Aircraft position estimation system and aircraft position estimation method

Abstract

To provide a system for estimating the position of a flying vehicle and a method for estimating the position of a flying vehicle, with which the estimation accuracy of a current position of a flying vehicle is improved so that the flying vehicle can fly stably for a long time.SOLUTION: A system 20 for estimating the position of a flying vehicle estimates a current position of a current flying vehicle 10 by executing a machine learning algorithm N, using time series data related to the flying vehicle 10 flying in a predetermined flight zone as input values. The machine learning algorithm N is composed of an input cross-attention layer C1, a self-attention layer S, and an output cross-attention layer C2, and uses first array data D1 as time series data being the input values. Further, the machine learning algorithm has a function for generating second array data D2 by reconstructing the first array data D1 based on a prescribed weighting factor.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a system and method for estimating the position of an aircraft, and more particularly to a system and method for estimating the position of an aircraft that improves the accuracy of estimating the current position of the aircraft and enables the aircraft to fly stably for a long period of time. [Background technology]

[0002] To ensure safety during autonomous flight, unmanned aerial vehicles such as unmanned airplanes and drones need to acquire their current position in real time during flight. To acquire the current position of an unmanned aerial vehicle, for example, a method is adopted in which a GPS device mounted on the unmanned aerial vehicle receives GPS signals sent from artificial satellites, or a method is adopted in which a flight route from a starting point is calculated by combining vectors of travel distance and direction for each subdivision point based on an inertial navigation device such as an acceleration sensor or gyro sensor mounted on the unmanned aerial vehicle.

[0003] However, GPS devices cannot obtain stable position information in environments where GPS signals are difficult to receive, such as under bridges or inside buildings. Inertial navigation systems also have a large accumulation of errors over the distance traveled, limiting the distance over which they can fly safely.

[0004] Therefore, to fly an unmanned aerial vehicle safely for a long period of time in such an environment, a human must visually monitor the unmanned aerial vehicle during flight. Furthermore, when autonomous flight or automatic control is difficult, it is necessary to switch to manual remote control to assist the flight. However, human monitoring and control imposes a heavy burden and can lead to serious accidents due to oversights or piloting errors caused by human error. Therefore, various technologies have been proposed to enable the accurate current position of an unmanned aerial vehicle to be acquired.

[0005] For example, Patent Document 1 discloses a system that determines the current flight position based on an object captured by an imaging device mounted on an unmanned aerial vehicle. Specifically, the system compares a reference object captured at a reference position with a measured object captured at a measured position, and estimates the current position of the vehicle by detecting three-dimensional displacement of the measured position from the reference position based on two-dimensional position deviation and three-dimensional angle deviation. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 2020-176961 Summary of the Invention [Problem to be solved by the invention]

[0007] According to Patent Document 1, it is possible to determine the aircraft's flight position even in an environment where GPS signals are difficult to receive, and since there is no error associated with the distance traveled, as occurs with inertial navigation systems, it is possible to always stably determine the aircraft's position even during long flights.

[0008] However, when acquiring the flight position of the drone based on images captured by an imaging device, calculations for image processing are required, which increases the calculation load and time, and also increases the battery consumption of the drone, limiting flight time.

[0009] The present invention was devised in consideration of the above points, and aims to provide an aircraft position estimation system and aircraft position estimation method that improve the accuracy of estimating the current position of an aircraft and enable the aircraft to fly stably for long periods of time. [Means for solving the problem]

[0010] In order to achieve the above-mentioned object, the position estimation system for an aircraft of the present invention comprises a communication unit capable of receiving time series data regarding an aircraft flying in a specified area via wireless communication, and an estimation unit that estimates the current flight position of the aircraft by executing a specified machine learning algorithm using the time series data received by the communication unit as input values.

[0011] Here, by providing a communication unit capable of receiving time series data relating to an aircraft flying in a specified area via wireless communication, it is possible to receive time series data indicating the flight status of the aircraft via wireless communication.

[0012] In addition, by providing an estimation unit that estimates the current flight position of the aircraft by executing a predetermined machine learning algorithm using time series data received by the communication unit as input values, the current position of the aircraft can be accurately determined with little computational load based on past learned data.

