Risk prediction method and system based on QAR data and convolutional neural network

Abstract

The invention discloses a risk prediction method and system based on QAR data and a convolutional neural network, and the method comprises the following steps: obtaining to-be-analyzed QAR data, and carrying out the preprocessing of the QAR data; carrying out clustering analysis on normal data and overrun data in the preprocessed QAR data by utilizing a K-means clustering algorithm, carrying out data mining by utilizing an FP-Growth algorithm, and extracting an association rule of an unstable approach event and an incentive of the aircraft; using rule samples consistent with the association rules as a sample set to train a convolutional neural network; and converting the QAR data into picture data, and predicting the unstable approach risk of the aircraft by adopting the trained convolutional neural network. According to the method, a K-means clustering algorithm is fused to improve an FP-Growth algorithm, multi-level association rule mining of QAR data and unstable approach risk inducements is carried out, and analysis and risk prediction of high-complexity unsafe events such as an unstable approach stage are realized by adopting a convolutional neural network.

Description

Technical Field

[0001] This invention relates to the field of flight risk prediction technology, specifically a risk prediction method and system based on QAR data and convolutional neural networks. Background Technology

[0002] Currently, studies have been conducted to predict the long-term trends and temporal patterns of flight accident symptoms. The main prediction methods include: grey prediction method, autoregressive moving average model (ARIMA), support vector machine model (SVM), and neural network model.

[0003] However, most research models only focus on linearly correlated trends and cannot accurately predict unsafe events with superposition, coupling, and time-varying characteristics. Neural networks have been increasingly used in research related to flight safety event prediction and analysis, and the establishment of aircraft performance parameter prediction models, achieving some progress. However, traditional neural networks suffer from two major problems: ① Fully connected neurons consume significant computational resources and cannot extract structured spatiotemporal information; ② Traditional training algorithms cannot effectively propagate training errors from deep networks, preventing them from reaching their full potential. This is unsuitable for the analysis and prediction of highly complex unsafe events. Therefore, it is necessary to develop a new risk prediction method to grasp the main causes and development trends of unsafe events during the approach phase of transport aircraft. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a risk prediction method based on QAR data and convolutional neural networks.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] This invention relates to a risk prediction method based on QAR data and convolutional neural networks, comprising the following steps:

[0007] Obtain the QAR data to be analyzed and preprocess the QAR data;

[0008] K-means clustering analysis was performed on the normal and out-of-limit data in the preprocessed QAR data, and data mining was performed using the FP-Growth algorithm to extract the association rules between aircraft unstable approach events and their causes.

[0009] A convolutional neural network is trained using rule samples consistent with the association rules as the sample set;

[0010] QAR data was converted into image data, and a trained convolutional neural network was used to predict the risk of unstable aircraft approach.

[0011] Furthermore, based on the FP-Growth algorithm, multi-layer association rules of event attributes are mined, and the correlation between operational behaviors (throttle, pitch, and gradient), environmental factors (wind direction, wind speed, ambient temperature, and terrain data), and QAR data (roll angle, vertical acceleration, pitch angle, airspeed, track angle, and rate of descent) is extracted, thus summarizing the occurrence patterns of events.

[0012] Furthermore, the steps of the FP-Growth algorithm for mining association rules are as follows:

[0013] Scan the transaction database to determine if the current item value exists in the object array. If it does, delete the item from the object array, increment count by 1, and check again.

[0014] If not, add the item to the object array and set count to 1;

[0015] Calculate the hash address;

[0016] Add a hash table entry where the key is the hash address and the value is the entry's value.

[0017] Construct an ordered FP-tree;

[0018] Determine whether the evaluation indicators support, implementation and imbalance are met. If so, generate the frequent rule; otherwise, delete the frequent rule.

[0019] Perform frequent itemset mining.

