CN114819111A - Focusing control neural network input sampling method of space synthesis laser explosive-removing system - Google Patents

Focusing control neural network input sampling method of space synthesis laser explosive-removing system Download PDF

Info

Publication number
CN114819111A
CN114819111A CN202210720172.0A CN202210720172A CN114819111A CN 114819111 A CN114819111 A CN 114819111A CN 202210720172 A CN202210720172 A CN 202210720172A CN 114819111 A CN114819111 A CN 114819111A
Authority
CN
China
Prior art keywords
input
neural network
class
sampling
input data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210720172.0A
Other languages
Chinese (zh)
Other versions
CN114819111B (en
Inventor
姜春福
王晓明
高学强
王书峰
闫梦龙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jigang Defense Technology Co ltd
Original Assignee
Jigang Defense Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jigang Defense Technology Co ltd filed Critical Jigang Defense Technology Co ltd
Priority to CN202210720172.0A priority Critical patent/CN114819111B/en
Publication of CN114819111A publication Critical patent/CN114819111A/en
Application granted granted Critical
Publication of CN114819111B publication Critical patent/CN114819111B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Neurology (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Feedback Control In General (AREA)

Abstract

The invention provides a focusing control neural network input sampling method of a space synthesis laser explosion elimination system, which comprises the following specific scheme: obtaining a target distance D; generating class I input data and class II input data according to a sampling method, and performing dimension expansion respectively; respectively carrying out random sampling on the data after the dimensionality expansion; carrying out 1 × 1 convolution calculation on the data after random sampling respectively; performing tiling processing on convolution calculation type I input data and convolution calculation type II input data to generate one-dimensional vectors of the type I input data and the type II input data; the combined one-dimensional vector is used as an input layer of the neural network, enters the trained neural network for calculation, and outputs a vector of the adjustment quantity of the focusing platform; and controlling the displacement platform of the system to complete respective focusing control. The invention is beneficial to improving the beam combination effect of the space combination laser at the target position and improving the striking performance of the space combination laser under the same condition.