[0013] In addition, when the estimation unit generates array data by reconstructing the time series data based on a predetermined weighting coefficient, it can reconstruct the time series data as array data that requires attention for estimating the current position of the aircraft.

[0014] Furthermore, when the estimation unit has an input layer that generates first array data in which input values ​​are compressed based on a predetermined weighting coefficient, the time series data of the aircraft can be compressed into array data in which features are extracted, so there is no need to calculate interactions between all times in the time series data, and the amount of memory and calculation required for calculation can be reduced.

[0015] Furthermore, when the estimation unit has an intermediate layer that reconstructs the first sequence data to generate second sequence data, it can further optimize noteworthy time series data from the first sequence data compressed in the input layer. More specifically, it can learn correlations between time series data included in the first sequence data and reconstruct them into time series data with high correlation.

[0016] Furthermore, when the estimation unit has an output layer that converts the second array data into output values ​​of the same size as the input values, it can output the time series data compressed by the input layer as time series data of the same size as the input values ​​and characterized by learning.The current position of the aircraft can then be accurately estimated based on the time series data output by the output layer.

[0017] Furthermore, when cross-attention processing using a cross-attention mechanism is performed once for each of the input layer and the output layer, the input layer can reduce the computational load of the machine learning algorithm by compressing the input values ​​to data of a predetermined size.Furthermore, the output layer can output the data learned by the machine learning algorithm as time-series data of the original size.

[0018] In addition, when the intermediate layer performs self-attention processing using a self-attention mechanism multiple times, it can optimize the time series data by repeating the self-attention processing and capture the characteristic data necessary to estimate the current position of the aircraft, thereby improving the accuracy of estimating the current position of the aircraft.

[0019] Furthermore, when the input layer generates first array data by mapping a byte array based on an input value with a latent array having a smaller number of data items than the byte array, the time series data can be compressed into array data characterized based on a latent array determined by random numbers.

[0020] Furthermore, when wireless communication is performed using the LPWA (Low Power Wide Area) method, LPWA is a long-distance data communication method with low power consumption, so battery consumption when sending and receiving data from the aircraft is low, and the aircraft's flight time and distance can be extended.

[0021] Furthermore, if the time series data includes the roll angle, pitch angle, yaw angle, and altitude of the aircraft, the current position of the aircraft can be accurately estimated based on these parameters.

[0022] Furthermore, when the machine learning algorithm is a perceiver or a transformer, it can learn long-term dependencies from long-term time-series data without being affected by past training data. Furthermore, by adopting a perceiver, it is possible to reduce the computational load.

[0023] In order to achieve the above-mentioned object, the position estimation method for an aircraft of the present invention comprises the steps of receiving time series data related to an aircraft flying in a specified area via wireless communication, and estimating the current position of the aircraft by executing a specified machine learning algorithm using the received time series data as input values.

[0024] Through the above process, the time series data of the aircraft can be used as input values, and the current position of the aircraft can be accurately determined based on the learned data with a small computational load.

[0025] Furthermore, the present position of the aircraft is compared with a predetermined flight route, and if the present position of the aircraft deviates from the flight route, the aircraft is controlled autonomously or by an external control signal to return to the flight route, thereby enabling the aircraft to always fly along the correct flight route. [Effects of the Invention]

[0026] The aircraft position estimation system and aircraft position estimation method according to the present invention improve the accuracy of estimating the current position of the aircraft, enabling the aircraft to fly stably for a long period of time. [Brief explanation of the drawings]

[0027] [Figure 1] 1 is a schematic diagram of an entire network configuration including an aircraft position estimation system according to an embodiment of the present invention. [Figure 2] 1 is a block diagram showing the hardware configuration of an aircraft according to an embodiment of the present invention. FIG. [Figure 3] FIG. 2 is a block diagram showing a hardware configuration of an estimation unit according to the embodiment of the present invention. [Figure 4] FIG. 1 is a schematic diagram of a machine learning model applied in a location estimation system. [Figure 5] 3 is a flow chart showing a position estimation flow of an aircraft according to an embodiment of the present invention. [Figure 6] In an embodiment of the present invention, (a) is an autopilot program for an aircraft, and (b) is time-series data acquired by the aircraft. [Figure 7] 10 shows the results of outputting arguments for an autopilot program from time-series data in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0028] Hereinafter, a position estimation system for an aircraft and a position estimation method for an aircraft according to an embodiment of the present invention will be described in detail with reference to the drawings etc. to facilitate understanding of the present invention.

