correlated, but also have nonlinear time series changes.

[0004] At present, the research on flight quality assessment based on time series analysis is still in the exploratory stage, especially in terms of how to effectively measure the similarity of flight data and evaluate the landing quality of flight trainees. Relevant research results are relatively scarce. In order to overcome the above problems, the present invention proposes a multidimensional time series similarity measurement method based on the dynamic time warping (DTW) algorithm combined with the Mahalanobis distance. This method uses the flight data of flight trainees and flight instructors for similarity comparison, and uses machine learning algorithms to objectively evaluate flight quality. It can extract time-series features from the flight data, thereby achieving accurate landing quality assessment of flight trainees.

[0005] Therefore, the purpose of the present invention is to develop a scientific, objective and efficient method for evaluating the quality of flight trainees, which uses a data-driven approach to conduct a comprehensive quantitative analysis of the pilot's landing process, provide a more accurate and personalized flight training program, and ultimately improve the flight level of flight trainees and reduce the risk of flight accidents. Summary of the Invention

[0006] The present invention aims to provide a flight training quality assessment method based on DTW and K-means clustering to solve the problems raised in the above background technology.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A flight training quality evaluation method based on DTW and K-means clustering includes the following steps:

[0009] S1. Data collection and processing: The flight data used comes from the Garmin G1000 integrated avionics system. The recorded flight data includes multiple parameters such as indicated airspeed, flight altitude, vertical speed, pitch angle, and track angle.

[0010] At the same time, the Lagrange interpolation method is used to interpolate data to ensure the integrity and accuracy of the data;

[0011] S2, data preprocessing and standard template generation, all flight trainees’ flight data are subjected to time series modeling through the Transformer encoder;

[0012] After being processed by the encoder, all flight instructors’ flight data will be converted into a standard multi-dimensional time series template;

[0013] By averaging and pooling all flight data, a standard time series is obtained, which serves as a benchmark template for subsequent flight data of flight trainees;

[0014] S3,DTW similarity metric and Mahalanobis distance weighting,In order to quantify the similarity between the flight data of flight students and flight instructors, the dynamic time warping algorithm, namely, DTW algorithm, is used;

[0015] The DTW algorithm can flexibly perform nonlinear alignment on time series, overcoming the limitation of traditional methods that strictly align the time axis;

[0016] Due to the multidimensionality of flight data and the strong correlation between parameters, this method introduces the covariance matrix based on DTW and combines it with Mahalanobis distance. By considering the correlation between various flight parameters, the accuracy of the measurement results can be improved.

[0017] S4. Similarity measurement and cluster analysis. After calculating the similarity between the flight trainee's landing time series and the standard template sequence using DTW and Mahalanobis distance, the K-means clustering algorithm is used to perform cluster analysis on the trainee's similarity measurement results.

[0018] Finally, the flight trainees' flying qualities are divided into five levels, including excellent, good, average, qualified, and unqualified;

[0019] S5. Generate personalized training plans, and provide personalized training suggestions based on each student's weak points during the landing process.

[0020] Preferably, the specific implementation steps of step S2 are as follows:

[0021] S201, perform Transformer encoder modeling to convert each time series into a matrix of dimension T×N, where T is the time step and N is the flight parameter dimension;

[0022] Set the number of attention heads H and the hidden layer dimension D, and calculate the query matrix Q h , key matrix K h , value matrix V h , the formula is as follows:

[0023]

[0024] S202, then calculate the output of each head by scaling the dot product attention, the formula is as follows:

[0025]

[0026] S203. Concatenate all attention head outputs and map them through a linear layer. The formula is as follows:

[0027] MultiHead(Q,K,V)=Concat(head1,head2,…,head H )W O

[0028] S204, the encoded time series Perform average pooling operation:

[0029]

[0030] S205, map the high-dimensional features back to the original dimensional space through the fully connected layer:

[0031] Y t′ =W′+b

[0032] S206. Finally, a standard multi-dimensional time series with a shape of T×N is obtained, which can be used as a benchmark template for subsequent DTW similarity measurement.

[0033] Preferably, the specific implementation steps of step S3 are as follows:

[0034] S301. Based on the multidimensional time series data of the flight instructor, calculate its covariance matrix S and find its inverse matrix S -1 Then define the student's multidimensional time series X and the standard template time series Y. i =(x 1i ,x 2i ,…,x di ) T represents the i-th column of X, that is, the value of each parameter at the i-th moment, Y j =(y 1j ,y 2j ,…,y dj ) T represents the jth column of Y, that is, the value of each parameter at the jth moment. Then the local distance between two sequences based on Mahalanobis distance in DTW is:

[0035]

[0036] S302. The DTW distance between the multidimensional time series X and Y based on the Mahalanobis distance can be:

[0037] D M (x,y)=d(X i ,Y j )+min{D(x-1,y),D(x,y-1),D(x-1,y-1)}

[0038] This distance represents the gap between each student's multidimensional time series and the standard multidimensional time series generated by excellent instructors. The smaller the distance value, the smaller the difference between the standard series and the flight student's landing level is.

