Patent Application
20180162531
The invention relates to air drops, and in particular, to determining when to release a payload.
In many cases, it is desirable to deliver a payload by dropping it from an airplane. A difficulty that arises is making sure that the payload lands where it is intended to. This requires consideration of when the delivery airplane should release the payload.
One approach to determining when to release the payload is to rely on ballistics. Given the airplane's speed and altitude, the acceleration due to gravity, and the target location, it is a relatively simple matter to calculate the correct release point.
Unfortunately, the presence of wind tends to make this calculation inaccurate.
Another approach is to use wind forecasts to predict the wind in the vicinity of the drop zone. However, this method assumes the forecast is accurate at the local level. In many cases, the forecast has difficulty predicting winds at a local level, particularly as one approaches the ground.
A suitable way to overcome this difficulty is to use the delivery airplane to drop a dropsonde. As the dropsonde falls, it collects data concerning the conditions that it encounters and sends that data back to a processor on board the delivery airplane. The processor then calculates the optimal release point based on this data and proceeds to drop the payload from the release point.
A disadvantage of the foregoing method is that the airplane must now make two passes: a first one to drop the dropsonde and a second one to drop the payload. Aside from requiring more fuel, in some circumstances, this exposes the airplane to hostile fire twice instead of once, and pinpoints release location of a payload in advance.
In one aspect, the invention features a method that includes delivering a payload to a drop zone. The process of delivering the payload includes selecting a remote location at which to measure a remote wind stick and doing so based at least in part on current weather conditions. This is followed by determining a release point, flying to it, and then dropping the payload from the release point. The process of determining the release point includes estimating a wind stick at the drop zone based at least in part on the measurement of the remote wind stick.
Among the practices of this method are those that also include selecting a remote location at which to measure the remote wind stick, and to do so at least in part on the basis of a confidence level of the estimate of the wind stick at the drop zone. This confidence level is determined at least in part on the basis of the current weather conditions.
Also among the practices of the invention are those that include estimating a confidence level of the estimate of the wind stick at the drop zone based on the current weather conditions.
Also among the practices of this method are those that include releasing a dropsonde at the remote location, collecting data from the dropsonde, and determining the release point based at least in part on the collected data.
In yet other practices, determining the release point comprises providing the remote wind stick to an operational model that has been configured to provide a transformation based on historical weather data and applying the transformation to the remote wind stick to obtain an estimate of a drop-zone wind stick at the drop zone.
Some practices further include obtaining the operational model by providing a training model with historical weather data and simulated wind sticks.
Other practices include obtaining the operational model by providing a training model with training data, and using boosted regression trees to identify transformations to transform a measurement of a wind stick at a first location into an estimate of a wind stick at a second location.
In another aspect, the invention features a method that includes obtaining a first value of a variable, providing the first value to an operational model that has been configured by a machine-learning algorithm to provide a transformation to yield an estimate of a value of the variable at a second location, and transforming the first value of the variable to obtain a second value of the variable. The second value of the variable is an estimate of a value of the variable at a remote location. This first value is one that is obtained by measuring at a measurement location that differs from the remote location.
In some embodiments, the remote location is below a region that contains the measurement location. In other embodiments, the remote location and the measurement location are separated along a circumferential direction that is perpendicular to a line extending to the center of the Earth.
As the dropsonde 16 falls, it periodically measures wind vectors. The result is a set of wind vectors along the dropsonde's path. This set of wind vectors will be referred to herein as a “wind stick.” As shown in the figure, the dropsonde 16 transmits a first wind stick back to a receiver 20 on the aircraft 10. This first wind stick consists of wind vectors measured in the remote location 18.
Aboard the aircraft 10, a processor 22 receives, from the receiver 20, data representative of the first wind stick. The processor 22 then retrieves, from a memory 24, a suitable transformation. This transformation is obtained during a training phase of a machine-learning procedure to be discussed below. The processor 22 then applies this transformation to the first wind stick to obtain a second wind stick. This second wind stick represents an estimate of the wind present at the drop zone 12. Based on this second wind stick, the processor 22 determines an appropriate release point 26 at which the aircraft 10 should release the payload 14.
The aircraft 10 then continues until it reaches the release point 26, as shown in
The distance between the remote location 18 and the drop zone 12 is essentially a race between the aircraft 10 and the processor. The shorter this distance is, the greater will be the accuracy of the estimate. On the other hand, sufficient time must be allotted for the processor 22 to actually determine the release point 26. Otherwise, the aircraft 10 have to circle around possibly hostile airspace while waiting for the processor 22 to complete its work. In a typical operating environment, a distance of approximately one-hundred kilometers has been found to be practical.
The process of providing a selection of transformations that can be applied to the first wind stick to estimate the second wind stick is the product of a training procedure in which a machine-learning algorithm uses historical data to identify a pattern of differences between a wind stick at a selected zone and a wind stick at a remote location 18.
