KR20190054327A - Method for predicting runway visual range by deep neural network - Google Patents

Abstract

The present invention relates to a method for predicting a runway visual range using a deep neural network. The method comprises the steps of: analyzing past weather observation data of a target airfield and extracting wind velocity, temperature, humidity, and runway visual range measurement values related to the visibility to input the wind velocity, temperature, humidity, and runway visual range measurement values into a deep neural network; learning the deep neural network based on the weather observation data inputted into the deep neural network; and measuring current wind velocity, temperature, and humidity of the target airfield and inputting the current wind velocity, temperature, and humidity into the deep neural network, and deriving a predicted visual range value.

Description

METHOD FOR PREDICTING RUNWAY VISUAL RANGE BY DEEP NEURAL NETWORK BACKGROUND OF THE INVENTION [0001]

The present invention relates to a runway visibility distance prediction method, and more particularly, to a runway visibility distance prediction method using a depth network.

Weather factors affecting aircraft safety include thunder storms, icing, turbulence, mountain waves, wind shear, microburst, visibility, etc. . Among these, visibility can be classified into flight visibility, runway visual range, etc., and runway visibility is a factor that directly affects the decision of takeoff and landing on the runway.

At the airports operated by the aircraft, the meteorological aerodrome report (METAR), which is issued every hour, and the terminal aerodrome forecast (TAF), which is issued every 6 hours, Ensure that airline personnel perform safe operations. The Korea Meteorological Agency office located in each airport observes the real-time weather information of the corresponding airport, provides the observation information and forecast information, and if the airport uses the military aerodrome jointly, it provides the weather information from the military weather station have. A privately owned airfield, such as Hanseo University, located in Taean, Chungcheongnam-do, observes real-time weather information to confirm whether a safe aircraft can land or take off. However, it does not provide any weather forecast.

The runway visibility distances of the airfield meteorological measurements indicate the maximum distances that can be seen on the runway at the farthest part of the runway unlike the visibility information of the nearby area including the airfield. And can be operated when the visibility is above the level. Visual flight rule (VFR) aircraft require visibility of at least 5,000 meters for recognition of the circumstances necessary to avoid collisions with other aircraft, and the same criteria are generally applied during takeoff and landing.

Unlike visibility, the runway visibility distance can be different from the visibility, as it is the ground clearance distance of an aircraft moving on the runway or an airport ground and a ground vehicle. Therefore, on the aircraft, runway visibility or runway visual range (RVR) is specified as a detailed measure of visibility. The runway visibility is important not only for the safety of the moving object operating on the ground but also for the safety of the aircraft in IFR (Instrument Flight Rule), which is taken off and landing at low visibility. . In particular, in order to generate relevant information at an aerodrome where resources are lacking to provide the weather forecast function, a method for generating information tailored to the area needs to be developed.

The meteorological phenomenon has a complicated relationship with many variables. Conventional weather forecasting methods require a large amount of data processing and calculation by using precise prediction models and meteorological measurements at various locations for weather forecasting, . This is not only costly but also uses a numerical forecasting model, so there is a limit to the ability to make appropriate weather forecasts for various meteorological changes occurring in small areas. In order to operate safe airplanes, weather forecasting of airports in take-off and landing areas is important. However, weather forecasting services can not be provided due to resource limitations such as cost or supercomputer in small airfields that are not operated by transportation airplanes such as international airports.

The object of the present invention is to solve the above problems by using a deep neural network (DNN) as a learning model for inputting time-based meteorological observation information such as wind speed, humidity, temperature, A method for predicting runway visibility using a depth neural network that generates runway visibility distance prediction information is provided.

In order to achieve the above object, the runway visibility distance prediction method using the depth neural network of the present invention analyzes historical weather observation data of a target airfield, extracts wind speed, temperature, humidity, runway visibility distance values related to visibility, Inputting to a neural network; Learning a depth neural network based on meteorological data input to the neural network; And measuring the current wind speed, temperature, and humidity of the target airfield, and inputting the current wind speed, temperature, and humidity into the in-depth neural network to derive a visible range predicted value.

The depth network is characterized by comprising an input layer, a hidden layer and an output layer.

The input layer is composed of nodes corresponding to the number of input information including previous data, and the hidden layer is composed of two nodes capable of optimizing the data processing, and the output layer is composed of one node.

As described above, according to the present invention, it is possible to predict the visibility distance in the local area using the past weather data of the target airfield, and it is possible to use only the data of the wind speed, temperature, Can be predicted.

Also, according to the present invention, it can be predicted in advance that a low visibility condition that may affect the safety operation of the airplane and the operational efficiency of the airports may occur.