Description

Focusing control neural network input sampling method of space synthesis laser explosive-removing system
Technical Field
The invention relates to a focusing control neural network input sampling method of a space synthesis laser explosive-removing system, and belongs to the technical field of laser explosive-removing.
Background
In weapon development and military combat training, obsolete ammunition, overdue ammunition storage, training unexploded ammunition, unexploded ammunition placed or left by violence and terrorism are frequently encountered, and under the condition that the state of the ammunition and explosion controllability is changed in an unknown manner, the ammunition has fatal and unstable risks. On the other hand, in social life and national economic construction, an event that a non-explosive bomb is left in war is often found. The waste cannonball is generally buried underground for a long time, the storage period is already exceeded, the source is unknown, the filling charge is unknown, and the variety is complicated and dangerous. How to safely dispose is also an examination to the emergency disposing capacity of the public security department.
At present, most of domestic laser explosive disposal systems are portable, the power is mostly below kilowatt (such as 500W), the system is more effective for small ammunition and thinner shell ammunition, but for unexploded ammunition with larger ammunition diameter and thicker ammunition body, a laser with hundreds of watts power cannot be effectively detonated, and if the system is improperly used, the uncertain danger of the unexploded ammunition is increased.
To effectively dispose of large unexploded ammunition, a high-power laser explosive disposal system is necessary. Due to the power limitation of the current single mode single laser, in order to obtain a laser with higher power (e.g. more than 10kW and higher), a technical scheme of spatial synthesis, spectral synthesis or spectral synthesis + spatial synthesis is usually adopted to obtain a sufficient thermal damage effect at a target point. When the explosion-eliminating system adopts space synthesis, multiple paths of laser beams of the laser need to be provided with relatively independent focusing control systems, so that each path of laser can be accurately focused at a target, the light spot is minimum, the energy is most concentrated, and the synthetic damage effect is maximum. During practical use, the focusing performance of the explosive disposal system can be affected by various errors, and the method mainly comprises the following steps:
firstly, the laser ranging error of a static target, or the radar ranging error of a moving target, and the like, wherein the errors are usually in the sub-meter level, the meter level or the dozen of meter levels;
secondly, surface shape errors, micro defects, thermal deformation caused by absorption of ambient temperature and laser energy and the like exist in optical components through which each path of light beam passes, the errors are usually in the micron order, the micro radian order or even smaller, but the errors are transmitted through a long distance of the light beam, when the light beam reaches a target point, a smaller error is amplified, and great influence is generated on energy focusing, alignment precision and the like of the target point;
thirdly, the coating of the optical element has performance difference of substrate materials, substrate fouling, coating mode difference and the like, which may influence the absorption of the optical element by laser, thereby causing thermal deformation of the optical element and also influencing the absorption loss and propagation error of laser beams;
mechanism errors, manufacturing and processing errors, motion control errors, repeated positioning errors, installation accuracy errors and the like exist in the mechanical structure, the errors are usually in a micron or millimeter level, and some errors also have great influence on light beam focusing and stable work due to the amplification effect of light beam propagation;
fifthly, low-frequency errors and other random errors caused by atmospheric disturbance in different seasons, different weather and different external field environments are obtained.
Therefore, the errors exist in large scale errors of more than tens of meters or meters and small scale errors of less than millimeters and micrometers, and the error factors with different magnitudes have great influence on the form, the quality and the focusing position of each beam of the laser, so that the focal positions of the system are different during working, and the light spot energy synthesis effect is poor. Meanwhile, the error influence is usually individually different, and each individual difference may be different for each specific hardware system, each specific working environment and each specific detonation target, and is difficult to describe, calculate, predict and control by using a definite expression.
The existing space synthesis laser generally uses an optical formula and a focusing control protocol to calculate a focusing parameter according to the radar or laser ranging result, uses the same parameter to control a plurality of focusing platforms, and performs a plurality of trial references according to experience to obtain a relatively ideal synthetic light spot, instead of setting respective focusing parameters for each path of laser respectively to perform targeted focusing. In this case, the emitting state and the focusing position of each light beam may be different, and thus, an ideal focusing effect and effect combining effect cannot be achieved.
Disclosure of Invention
The invention aims to provide a focusing control neural network input sampling method of a space synthesis laser explosion elimination system, which is beneficial to improving the beam synthesis effect of a space synthesis laser at a target and improving the striking performance of the space synthesis laser under the same condition.
In order to achieve the purpose, the invention is realized by the following technical scheme:
step 1: the focusing platform is moved to obtain a target distance D through laser ranging;
step 2: respectively generating class I input data and class II input data according to the target distance D and the class I key point sampling method and the class II key point sampling method, wherein the method specifically comprises the following steps:
2-1) sampling I key points; class I expansion of the input layer, spaced by distance in the region where the beam can be normally focusedL 1 Uniform sampling and co-selectionPThe key distance value is used as the I-type input of the input layer of the neural network, also called as the large-scale key point characteristic, and the value rule is as follows:
Figure 617747DEST_PATH_IMAGE001
