MISSION PLANNING FOR WEAPONS SYSTEMS

Abstract

A mission planning method for use with a weapon is disclosed. The method comprises: obtaining a first training data set describing the performance of the weapon; using the first training data set and a Gaussian Process (GP) or Neural Network to obtain a first surrogate model giving a functional approximation of the performance of the weapon; and providing the first surrogate model to a weapons system for use in calculating a performance characteristic of the weapon during combat operations.

Claims

1. A mission planning method for use with a weapon, the method comprising the steps of: obtaining a first training data set describing the performance of the weapon; using the first training data set and a Gaussian Process (GP) or Neural Network to obtain a first surrogate model giving a functional approximation of the performance of the weapon; providing the first surrogate model to a weapons system for use in calculating a performance characteristic of the weapon during combat operations.

2. A mission planning method according to claim 1, the method comprising using the training data set and a Gaussian Process (GP) to obtain a surrogate model comprising a covariance function, a set of hyper-parameters and a set of weighted values.

3. A mission planning method according to claim 2, wherein the surrogate model further comprises a set of inducing points.

4. A mission planning method according to claim 2, wherein the Gaussian Process algorithm used is the Fully Independent Training Conditional (FITC) algorithm.

5. A mission planning method according to claim 1, the method comprising using the training data and a Neural Network to obtain a surrogate model comprising an activation function or a basis function, and a set of Neural Network parameters.

6. A mission planning method according to claim 1, further comprising calculating a performance characteristic of the weapon during combat operations using the surrogate model.

7. A mission planning method according to claim 6, further comprising initiating launch of the weapon in dependence on the performance characteristic so calculated.

8. A mission planning method according to claim 1, the method comprising the steps of: obtaining a second training data set describing the performance of the weapon in a second, different, parameter space to the first training data set; using the second training data set and a Gaussian Process (GP) or Neural Network to obtain a second, different, surrogate model giving a functional approximation of the performance of the weapon in the second parameter space; providing the first and second surrogate models to a weapons system for use in calculating a performance characteristic of the weapon during combat operations.

9. A mission planning method according to claim 8, further comprising, during combat operations, selecting the first or second surrogate model in dependence on the current parameters and using the surrogate model so selected to calculate a performance characteristic of the weapon.

10. A mission planning method according to claim 1, wherein the weapon is a missile.

11. A mission planning method according to claim 1, wherein the weapon system comprises a weapons platform and the weapons platform is an aircraft, ship or land vehicle.

12. A mission planning method according to claim 1, wherein the performance characteristic is the Launch Success Zone (LSZ), the Launch Acceptable Region (LAR), the footprint, the aerodynamic drag of the weapon and/or the trajectory of an enemy weapon.

13. A weapons system comprising a processor programmed with software configured to calculate a performance characteristic of a weapon of the weapons system during combat operations using a functional approximation of the performance of the weapon, said functional approximation comprising a surrogate model produced using a Gaussian Process or neutral network.

14. A weapons system according to claim 13, further comprising a launcher, and wherein the launcher comprises the processor.

15. A weapon configured for use as the weapon of the weapons system of claim 13, wherein the weapon comprises the processor.

16. A weapons system according to claim 13, further comprising a weapons platform, wherein the processor is part of the command and control system of the weapons platform.

17. A computer software product for loading onto a processor associated with a weapons system, wherein the software product is configured to calculate a performance characteristic of a weapon of the weapons system during combat operations using a functional approximation of the performance of the weapon, said functional approximation comprising a surrogate model produced using a Gaussian Process or neutral network.

18. A computer software product according to claim 17, wherein the surrogate model is produced using a Gaussian Process and the surrogate model comprises a covariance function, a set of hyper-parameters and a set of weighted values.

Description

DESCRIPTION OF THE DRAWINGS

[0032] Embodiments of the present invention will now be described by way of example only with reference to the accompanying schematic drawings of which:

[0033] FIG. 1 shows a mission planning process according to a first embodiment of the invention;

[0034] FIG. 2 shows part of a weapons platform configured for use with the process of the first embodiment;

[0035] FIG. 3 shows part of a weapons platform configured for use with the process of the first embodiment;

[0036] FIG. 4 shows a schematic view of a parameter space for use in a method according to a second embodiment of the invention; and

[0037] FIG. 5 shows a schematic view of a parameter space for use in a method according to a third embodiment of the invention.