[0029] 1 is a schematic diagram of an overall network configuration according to an embodiment of the present invention. In this embodiment, an aircraft 10 flies within an LPWA (Low Power Wide Area) network, and communication between the aircraft 10 and a position estimation system 20 is possible via a network 40 via an LPWA gateway 30.

[0030] An LPWA network is a network that uses specified low-power radio stations as base stations, and can appropriately apply standards such as SIGFOX, LoRa (LoRa WAN), Wi-Fi HaLow, Wi-SUN, RPMA, Flexnet, IM920, Cat. M1, and Cat. NB1.

[0031] The communication method between the aircraft 10 and the position estimation system 20 is not necessarily limited to LPWA, and can be selected from known communication methods as appropriate. However, LPWA is a low-power, long-distance data communication method, and since the aircraft 10 consumes little battery power when transmitting and receiving data between the aircraft 10 and the position estimation system 20, the flight time and flight distance of the aircraft 10 can be extended.

[0032] [Flying object] The aircraft 10 of this embodiment is an unmanned aircraft that flies by remote control or automatic piloting, such as a drone, multicopter, unmanned aerial vehicle, etc. As shown in the block diagram of Figure 2, the aircraft 10 is composed of a communication unit 11, a data processing unit 12, a sensor unit 13, and a flight control unit 14.

[0033] The communication unit 11 has the function of connecting the aircraft 10 to an LPWA network and transmitting and receiving data and control signals to and from the position estimation system 20 via the network 40.

[0034] The sensor unit 13 is a sensor for detecting the flight state of the flying object 10, and mainly has a roll sensor 131 that measures angular velocity in the roll direction, a pitch sensor 132 that measures angular velocity in the pitch direction, a yaw sensor 133 that measures angular velocity in the yaw direction, and an altitude sensor 134 that measures altitude, and while the flying object 10 is flying, the sensor unit 13 measures time series data related to the flight of the flying object 10. The time series data measured by the sensor unit 13 is input to the data processing unit 12, where various processes are executed and the data is output from the communication unit 11 to the position estimation system 20.

[0035] The flight control unit 14 outputs a control signal to the power mechanism of the aircraft 10 in order to fly the aircraft 10 according to a predetermined flight route. In addition, the flight control unit 14 has a function of outputting a control signal to the power mechanism based on an external control signal in the event of an abnormal situation that requires the aircraft 10 to make an emergency landing due to a system failure or a cyber attack, or in the event that the aircraft 10 deviates from the flight route and needs to be returned to the correct flight route.

[0036] [Location estimation system] The position estimation system 20 has a function of estimating the current position of the aircraft 10 based on time-series flight data transmitted from the aircraft 10 while it is flying in the flight area. The position estimation system 20 is composed of a computer device having a CPU, memory, etc., and as shown in the block diagram of Figure 3, is composed of a communication unit 21 that receives radio waves, an estimation unit 22 that estimates the current position of the aircraft 10 based on a machine learning model using the time-series flight data received by the communication unit 21 as input values, and a learning unit 23 that learns the machine learning model based on training data stored in a memory unit 24.

[0037] The machine learning model N applied in the estimation unit 22 of an embodiment of the present invention is a mathematical model that predicts the current position of the aircraft 10 using time series data of the aircraft 10 as input values, and specifically, the machine learning model uses a perceiver.

[0038] Here, the machine learning model N applied in the estimation unit 22 is not necessarily limited to a perceiver, but can be appropriately selected from known machine learning models such as a long-short-term memory model (LSTM), a convolutional neural network (CNN), or a transformer.

[0039] However, when time series data of the aircraft 10 is used as input values, as in the present invention, it is possible that the flight direction of the aircraft 10 may change after a long period of time has passed, so it is preferable to select a machine learning model that can capture long-term dependencies and has a low computational load.