in the above formula, the first and second carbon atoms are,L 1 the size of the separation distance between the fingers,D near representing the closest distance between which the beam can travel,D far represents the farthest distance between which the light beam can move; to obtainPVector composed of I-type key points
Figure 507075DEST_PATH_IMAGE002
；
2-2) sampling II key points; performing class II expansion on the input layer at the target distanceDNear, at intervalsL 2 Uniformly sampling and selecting the [ alpha ], [ beta ] -aD-L ’ /2，D+L ’ /2]Within a domainQThe key distance value is used as the input of class II of the input layer of the neural network, also called the small-scale key point characteristic, and the value rule is as follows:
Figure 524578DEST_PATH_IMAGE003
in the above formula, the first and second carbon atoms are,L ’ by finger andDthe distance is symmetrically selected from the left side and the right side as the value selection range of the class II input,L 2 by finger andDis a central termD-L ’ /2，D+L ’ /2]Symmetrically selecting left and right sidesQA smaller distance apart; to obtainQVector composed of class II key points
Figure 173866DEST_PATH_IMAGE004
；
And step 3: performing dimension expansion on the class I input data and the class II input data respectively;
will vector
Figure 700705DEST_PATH_IMAGE005
Sum vector
Figure 495355DEST_PATH_IMAGE006
Respectively expand intom×m×PAndn×n×Qof a three-dimensional matrix of (2), whereinm、nIs a positive integer not less than 1;
and 4, step 4: respectively randomly sampling I-type input data and II-type input data after dimensionality extension;
for each class I key pointP i Or class II key pointsQ i Randomly sampling within a certain range on the left and right sides
Figure 750886DEST_PATH_IMAGE007
(M =2, 3, 4, 5, … …) points, expanded with key pointsm×m×POrn×n×QThe matrix of (a) is,mandnthe values of (A) may be equal or unequal;
and 5: respectively carrying out 1 × 1 convolution calculation on the I-type input data and the II-type input data which are randomly sampled;
step 6: performing convolution calculation on the I-type input data and the II-type input data to generate one-dimensional vectors of the I-type input data and the II-type input data, and performing convolution calculation on the I-type input data and the II-type input data to generate the one-dimensional vectorsm×m×PAndn×n×Qbecomes a vector which contains
Figure 921974DEST_PATH_IMAGE008
An input element;
and 7: the combined one-dimensional vector is used as an input layer of the neural network, enters the trained neural network for calculation, and outputs a vector containing all the adjustment quantities of the focusing displacement platform
Figure 563039DEST_PATH_IMAGE009
，fnThe number of focusing displacement platforms in the system;
and 8: according tod' 1 、d' 2 ……d' fn And respectively controlling the focusing displacement platform of the system to complete respective focusing control.
Preferably, the specific scheme of the step 5 is as follows:
in thatn×n×QIn the matrix of (2), hasQGroup ofn×nEach of whichn×nComprises a matrix ofQ i And randomly sampled around itn 2 1 number of data, thisn 2 Multiplying the data by a 1 × 1 filter, respectively, to output a new onen×nMatrix, each data is reduced or enlarged according to the same proportion;Qgroup ofn×nMatrix mappingQGroup 1X 1 filters, outputting onen×n×QA matrix of (a);
in thatm×m×PIn the matrix of (2), hasPGroup ofm×mEach of whichm×mComprises a matrix ofP i And randomly sampled around itm 2 1 number of data, thism 2 Multiplying the data by a 1 × 1 filter, respectively, to output a new onem×mMatrix, each data is reduced or enlarged according to the same proportion;Pgroup ofm×mMatrix mappingPGroup 1X 1 filters, outputting onem×m× PA matrix of (a);
preferably, the neural network samples the number of the two types of inputsQIs greater thanPIn (1).
Preferably, the focusing stage is movedD near OrD far When the mechanical limit or the software limit is reached, the values are taken according to the size of the intervalL ’ /QSelecting [2 ]D，D+L ’ ]Or [ alpha ], [ alphaD-L ’ ，D]And taking Q key distance values in the domain as input of a neural network input layer II type.
Preferably, the focusing stage is movedD near OrD far Not reaching the mechanical limit or the software limit but providing a control marginl ’ Is not sufficient to satisfyD near -L ’ /2 orD far +L ’ /In case 2, the value is taken according to the interval sizeL ’ /QSelecting [2 ]D-l ’ ，D+L ’ -l ’ ]Or [ alpha ], [ alphaD-L ’ +l ’ ，D+l ’ ]And taking Q key distance values in the domain as input of a neural network input layer II type.
Preferably, the focusing stage is movedD near OrD far The mechanical limit or the software limit is not reached, and the control margin isl ’ Can be full ofFootD near -L ’ /2 orD far +L ’ /In case 2, the [2 ] is selected symmetricallyD-L ’ /2，D+L ’ /2]Left and right sidesQEach interval has a size ofL ’ /QAs the neural network input layer class II input.
The invention has the advantages that: the invention increases the input information of the neural network through the sampling expansion design of the input layer of the neural network, and provides an application method for simultaneously outputting a plurality of focusing control parameters and respectively controlling each path of light beam focusing platform for a high-power laser adopting a space synthesis mode. The method comprehensively considers the influences of various nonlinear factors which cannot be expressed by definite expressions, such as ranging errors, the difference of the structure and the performance of each path of light beam component, the difference of mechanical precision, atmospheric disturbance and the like, outputs the focusing control quantity of the multiple paths of light beams through the artificial neural network, realizes the simultaneous and respective control of the multiple paths of light beams, is favorable for improving the light beam synthesis effect of the space synthesis laser at a target position, and improves the striking performance of the space synthesis laser under the same condition.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention.
FIG. 1 is a schematic diagram of the optical path spatial synthesis according to the present invention.
FIG. 2 is a schematic diagram of a single input neural network.
FIG. 3 is a schematic diagram of class I input sampling according to the present invention.
FIG. 4 is a schematic diagram of class II input sampling according to the present invention.
Fig. 5 is a multiple input neural network architecture.
FIG. 6 is a schematic diagram of the dimension expansion of the class I and class II keypoint vectors of the present invention.
FIG. 7 is a schematic diagram of class I and class II keypoint expansion according to the present invention.
FIG. 8 is a diagram of the random sampling spreading according to the present invention.