DETAILED DESCRIPTION

[0038] FIG. 1 shows a process for calculating the Launch Acceptability Region (LAR) of a missile in accordance with a first example embodiment of the invention. At the highest level the process comprises three stages (in order); training data generation 1; determining a surrogate model 2 for calculating LAR; and an operational step 3, where the surrogate model produced in step 2 is used in deciding whether to launch the weapon at a target.

[0039] In order to calculate the LAR of a missile it may be necessary to approximate four functions associated with a given engagement situation: IR-Outer, IR-Inner, IZ-Outer and IZ-Inner. IR refers to in-range and denotes the weapon attainability boundary for an engagement with no explicit user specified constraints. IZ refers to in-zone which may further include user specified constrains such as demanded missile impact heading, cruise altitude, specified way-points and run-in distance. This example will consider the calculation of one of these functions, but it will be appreciated that a similar process may be applied to the other functions. It will be appreciated that different parameters may be used in the calculation of different functions. The parameter R to be approximated may be formulated as a function LAR of the parameters , H, v, as follows:

R=LAR(,H,v,)

[0040] Where is the angle of launch position with respect to the target (deg), H is the launch altitude (m), v is the launch speed (m/s) and is the pitch/dive angle at impact (deg). In the training data generation step 1, a range of values for each of the parameters , H, v, are input to a kinematic model. The kinematic model is then run multiple times 4 with different combinations of parameter values to produce a set of training data 6 and a set of validation data 8 describing the variation of R over the parameter space.

[0041] In the surrogate model production step 2, the training data 6 is prepared 10. This comprises formatting the functional data from the kinematic model into pairs of input parameters (i.e. one combination of inputs X=(, H, v, ) and the corresponding function value R(X)). This data sets represents noisy and sparse observations of the true continuous underlying LAR function. After preparation the training data is input into a FITC algorithm (Fully Independent Training Conditional approximation as described in A unifying View of Sparse Approximate Gaussian Process Regression by Quinonero-Candela J. & Rasmussen C. E., Journal of Machine Learning Research, Vol. 6, pp 1939-1959, 2005, available as part of GPML Matlab Code version 4.0). In the FITC approach the pseudo or inducing-points u are treated as hyper-parameters to be optimised. Thus, the LAR approximation requires the following hyper-parameters 14 to be generated;

.sub.,.sub.H,.sub.v,.sub.,f,X.sub.u,w

[0042] Where .sub.,.sub.H,.sub.v,.sub., are length-scale parameters learned during training, .sub.f is an overall scale factor determined from training, X.sub.u represents the induction points determined in training and w represents a weighted output value, one per induction point, derived from the covariance function (see below) and .sub.n (the noise parameter). These hyper-parameters 14 are calculated 12 using the FITC algorithm and a squared exponential covariance function 15 with Automatic Relevance Detection (ARD). Once calculated 12, the hyper-parameters 14 are passed to an evaluation step 18 which compares the predicted values calculated using a covariance function employing those parameters 14 with the validation data 8 to verify that the resulting surrogate model is sufficiently accurate. The covariance function 15 corresponding to the GP and hyper-parameters 14 are then incorporated 16 into a playback algorithm 19, for use in stage 3. Stages 1 and 2 of the method are carried out off-line, and separate from any weapons platform.

[0043] To calculate R the following covariance function is used:

R*=K(x.sub.u,x*).Math.w

Where K( ) is the squared-exponential covariance function:

[00001]$\begin{array}{c}\mathrm{cov}\left(f\left({\left(x_u\right)}_i\right),f\left(x_i^{*}\right)\right)=.Math.K\left({\left(x_u\right)}_i,x_j^{*};\right)\\ =.Math.{}_f^2.Math.\exp \left(-\frac{1}{2}.Math.{\left({\left(x_u\right)}_i-x_j^{*}\right)}^T.Math.{\mathrm{diag}\left({}^2\right)}^{-1}.Math.\left({\left(x_u\right)}_i-x_j^{*}\right)\right)\\ .Math.\mathrm{for}.Math..Math.\left\{i,j\right\}\end{array}$

[0044] and ={.sub.f,.sub.1,.sub.2, . . . } are the learned amplitude and length-scale hyper-parameters, (x.sub.u).sub.i 1im is the i.sup.th induction point, x.sub.j* 1jp is the j.sup.th input/test point, p is the number of test points, .sub.f is the scale factor parameter determined from training, and .sup.2=(.sub..sup.2,.sub.H.sup.2,.sub.v.sup.2,.sub..sup.2).

