Flight data clustering and classification method based on deep learning

Abstract

The invention relates to a flight data clustering and classification method based on deep learning. The method comprises the following steps: 1) flight data potential representation feature extraction: constructing a deep learning model of a convolutional neural network, a bidirectional gating circulation unit and an attention mechanism; and 2) clustering and classification decisions combined with dynamic optimization are used for realizing automatic classification and optimization analysis of flight data. According to the deep learning technology, high-order features in the data can be automatically extracted, the time dynamics and structural features of the data can be mined, the nonlinear relation, the long and short term dependency relation and the hidden mode features in the flight data can be effectively captured by constructing the multi-layer neural network model, and the accuracy of the flight data is improved. Therefore, the defects of a traditional method in the aspects of feature extraction and dynamic modeling are overcome. Besides, the deep learning framework has good expandability, large-scale and high-dimensional time series data can be rapidly processed, and an efficient and accurate solution is provided for clustering and classification of flight data.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and civil aviation data analysis, and specifically to a flight data clustering and classification method based on deep learning. Background Art

[0002] With the rapid development of the global air transport industry, airlines have accumulated a large amount of flight operation data, including multi-dimensional information such as passenger numbers, fares, seat utilization, flight revenue, etc. These data contain rich regularities and structural characteristics. Reasonable data analysis and classification will not only help airlines improve their operational management level, but also help to explore flight operation rules, thereby optimizing flight scheduling and resource allocation, and improving the overall revenue and passenger satisfaction of flights.

[0003] However, the clustering and classification methods commonly used in the industry are still at a relatively basic technical level, mainly relying on rule-driven or statistical modeling methods, such as linear regression models, K-means clustering and other traditional machine learning algorithms. These methods can still achieve certain results when processing single-dimensional or small amounts of flight data, but in actual applications, flight data usually has complex characteristics such as high dimensionality, nonlinearity, and strong time dependence, and traditional methods are unable to cope with them. For example, although the linear regression model is simple and intuitive, it is difficult to fit nonlinear data patterns, and the K-means clustering algorithm is prone to unstable clustering results in high-dimensional space. These deficiencies directly lead to rough clustering results and low classification accuracy, making it difficult to provide accurate decision support for airlines.

[0004] At the same time, existing flight data analysis methods rely too much on manually set rules or simple statistical models, and lack the ability to automatically learn and model the deep characteristics of the data. This approach that relies on prior knowledge not only has subjective biases, but may also ignore important information hidden in the data, resulting in poor robustness and adaptability of the results. In addition, with the rapid growth of data scale, the dimension and magnitude of flight time series data continue to increase. Traditional algorithms have high computational overhead when processing large-scale data, and it is difficult to obtain efficient and accurate results within a limited time. Therefore, in actual operations, clustering and classification results are often difficult to adapt to the changing market demands in flight operations.

[0005] To this end, a flight data clustering and classification method based on deep learning is designed to overcome the above problems. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a flight data clustering and classification method based on deep learning with high efficiency and accuracy.

[0007] The present invention is implemented by the following technical solution: a flight data clustering and classification method based on deep learning, the method comprising: 1) Extraction of latent representation features of flight data: Deep learning models used to build convolutional neural networks, bidirectional gated recurrent units, and attention mechanisms; 2) Clustering and classification decisions combined with dynamic optimization: used to achieve automatic classification and optimization analysis of flight data.

[0008] As a preferred embodiment, the specific steps of the flight data potential representation feature extraction module are as follows: Step 1: Collect and clean the historical flight sales process data, including core indicators such as total fare revenue, total number of passengers, economy class and business class revenue and number of passengers at key time points; Step 2: Combine the cleaned flight data into time series data according to flight date, flight number, and flight segment identifier, and input it into the deep learning model; Step 3: Use a one-dimensional convolutional network to extract features from the flight time series. The convolutional layer uses a sliding window method to capture the short-term changes in flights, and the pooling layer reduces the data dimension to compress information redundancy. Step 4: The features extracted by convolution are input into the bidirectional gated recurrent unit. Through joint learning in the two time directions, the long-term dependency of flight time series data is modeled to generate a high-dimensional feature representation for each time step. Step 5: Add an attention mechanism based on GRU to calculate the importance weight of each time step. Use the attention mechanism to dynamically focus on the time steps that have a greater impact on the change trend of flight data, and compress the time series features into a global potential representation. Step 6: Use the autoencoder structure to reconstruct the flight data to ensure that the extracted latent representation can accurately retain the original information.