FIG. 9 is a schematic view of class I keypoint processing according to the present invention.
FIG. 10 is a schematic diagram of the 1 × 1 convolution process of the present invention.
FIG. 11 is a diagram illustrating a neural network process according to the present invention.
FIG. 12 is a flow chart of a neural network construction method according to the present invention.
Fig. 13 is a schematic diagram of an application flow structure of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The conditions to which the calculation method is applied are:
the multiple emergent beams are combined at the target, namely a spatial combination mode: the laser adopts multi-channel laser beams to independently emit, and is synthesized at a target point in a space synthesis mode. The single-path laser can be a laser beam only containing one wavelength, or a laser beam obtained by performing spectrum synthesis by adopting a plurality of single-mode single-fiber lasers, as long as emergent light is a beam;
the focusing platform of each laser beam can be independently controlled: the focusing of each path of laser has an independent movable focusing platform which can be independently controlled, but the selection type, the performance and the like of the movable focusing platform are not required to be completely consistent.
One, construct the neural network of single input
Generally, laser striking has a certain effective working distance, which is related to a working scene and working requirements, for example, a laser with light output power of 6kW has a maximum effective distance of 1000m when striking a target, but can emit light at longer distances of 1100m, 1200m, and the like, and the striking time is slightly longer due to some influences and attenuations in the atmospheric transmission process. The general reference to 1000m is a technical indicator of the system and not an inexorable capability indicator. In consideration of the acquirement and consistency of training samples, the effective working distance mentioned herein is subject to the technical index value.
Is provided with a systemN（NNot less than 1) paths of light beams are synthesized in space, and the moving range of the movable focusing platform of each path of light beams is controlledd si ,d ei ]（i=1，2，……N) Moving the focusing platform in [2 ]d si ,d ei ]The movement in the interval can adjust each light beam in the rangeD near ，D far ]Is effectively focused within a range of distances. When the system works, all the light beams are in the positionD near ，D far ]Focusing and synthesizing the inner target points by the distance of the target pointsD. By constructing neural networks, creatingDAndNfocusing position of movable focusing platformd i Effectively nonlinear mapping relationship between them.
In general, the focus control of the space-combining laser is performed at a target distanceDAnd focal length of lensfThe required focusing distance is solved by the relation formula between the two, and the required focusing distance is converted into the input control parameter of the movable focusing platform to realize focusing. For easy understanding, let us set the input control parameter of the movable focusing platform as a certain position of the movable focusing platformdI.e. the range of movement of the platformd si ,d ei ]A value of (1). In the actual control, the control system is used,dthe real control parameters (such as pulse number and the like) of the movable focusing platform have a definite calculation relation, and the calculation relation is converted by using an interface or the calculation relation, so that the algorithm is not substantially influenced.
For a hit target, the input of the conventional neural network only needs one distanceD，Output is asN（NNot less than 1) the position of the movable focusing platform, as shown in FIG. 2. Theoretically, the single-input neural network structure can also be completedD→（d 1 ,d 2 ,d 3 ) To establish bothThe input and output relationship between the two networks is limited, but the learning speed and universality of the neural network are limited due to too little input information. When an actual laser works, the existing distance measurement errors, optical components, thermal deformation, mechanism structures, control errors and the like exist all the time in the working process of the focusing platform, and even if the laser is at a fixed position of the focusing platform, the random errors of the laser also change and cannot be measured. If the information of the input layer is too single and rare, it is difficult to fully considerD near ，D far ]The total comprehensive error influence on the focusing platform under all target distances in the interval is difficult to highlight the focusing positiondThe influence of local random errors nearby is limited by the structure of the neural network and the time consumption of calculation, and the method cannot be used for solving the problem that the error of the local random errors nearby is influencedD near ，D far ]All values within the interval are used as input values for the neural network. Therefore, further improvement and expansion on the basis of a single-input neural network structure are needed.
Class II and I keypoint sampling
First, class I extensions are made to the input layer. Is arranged in the adjustable interval of the movable focusing platformd si ,d ei ]In the interior of said container body,Nthe road beam can be inD near ，D far ]Normal focus within the interval.
In the intervalD near ，D far ]Uniformly sampling according to a certain distance interval, and selectingPThe key distance value is used as the I-type input of the input layer of the neural network and is also called as the characteristic of the large-scale key point. The value rule is as follows:
Figure 279323DEST_PATH_IMAGE010
in the above formula, the first and second carbon atoms are,L 1 refers to the size of the separation distance. For example, in a certain practical system, theD near ，D far ]Is taken to be [200, 1000 ]]I.e. the working distance of the laser is between 200m and 1000m, the separation distance is large or smallL 1 Can get100m, namely taking one point every 100m as the input of the neural network class I, and totally 9 points.
Class III, II keypoint sampling
Second, class II extensions are made to the input layer. At the target distanceDNearby, uniformly sampling according to a certain interval, selecting the [2 ]D-L ’ /2，D+L ’ /2]Within a domainQThe key distance value is used as input of a neural network input layer class II, and is also called as a small-scale key point characteristic. The value rule is as follows:
Figure 5839DEST_PATH_IMAGE011
in the above formula, the first and second carbon atoms are,L ’ by finger andDis selected symmetrically from the left and the right sides as a center
Figure 980617DEST_PATH_IMAGE012