[0040] In this regard, LSTM is heavily influenced by past events, making it unable to capture long-term dependencies, resulting in large errors in output values, and also requiring a huge amount of calculations, which requires a large amount of memory. CNN, on the other hand, captures local events and is not suitable for capturing long-term dependencies. On the other hand, transformers are less likely to produce errors like LSTMs and CNNs, but because they calculate all interactions in the input time series data, they require a large amount of memory and calculations when dealing with long-term time series data.

[0041] In contrast, the basic concept of a machine learning model for a perceiver is similar to that of a transformer, but because a perceiver calculates interactions between compressed time-series data, it does not require as much memory or computational overhead as a transformer, even if the time-series data is long. Therefore, it is preferable to use a perceiver in order to capture long-term dependencies and reduce the computational load.

[0042] 4 is a schematic diagram illustrating a machine learning model N applied in an embodiment of the present invention. The machine learning model N has a function of using time series data measured by the sensor unit 13 of the aircraft 10 as input values ​​and reconstructing this time series data based on predetermined weighting coefficients to generate sequence data, and is composed of an input cross-attention layer C1 as an input layer, a self-attention layer S as an intermediate layer, and an output cross-attention layer C2 as an output layer.

[0043] The input cross-attention layer C1 is an input layer where a known cross-attention mechanism is implemented, and functions as an encoder that encodes input data. As described above, the input cross-attention layer C1 receives time-series data measured by the sensor unit 13 of the aircraft 10 as input values. This time-series data is data arranged in time series, extracted from data over a certain time interval using a window function. In this embodiment, for example, 30-minute time-series data can be represented as 18,000 x 37 matrix data.

[0044] When time-series data is input to the input cross-attention layer C1, the layer calculates interactions between the time-series data based on a latent array L (an abstract array of random numbers that embeds the features of the time-series data) to extract features from the time-series data. The input cross-attention layer C1 repeatedly extracts time-series data and stores it in the latent array L as needed, thereby learning the latent array L. In this embodiment, the latent array L is a 256 x 128 matrix of data, but the size of the latent array L can be changed as needed.

[0045] Then, first sequence data D1 is generated that is compressed to a predetermined extent based on the extracted feature amount. That is, time-series data consisting of a predetermined byte sequence is mapped with a latent sequence L that has a smaller number of data than this byte sequence, thereby generating first sequence data D1 in which the time-series data is compressed.

[0046] The self-attention layer S is an intermediate layer where a known self-attention mechanism is implemented, and further optimally encodes the data of interest in the first array data D1. Specifically, it calculates the interactions of the first array data D1, calculates weighting coefficients based on the similarity between which parts need to be correlated to obtain correct output data, and derives the optimal solution as the weighted average.

[0047] Then, the second sequence data D2 is generated by optimizing the first sequence data D1 in the self-attention layer S. Note that in order to estimate the current position of the aircraft 10, it is necessary to calculate the direction of movement of the aircraft 10 and the flight distance from the starting point. As a result of self-attention, the correlation between the time series data at the starting point and the end point of the traveling direction of the aircraft 10 is generally high.

[0048] The output cross-attention layer C2 is an output layer that executes a known cross-attention mechanism, similar to the input cross-attention layer C1, and functions as a decoder that outputs the second array data D2 generated by the self-attention layer S as output data of the original size. That is, the output cross-attention layer C2 decodes the second array data D2, in which the features obtained by learning have been emphasized, into the same size as the time-series data that is the input value.

[0049] In the embodiment of the present invention, the cross-attention processing in the input cross-attention layer C1 and the output cross-attention layer C2 is performed once each, and the self-attention processing in the self-attention layer S is performed 12 times.

[0050] The number of times each calculation process is performed is not necessarily limited to the above. In particular, the number of times the self-attention mechanism processes the time-series data to capture features can be changed as appropriate depending on the type and size of the time-series data from which features should be captured.

[0051] The learning unit 23 learns the machine learning model using training data. Specifically, the learning unit 23 repeatedly updates the weighting coefficients in the machine learning model using a large amount of training data, for example, by a known backpropagation method, so that the difference between the output value of the machine learning model and the actual measured value of the output parameter becomes smaller. As a result, the machine learning model is trained, and a trained machine learning model is generated. Information about the trained machine learning model (such as the model structure and weighting coefficients) is stored in the storage unit 24 in the position estimation system 20, and is used as the machine learning model in the estimation unit 22 from the next time onwards.

