A symbolic classification turbulence detection method based on aircraft onboard data

Abstract

A symbolic classification method for turbulence detection based on aircraft airborne data is disclosed. The method includes steps such as acquiring airborne QAR data, processing it, and dividing it into training and test sets; obtaining labeled data; selecting a suitable symbol set; constructing a symbolic classification model based on genetic programming; training the model using the labeled data and symbol set; and outputting turbulence detection results from the trained model. This invention achieves good model accuracy and good results in data detection independent of the training data. The process is cost-effective and scientifically sound. Using genetic programming, the entire process is relatively intuitive, avoiding the "black box" drawbacks of traditional machine learning. Turbulence anomaly information is directly obtained from airborne QAR data, and the computer autonomously learns the correlation between QAR data and turbulence anomaly information, achieving relatively accurate acquisition of turbulence anomaly information without the need for manually defined calculation methods. This method demonstrates innovation and broad application prospects.

Description

Technical Field

[0001] This invention belongs to the field of aviation meteorology technology, and specifically relates to a symbol classification turbulence detection method based on aircraft airborne data. Background Technology

[0002] With global warming and the development of the global civil aviation industry, civil flights will encounter turbulence more frequently. Turbulence not only affects flight comfort but also poses a threat to flight safety, thus attracting more attention in recent years. Traditional methods for detecting turbulence by calculating eddy current dissipation rate (EDR) are limited in their widespread application due to complex algorithms, numerous parameters, and the fact that accurate calculation methods are monopolized by a few institutions. Summary of the Invention

[0003] To address the aforementioned problems, the present invention aims to provide a symbolic classification turbulence detection method based on aircraft airborne data.

[0004] To achieve the above objectives, the symbolic classification turbulence detection method based on aircraft airborne data provided by the present invention includes the following steps performed in sequence:

[0005] 1) Acquire airborne QAR data with historical turbulence anomaly information, then extract data related to turbulence occurrence and normalize it, and then divide the processed data into training set and test set according to the proportion;

[0006] 2) Use the eddy current dissipation rate to classify and label the processed data in the training set to obtain labeled data;

[0007] 3) Select a suitable symbol set;

[0008] 4) Construct a symbol classification model based on genetic programming, and then use the labeled data and symbol set mentioned above to train the symbol classification model based on genetic programming to obtain the trained symbol classification model based on genetic programming.

[0009] 5) Input the processed data from the test set or the airborne QAR data to be detected, which is acquired in real time by the airborne computer and processed according to the method in step 1), into the trained genetic programming-based symbolic classification model. The output of the model is the turbulence detection result.

[0010] In step 1), the method for acquiring airborne QAR data with historical turbulence anomaly information, extracting data related to turbulence occurrence, normalizing the data, and then dividing the processed data into training and testing sets proportionally is as follows:

[0011] The processed data contains data for 24 normalized parameters, namely: left angle of attack, right angle of attack, flight number, total weight, altitude, latitude, longitude, radio altitude, total air temperature, static air temperature, aircraft display heading, angle of attack (drift angle), pitch angle, roll angle, wind direction, ground speed, indicated airspeed, Mach number, vacuum speed, wind speed, vertical acceleration, lateral acceleration, longitudinal acceleration, and vertical velocity.

[0012] In step 2), the method for classifying and labeling the processed data in the training set using eddy current dissipation rate to obtain labeled data is as follows:

[0013] Data with an eddy current dissipation rate greater than or equal to 0.1 is marked as 1, indicating that turbulence has occurred; the rest of the processed data is marked as 0, indicating that no turbulence has occurred.

[0014] In step 3), the symbol set contains 7 operators: addition, subtraction, multiplication, division, square root, logarithmic function, and absolute value.

[0015] In step 4), the method for constructing a genetic programming-based symbol classification model and then training the genetic programming-based symbol classification model using the aforementioned labeled data and symbol set to obtain the trained genetic programming-based symbol classification model is as follows:

[0016] The labeled data is used as the terminal set of the model, and the symbol set is used as the function set. The terminal set and function set are used as input data to train a genetic programming-based symbolic classification model. The model randomly generates a given number of formulas from the terminal set and function set to express the relationship between the input and output data, and selects the most suitable formula for crossover, replication, and mutation operations to generate the next generation of individuals. During training, the loss function of each generation of individuals is calculated and a new generation is continuously generated to continuously improve the model's accuracy in turbulence detection until the set number of iterations is reached, resulting in a trained genetic programming-based symbolic classification model for turbulence detection.

[0017] The loss function is a logarithmic loss function. This model can select the best individual based on the magnitude of the logarithmic loss function of the individual on the input data, as shown in the following expression:

[0018]

[0019] Among them, y i Input data x i The true category, p i To predict input data x i The probability that the input category belongs to 1.