The distance is used as the selection range of the class II input. For example, in a practical system, the target distance is determined by the distance of the target when hittingD=700m, it is totally around 700m
Figure 758080DEST_PATH_IMAGE013
I.e. the range [650, 750 ]]Selecting againQPoints are input as neural network class II, as in fig. 4.
In the above formula, the first and second carbon atoms are,L 2 by finger andDis a central termD-L ’ /2，D+L ’ /2]Symmetrically selecting left and right sidesQSpaced apart by a smaller distance. For example, in a practical system, the target distance is determined by the distance of the target when hittingD=700m, then [650, 750 ] within a total distance of 100m around 700m]Interval subdividing fetchQPoints, hereL 2 May take 1m, i.e. in the interval 650, 750]And taking points every 1m, wherein the total number of the points is 101, and the points are used as class II input of the neural network input layer.
The design is such that the input of the neural network comprises the laser in its effective working rangeD near ，D far ]Inner averageThe error features of the distributed global keypoints (called large-scale keypoint error features) also contain some actual distanceDLocal keypoint error features (called small-scale keypoint error features) in a certain range on the left side and the right side, and thereforeDDot，The neural network inputs the global error characteristics of large scale and the local error characteristics of small scale.
For a specific multi-beam space synthesis laser system, by adding input information of the neural network, the input of the neural network comprises error characteristic information which shows that the system is relatively fixed, and also comprises relatively different error characteristic information which shows different positions, so that the approximation of the neural network to the input-output relationship of an actual system is facilitated. Thus, the basic structure of a neural network, now expanded from FIG. 2 to FIG. 4, has its inputs from only one pointDTwo categories have been expanded: class I large-scale key point features and class II small-scale key point features, the number of which are respectivelyPAnQAnd (4) respectively.
Sampling number of two kinds of input of neural networkPAndQmay or may not be equal, in most casesQIs greater thanPBecause, in general, the distance to a target isDThe class II keypoint features are more closely associated with focusing, and relatively more affected than the class I keypoint features, and therefore need to be considered inDMore sampling values are selected nearby, so that the neural network can acquire more characteristic information nearby the focusing position under the distance, and more accurate focusing amount is output through learning.
Class IV and III keypoint sampling
Again, class I and class II keypoint sampling is further extended. The class I key points and the class II key points are sampled averagely in a certain range, and for actual systems of the same model and the same batch, if only class I information and class II information are input, the input structure and the input information of the neural network are fixed under the same sampling rule. In order to improve the flexibility of a network structure and improve the adaptability of the network structure to different personalized difference systems, random sampling information is further added on the basis of average sampling, input information of a network is enriched, and the purposes of improving the complexity of the network and enhancing the learning capacity of the network are achieved. The method comprises the following steps:
1) dimension extension
The inputs to the neural network in fig. 5 are, in effect, two vectors: firstly, is composed ofPVector composed of I-type key points
Figure 160112DEST_PATH_IMAGE005
And the second is composed ofQVector composed of class II key points
Figure 390236DEST_PATH_IMAGE006
. We first fit the vectors
Figure 434284DEST_PATH_IMAGE005
Sum vector
Figure 247345DEST_PATH_IMAGE006
Respectively expand intom×m×PAndn×n×Qthree-dimensional matrix of (A), (B)m、nAn integer not less than 1) as shown in fig. 6.
In FIG. 6m、nAll take the value of 3. Thus, vector
Figure 820278DEST_PATH_IMAGE005
Sum vector
Figure 272119DEST_PATH_IMAGE006
Is expanded from one-dimensional vector to 3 x 3 inPAnd 3 x 3QOf the three-dimensional matrix of (a).
2) Random sampling spreading
On the basis of dimension extension, corresponding to eachP i 、Q i And further expanding input information of the neural network, namely sampling expansion.
The random sampling spreading method comprises the following steps:
for each class I key pointP i Or class II key pointsQ i Randomly sampling within a certain range on the left and right sides
Figure 119858DEST_PATH_IMAGE007
(M =2, 3, 4, 5, … …) points, expanded with key pointsM 1 ×M 1 ×POrM 2 ×M 2 ×QThe matrix of (a) is,M 1 andM 2 may or may not be equal.
With class II key pointsQ i For example.Q i The sampling ranges on both sides are generally setQ i-1 ，Q i+1 ]However, as an experience value, the interval is not strictly limited, and the interval is selected according to specific conditions, if the distance between each point is larger, the value range can be reduced, the distance is smaller, and the value range can be expanded, for example, the value range of the key point in class I does not need to be followed by the value range of the termP i-1 ，P i+1 ]Interval limitation, can be reduced toP i Within a certain range of the surroundings. And for the key point of class II, ifQ i-1 ，Q i ]Is relatively close to itself and can be further expanded toQ i-2 ，Q i+2 ]And adjusting by combining the whole learning efficiency of the network.
Thus, the key points in class II are expanded into 3 x 3 and a databaseQAfter the matrix ofQ i Randomly selecting 3 again in the interval determined by the left side and the right side 2 -1=8 sampling points, then andQ i together, randomly filled in the corresponding 3 × 3 matrix. Class I keypoints are treated the same as in fig. 9.
3) 1 x 1 convolution processing
The random sampling is to improve the robustness of the neural network, so that after the neural network is subjected to learning of enough samples, the adaptability of the neural network under different system hardware, different system indexes and different working environments can be improved. The random sample value is not directly learned as an input to the network, but is subjected to a 1 × 1 convolution process once. The purpose of the 1 × 1 convolution is to reacquire the original sampled information, and enhance the learning efficiency and ability of the whole network by adjusting the contribution of each sampling point to the input. As shown in fig. 10:
the calculation rule of the operator "" in fig. 10 is:
in 3 x 3QIn the matrix of (2), hasQA 3 x 3 matrix. Each 3 x 3 matrix comprises oneQ i And 8 data sampled randomly around it, these 9 data are multiplied by a 1 x 1 filter respectively to output a new 3 x 3 matrix, each data is scaled down or up in the same scale.QGroup 3 x 3 matrix correspondencesQGroup of 1 × 1 filters, convolved according to the above method, and finally still output a 3 × 3 templateQOf the matrix of (a).
In the same way, the method for preparing the composite material,Pgroup 3 x 3 class I element matrix andPgroup 1X 1 filter, after convolution calculation, output a 3X 3 filterPOf the matrix of (a).
Fifthly, constructing a multi-input neural network
After the above processing, the single-input neural network structure is developed into a multi-input neural network structure, and the whole neural network processing process is shown in fig. 5:
the 1 x 1 convolution processing improves the flexibility of the whole neural network, different convolution calculations can be carried out on different sampling values during learning, and the nonlinear mapping capability of the whole network to an actual system is improved.
After 1 × 1 convolution is performed on the input, the output matrix is sequentially tiled, and 3 × 3 is used as a templatePIs 3 x 3 inQBecomes a vector which contains
Figure DEST_PATH_IMAGE014
An input element. After this layer, several (≧ 3) fully-connected layers are connected as hidden layers, and finally a vector containing all focusing stage adjustments is output, e.g., (c: ≧ 3)d 1 ,d 2 ,d 3 ) As the final output of the neural network.
The random sampling expansion processing enables the input of the neural network to have flexible changing capability and can adapt to different individual characteristics of products of the same model and the same batch. The random sampling part surrounds the I-type sampling point and the II-type sampling point, the sampling range is fine, the convolution processing is individualized, the input of the neural network can be adjusted in structure and input according to each specific system, the selection of the input layer has flexibility and variability, the sampling value of the random sampling part can be continuously adjusted according to the learning and training effect, and the input information of the neural network is optimized.
Once the optimized neural network input structure is established, itP、Q、m、nAfter the parameters are determined, the parameters do not change any more when the subsequent samples are learned and actually used, the parameters are used as a fixed neural network model and neural network input to be applied to a specific system, and the final input layer comprises (A)
Figure 730968DEST_PATH_IMAGE015
) An input element.
The neural network construction method is shown in fig. 11:
sixth, boundary processing
The boundary processing is mainly for the selection of local features. When the actual distance isDIs set in the definition domainD near ，D far ]At or near the boundary of (a),D near -L ’ /2 orD far +L ’ /2 may exceed [2 ]D near ，D far ]And becomes an invalid value. There are several cases depending on the specific design of the system:
firstly, the focusing platform is movedD near OrD far Mechanical limit or software limit is achieved. In this case, if the limit value is exceeded, the system is ineffective to learn the features other than the limit, and the control signal output by the system may not be executed due to the limit, so that the system can only get back to the next step at this time, and no longer needs to learn the features due to the limitDAre taken as central symmetry, but are taken asDThe domain side value of (2). When taking value, select the value according to a certain intervalD，D+L ’ ]Or [ alpha ], [ alphaD-L ’ ，D]Q key distance values in the domain as the input layer of the neural networkAnd (4) inputting in a II type. The value can be slightly denser than other conditions so as to ensure that the neural network has enough information input.
Secondly, the focusing platform is movedD near OrD far Not reaching the mechanical limit or the software limit but providing a control marginl ’ Is not sufficient enough to satisfyD near -L ’ /2 orD far +L ’ /2. In this case, the distance is ensured as much as possibleDFor feature learning, the value is selected at certain intervalsD-l ’ ，D+L ’ -l ’ ]Or [ alpha ], [ alphaD-L ’ +l ’ ，D+l ’ ]And taking Q key distance values in the domain as input of a neural network input layer II type.
Thirdly, the movable focusing platform is arrangedD near OrD far The mechanical limit or the software limit is not reached, and the control margin isl ’ Can satisfyD near -L ’ /2 orD far +L ’ /2. In this case, the distance can be effectively ensuredDFeature learning centered on can be selected symmetricallyD-L ’ /2，D+L ’ /2]Left and right sidesQThe segments are separated by a smaller distance and used as input of class II of the input layer of the neural network.
And fourthly, any combination of the three possible combinations is realized. In this case, the three value-taking rules are respectively adopted according to specific limit constraints. Generally, the technical state of a certain type of product is consistent and fixed, and during system design, the consideration of some limit constraints is also uniform and consistent, which also determines that the model structure when the algorithm is applied is also specific and definite, and the structural change of the neural network model in the application process can not occur.
In a real system, canBy selecting an optical lens with appropriate parameters and combining the stroke of the movable focusing platform for unified design, the movable focusing platform is ensured to have enough stroke allowance within a normal working range, and a laser light path is supported to cover a working distance determined by technical indexes. Therefore, we have a third case in the discussion, namely that the moving focus stage has enough margin to supportD near OrD far And the value when the boundary position is equal. Thus, the neural network model can learn the limit constraintD near OrD far And (4) system error characteristics of equal boundary positions.
Seven, network training
After the neural network is constructed, for each specific laser explosive ordnance disposal system, the number of neurons of each layer such as an input layer, a hidden layer and an output layer is determined, and then the next step is to provide an enough training set to train the system until the output meets the application requirements. The training method is the same as the BP network training.
Eighthly, application process
After the neural network training is finished, the whole neural network can be used as a functional black box, is embedded into the control software of the laser explosion-removing system and is used as a functional module with an input and output function. The flow of its use is shown in fig. 12.