[0045] During flight operations 3, the playback algorithm 19 embodying the covariance function 15 and hyper-parameters 14 is used to calculate 20 the function R at any given instant. The other functions required to calculate the LAR are similarly calculated. The prediction of the LAR is continually updated as engagement conditions change and this information is provided to the pilot who uses that information to decide 22 whether to launch 24 the missile against a given target.

[0046] In testing the FITC algorithm was found to give 50 m Root Mean Square (RMS) errors (with all better than 400 m absolute error) when the number of induction points is 10% of the number of training data points, and 330 m RMS (with all better than 2 km absolute worst error) when the number of induction points is 2.5% of the number of training data points. Depending on where the acceptable accuracy was defined, this allows a trade-off in playback speed in the range 20 kHz-88 kHz for estimation of the LAR vertices (equivalent to 1 to 4 Kilo-LARs/second) when using MATLAB 2012b on an HP840 Laptop equipped with an intel core i5-4300U@1.9/2.9 GHz-Boost CPU and executing on a single thread with no other applications running.

[0047] FIG. 2 shows part of an aircraft 100 for use with the method of the first embodiment. The aircraft 100 has a fuselage 102 and a wing 104, and a missile 108 mounted on a launcher 106 located on the underside of the wing 104. A processor 110 programmed with the playback algorithm 19 embodying the covariance function 15 and having access to the hyper-parameters 14 is located within the fuselage 102 and forms part of the command and control (C2) system (not shown) of the aircraft 100. In use, data representing the current conditions and location of the aircraft 100 and a target (not shown) is provided to the processor 110 which uses the covariance function 15 and hyper-parameters 14 to calculate the LAR for that target. The LAR is provided to the pilot who then uses that information in deciding whether to launch the missile 108 as discussed above. In other embodiments, the missile may be located in a bomb bay, internal to the aircraft.

[0048] FIG. 3 shows a variation of the arrangement of FIG. 2. The same reference numerals denote substantially similar elements. Only those aspects of FIG. 3 which differ significantly from the FIG. 2 arrangement will be discussed. In the arrangement of FIG. 3 the processor 110 is located within the missile 108 and provides a LAR to the aircraft command and control (C2) system (not shown) which relays this information to the pilot. As the covariance function 15 and hyper-parameters 14 are provided with the missile 108, weapons systems in accordance with the present example embodiment may facilitate interoperability and maintenance as there is no need to update on-board software to reflect changes in missile performance; this information is provided as part of the missile itself through the hyper-parameters 14 and covariance function 15.

[0049] In a variation of the process of FIG. 1 more than one set of training data may be generated at step 4; each set of training data corresponding to a different zone within the engagement parameter space. FIG. 4 shows a schematic depiction of a cuboidal parameter space 200, with a first parameter A increasing along the x-axis, a second parameter B increasing along the y-axis and a third parameter C increasing along the z-axis. The parameter space 200 has been divided into four zones 208a, 208b, 208c and 208d. Each of the zones 208a, 208b, 208c occupies a separate volume of the cuboidal space 200, with the fourth zone 208d representing the space not falling within the first three zones 208a, 208b, 208c. The first zone 208a is immediately adjacent to the second zone 208b. The third zone is spaced apart from both the first zone 208a and the second zone 208b. The training data for a given zone is generated using combinations of parameters falling within, and immediately adjacent to, the zone. Thus, a set of training data 6a and verification data 8a is obtained for zone 208a and so on for each of zones b to d. Each set of training data is then prepared and the FITC algorithm used to produce a set of hyper-parameters (including inducing points) for each zone. If the behaviour of the missile is different between different zones then the variables describing the LAR may differ leading to different sets of hyper-parameters for each zone. A different covariance function may also be used for each zone. Each covariance function 15 and set of hyper-parameters 14 may then be passed to the aircraft 100 for use in operations 3. In flight, the step of predicting the LAR using the covariance function 15 may first comprise identifying which zone the currently observed parameters are located in. The covariance function 15 and hyper-parameters 14 are then used to predict the LAR. Methods in accordance with the present embodiment may further reduce the amount of computation that must be carried out by the weapons system as only the inducing-points u relating to the current zone need be considered during the playback calculation.