[0020] In step 5), the method of inputting the processed data from the test set or the airborne QAR data to be detected, acquired in real time by the airborne computer and processed according to the method in step 1), into the trained genetic programming-based symbolic classification model, and the model output being the turbulence detection result, is as follows:

[0021] The trained genetic programming-based symbolic classification model will determine whether turbulence occurs at each data point, marking the location where turbulence is considered to occur as 1, and otherwise as 0, and then output the turbulence detection result.

[0022] The symbol classification turbulence detection method based on aircraft airborne data provided by this invention has the following beneficial effects:

[0023] 1. It has good model accuracy and achieves good results in data detection independent of the training data.

[0024] 2. Low process costs and a scientific system.

[0025] 3. Using genetic programming makes the whole process more intuitive and avoids the shortcomings of the traditional machine learning 'black box'.

[0026] 4. Turbulence anomaly information can be obtained directly from airborne QAR data. The computer can autonomously learn the correlation between QAR data and turbulence anomaly information. It can obtain turbulence anomaly information relatively accurately without the need for manual calculation methods. It has certain innovation and broad application prospects. Attached Figure Description

[0027] Figure 1 This is a flowchart of a symbol classification turbulence detection method based on aircraft onboard data.

[0028] Figure 2 This is a schematic diagram illustrating the training iteration of a symbolic classification model based on genetic programming.

[0029] Figure 3 The diagram shows the training results of a symbol classification model based on genetic programming. (a) shows the model's accuracy on the training set, (b) shows the iterative process of the symbol classification model (optimal individual), (c) shows the model's accuracy on the test set, and (d) shows the iterative process of the symbol classification model (population average).

[0030] Figure 4 This is a diagram illustrating the mutation. Detailed Implementation

[0031] The present invention will be further described below with reference to specific embodiments.

[0032] like Figure 1As shown, the symbol classification turbulence detection method based on aircraft airborne data provided in this embodiment includes the following steps performed in sequence:

[0033] 1) Acquire airborne QAR data with historical turbulence anomaly information, then extract data related to turbulence occurrence and normalize it, and then divide the processed data into training set and test set according to the proportion;

[0034] Since airborne QAR data contains thousands of aircraft-related parameters, most of which are significantly unrelated to turbulence, it is necessary to delete data for parameters significantly unrelated to turbulence occurrence and retain data for parameters related to turbulence occurrence to avoid redundant information interfering with model judgments. Furthermore, because different parameters in the airborne QAR data have different scales and units, normalization processing is required to eliminate the influence of different dimensions between parameters. The processed data contains data for 24 normalized parameters, as detailed in Table 2.

[0035] In this embodiment, a total of 26,069 airborne QAR data points were acquired from three flights and preprocessed. After normalization, 11,140 processed data points were obtained. The processed data from one flight was used as the training set, and the processed data from the other two flights were used as the test set, with a training set to test set ratio of 7:3.

[0036] 2) Use the eddy current dissipation rate to classify and label the processed data in the training set to obtain labeled data;

[0037] In the aviation field, eddy dissipation rate is widely used to assess the size of eddies generated by an aircraft in ambient atmospheric turbulence. It can be used to reflect whether the aircraft has encountered turbulence. Therefore, eddy dissipation rate can be used as a criterion for classifying and labeling training data.

[0038] Data with an eddy current dissipation rate greater than or equal to 0.1 is marked as 1, indicating that turbulence has occurred; the rest of the processed data is marked as 0, indicating that no turbulence has occurred.

[0039] 3) Select a suitable symbol set;

[0040] The symbol set contains seven operators: addition, subtraction, multiplication, division, square root, logarithmic function, and absolute value.

[0041] 4) Construct a symbol classification model based on genetic programming, and then use the labeled data and symbol set mentioned above to train the symbol classification model based on genetic programming to obtain the trained symbol classification model based on genetic programming.

[0042] The symbolic classification model based on genetic programming is a machine learning algorithm that uses operators to describe the functional relationships between input variables in order to predict the values ​​of output variables.

[0043] like Figure 2 As shown, the labeled data is used as the terminal set of the model, and the symbol set is used as the function set. The terminal set and function set are used as input data to train a genetic programming-based symbol classification model. The model randomly generates a given number of formulas from the terminal set and function set to express the relationship between the input and output data, and selects the most suitable formula for crossover, replication, and mutation operations to generate the next generation of individuals. During training, the loss function of each generation of individuals is calculated and a new generation is continuously generated to continuously improve the model's accuracy in turbulence detection until the set number of iterations is reached, resulting in a trained genetic programming-based symbol classification model for turbulence detection.

[0044] The loss function is a logarithmic loss function. This model can select the best individual based on the magnitude of the logarithmic loss function of the individual on the input data, as shown in the following expression:

[0045]

[0046] Among them, y i Input data x i The true category, p i To predict input data x i The probability that the input category belongs to 1.

[0047] In this embodiment, the initial number of generated individuals is set to 5000, the number of iterations is 30, the number of competitive winners is 20, the range of the terminal set constant is (-1.0, 1.0), and the total probability of mutation is 0.03 (the three mutation probabilities are each 0.01).