Claims (6)

1. A focusing control neural network input sampling method of a space synthesis laser explosive-removing system is characterized by comprising the following steps:
step 1: the focusing platform is moved to obtain a target distance D through laser ranging;
step 2: respectively generating class I input data and class II input data according to the target distance D and the class I key point sampling method and the class II key point sampling method, wherein the method specifically comprises the following steps:
2-1) sampling I key points; class I expansion of the input layer, spaced by distance in the region where the beam can be normally focusedL 1 Uniform sampling and co-selectionPA key distance value as a neural networkThe class I input of the input layer is also called large-scale key point characteristics, and the value-taking rule is as follows:
Figure 710005DEST_PATH_IMAGE001
in the above formula, the first and second carbon atoms are,L 1 the size of the separation distance between the fingers,D near representing the closest distance between which the beam can travel,D far represents the farthest distance between which the light beam can move; to obtainPVector composed of I-type key points
Figure 949357DEST_PATH_IMAGE002
；
2-2) sampling II key points; performing class II expansion on the input layer at the target distanceDNear, at intervalsL 2 Uniformly sampling and selecting the [ alpha ], [ beta ] -aD-L ’ /2，D+L ’ /2]Within a domainQThe key distance value is used as the input of class II of the input layer of the neural network, also called the small-scale key point characteristic, and the value rule is as follows:
Figure 496882DEST_PATH_IMAGE003
in the above formula, the first and second carbon atoms are,L ’ by finger andDthe distance is symmetrically selected from the left side and the right side as the value selection range of the class II input,L 2 by finger andDis a central termD-L ’ /2，D+L ’ /2]Symmetrically selecting left and right sidesQA smaller distance apart; to obtainQVector composed of class II key points
Figure 821684DEST_PATH_IMAGE004
；
And step 3: performing dimension expansion on the class I input data and the class II input data respectively;
will vector
Figure 644015DEST_PATH_IMAGE005
Sum vector
Figure 206715DEST_PATH_IMAGE006
Respectively expand intom×m×PAndn×n×Qof a three-dimensional matrix of (2), whereinm、nIs a positive integer not less than 1;
and 4, step 4: respectively randomly sampling I-type input data and II-type input data after dimensionality extension;
for each class I key pointP i Or class II key pointsQ i Randomly sampling within a certain range on the left and right sides
Figure 456299DEST_PATH_IMAGE007
(M =2, 3, 4, 5, … …) points, expanded with key pointsm×m×POrn×n×QThe matrix of (a) is,mandnthe values of (A) may be equal or unequal;
and 5: respectively carrying out 1 × 1 convolution calculation on the I-type input data and the II-type input data which are randomly sampled;
step 6: performing convolution calculation on the I-type input data and the II-type input data to generate one-dimensional vectors of the I-type input data and the II-type input data, and performing convolution calculation on the I-type input data and the II-type input data to generate the one-dimensional vectorsm×m×PAndn×n×Qbecomes a vector which contains
Figure 268397DEST_PATH_IMAGE008
An input element;
and 7: the combined one-dimensional vector is used as an input layer of the neural network, enters the trained neural network for calculation, and outputs a vector containing all the adjustment quantities of the focusing displacement platform
Figure 645152DEST_PATH_IMAGE009
，fnThe number of focusing displacement platforms in the system;
and 8: according tod' 1 、d' 2 ……d' fn And respectively controlling the focusing displacement platform of the system to complete respective focusing control.
2. The method for inputting and sampling the focusing control neural network of the space synthesis laser explosion elimination system according to claim 1, wherein the specific scheme of the step 5 is as follows:
in thatn×n×QIn the matrix of (2), hasQGroup ofn×nEach of whichn×nComprises a matrix ofQ i And randomly sampled around itn 2 1 number of data, thisn 2 Multiplying the data by a 1 × 1 filter, respectively, to output a new onen×nMatrix, each data is reduced or enlarged according to the same proportion;Qgroup ofn×nMatrix mappingQGroup 1X 1 filters, outputting onen×n×QA matrix of (a);
in thatm×m×PIn the matrix of (2), hasPGroup ofm×mEach of whichm×mComprises a matrix ofP i And randomly sampled around itm 2 1 number of data, thism 2 Multiplying the data by a 1 × 1 filter, respectively, to output a new onem×mMatrix, each data is reduced or enlarged according to the same proportion;Pgroup ofm×mMatrix mappingPGroup 1X 1 filters, outputting onem×m×PA matrix of (c).
3. The method for sampling the input of the neural network for focus control of the space synthesis laser explosion elimination system according to claim 1, wherein the neural network has two types of input and the number of the samples is the sameQIs greater thanPIn (1).
4. The method for neural network input sampling for focus control of a space-synthesized laser explosion suppression system as claimed in claim 1, wherein the moving focus is adjustedPlatform is onD near OrD far When the mechanical limit or the software limit is reached, the values are taken according to the interval sizeL ’ /QSelecting [2 ]D，D+L ’ ]Or [ alpha ], [ alphaD-L ’ ，D]And taking Q key distance values in the domain as input of a neural network input layer II type.
5. The method for sampling the input of the neural network for focus control of the space-synthesis laser explosion-elimination system according to claim 1, wherein a focus platform is movedD near OrD far Not reaching the mechanical limit or the software limit but providing a control marginl ’ Is not sufficient to satisfyD near -L ’ /2 orD far +L ’ /In case 2, the value is taken according to the interval sizeL ’ /QSelecting [2 ]D-l ’ ，D+L ’ -l ’ ]Or [ alpha ], [ alphaD-L ’ +l ’ ，D+l ’ ]And taking Q key distance values in the domain as input of a neural network input layer II type.
6. The method for sampling the input of the neural network for focus control of the space-synthesis laser explosion elimination system according to claim 1, wherein the focus platform is movedD near OrD far The mechanical limit or the software limit is not reached, and the control margin isl ’ Can satisfyD near -L ’ /2 orD far +L ’ /In case 2, the [2 ] is selected symmetricallyD-L ’ /2，D+L ’ /2]Left and right sidesQEach interval has a size ofL ’ /QAs the neural network outputEntering layer II type input.
CN202210720172.0A 2022-06-24 2022-06-24 Focusing control neural network input sampling method of space synthesis laser explosive-removing system Active CN114819111B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210720172.0A CN114819111B (en) 2022-06-24 2022-06-24 Focusing control neural network input sampling method of space synthesis laser explosive-removing system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210720172.0A CN114819111B (en) 2022-06-24 2022-06-24 Focusing control neural network input sampling method of space synthesis laser explosive-removing system