[0009] As a preferred embodiment, the specific steps of the clustering and classification decision module combined with dynamic optimization are as follows: Step 1: pre-train the potential representation features and initialize the cluster center using the KMeans algorithm; Step 2: Add KL divergence loss to model training to minimize the distribution difference between the potential representation and the cluster center. The loss function includes reconstruction error and clustering loss. Step 3: During the model training process, the KMeans algorithm is rerun periodically using the latent representation features to dynamically adjust the location of the cluster center to adapt to changes in feature distribution and improve the accuracy of the final clustering results. Step 4: After training is completed, the latent representation of each flight sample is mapped to the corresponding cluster label to generate the classification result.

[0010] The beneficial effects of the present invention are as follows: Compared with the prior art, the flight data clustering and classification method based on deep learning provided by the present invention, by introducing deep learning into the flight classification problem, can achieve the following positive effects: (1) This method extracts short-term key features through CNN, combines Bi-GRU to capture the long-term dependencies of time series, and then introduces the attention mechanism to dynamically focus on key time steps, so that the model can extract the potential rules of complex flight data. Compared with traditional static clustering methods, this method significantly improves the modeling ability of dynamic time series features.

[0011] (2) By adding KL divergence loss to the latent representation space to optimize the location of cluster centers and combining it with the dynamic KMeans update mechanism, the dependence of clustering results on the initial centers is significantly reduced.

[0012] (3) The present invention constructs an end-to-end clustering model with the autoencoder as the core, realizes automatic feature extraction and clustering optimization of data, improves analysis efficiency, and adapts to the needs of real-time flight data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the flight data clustering and classification model structure based on deep learning.

[0014] Figure 2 Schematic diagram of flight data clustering and classification algorithm based on deep learning. DETAILED DESCRIPTION

[0015] In order to enable those skilled in the art to more clearly understand the objectives, technical solutions and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0016] In the description of the present invention, it is necessary to understand that the orientation or position relationship indicated by the terms such as "up", "down", "left", "right", "inside", "outside", "horizontal" and "vertical" is based on the orientation or position relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention, rather than indicating or implying that the device or original referred to must have a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0017] The present invention will be described in detail below with reference to the accompanying drawings: Figure 1-2 As shown, a flight data clustering and classification method based on deep learning, the method includes: 1) Extraction of latent representation features of flight data: Deep learning models used to build convolutional neural networks, bidirectional gated recurrent units, and attention mechanisms; 2) Clustering and classification decisions combined with dynamic optimization: used to achieve automatic classification and optimization analysis of flight data.

[0018] The specific steps of the flight data potential representation feature extraction module are as follows: Step 1: Collect and clean the historical flight sales process data, including core indicators such as total fare revenue, total number of passengers, economy class and business class revenue and number of passengers at key time points; Step 2: Combine the cleaned flight data into time series data according to flight date, flight number, and flight segment identifier, and input it into the deep learning model; Step 3: Use a one-dimensional convolutional network to extract features from the flight time series. The convolutional layer uses a sliding window method to capture the short-term changes in flights, and the pooling layer reduces the data dimension to compress information redundancy. Step 4: The features extracted by convolution are input into the bidirectional gated recurrent unit. Through joint learning in the two time directions, the long-term dependency of flight time series data is modeled to generate a high-dimensional feature representation for each time step. Step 5: Add an attention mechanism based on GRU to calculate the importance weight of each time step. Use the attention mechanism to dynamically focus on the time steps that have a greater impact on the change trend of flight data, and compress the time series features into a global potential representation. Step 6: Use the autoencoder structure to reconstruct the flight data to ensure that the extracted latent representation can accurately retain the original information.