1 is a conceptual diagram for explaining a general deep-layer neural network structure.
FIG. 2 is a conceptual diagram for explaining a depth-of-field-based runway visibility distance prediction model structure according to the present invention.
FIG. 3 is a flowchart illustrating a runway visibility distance prediction method using a depth-of-field network according to the present invention.
FIG. 4 is a graph showing comparison of observed values and predicted values 1 before settlement of the runway visibility distance prediction model based on the depth-of-field network according to the present invention.
FIG. 5 is a graph showing comparison of observed values and predicted values 1 after settlement of the runway visibility distance prediction model based on the depth-of-field network according to the present invention.
FIG. 6 is a graph showing comparison of observed values and predicted values after settlement of a runway visibility distance prediction model based on a depth-of-field neural network according to the present invention.
FIG. 7 is a graph showing a comparison of observed values and predicted values of 4-month values after settlement of the runner visibility distance prediction model according to the present invention.
8 is a graph showing comparison of observed values and predicted values of 4-month values of the runway visibility distance prediction model according to the present invention.
9 is a graph showing detailed comparisons of observation values and predicted values according to the depth-of-field-based runway visibility distance prediction model of the present invention.
10 is a graph showing further detailed comparisons of observations and predicted values according to the depth-of-field neural network-based runway visual prediction model of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein.

First, a general deep-layer neural network structure will be described.

1 is a conceptual diagram for explaining a general deep-layer neural network structure.

Referring to FIG. 1, when two or more hidden layers are referred to as a deep neural network, a learning can be implemented by constructing a deep neural network structure having two hidden layers. In FIG. 1, a round circle means a node, and each node includes an active function such as a sigmoid function or a ReLU (rectified linear unit) function. The arrows connecting the nodes of each layer represent the flow of the signal, the data of the previous node is multiplied by the weight, and the arrows are connected to the next node. The input data multiplied by the transmitted weight are all summed up, that is, weighted sum, and an output value is generated through the active function. In order to minimize the error between the output value and the actual measured value, each weight is updated through learning.

Next, the structure of the runway visual-distance prediction model using the depth-based neural network according to an embodiment of the present invention will be described.

FIG. 2 is a conceptual diagram for explaining a depth-of-field-based runway visibility distance prediction model structure according to the present invention.

Referring to FIG. 2, the depth-based neural network-based runway visual prediction model according to the present invention is a model in which the input of the depth-of-field neural network includes a wind speed WS _n , a temperature Temp _n , a humidity Hum _n , Average data of distance (RVR _n ) measurements or runway visibility (RVR _n ) measurements can be used. Here, the subscript n means current, and the subscript nm means meteorological data before m units.

In order to predict the correction, instead of inputting only the meteorological observation information of only the measurement point, the past change rate is included by using the past meteorological information together for a predetermined period. For this, wind speed, temperature, humidity, and corrective data can be input from past data before a certain period (m unit period) to present time.

Since the weather observation information is measured every predetermined unit time, the amount of data between pieces of information in the m period can be input to the input layer using 4 m pieces of data (m Ticks x 4 pieces of weather information) as shown in Fig. Here, the input layer is composed of nodes corresponding to the number of input data, the hidden layers 1 and 2 are nodes capable of optimizing the data processing, and the output layer is composed of one node. This configuration can be used to identify the structure from a relatively complex structure to a simple structure and to identify the characteristics of the empirical data in order to find an appropriate condition structure for model development. At this time, a sigmoid function was used as an activation function, and weights were updated by an error back propagation method using stochastic gradient descent (SGD).

And, in order to derive the appropriate learning frequency and the learning rate, we repeatedly simulated the variables while changing them and selected the appropriate learning frequency and learning rate. In the present invention, initial values of a model are initialized through a random function.

Next, a description will be made of a runway visibility distance prediction method using a depth-of-field network according to an embodiment of the present invention.

FIG. 3 is a flowchart illustrating a runway visibility distance prediction method using a depth-of-field network according to the present invention.

Referring to FIG. 3, in the runway visibility distance prediction method using the depth network of the present invention, meteorological information related to visibility is extracted and input to a depth neural network (S100). That is, the average data of wind speed, temperature, humidity, runway visibility (RVR) measurement value or runway visibility distance (RVR) measurement values of the past weather field of the target airfield are inputted into the deep neural network.

Next, the depth neural network is learned based on the selected past meteorological observation data (S110).

Finally, the current wind speed, temperature, and humidity of the target airfield are measured and input to the depth-of-field network to derive a visible range predicted value (S120).

Next, in order to confirm and verify the runway visibility distance prediction model based on the depth information network according to an embodiment of the present invention, depth learning of the neural network (200 times) was performed using the one-match data of the corresponding period. For the learned model, the model was calibrated so that the observed data and the model predicted data could be visually compared to ensure similarity or consistency.