Referring now to
In the same particular embodiment, the expected output 34 is represented through analysis data that best estimates the actual wind stick over the drop zone 12. This will make it possible to estimate the error between the wind stick data from the historical data 32 and the target wind stick at the drop zone 34. This estimate in the error provides a basis for estimating uncertainty in the prediction at the drop zone. In general, this uncertainty varies with the weather conditions that prevailed at the time that the estimate was made.
The training proceeds with the application of any of a variety of data-mining and machine-learning algorithms 36. However, a particularly useful machine-learning algorithm 36 relies on boosted regression-trees. Such an algorithm has been found to accurately assimilate data obtained by a dropsonde from a remote location 18.
Accordingly, unlike conventional assimilation methods, which rely exclusively on static data, the method described in
The ability to estimate the extent of this match also provides a way to choose an optimal remote location 18 at which to measure a remote wind stick and to do so as a function of the weather conditions at the time of delivery of the payload to the drop zone 12. This differs from the conventional approach in which the optimal remote locations 18 are given in advance without having considered the actual weather conditions at the time of the delivery.
This distinction arises in part because the use of machine learning is able to avoid predictions based on static data that do not consider current weather conditions. As a result, the transformation that is applied to the measured wind stick at the remote location 18 need not be a static transformation that would be the same regardless of current weather conditions. Instead, it is a dynamic transformation that responds to current weather conditions. This offers a distinct advantage in environments in which local weather conditions exhibit high variance.
The model is trained on a pressure-level basis, with each pressure level generally corresponding to an altitude at which data is collected. It has been found that thirty pressure levels is adequate to for normal operation.
The data relied upon is by no means restricted to wind speed and direction. Any observable data is fair game for use in the training procedure. These observables include wind, temperature, and humidity forecasts in the vicinity of the drop zone 12 and the remote location 18, as well as similar features extracted from dropsonde data. A suitable and widely available source of data is that provided by NOAA based on historical weather maps. In addition, numerous other forms of data can be created using principal component analysis or other feature development methods. This uses data over an extended wind field evaluated within an area surrounding the drop zone 12. The extent of this area varies with local factors and desired performance. However, a suitable area is a circular area having a diameter of approximately one hundred kilometers.
Examples of data that have been found useful are listed below. In the following list, the first and second components of the wind vector are orthogonal and lie in a plane parallel to the Earth. Any component of the wind vector perpendicular to this plane is ignored, but in principle, need not be. The wind stick, being a set of wind vectors, is regarded as having a component that depends on its constituent wind vectors. Since each wind vector is a vector, it can be resolved into components that point in orthogonal directions. For example, a wind vector may have a component along a north-south direction and an orthogonal component along an east-west direction. As used herein, a “vector-component” is used to refer to any one of the constituent orthogonal components of a wind vector.
A humidity stick is a similar set of measurements of humidity, which, being a scalar, makes the humidity stick a scalar. The data that has been found useful for training includes:
The outcome of the training procedure is an operational model 30 that can be carried with the aircraft 10 and used for near-real time decision-making 38 based on operational measurements 40 that would include the dropsonde data referred to in connection with
The preceding method thus amounts to spatial rather than temporal weather forecasting. In temporal weather forecasting, the question is usually, “Given the weather today, what will the weather be tomorrow?” In the method described herein, the question instead becomes, “Given the weather over here, what is the weather over there?”
The preceding method can also be regarded as a form of computational remote-sensing. In conventional remote-sensing, one makes a physical measurement at a remote location, typically by observing a disturbance to a wave that has passed through that location. Examples include the use of Doppler radar, for electromagnetic waves, and the sensing of various subsurface structures of geological interest, using acoustic waves.
Rather than attempt to make such measurements, the method described herein achieves a similar result by carrying out local sensing and using the results of such local sensing, together with historical data, to computationally infer or estimate conditions at a remote location.
The methods described herein find particular application in weather-related phenomena. However, the techniques are general enough to be used in other applications. For example, in principle machine learning of the same type can be used to estimate ocean currents at otherwise inaccessible locations, or to infer the existence of subsurface structures based on measurements closer to the surface.
In the application described in connection with
In particular, as the parafoil descends, the Earth's atmosphere is continuously being divided into two parts: an upper part through which the parafoil has already descended, and a lower part through which the parafoil has yet to descend. As the parafoil descends, it makes measurements similar to those made by the dropsonde. In effect, the parafoil is acting as its own dropsonde.
The measurements made by the parafoil thus provide a growing body of measurements for the upper part. These measurements can then be used to predict conditions within the lower part. In anticipation of these conditions, the parafoil's control surfaces can be adjusted to adaptively guide it towards the drop zone based at least in part on its observations of atmospheric conditions in the upper part.
This embodiment thus carries out a procedure similar to that discussed in connection with
Additional details and experimental results are provided in the attached Appendix, which is incorporated herein by reference.
This application claims the benefit of the Dec. 9, 2016 priority date of U.S. Provisional Application 62/432,187, the contents of which is incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 62432187 | Dec 2016 | US |