[0019] The specific steps of the clustering and classification decision module combined with dynamic optimization are as follows: Step 1: pre-train the potential representation features and initialize the cluster center using the KMeans algorithm; Step 2: Add KL divergence loss to model training to minimize the distribution difference between the potential representation and the cluster center. The loss function includes reconstruction error and clustering loss. Step 3: During the model training process, the KMeans algorithm is rerun periodically using the latent representation features to dynamically adjust the location of the cluster center to adapt to changes in feature distribution and improve the accuracy of the final clustering results. Step 4: After training is completed, the latent representation of each flight sample is mapped to the corresponding cluster label to generate the classification result. Example

[0020] Take an airline's flight from city A to city B as an example. This flight is a direct flight with no stops and a total of 180 seats. The airline hopes to analyze passenger behavior patterns based on aviation data and classify the demand characteristics of this flight. First, obtain the flight's operating data from the operation database or predict it based on the model, including total fare revenue, total number of passengers, economy class and business class fare revenue and number of passengers, etc., and clean and standardize the data to construct a time series sample. Each sample contains the key features of the past 15 days, forming a two-dimensional matrix of shape (15, 7), which is input into the model for analysis.

[0021] When using the trained model, the local features of the time series are first extracted through a convolutional neural network to capture short-term changes in ticket prices and passenger flow over time, such as significant fluctuations during holidays or promotions. Subsequently, the long-term dependencies of the time series are modeled using a bidirectional gated recurrent unit (Bi-GRU) to extract global change trends. On this basis, an attention mechanism is added to dynamically assign importance weights to different time steps, such as focusing on the key time period before holidays to generate a global potential feature representation of flights. Finally, the model outputs cluster labels in digital form (such as 0, 1, 2, etc.), which represent the categories of flights in the feature space. Through subsequent business analysis, these labels can be mapped to specific flight categories, such as "high passenger flow and high revenue" or "large demand fluctuations" and other business meanings, thereby providing data support for airlines' operational decisions.

[0022] Anyone familiar with the technology may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A flight data clustering and classification method based on deep learning, characterized in that: The method comprises: 1) Extraction of latent representation features of flight data: Deep learning models used to build convolutional neural networks, bidirectional gated recurrent units, and attention mechanisms; 2) Clustering and classification decisions combined with dynamic optimization: used to achieve automatic classification and optimization analysis of flight data.

2. The flight data clustering and classification method based on deep learning according to claim 1, characterized in that: The specific steps of extracting the potential representation features of the flight data are as follows: Step 1: Collect and clean the historical flight sales process data, including core indicators such as total fare revenue, total number of passengers, economy class and business class revenue and number of passengers at key time points; Step 2: Combine the cleaned flight data into time series data according to flight date, flight number, and flight segment identifier, and input it into the deep learning model; Step 3: Use a one-dimensional convolutional network to extract features from the flight time series. The convolutional layer uses a sliding window method to capture the short-term changes in flights, and the pooling layer reduces the data dimension to compress information redundancy. Step 4: The features extracted by convolution are input into the bidirectional gated recurrent unit. Through joint learning in the two time directions, the long-term dependency of flight time series data is modeled to generate a high-dimensional feature representation for each time step. Step 5: Add an attention mechanism based on GRU to calculate the importance weight of each time step. Use the attention mechanism to dynamically focus on the time steps that have a greater impact on the change trend of flight data, and compress the time series features into a global potential representation. Step 6: Use the autoencoder structure to reconstruct the flight data to ensure that the extracted latent representation can accurately retain the original information.

3. The flight data clustering and classification method based on deep learning according to claim 1, characterized in that: The specific steps of clustering and classification decision-making combined with dynamic optimization are as follows: Step 1: pre-train the potential representation features and initialize the cluster center using the KMeans algorithm; Step 2: Add KL divergence loss to model training to minimize the distribution difference between the potential representation and the cluster center. The loss function includes reconstruction error and clustering loss. Step 3: During the model training process, the KMeans algorithm is rerun periodically using the latent representation features to dynamically adjust the location of the cluster center to adapt to changes in feature distribution and improve the accuracy of the final clustering results. Step 4: After training is completed, the latent representation of each flight sample is mapped to the corresponding cluster label to generate the classification result.

Images