[0039] S303. Finally, calculate the contribution of each parameter to the DTW distance by calculating the difference between each pair of matching points, squaring the difference, and summing up the contributions of all time points to obtain the total contribution of each flight parameter. Then, normalize the contribution of each flight parameter so that they can be compared on the same scale. Finally, find the student's weakest flight parameter and, combined with the visualization curve, provide targeted training suggestions.

[0040] Preferably, step S4 uses K-means cluster analysis to classify the flight quality of the flight trainees to form an accurate landing level assessment. The specific implementation method is as follows:

[0041] K-means clustering analysis is performed based on the similarity metric. The most important parameter of the K-means clustering algorithm is the K value. Before clustering the student data, the elbow criterion method is used to select the optimal cluster number K value.

[0042] Based on the relationship curve between the WCSS value and the K value of the elbow criterion method, an inflection point on the curve is found, that is, the point where the inertia reduction brought about by increasing the number of clusters begins to slow down, which is taken as the optimal cluster number K. Then, cluster analysis is performed based on the similarity measurement value, and the landing level of the flight trainees is divided into K categories to complete the overall evaluation of the landing level of the flight trainees.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. Improve the objectivity and accuracy of assessments: Traditional flight student assessment methods often rely on subjective judgment. However, this invention uses the DTW time series similarity measurement method based on the Mahalanobis distance to objectively and quantitatively evaluate the similarity between flight student data and flight instructor standard data. This invention avoids interference from human factors, making flight student assessment results more accurate and reliable.

[0045] 2. Comprehensive Flight Quality Assessment: By using multiple flight parameters (such as indicated airspeed, altitude, vertical speed, pitch angle, and track angle) to evaluate landing quality, this system comprehensively considers the performance of each key parameter during flight, providing a comprehensive reflection of the student's flying proficiency. This comprehensive assessment helps identify potential flight performance issues and provides more detailed improvement suggestions.

[0046] 3. Provide personalized training plans: By identifying students' weaknesses, the present invention can provide each student with a personalized training plan. For example, if a student's vertical speed is too high, the system will provide corresponding airspeed and pitch control suggestions and conduct targeted training. This personalized plan can effectively improve students' flying skills and help them overcome technical difficulties during flight.

[0047] 4. Improve flight training efficiency: Through data-driven flight quality assessment methods, flight instructors can more accurately grasp students' flight conditions, rationally arrange training content, avoid repetitive training or waste of resources, and thus improve the efficiency and effectiveness of flight training.

[0048] 5. Improved flight safety: The assessment of a student's landing quality is directly related to flight safety. The technical solution of this invention can promptly identify risk factors during flight and proactively correct them, reducing the probability of flight accidents and enhancing the safety of flight training. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of a flight training quality assessment method based on DTW and K-means clustering; DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0051] like Figure 1 As shown in Figure 1, a flight training quality evaluation method based on DTW and K-means clustering is shown in Figure 2. The steps are as follows:

[0052] 1. Data Collection and Processing. The flight data used in this paper comes from the Garmin G1000 integrated avionics system. The recorded flight data includes multiple parameters such as indicated airspeed, altitude, vertical speed, pitch angle, and track angle. These parameters are recorded at a frequency of 1 Hz, and the data volume is large, sparse, and sometimes missing. To address the missing data issue, this paper uses Lagrange interpolation to ensure data integrity and accuracy.

[0053] 2. Data preprocessing and standard template generation: All flight trainee flight data is passed through a Transformer encoder for time series modeling. The Transformer model uses a self-attention mechanism to extract time series features from the flight data and learn the dynamic changes in the flight data. After processing by the encoder, all flight instructor flight data is converted into a standard multidimensional time series template. By performing average pooling on all flight data, a standard time series is generated, which serves as a baseline template for subsequent flight trainee flight data.

[0054] 3. DTW similarity metric and Mahalanobis distance weighting. To quantify the similarity between flight trainee and flight instructor flight data, the present invention employs the Dynamic Time Warping (DTW) algorithm. The DTW algorithm can flexibly perform nonlinear alignment of time series, overcoming the limitations of traditional methods that strictly align time axes. Due to the multidimensionality of flight data and the strong correlation between parameters, the present invention combines DTW with the Mahalanobis distance. This method introduces the covariance matrix and improves the accuracy of the measurement results by considering the correlation between various flight parameters.