[0020] Furthermore, the QAR data is QAR time-series data, and the preprocessing process includes:

[0021] Extract QAR data from the initial approach, intermediate approach, and final approach segments;

[0022] Based on the aircraft's flight altitude, the flight decoding results are extracted to obtain the decoding timing data;

[0023] The recording frequency of QAR data ranges from once per second to eight times per second. For parameter fields with low recording frequency, multinomial fitting is used to fill in missing values.

[0024] After removing exceptions, error messages, and fields, usable samples are obtained.

[0025] Furthermore, before inputting the QAR data into the convolutional neural network, the corresponding airport elevation model data provided by the geospatial data cloud website and the corresponding airport meteorological information provided by the China Weather Network are used to achieve spatiotemporal weighting of the QAR data.

[0026] Furthermore, convolutional neural networks have densely connected convolutional units and a CA (Channel Attention) module.

[0027] A risk prediction system based on QAR data and convolutional neural networks, characterized in that it includes:

[0028] The acquisition module is used to acquire the QAR data to be analyzed.

[0029] Preprocessing module: Used to preprocess QAR data to obtain QAR time series data and convert it into an image;

[0030] Output module: Used to feed the images transformed from QAR time-series data into the trained convolutional neural network and output the prediction results.

[0031] Another object of the present invention is to provide an electronic device comprising: a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned risk prediction method based on QAR data and convolutional neural networks.

[0032] The beneficial effects of this invention are:

[0033] This invention integrates the K-means clustering algorithm with the improved FP-Growth algorithm to mine multi-level association rules between QAR data and unstable approach risk factors, and uses convolutional neural networks to analyze and predict the risks of highly complex unsafe events such as unstable approach phases. Attached Figure Description

[0034] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0035] In the attached diagram:

[0036] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0037] Figure 2 This is a schematic diagram of the improved FP-Growth algorithm flow architecture of this invention;

[0038] Figure 3 This is a schematic diagram of the improved convolutional neural network and risk prediction architecture of the present invention;

[0039] Figure 4 This is a schematic diagram of the model part code implementation in an embodiment of the present invention;

[0040] Figure 5 This is a comparison chart of the true and predicted values ​​of the training and test sets during the model training process of this embodiment of the invention;

[0041] Figure 6 This is a schematic diagram of data output during the model training process in an embodiment of the present invention;

[0042] Figure 7 This is a schematic diagram of the neural network structure according to an embodiment of the present invention. Detailed Implementation

[0043] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0044] This invention provides a risk prediction method based on QAR data and convolutional neural networks, referring to... Figure 1 As shown, it includes the following steps:

[0045] Obtain the QAR data to be analyzed and preprocess the QAR data;

[0046] Specifically, QAR data records thousands of flight data points, including engine operating parameters, environmental parameters, and pilot actions. However, due to discrepancies in the recording frequency of different data points, some frequency ranges may have missing data. Furthermore, due to decoding errors and sensor mis-collection, some parameters and information in the data may have missing values ​​(data loss) or incorrect labeling. Therefore, preprocessing of the raw QAR data is necessary, primarily including:

[0047] ① Extract QAR data during the approach phase (mainly consisting of three parts: initial approach, intermediate approach, and final approach);

[0048] ②Based on the aircraft's flight altitude, the flight decoding results are extracted to obtain the decoding timing data;

[0049] ③ The recording frequency of QAR data varies from once per second to eight times per second. For parameter fields with low recording frequency, missing values ​​are filled by using polynomial fitting.

[0050] ④ Remove abnormal and error messages and fields to obtain usable samples.

[0051] K-means clustering analysis was performed on the normal and out-of-limit data in the preprocessed QAR data, and data mining was performed using the FP-Growth algorithm to extract the association rules between aircraft unstable approach events and their causes.