[0052] Next, the flow of estimating the position of the flying object 10 executed by the position estimation system 20 will be described with reference to FIG.

[0053] First, when the position estimation system 20 receives time series data acquired by the sensor unit 13 of the aircraft 10, the estimation unit 22 calculates the features of the time series data from the interactions of the time series data based on the latent array L in the input cross-attention layer C1 (STEP 1), and generates first array data D1 compressed to a predetermined size (STEP 2).

[0054] Next, the first sequence data D1 is input to the self-attention layer S, and the self-attention mechanism is executed (STEP 3). As described above, the calculation in this self-attention layer calculates the interactions of the first sequence data D1, calculates weighting coefficients based on similarity, and derives the optimal solution as the weighted average. Then, the second sequence data D2 is generated by optimizing the first sequence data D1 (STEP 4).

[0055] The second array data D2 generated in STEP 4 is decoded in the output cross-attention layer C2 to the same size as the time-series data of the input values ​​(STEP 5), and the current position of the flying object 10 is estimated based on the output data.

[0056] The current position of the aircraft 10 estimated by the estimation unit 22 is then compared with the predetermined flight route, and if the current position of the aircraft 10 is flying along the flight route, no special instructions are given to the aircraft 10, allowing it to continue flying. On the other hand, if the aircraft 10 has deviated from the flight route, a control signal is sent to the aircraft 10 to return to the flight route. Upon receiving the control signal, the aircraft 10 outputs a control signal from the flight control unit 14 to the power mechanism of the aircraft 10, allowing the aircraft 10 to autonomously return to the flight route.

[0057] Next, an embodiment of the present invention will be described. First, Fig. 6(a) shows a Python program for autopiloting an aircraft 10. This program (1) ascends 5 meters after takeoff, (2) remains stationary for 10 seconds, and (3) moves west at a speed of 5 meters per second for 2 seconds.

[0058] Figure 6(b) shows time series data measured by the aircraft 10 flying according to this program. In this embodiment, to estimate the current position of the aircraft 10, a perceiver is used as a machine learning model to infer the values ​​of the arguments of the autopilot program from the time series data. To achieve this, the output of the machine learning model in Figure 4 is configured to be a multivariate linear regression model.

[0059] Figure 7 shows the results of outputting arguments for the autopilot program from time-series data using Hugging Face's perceiver code. The values ​​enclosed in white boxes are the correct value (top) and the calculated result (bottom). We can see that the error in the calculated result is a few percent.

[0060] As described above, the aircraft position estimation system and aircraft position estimation method to which the present invention is applied improve the accuracy of estimating the aircraft's current position, allowing the aircraft to fly stably for long periods of time. [Explanation of symbols]

[0061] 10 Flying Objects 11 Communications Department 12 Data processing section 13 Sensor section 131 Roll sensor 132 Pitch sensor 133 Yaw sensor 134 Altitude Sensor 14 Flight control unit 20 Location Estimation System 21 Communications Department 22 Estimation part 23 Learning Department 24 Memory section 30 LPWA gateways 40 Network N machine learning models D1 First array data D2 Second array data L latent array C1 Input Cross Attention Layer C2 Output Cross Attention Layer S Self-Attention Layer

Claims

1. a communication unit capable of receiving, via wireless communication, time series data relating to an aircraft flying in a predetermined area, the time series data including a roll angular velocity, a pitch angular velocity, a yaw angular velocity, and an altitude of the aircraft, but not including position information data based on a GPS signal; an estimation unit that estimates the current flight position of the aircraft by executing a predetermined machine learning algorithm using the time series data received by the communication unit as an input value. Aircraft position estimation system.

2. A step of receiving, via wireless communication, time series data relating to an aircraft flying in a predetermined area, the time series data including the roll angular velocity, pitch angular velocity, yaw angular velocity, and altitude of the aircraft, and not including location information data based on GPS signals; and executing a predetermined machine learning algorithm using the received time series data as an input value to estimate the current position of the aircraft. A method for estimating the position of an aircraft.

3. Comparing the current position of the aircraft with a predetermined flight route; and when the current position of the flying object deviates from the flight route, controlling the flying object by an external control signal so as to return to the flight route. The method for estimating the position of an aircraft according to claim 2.

Images