FIG. 4 is a schematic diagram of the runway visibility distance model of the present invention of the present invention compared with the actual runway time constant value and the model predicted value prior to settlement, and FIG. 5 is a graph showing the runway visibility distance prediction model And then plotted the corresponding values. Here, the solid line is the actual observation value, and the dotted line is the predicted value in the model. As can be seen from FIG. 5, it can be seen that the similarity of data after the model has been increased.

FIG. 6 illustrates the observed value and the model predicted value of FIG. 5 as 1: 1. Here, when the data point is on a straight line having the slope 1, it means that the observed value and the model predicted value are the same, and the closer the data point is to the straight line, the higher the similarity of data. The quantitative value of this is confirmed by the difference between the straight line and the data point. The data correlation coefficient (R ² ) value is 0.9981 and root mean square error (RMSE) value is 31.3123.

Next, in order to predict the actual runway visual distance using the runway visual range prediction model using the depth-of-field neural network according to an embodiment of the present invention, the predicted value of the model result Respectively. In order to confirm the appropriateness of the predicted weather data, comparative analysis with the measured runway visibility distance was performed. At this time, the input data of the forecasting model is Taean Airfield of Hanseo University as a regional target, and weather information for 4 months from April to July 2015 is used. Since each input data is different in wind speed, temperature, humidity, and runway visibility, there is no similarity between the data units and the measurement range. Therefore, it is necessary to normalize the input data for each input data. Respectively.

,

,

,

는 i번째 측정된 기상 정보, i번째의 정규화된 기상 정보, 측정된 기상 정보의 최대값과 최소값을 의미한다.here,

,

,

,

Means the maximum value and the minimum value of the i-th measured weather information, the i-th normalized weather information, and the measured weather information.

Through the normalization process, each meteorological data is converted into a value between 0 and 1, so that the weight is not concentrated on a specific input variable.

The foggy one - match weather data was selected from the four - month weather data, and the model results were studied through the depth neural network proposed in the present invention.

Figure 7 compares runway visibility and predicted values over a 4 month period using a runway visibility prediction model using a depth neural network in accordance with an embodiment of the present invention. Here, the solid line represents the measured value of the runway visibility distance, and the dotted line represents the predicted value. The total number of data is 921,080, and it can be confirmed that relatively similar data form is shown except that the lowest values in the low visibility state observed with a viewing distance of 400 m or less are not perfectly matched.

FIG. 8 is a graph showing the observed value and the predicted value of FIG. 7 in a 1: 1 ratio. As shown in FIG. 7, the data immediately preceding the slope 1 means that the observed value and the predicted value are exactly the same. It can be seen that the overall data format of FIG. 8 is more scattered on a straight line of slope 1 than the data of FIG. 7, but this is because the number of data has increased.

The data corresponding to the point on the left side of FIG. 8, in which the actual observation value is as low as 200 m or less in the data, but the error of the predicted value is very large, is that the observation data such as the wind speed, temperature, It is. This can be handled by using a method of disregarding a large error value compared to the existing value trend at the step of displaying or utilizing data in a unit of 10 seconds as a very small number of data when actual predicted value is utilized.

9 is a detailed comparison between the observed value and the predicted value. The solid line indicates the observed value and the dotted line indicates the predicted value. In FIG. 9, it can be seen that a part of the dotted line has a lower predicted value than the solid line, because the developed prediction model is sensitive to the derivation of the predicted value using the observed value. This feature can also be used to predict trends in which the visibility may be lowered before the actual visibility goes down, and the corresponding characteristics can be seen in the second data descent part of Figure 9 (a).

FIG. 10 is a graph showing a tendency of the developed model to predict trends in which a short-term visibility is improved and then decreased again in a state where the runway visibility distance is lowered. As can be seen from the data of FIG. 9 and the data front part of FIG. 10, the developed model is sensitive to the measured data, and it can be confirmed that the characteristic of the lowered visible distance after the improvement is lowered Can be predicted in advance.

According to the present invention, it is possible to predict the runway visibility distance using only the data of the wind speed, temperature and humidity among the aerodynamic measurement variables through the developed prediction model, It is possible to predict in advance that a low-correction state such as the one shown in Fig.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is clearly understood that the same is by way of illustration and example only and is not to be construed as limiting the scope of the invention as defined by the appended claims. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention.

Claims (3)

Analyzing the past meteorological data of the target aerodrome and extracting wind speed, temperature, humidity, runway visibility distance related to visibility, and inputting to the depth neural network;
Learning a depth neural network based on meteorological data input to the neural network; And
Measuring a current wind speed, a temperature and a humidity of the target airfield, and inputting the current wind speed, the temperature and the humidity into a neural network to derive a visible range predicted value. The method according to claim 1,
Wherein the depth network comprises an input layer, a hidden layer and an output layer. 3. The method of claim 2,
Wherein the input layer includes nodes corresponding to the number of input information including previous data, two hidden layers, and a node capable of optimizing the data processing, and the output layer includes one node.

Images