[0052] Traditional FP-Growth algorithms require recursively generating conditional databases and conditional FP-trees, resulting in high memory overhead and unsuitability for analyzing the causes of insecure events involving multiple levels and dimensions. Therefore, this project proposes an improvement to the FP-Growth algorithm using K-means clustering. The main idea is as follows:

[0053] ① Replace the tree structure of FP-tree with ordered FP-tree to reduce memory resource consumption;

[0054] ② Based on the conceptual hierarchy of attributes, different support levels are set to generate FP-Trees containing items at different levels;

[0055] ③ Traverse the FP-Tree to generate frequent itemsets, and generate association rules from the frequent itemsets;

[0056] ④ An improved FP-growth algorithm, which integrates the K-means clustering algorithm, is used to perform clustering analysis and data mining analysis on normal and out-of-limit data in the approach phase.

[0057] ⑤ Based on the improved FP-Growth algorithm, we mine multi-layer association rules of event attributes, extract the associations between operational behaviors (throttle, pitch, and gradient), environmental factors (wind direction, wind speed, ambient temperature, and terrain data), and QAR data (roll angle, vertical acceleration, pitch angle, airspeed, track angle, and rate of descent), and summarize the occurrence patterns of events.

[0058] A convolutional neural network is trained using rule samples consistent with the association rules as the sample set;

[0059] For QAR data, we will use the corresponding airport elevation model data provided by the geospatial data cloud website and the corresponding airport meteorological information provided by China Weather Network to realize the spatiotemporal weighting of QAR data. We will also convert the time series data into images and feed them into a deep convolutional neural network to realize mining analysis.

[0060] QAR data was converted into image data, and a trained convolutional neural network was used to predict the risk of unstable aircraft approach.

[0061] Convolutional Neural Network (CNN) models, with their shared convolutional kernels, exhibit significant advantages when processing large amounts of flight data. However, the presence of pooling layers leads to the loss of much valuable information. Therefore, this project proposes an improved CNN model to establish an approach phase risk prediction model based on QAR data while avoiding information loss. The main idea is as follows:

[0062] ① A novel deep prediction network with densely connected convolutional units and a CA (Channel Attention) module is proposed to extract hierarchical features of insecure events and reduce information loss;

[0063] ② Design a relatively deep architecture to achieve higher accuracy by overcoming the randomness and instability of insecure events;

[0064] ③ Considering that the risk factors in the unstable approach phase are significantly influenced by geographical features and exhibit substantial seasonal and cyclical trends, data preprocessing is performed when transforming connection depth features to improve the model's learning ability in terms of temporal and spatial features. This improves the reliability and robustness of network training.

[0065] ④ Process the preprocessed QAR time series data to convert the time series data into images;

[0066] ⑤ Input the image into the trained convolutional neural network and output the unstable approach type and prediction result;

[0067] ⑥ Complete the reliability verification of the prediction model.

[0068] The specific process is as follows: compiling Matlab code, importing data, constructing a randomized array, dividing the training and test sets, normalizing the imported data, tiling the data, and constructing the network structure. This code is a Matlab script for training a Convolutional Neural Network (CNN) model, along with a series of parameter and hyperparameter settings. The established code implements the training process of a CNN model for time series data prediction that can accept multiple feature inputs, including model initialization, convolutional kernel size, Batch Normalization (BN) layers, ReLU activation layers, pooling layers, regression layers, loss calculation, and other operations. Partial code implementation is shown below. Figure 4 As shown;

[0069] Compile Matlab code to calculate and plot a comparison graph of the true and predicted values ​​on the training and test sets during the training process (e.g., Figure 5 This code takes four feature values ​​as input: altitude, indicated airspeed, current deviation, and glide slope deviation. The output value (predicted term) is the course deviation. This code is primarily used to visualize and evaluate the performance of the trained LSTM model, as well as to inverse normalize the model's predictions and calculate some performance metrics.