[0055] 4. Similarity Measurement and Cluster Analysis: After calculating the similarity between the pilot trainee's landing time series and the standard template sequence using DTW and Mahalanobis distance, the K-means clustering algorithm is used to perform cluster analysis on the pilot's similarity measurement results. Ultimately, the pilot's flight performance is classified into five levels (excellent, good, fair, qualified, and unqualified). This clustering method not only quantifies the pilot's flying level but also supports subsequent personalized training plans.

[0056] 5. Personalized training plan generation: This system can provide personalized training recommendations based on each student's weaknesses during landing. For example, if a student improperly controls vertical speed during landing, resulting in excessive impact force, the student will be recommended to focus on airspeed and pitch angle control for approximately 15 seconds before touchdown and receive relevant training.

[0057] The technical points in the above method include:

[0058] 1. Transformer Encoder-Based Flight Data Standard Template Generation Method: By modeling the time series of flight instructor data, a standard multi-dimensional flight time series template is generated, providing a benchmark for subsequent similarity measurement. The specific implementation steps and details are as follows.

[0059] Collect the landing phase flight data (indicated airspeed, altitude, vertical speed, pitch angle, track angle) of excellent flight instructors, with a time span from the maximum pitch angle to the maximum elevation angle, and a sampling frequency of 1Hz. Perform Lagrange interpolation on the original data to fill in missing values ​​and ensure the continuity of the time series. Then perform Transformer encoder modeling to convert each time series into a matrix of dimension T×N (T is the time step and N is the flight parameter dimension). Set the number of attention heads H and the hidden layer dimension D, and calculate the query matrix Q h , key matrix K h , value matrix V h .

[0060]

[0061] The output of each head is then calculated using scaled dot product attention:

[0062]

[0063] Concatenate all attention head outputs and map them through a linear layer:

[0064] MultiHead(Q,K,V)=Concat(head1,head2,…,head H )W O

[0065] The encoded time series Perform average pooling operation:

[0066]

[0067] Map the high-dimensional features back to the original dimensional space through the fully connected layer:

[0068] Y t′ =W′+b

[0069] Finally, a standard multidimensional time series with a shape of T×N is obtained, which can be used as a benchmark template for subsequent DTW similarity measurement.

[0070] 2. DTW Multidimensional Time Series Similarity Measurement Method Based on Mahalanobis Distance: This method combines DTW and Mahalanobis distance techniques to objectively evaluate student pilots' flight performance through multidimensional time series similarity measurement. The specific implementation steps and details are as follows.

[0071] First, based on the multidimensional time series data of the flight instructor, the covariance matrix S is calculated, and its inverse matrix S is obtained. -1 Then define the student's multidimensional time series X and the standard template time series Y. i =(x 1i ,x 2i ,…,x di ) T represents the i-th column of X, that is, the value of each parameter at the i-th moment, Y j =(y 1j ,y 2j ,…,y dj ) T represents the jth column of Y, that is, the value of each parameter at the jth moment. Then the local distance between two sequences based on Mahalanobis distance in DTW is:

[0072]

[0073] Finally, the DTW distance between the multidimensional time series X and Y based on the Mahalanobis distance is:

[0074] D M (x,y)=d(X i ,Y j )+min{D(x-1,y),D(x,y-1),D(x-1,y-1)}

[0075] This distance represents the gap between each student's multidimensional time series and the standard multidimensional time series generated by excellent instructors. The smaller the distance value, the smaller the difference between it and the standard series, and the better the landing level of the flight student.

[0076] Finally, the contribution of each parameter to the DTW distance is calculated. The difference between each pair of matching points is calculated and squared. The contribution of all time points is summed up to obtain the total contribution of each flight parameter. The contribution of each flight parameter is then normalized so that they can be compared on the same scale. Finally, the student's weakest flight parameter is found, and combined with the visualization curve, targeted training suggestions are given.

[0077] 3. Pilot student similarity cluster analysis method: Using K-means cluster analysis, we grade pilot student flight performance and provide an accurate landing performance assessment. The specific implementation steps and details are as follows.

[0078] K-means cluster analysis was performed based on the similarity metric. The most important parameter in the K-means clustering algorithm is the K value. Before clustering the trainee data, the elbow criterion was used to select the optimal number of clusters, K. Based on the curve of the relationship between the WCSS value and the K value obtained using the elbow criterion, an inflection point was found on the curve where the reduction in inertia with increasing the number of clusters began to slow down, defining it as the optimal number of clusters, K. Cluster analysis was then performed based on the similarity metric, classifying the flight trainees' landing performance into K categories, completing a comprehensive assessment of their landing performance.