[0048] Figure 3 This is a diagram showing the training results of a symbolic classification model based on genetic programming. Figure 4 This is a diagram illustrating the mutation.

[0049] 5) Input the processed data from the test set or the airborne QAR data to be detected, which is acquired in real time by the airborne computer and processed according to the method in step 1), into the trained genetic programming-based symbolic classification model. The output of the model is the turbulence detection result.

[0050] The trained genetic programming-based symbolic classification model will determine whether turbulence occurs at each data point, marking the location where turbulence is considered to occur as 1, and otherwise as 0, and then output the turbulence detection result.

[0051] In this embodiment, the processed data from two flights are used as a test set and input into the trained genetic programming-based symbolic classification model to verify the model's generalization and robustness. Table 1 shows the model detection results compared with the true values.

[0052] In comparison, the genetic programming-based symbolic classification model trained in this invention achieves an accuracy of 98% on the test set. Compared with traditional methods for calculating eddy current dissipation rate (EDR), it achieves 93.9% and 95.97% respectively, demonstrating a level comparable to traditional methods.

[0053] Table 1

[0054]

[0055] Table 2

[0056]

[0057]

Claims

1. A symbolic classification turbulence detection method based on aircraft airborne data, characterized in that: The symbol classification turbulence detection method based on aircraft onboard data includes the following steps performed in sequence: 1) Acquire airborne QAR data with historical turbulence anomaly information, then extract data related to turbulence occurrence and normalize it, and then divide the processed data into training set and test set according to the proportion; 2) Use the eddy current dissipation rate to classify and label the processed data in the training set to obtain labeled data; 3) Select a suitable symbol set; 4) Construct a symbol classification model based on genetic programming, and then use the labeled data and symbol set mentioned above to train the symbol classification model based on genetic programming to obtain the trained symbol classification model based on genetic programming. 5) Input the data from the test set or the airborne QAR data to be detected, which has been acquired in real time by the airborne computer and processed according to the method in step 1), into the trained genetic programming-based symbolic classification model. The output of the model is the turbulence detection result.

2. The symbolic classification turbulence detection method based on aircraft airborne data according to claim 1, characterized in that: In step 1), the method for acquiring airborne QAR data with historical turbulence anomaly information, extracting data related to turbulence occurrence, normalizing the data, and then dividing the processed data into training and testing sets proportionally is as follows: The processed data contains data for 24 normalized parameters, namely: left angle of attack, right angle of attack, flight number, total weight, altitude, latitude, longitude, radio altitude, total air temperature, static air temperature, aircraft display heading, angle of attack (drift angle), pitch angle, roll angle, wind direction, ground speed, indicated airspeed, Mach number, vacuum speed, wind speed, vertical acceleration, lateral acceleration, longitudinal acceleration, and vertical velocity.

3. The symbol classification turbulence detection method based on aircraft airborne data according to claim 1, characterized in that: In step 2), the method for classifying and labeling the processed data in the training set using eddy current dissipation rate to obtain labeled data is as follows: Data with an eddy current dissipation rate greater than or equal to 0.1 are marked as 1, indicating that turbulence has occurred; the rest of the data are marked as 0, indicating that no turbulence has occurred.

4. The symbol classification turbulence detection method based on aircraft airborne data according to claim 1, characterized in that: In step 3), the symbol set contains 7 operators: addition, subtraction, multiplication, division, square root, logarithmic function, and absolute value.

5. The symbol classification turbulence detection method based on aircraft airborne data according to claim 1, characterized in that: In step 4), the method for constructing a genetic programming-based symbol classification model and then training the genetic programming-based symbol classification model using the aforementioned labeled data and symbol set to obtain the trained genetic programming-based symbol classification model is as follows: The labeled data is used as the terminal set of the model, the symbol set is used as the function set of the model, and the terminal set and function set are used as input data to train the genetic programming-based symbol classification model. The model will randomly generate a given number of formulas from the terminal set and function set to express the relationship between the input data and the output data, and will select the most suitable formula to perform crossover, replication and mutation operations to generate the next generation of individuals. During training, the loss function of each generation of the model is calculated and a new generation is continuously generated to continuously improve the accuracy of the model in turbulence detection until the number of iterations reaches the set number of iterations, and a trained symbolic classification model based on genetic programming is obtained for turbulence detection. The loss function is a logarithmic loss function. This model can select the best individual based on the magnitude of the logarithmic loss function of the individual on the input data, as shown in the following expression: Among them, y i Input data x i The true category, p i To predict input data x i The probability that the input category belongs to 1.

6. The symbol classification turbulence detection method based on aircraft airborne data according to claim 1, characterized in that: In step 5), the method of inputting the data from the test set or the airborne QAR data to be detected, which has been acquired in real time by the airborne computer and processed according to the method in step 1), into the trained genetic programming-based symbolic classification model, and the model output being the turbulence detection result, is as follows: The trained genetic programming-based symbolic classification model will determine whether turbulence occurs at each data point, marking the location where turbulence is considered to occur as 1, and otherwise as 0, and then output the turbulence detection result.

Images