[0070] Modify the settings of relevant parameters in the model to find relatively better model parameters for subsequent predictions. Parameters such as kernel size, batch size, learning rate, descent factor, and number of training iterations can be modified. Then, plot the comparison between predicted and actual values, as well as data such as R² of the training and test sets. Figure 6It is a data output during the training process;

[0071] The network structure was established and initial parameters were determined for subsequent model optimization and comparison. The structure of the convolutional neural network was determined. Therefore, a set of pre-trained model structures was temporarily set to predict course deviation data. Figure 7 It is a neural network structure.

[0072] This embodiment also proposes a risk prediction system based on QAR data and convolutional neural networks, including: an acquisition module for acquiring QAR data to be analyzed;

[0073] Preprocessing module: Used to preprocess QAR data to obtain QAR time series data and convert it into an image;

[0074] Output module: Used to feed the images transformed from QAR time-series data into the trained convolutional neural network and output the prediction results.

[0075] This embodiment also proposes an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned risk prediction method based on QAR data and convolutional neural networks.

[0076] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A risk prediction method based on QAR data and convolutional neural networks, characterized in that: Includes the following steps: Obtain the QAR data to be analyzed and preprocess the QAR data; K-means clustering analysis was performed on the normal and out-of-limit data in the preprocessed QAR data, and data mining was performed using the FP-Growth algorithm to extract the association rules between aircraft unstable approach events and their causes. A convolutional neural network is trained using rule samples consistent with the association rules as the sample set; QAR data was converted into image data, and a trained convolutional neural network was used to predict the risk of unstable aircraft approach.

2. The risk prediction method based on QAR data and convolutional neural networks according to claim 1, characterized in that, Based on the FP-Growth algorithm, multi-level association rules of event attributes are mined, and the following are extracted: operational behaviors: throttle stick, pitch stick, and bank stick; environmental factors: wind direction, wind speed, ambient temperature, and terrain data; and the association between these and QAR data: roll angle, vertical acceleration, pitch angle, airspeed, track angle, and rate of descent attributes. The occurrence patterns of events are then summarized.

3. A risk prediction method based on QAR data and convolutional neural networks according to claim 2, characterized in that: The steps for mining association rules based on the FP-Growth algorithm are as follows: Scan the transaction database to determine if the current item value exists in the object array. If it does, delete the item from the object array, increment count by 1, and check again. If not, add the item to the object array and set count to 1; Calculate the hash address; Add a hash table entry where the key is the hash address and the value is the entry's value. Construct an ordered FP-tree; Determine whether the evaluation indicators support, implementation and imbalance are met. If so, generate the frequent rule; otherwise, delete the frequent rule. Perform frequent itemset mining.

4. A risk prediction method based on QAR data and convolutional neural networks according to claim 4, characterized in that, QAR data is QAR time-series data, and the preprocessing process includes: Extract QAR data from the initial approach, intermediate approach, and final approach segments; Based on the aircraft's flight altitude, the flight decoding results are extracted to obtain the decoding timing data; The recording frequency of QAR data ranges from once per second to eight times per second. For parameter fields with low recording frequency, multinomial fitting is used to fill in missing values. After removing exceptions, error messages, and fields, usable samples are obtained.

5. A risk prediction method based on QAR data and convolutional neural networks according to claim 5, characterized in that, Before being input into the convolutional neural network, the QAR data is spatiotemporally weighted using the corresponding airport elevation model data provided by the geospatial data cloud website and the corresponding airport meteorological information provided by the China Weather Network.

6. A risk prediction method based on QAR data and convolutional neural networks according to claim 6, characterized in that, Convolutional neural networks have densely connected convolutional units and a CA (Channel Attention) module.

7. A risk prediction system based on QAR data and convolutional neural networks, characterized in that, include: The acquisition module is used to acquire the QAR data to be analyzed. Preprocessing module: Used to preprocess QAR data to obtain QAR time series data and convert it into an image; Output module: Used to feed the images transformed from QAR time-series data into the trained convolutional neural network and output the prediction results.

8. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the risk prediction method based on QAR data and convolutional neural networks as described in claims 1-7.

Images