[0079] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A flight training quality assessment method based on DTW and K-means clustering, characterized by: The method comprises the following steps: S1. Data collection and processing. The flight data used comes from the Garmin G1000 integrated avionics system. The recorded flight data includes multiple parameters such as indicated airspeed, flight altitude, vertical speed, pitch angle, and track angle. The time span is from the maximum pitch angle to the maximum elevation angle, and the sampling frequency is 1 Hz. At the same time, Lagrange interpolation is performed on the original data to fill in missing values ​​and ensure the continuity of the time series; S2, data preprocessing and standard template generation, all flight trainees’ flight data are subjected to time series modeling through the Transformer encoder; After being processed by the encoder, all flight instructors’ flight data will be converted into a standard multi-dimensional time series template; By averaging and pooling all flight data, a standard time series is obtained, which serves as a benchmark template for subsequent flight data of flight trainees; S3,DTW similarity metric and Mahalanobis distance weighting,In order to quantify the similarity between the flight data of flight students and flight instructors, the dynamic time warping algorithm, namely, DTW algorithm, is used; The DTW algorithm can flexibly perform nonlinear alignment on time series, overcoming the limitation of traditional methods that strictly align the time axis; Due to the multidimensionality of flight data and the strong correlation between parameters, this method introduces the covariance matrix based on DTW and combines it with Mahalanobis distance. By considering the correlation between various flight parameters, the accuracy of the measurement results can be improved. S4. Similarity measurement and cluster analysis. After calculating the similarity between the flight trainee's landing time series and the standard template sequence using DTW and Mahalanobis distance, the K-means clustering algorithm is used to perform cluster analysis on the trainee's similarity measurement results. Finally, the flight trainees' flying qualities are divided into five levels, including excellent, good, average, qualified, and unqualified; S5. Generate personalized training plans, and provide personalized training suggestions based on each student's weak points during the landing process.

2. The flight training quality evaluation method based on DTW and K-means clustering according to claim 1, characterized in that: The specific implementation steps of step S2 are as follows: S201, perform Transformer encoder modeling to convert each time series into a matrix of dimension T×N, where T is the time step and N is the flight parameter dimension; Set the number of attention heads H and the hidden layer dimension D, and calculate the query matrix Q h , key matrix K h , value matrix V h , the formula is as follows: S202, then calculate the output of each head by scaling the dot product attention, the formula is as follows: S203. Concatenate all attention head outputs and map them through a linear layer. The formula is as follows: MultiHead(Q,K,V)=Concat(head1,head2,…,head H )W O S204, the encoded time series Perform average pooling operation: S205, map the high-dimensional features back to the original dimensional space through the fully connected layer: Y t′ =W′+b S206. Finally, a standard multi-dimensional time series with a shape of T×N is obtained, which can be used as a benchmark template for subsequent DTW similarity measurement.

3. The flight training quality evaluation method based on DTW and K-means clustering according to claim 1, characterized in that: The specific implementation steps of step S3 are as follows: S301. Based on the multidimensional time series data of the flight instructor, calculate its covariance matrix S and find its inverse matrix S -1 Then define the student's multidimensional time series X and the standard template time series Y. i =(x 1i ,x 2i ,…,x di ) T represents the i-th column of X, that is, the value of each parameter at the i-th moment, Y j =(y 1j ,y 2j ,…,y dj ) T represents the jth column of Y, that is, the value of each parameter at the jth moment. Then the local distance between two sequences based on Mahalanobis distance in DTW is: S302. The DTW distance between the multidimensional time series X and Y based on the Mahalanobis distance can be: D M (x,y)=d(X i ,Y j )+min{D(x-1,y),D(x,y-1),D(x-1,y-1)} This distance represents the gap between each student's multidimensional time series and the standard multidimensional time series generated by excellent instructors. The smaller the distance value, the smaller the difference between the standard series and the flight student's landing level is. S303. Finally, calculate the contribution of each parameter to the DTW distance by calculating the difference between each pair of matching points, squaring the difference, and summing up the contributions of all time points to obtain the total contribution of each flight parameter. Then, normalize the contribution of each flight parameter so that they can be compared on the same scale. Finally, find the student's weakest flight parameter and, combined with the visualization curve, provide targeted training suggestions.

4. The flight training quality evaluation method based on DTW and K-means clustering according to claim 1, characterized in that: Step S4 uses K-means cluster analysis to classify the flight quality of the flight trainees and form an accurate landing level assessment. The specific implementation method is as follows: K-means clustering analysis is performed based on the similarity metric. The most important parameter of the K-means clustering algorithm is the K value. Before clustering the student data, the elbow criterion method is used to select the optimal cluster number K value. Based on the relationship curve between the WCSS value and the K value of the elbow criterion method, an inflection point on the curve is found, that is, the point where the inertia reduction brought about by increasing the number of clusters begins to slow down, which is taken as the optimal cluster number K. Then, cluster analysis is performed based on the similarity measurement value, and the landing level of the flight trainees is divided into K categories to complete the overall evaluation of the landing level of the flight trainees.

Images