[0050] In a further variation, different correction factors may be applied to each of the different zones 208. For example, if in use, the missile performance is found to be different from that predicted in a given zone 208, the results produced by the covariance function 15 corresponding to that zone may be scaled accordingly. In contrast to prior art methods where this would have required a reworking of the kinematic model and consequently significant reprogramming of the weapons system, the present embodiment allows such scaling to be carried out by varying a single correction parameter. Accordingly, systems using the present embodiment may be more flexible and easier to update than prior art systems.

[0051] FIG. 5 shows a variation of the parameter space 200 and zones 208 of FIG. 4. The same reference numerals denote substantially similar elements. Only those aspects of FIG. 5 which differ significantly from FIG. 4 will be discussed. In FIG. 4, each of the four zones 208 occupies a different region of the parameter space 200. In FIG. 5, three zones 208a to c are shown, and the first zone 208a overlaps with and is contained completely within a second zone 208b. A third zone 208c partially overlaps with zone 208b at a location spaced apart from zone 208a. Use of such zones may allow performance of the missile to be limited in a particular region, for example if missile launch in zone 208a posed unacceptable risks, the output of the covariance function 15a could be scaled such that a LAR is rarely achievable within this zone. In use, a list giving the order in which zones are to be investigated in order to identify whether a given parameter combination is located within that zone is passed to the weapon system along with the covariance function 15 and hyper parameters 14. For example, a point in zone 208a is also geometrically in zone 208b, thus the algorithm must check whether a point is in zone 208a before considering if the point is in zone 208b in order for the zone 208a model to be used.

[0052] Whilst the present invention has been described and illustrated with reference to particular embodiments, it will be appreciated by those of ordinary skill in the art that the invention lends itself to many different variations not specifically illustrated herein. By way of example only, certain possible variations will now be described. The above example has been described in the context of a missile mounted on an aircraft, it will be appreciated that the systems and methods described above are equally applicable to sea or land based systems, for example to ships and/or land vehicles and other weapons types. The FITC algorithm discussed above has been found particularly advantageous as it allows the generation of an approximation to full covariance based on m optimised pseudo- or inducing-points u, where m<N (and frequently mN), where N is the number of points in the training data set. With FITC the training complexity is of O(N.Math.m.sup.2) and playback scales with O(m), this is in contrast with exact inference where the training complexity is of O(N.sup.3) and playback scales with O(N). However it will be appreciated that other GP algorithms may also be used. For example the Subset of Data (SD), Fast-Forward Selection (FFS) and Nystrom algorithms may, in some circumstances, be useful. These algorithms are also described in A unifying View of Sparse Approximate Gaussian Process Regression by Quinonero-Candela J. & Rasmussen C. E., Journal of Machine Learning Research, Vol. 6, pp 1939-1959, 2005. Finally, the applicability zones are discussed above in the context of a three-dimensional space, it will be appreciated that the parameter space, and therefore the applicability zones, may be of a higher dimensionality.

[0053] Where in the foregoing description, integers or elements are mentioned which have known, obvious or foreseeable equivalents, then such equivalents are herein incorporated as if individually set forth. Reference should be made to the claims for determining the true scope of the present invention, which should be construed so as to encompass any such equivalents. It will also be appreciated by the reader that integers or features of the invention that are described as preferable, advantageous, convenient or the like are optional and do not limit the scope of the independent claims. Moreover, it is to be understood that such optional integers or features, whilst of possible benefit in some embodiments of the invention, may not be desirable, and may therefore be absent, in other embodiments.