Publications (2)

Publication Number Publication Date
CN114819111A true CN114819111A (en) 2022-07-29
CN114819111B CN114819111B (en) 2022-09-02

Family

ID=82520919

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210720172.0A Active CN114819111B (en) 2022-06-24 2022-06-24 Focusing control neural network input sampling method of space synthesis laser explosive-removing system

Country Status (1)

Country Link
CN (1) CN114819111B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104266605A (en) * 2014-06-27 2015-01-07 西北工业大学 Imaging method for three-dimensional laser scanning imager
WO2018107307A1 (en) * 2016-12-14 2018-06-21 徐海军 Ultra-high power laser spatial combining system and systems related thereto
CN108287412A (en) * 2017-12-30 2018-07-17 湖北航天技术研究院总体设计所 A kind of laser space combination Transmission system based on optics Wei Sao mechanisms
CN109654945A (en) * 2018-11-22 2019-04-19 徐珂 With trajectory expressive ability and injure multifarious confrontation fire analogue technique
CN109802291A (en) * 2019-02-18 2019-05-24 中国人民解放军国防科技大学 Optical fiber laser aperture coherent synthesis phase control method based on deep learning
CN112462600A (en) * 2020-11-20 2021-03-09 华中光电技术研究所（中国船舶重工集团公司第七一七研究所） High-energy laser control method and system, electronic equipment and storage medium

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104266605A (en) * 2014-06-27 2015-01-07 西北工业大学 Imaging method for three-dimensional laser scanning imager
WO2018107307A1 (en) * 2016-12-14 2018-06-21 徐海军 Ultra-high power laser spatial combining system and systems related thereto
CN108287412A (en) * 2017-12-30 2018-07-17 湖北航天技术研究院总体设计所 A kind of laser space combination Transmission system based on optics Wei Sao mechanisms
CN109654945A (en) * 2018-11-22 2019-04-19 徐珂 With trajectory expressive ability and injure multifarious confrontation fire analogue technique
CN109802291A (en) * 2019-02-18 2019-05-24 中国人民解放军国防科技大学 Optical fiber laser aperture coherent synthesis phase control method based on deep learning
CN112462600A (en) * 2020-11-20 2021-03-09 华中光电技术研究所（中国船舶重工集团公司第七一七研究所） High-energy laser control method and system, electronic equipment and storage medium

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
孙非 等: ""低慢小"无人飞行器反制技术综述", 《中国安全防范技术与应用》 *
许栩 等: "未爆弹排除技术现状及发展趋势", 《机电技术》 *

Also Published As

Publication number Publication date
CN114819111B (en) 2022-09-02

Similar Documents

Publication Publication Date Title
Mohamed et al. Optimal power flow using moth swarm algorithm
Xu et al. Optimized multi-UAV cooperative path planning under the complex confrontation environment
CN102506863B (en) Universal gravitation search-based unmanned plane air route planning method
CN109962688A (en) The quick predict and reverse geometry design method of all dielectric Meta Materials filter transfer characteristic based on deep learning neural network
CN107976899A (en) A kind of precision target positioning and striking method based on someone/unmanned plane cooperative engagement systems
CN111370080B (en) Radiation temperature inversion method based on artificial neural network algorithm
Zohdi The Game of Drones: rapid agent-based machine-learning models for multi-UAV path planning
CN106199800A (en) A kind of integrated approach of the three-dimensional vortex array of spatial distribution
CN111639767A (en) Optical orbit angular momentum machine learning identification method based on turbulence effect
Li et al. Sand cat swarm optimization based on stochastic variation with elite collaboration
Ragazzoni et al. Ingot laser guide stars wavefront sensing
CN114819111B (en) Focusing control neural network input sampling method of space synthesis laser explosive-removing system
Digani et al. Framework for expediting discovery of optimal solutions with blackbox algorithms in non-topology photonic inverse design
CN113063419B (en) Unmanned aerial vehicle path planning method and system
CN116882147A (en) Array antenna synthesis method based on mixed sparrow search algorithm
CN117195332A (en) Method and device for designing optical system, and computer-readable storage medium
CN116933948A (en) Prediction method and system based on improved seagull algorithm and back propagation neural network
CN109933895B (en) Array directional diagram comprehensive method based on dynamic grid multi-target particle swarm optimization method
CN109185731A (en) A kind of design method of Image processing spectroscopic light source
CN113849953B (en) Micro focal spot device design optimization method for space X-ray communication
CN112906308A (en) Method, equipment and storage medium for quickly predicting radiation characteristic of antenna housing
Bender et al. Satellite Formation Design to Enhance Passive Millimeter Wave Imaging Mission Performance
Komorowski et al. THz diffractive optics designed with neural networks
An et al. Research on light-small lens structure design and weight reduction optimization based on neural network
CN113067615B (en) Near-field array beam control method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant