CN112232484B - Space debris identification and capture method and system based on brain-like neural network - Google Patents
Space debris identification and capture method and system based on brain-like neural network Download PDFInfo
- Publication number
- CN112232484B CN112232484B CN202011104427.8A CN202011104427A CN112232484B CN 112232484 B CN112232484 B CN 112232484B CN 202011104427 A CN202011104427 A CN 202011104427A CN 112232484 B CN112232484 B CN 112232484B
- Authority
- CN
- China
- Prior art keywords
- network
- space debris
- entropy
- complexity
- space
- 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.)
- Active
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
Abstract
The invention discloses a space debris identification and capture method and a system based on a brain-like neural network, wherein the method comprises the following steps: acquiring image information of the space debris; constructing a main network, and optimizing the main network into a target identification network; introducing an environment entropy calculation network to describe the distribution complexity of the space debris; introducing a network entropy calculation network to describe the network complexity of the target identification network; obtaining an entropy balance driving factor according to the complexity of the network and the complexity of the distribution of the space debris; under the guidance of a game theory framework, adaptively adjusting the network structure of the target identification network by utilizing an entropy balance driving factor; and establishing an information closed loop between the space debris and the space-based detector, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured. The invention has the characteristics of low power consumption, strong generalization capability and high identification precision.
Description
Technical Field
The invention relates to the technical field of space debris identification, in particular to a space debris identification and capture method and system based on a brain-like neural network.
Background
At present, the space debris is usually detected by adopting a machine learning and brain-like calculation means, and the space debris detection is to detect the space debris from the background so as to facilitate the development of subsequent work. The traditional machine learning has a large demand on data, depends on massive training, and can complete training only by massive sample data (hundreds of M, G and even T data); in addition, accurate feedback is needed in the learning and training process, and for many problems in the real world, such as the aerospace field or the robot field, good feedback is not available, a large number of simulation experiments cannot be performed to generate a large number of samples for training, or the cost required for generating the samples is too high, so that the machine learning application becomes difficult. Meanwhile, the traditional machine learning model has poor generalization capability and high energy consumption; when a system meets a new condition to obtain a new sample, training from 0 to the trained and modeled model on a data set containing the new sample is often needed, otherwise, the trained model cannot be applied to the new sample; a standard computer would need to consume 250 watts of energy to identify only 1000 different objects.
The traditional brain-like calculation has insufficient analysis on brain functional structures, and a brain research tool cannot integrate details and integrity and perform global imaging on the brain under high spatial and temporal resolution. Meanwhile, the brain function is abstracted into a mathematical model with great difficulty, the brain atlas is highly complex and dynamically changes, the functional division difference of brain neurons under different spatial scales and spatial distributions is obvious, the algorithm for cataloging spatial fragments is very complex, calculation must be carried out on the ground, and the requirement of real-time property cannot be met.
Therefore, it is an urgent need to solve the problem of the art to provide a method and a system for identifying and capturing space debris based on a neural network with low power consumption, strong generalization capability and high identification accuracy.
Disclosure of Invention
In view of this, the invention provides a method and a system for identifying and capturing space debris based on a brain-like neural network, which obtains a neural network model sensitive to the space debris by learning a space debris sample, and captures the space debris by planning a track according to a relative position between a space-based detector and the debris. When space debris which possibly generates threats is encountered, the space debris can be identified, tracked and processed in real time through quick and efficient identification.
In order to achieve the purpose, the invention adopts the following technical scheme:
a space debris identification and capture method based on a brain-like neural network comprises the following steps:
s1, acquiring image information of the space debris observed by the space-based detector;
s2, constructing a main network, analyzing the connection and disconnection strategy of each network node in the main network, modeling each network node according to the strategy, and optimizing the main network into a target identification network;
s3, introducing an environment entropy calculation network, and describing the distribution complexity of the space debris by using the environment entropy calculation network;
s4, introducing a network entropy calculation network, and describing the network complexity of the target identification network by using the network entropy calculation network;
s5, obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
s6, introducing a game theory framework, and under the guidance of the game theory framework, adaptively adjusting the network structure of the target recognition network by using the entropy balance driving factor to enable the network complexity of the target recognition network to be matched with the distribution complexity of the space debris;
and S7, establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
Preferably, in the above method for recognizing and capturing space debris based on a brain-like neural network, in S1, one or more feature vectors are extracted from the image of the space debris by using an image feature extraction tool, and are merged into an overall vector x.
Preferably, in the above method for capturing spatial debris based on the brain-like neural network, each network node in S2 is modeled as follows:
wherein V represents a set of network nodes; s i Represents each node v i The strategy to be adopted; p is a radical of 1 Representing the probability of each network node adopting a "growth" policy; p is a radical of 2 Representing the probability of each network node adopting a "fallback" policy; s denotes a mixing strategy.
Preferably, in the above method for identifying and capturing spatial debris based on a cranial nerve network, in S2, the network cost and the network error of the subject network are calculated according to the policy; obtaining an optimization objective function by using the network cost and the network error, and adjusting the connection structure and the network precision of the main network by using the optimization objective function;
the calculation formula of the network cost is as follows:
in the above formula, M i Representing a weight matrix, which is a variable quantity; t represents matrix transposition; a is ij Representing the connection condition of the ith node and the jth node in the connection matrix A; theta j Representing the weight of node i to node j; u shape i Represents each node v i Network cost under current policy; a' represents a connection matrix of the modeled subject network; l is 2，t Representing a total network cost of the subject network;
the calculation formula of the network error is as follows:
L 1，t ＝||X t -Y t || 2 ；
in the above formula, X t Image features, Y, representing input spatial patches t Is the output of the host network; l is 1，t Representing the accuracy of the subject network;
the calculation formula of the optimization objective function is as follows:
in the above formula, λ g Representing a weight coefficient.
Preferably, in the above method for identifying and capturing spatial debris based on a brain-like neural network, the calculation formula of the distribution complexity of the spatial debris in S3 is as follows:
H u ＝f(x)；
wherein f (-) represents the ambient entropy calculation network, H u Representing the ambient entropy.
Preferably, in the above method for identifying and capturing spatial debris based on a cranial nerve network, the calculation formula of the network complexity in S4 is as follows:
H n ＝g(H graph )；
wherein, the network entropy calculation network is regarded as a graph G ═ (V, E), and a node set V ═ { V } of the network entropy calculation network i I 1, 2.. k, each node v i And d i The nodes are connected with each other with the connection strength of
H graph Representing graph entropy for calculating network connection complexity under graph theory representation; g (-) represents a multi-layer perceptron, entropy H of the graph graph Mapping to high dimensionIn the feature space, the network entropy H is obtained n 。
Preferably, in the above method for identifying and capturing spatial debris based on a brain-like neural network, the calculation formula of the driving factor using entropy balance in S5 is as follows:
Q＝-λ(H u -H n )；
wherein Q is an entropy balance driving factor, represents the information difference between the environment and the network, and is used for adjusting the complexity of the network; λ is a proportionality coefficient.
Preferably, in the above-mentioned method for capturing spatial debris recognition based on a cranial neural network, the entropy balance driving factor is used in S6 to adjust the network cost of the target recognition network, and the "increase" and "decrease" strategies of the target recognition network are changed to change the network structure of the target recognition network.
Preferably, in the above method for recognizing and capturing spatial debris based on a cranial neural network, in S7, the space-based probe includes a controller and a capture device; the target identification network catalogs and positions the space debris and outputs the space debris to the controller in real time; the controller controls the capturer to capture the corresponding space debris according to the position information of the space debris; and the space-based detector freezes the target identification network in the moving process until the acquisition task is completed, and then continues to adjust the structure of the target identification network.
According to the technical scheme, compared with the prior art, the invention has the following advantages:
1. the invention adopts the length elimination strategy for each node of the main network to evolve into an effective and low-power-consumption network, which is a small-world network due to the characteristics of network connection, the connection between each node is sparser, but the function of the network is not inferior to that of a complex connection or even a fully-connected network, and the energy consumption is greatly reduced. The problems of fixed model and poor generalization capability of the traditional machine learning method are solved.
2. The invention respectively calculates the environment entropy and the network entropy by introducing an environment entropy calculation network and a network entropy calculation network, respectively describes the distribution complexity of space debris and the network complexity of a target identification network, maps the environment entropy and the network entropy into the same space to enable the environment entropy and the network entropy to have the same physical meaning, calculates the difference value of the two entropies by utilizing an entropy difference network, namely an entropy balance driving factor, and changes the node and the cost function of the target identification network under the driving of the entropy balance driving factor, thereby changing the length elimination strategy of the network, further changing the network structure and realizing the self-adaptive adjustment of the network structure along with the time-space change. Meanwhile, the evolution of the network is guided under the game theory framework, the network can continuously change the connection strength through continuous learning and the difference between external information and the network complexity, so that the network can continuously learn, the spatial debris with different forms can be better learned, and the accuracy and the speed of the spatial debris identification are greatly improved.
The invention also provides a space debris identification and capture system based on the brain-like neural network, which is suitable for the space debris identification and capture method based on the brain-like neural network, and comprises the following steps:
an input module for obtaining image information of space debris observed by the space-based detector
The network optimization module is used for constructing a main network, analyzing a strategy of connection and disconnection of each network node in the main network, modeling each network node according to the strategy and optimizing the main network into a target identification network;
the environment entropy calculation module is used for calculating the distribution complexity of the network description space debris by utilizing the environment entropy;
a network entropy calculation module for describing the network complexity of the target identification network by using the network entropy calculation network
The entropy balance driving factor calculation module is used for obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
the self-adaptive adjusting module is used for self-adaptively adjusting the network structure of the target identification network by utilizing the entropy balance driving factor under the guidance of the game theory framework so as to enable the network complexity of the target identification network to be matched with the distribution complexity of the space debris; and
and the capturing module is used for establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
According to the technical scheme, the input of the space debris capturing device is the image of the space debris input by the optical camera of the space-based detector, the network optimization module, the environment entropy calculation module, the network entropy calculation module, the entropy balance driving factor calculation module and the self-adaptive adjustment module form a brain-like neural network, and the image of the space debris passes through the brain-like neural network, is output for cataloging and positioning the space debris, and is transmitted to a downstream capturing module, namely a control system, so as to capture the space debris. The controller of the control system is used for adjusting engine parameters and the like of the space-based detector and controlling the relative position of the space-based detector and the space debris; the controlled object is a catcher, can be selected from a catching claw or a catching net according to actual requirements, and can be used for capturing the target by adjusting the posture of the catcher.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a flow chart of a method for recognizing and capturing space debris based on a neural network of brain-like according to the present invention;
FIG. 2 is a schematic diagram of adaptive adjustment of a target recognition network under the guidance of a game theory framework according to the present invention;
FIG. 3 is a schematic diagram illustrating an optimization process of a subject network according to the present invention;
fig. 4 is a structural block diagram of a space debris recognition and capture system based on a brain-like neural network provided by the invention;
fig. 5 is a flowchart illustrating the operation of a space debris recognition and capture system based on a brain-like neural network according to 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.
As shown in fig. 1-2, the embodiment of the invention discloses a space debris identification and capture method based on a brain-like neural network, which is characterized by comprising the following steps:
s1, acquiring image information of the space debris observed by the space-based detector;
s2, constructing a main network, analyzing the connection and disconnection strategy of each network node in the main network, modeling each network node according to the strategy, and optimizing the main network into a target identification network;
s3, introducing an environment entropy calculation network, and describing the distribution complexity of the space debris by using the environment entropy calculation network;
s4, introducing a network entropy calculation network, and describing the network complexity of the target identification network by using the network entropy calculation network;
s5, obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
s6, introducing a game theory framework, and under the guidance of the game theory framework, adaptively adjusting the network structure of the target recognition network by using the entropy balance driving factor to enable the network complexity of the target recognition network to be matched with the distribution complexity of the space debris;
and S7, establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
The above steps are specifically described below.
S1, acquiring image information of the space debris observed by the space-based detector:
in order to better describe the information content of the space debris image, the acquired image needs to be converted into a vector form, namely, feature extraction is carried out. For this purpose, one or more feature vectors are extracted from the image by using an existing image feature extraction tool, and the extracted feature vectors are combined into an overall vector x (x ═ x) 1 ，x 2 ，x 3 ...x n ) Wherein x is i The feature vector of the i-th kind is represented,
representing an m-dimensional vector space.
S2, constructing a main network, analyzing the connection and disconnection strategy of each network node in the main network, modeling each network node according to the strategy, and optimizing the main network into a target identification network.
As shown in fig. 3, in the graph theory, the target recognition network is regarded as a graph G ═ (V, E), and the network nodes are modeled, where V ═ V { (V, E) i I 1, 2.. k }, E represents a set of edges in the graph G, and k represents the number of nodes, i.e., the number of neurons, in the graph G; the strategy that each node can adopt is 'increase' and 'fade', and the probability is p respectively 1 ，p 2 The strategy made by each node i after the strategy is adopted by the jth node is s ij ：
S i ＝(s i1 ，s i2 ，...，s ik )；S i Indicating a hybrid strategy.
The evolution matrix formed by the growth and the regression is:
if the connection matrix of the network is A at this time, that is
The connection condition of the network is shown, and the new connection matrix is
In order to better describe the effect of the whole network after the network nodes of the main network adopt the strategies under the game theory framework, the network cost U needs to be selected according to the strategies i A calculation is made, at which point the network cost U is made according to the new connection i Calculating, and assuming that the weight of the node i to the node j is theta j Write it into the weight matrix M i ＝(Θ 1 ，Θ 2 ，...，Θ k )：
Weight matrix M i The network cost under the current connection strategy is measured by adjusting the weight matrix, and the further adjustment of the network structure is guided. T represents matrix transposition; a is ij Showing the connection condition of the ith node and the jth node in the connection matrix A.
The invention focuses on the problem of adjusting the network structure, which is a dynamic process, so that it is only meaningful to calculate the loss function at a specific time t. Under this problem, we need to calculate the cost of the whole network, namely:
the above equation represents the sum of the node costs of each node after the policy is taken, i.e. the total cost of the network. In a specific task, the accuracy of the network also needs to be adjusted by using a network error function, namely:
L 1，t ＝||X t -Y t || 2 ；
wherein, X t For image features of input space debris, Y t Is the output of the subject network, L 1，t For measuring the accuracy of the network.
Obtaining an optimized objective function according to the network cost function and the network error function, wherein the calculation formula is as follows:
wherein λ is g Is a weight coefficient which represents the proportion of the network complexity in the optimization process.
When the main network works, the back propagation is not executed any more, but the accuracy of the network can still change along with the change of the network structure, so the loss at the time t-1 is input into the network along with the input at the next time, the dual adjustment of the network accuracy and the connection structure is realized, and the final target identification network is obtained.
As shown in fig. 2, S3, an environment entropy calculation network is introduced, and the distribution complexity of the space debris is described by using the environment entropy calculation network.
After obtaining the feature vector of the space debris, the feature vector x is equal to (x) 1 ，x 2 ，x 3 ...x n ) And inputting the environment entropy calculation network, and calculating the environment entropy, wherein the environment entropy is used for describing the distribution complexity of the space debris. The specific calculation method is as follows:
H u ＝f(x)
wherein f (-) represents an environment entropy calculation network, and only a simple multilayer perceptron is adoptedThen the method is finished; h u Representing the ambient entropy. The multi-layered perceptron is a feedforward artificial neural network model that can map multiple data sets of an input onto a single data set of an output.
And S4, introducing a network entropy calculation network, and describing the network complexity of the target identification network by using the network entropy calculation network.
The network entropy is calculated for the target identification network by using the network entropy calculation network, and the target identification network can be regarded as a graph G (V, E). Set of nodes V ═ { V ═ V i I 1, 2.. k, each node i and d i The nodes are connected with each other with the connection strength of
H n ＝g(H graph )；
In the above formula, H graph The method is a graph entropy, and the network connection complexity of the target recognition network under the graph theory representation is directly calculated. g (-) is also a multilayer perceptron, and can entropy H the graph graph Mapping to a high-dimensional feature space to obtain a network entropy H n 。
And S5, obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris.
The information entropy after passing through the entropy calculation network has the same physical significance as the network entropy, so that subtraction can be performed to drive the change of the network structure, and the entropy balance driving factor is calculated by using the entropy difference network.
Q＝-λ(H u -H n )
Where Q is the entropy balance driving factor, representing the difference in information between the environment and the network, and λ is the scaling factor.
S6, introducing a game theory framework, and under the guidance of the game theory framework, adaptively adjusting the network structure of the target recognition network by using the entropy balance driving factor to enable the network complexity of the target recognition network to be matched with the distribution complexity of the space debris. Specifically, the entropy balance driving factor is used for adjusting the connection strategy of the target identification network, namely changing the 'increase' strategy and the 'fade' strategy of the target identification network so as to change the network structure. The network cost is adjusted through the connection strategy of the target identification network, and meanwhile, the accuracy of the target identification network also needs to be adjusted through calculating network errors. And the network entropy calculation network recalculates the network complexity of the target identification network through the change of the network structure, matches the network complexity with the distribution complexity of the space debris again, and repeats the steps until the network complexity of the target identification network is matched with the distribution complexity of the space debris. And further, the network structure of the target identification network is adaptively adjusted by utilizing the entropy balance driving factor.
In the following, how the structure of the target recognition network is adjusted by using the entropy balance factor is described in detail.
To facilitate explicit time t, the entropy balances the drive factor Q t Influence on network structure adjustment of target identification network, time t, U i，t ＝U i，t (S i，t ，Q t ) Is node v i In a hybrid strategy S i，t The cost function of. Weight matrix M for calculating network cost i，t (Q t ) The evolution process of (2) is as follows:
M i，t ＝M i，t -1+f M (Q t )；
wherein f is M (. cndot.) is a directed polynomial that directs the change in the weight matrix.
Is fitted by a k-th order polynomial:
in the above formula, γ i Is a coefficient of a polynomial and is a fixed value.
Policy set S of network nodes i，t The evolution process with entropy-balanced driving factors is as follows, where g S (. cndot.) is a polynomial function that directs the change in connection policy.
S i，t ＝S i，t-1 +g S (Q t )
The following equation represents the connection policy writing component form of the node i and the node j:
wherein, g S (Q t )| i [j]Denotes g S (Q t )| i The jth component of (a).
Wherein, in order to ensure
Of significance, for g S (Q t )| i [j]The following constraints need to be made:
therefore, g S (Q t )| i The following can be written:
And S7, establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured. After the target recognition network performs the adaptive adjustment,
after the lightweight and effective target identification network is realized, an information closed loop between the space debris and the space-based detector is established, so that the network continuously learns, the relative positions of the space-based detector and the space debris are output in real time, and a control closed loop is formed until the space debris is successfully captured. In the process, the surrounding environment changes continuously along with the continuous movement of the space-based detector, and the target recognition network can continuously receive the images of the space debris observed by the space-based detector and continuously learn the images.
In order to ensure that the identification is stable during the movement of the space-based probe, the target identification network needs to be frozen until the network structure can not be adjusted continuously after the acquisition task is realized. After the relative position information and the self attitude information are obtained, the space-based detector can perform attitude regulation and control and target capture through a traditional control method such as Kalman filtering.
The steps are as follows:
1. the invention adopts the length elimination strategy for each node of the main network to evolve into an effective and low-power-consumption network, which is a small-world network due to the characteristics of network connection, the connection between each node is sparser, but the function of the network is not inferior to that of a complex connection or even a fully-connected network, and the energy consumption is greatly reduced. The problems of fixed model and poor generalization capability of the traditional machine learning method are solved.
2. The invention respectively calculates the environment entropy and the network entropy by introducing an environment entropy calculation network and a network entropy calculation network, respectively describes the distribution complexity of space debris and the network complexity of a target identification network, maps the environment entropy and the network entropy into the same space to enable the environment entropy and the network entropy to have the same physical meaning, calculates the difference value of the two entropies by utilizing an entropy difference network, namely an entropy balance driving factor, and changes the node and the cost function of the target identification network under the driving of the entropy balance driving factor, thereby changing the length elimination strategy of the network, further changing the network structure and realizing the self-adaptive adjustment of the network structure along with the time-space change. Meanwhile, the evolution of the network is guided under the game theory framework, the network can continuously change the connection strength through continuous learning and the difference between external information and the network complexity, so that the network can continuously learn, the spatial debris with different forms can be better learned, and the accuracy and the speed of the spatial debris identification are greatly improved.
As shown in fig. 4 to 5, an embodiment of the present invention further provides a system for identifying and capturing spatial debris based on a cranial neural network, which is suitable for the above method for identifying and capturing spatial debris based on a cranial neural network, and includes:
an input module for acquiring image information of space debris observed by the space-based detector
The network optimization module is used for constructing a main network, analyzing the connection and disconnection strategy of each network node in the main network, modeling each network node according to the strategy and optimizing the main network into a target identification network;
the environment entropy calculation module is used for calculating the distribution complexity of the network description space debris by utilizing the environment entropy;
a network entropy calculation module for calculating the network complexity of the network description target recognition network by using the network entropy
The entropy balance driving factor calculation module is used for obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
the self-adaptive adjusting module is used for self-adaptively adjusting the network structure of the target recognition network by utilizing the entropy balance driving factor under the guidance of a game theory framework so as to enable the network complexity of the target recognition network to be matched with the distribution complexity of the space debris; and
and the capturing module is used for establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
The network optimization module, the environment entropy calculation module, the network entropy calculation module, the entropy balance driving factor calculation module and the self-adaptive adjustment module form a brain-like neural network, the input of the brain-like neural network is an image or video of a space debris input by an optical camera of a space-based detector, the image or video of the space debris passes through the brain-like neural network, then cataloguing and positioning of the space debris are detected and output, and position information of the space debris is transmitted to a downstream capture module. The capturing module comprises a controller and a capturing device, and the controller adjusts the space posture of the capturing device according to the position information of the space debris to enable the space attitude to be close to the space debris and capture the space debris.
And freezing the parameters of the target identification network by the space-based detector in the moving process until the acquisition task is completed, and then continuously adjusting the structure of the target identification network.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (10)
1. A space debris identification and capture method based on a brain-like neural network is characterized by comprising the following steps:
s1, acquiring image information of the space debris observed by the space-based detector;
s2, constructing a main network, analyzing the connection and disconnection strategy of each network node in the main network, modeling each network node according to the strategy, and optimizing the main network into a target identification network;
s3, introducing an environment entropy calculation network, and describing the distribution complexity of the space debris by using the environment entropy calculation network;
s4, introducing a network entropy calculation network, and describing the network complexity of the target identification network by using the network entropy calculation network;
s5, obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
s6, introducing a game theory framework, and under the guidance of the game theory framework, adaptively adjusting the network structure of the target recognition network by using the entropy balance driving factor to enable the network complexity of the target recognition network to be matched with the distribution complexity of the space debris;
the game theory framework guides that:
adjusting the connection strategy of the target identification network by using the entropy balance driving factor so as to change the network structure of the target identification network; through the change of the network structure of the target identification network, the network entropy calculation network recalculates the network complexity of the target identification network, matches the network complexity of the space debris again, and repeats the steps until the network complexity of the target identification network matches the distribution complexity of the space debris;
and S7, establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
2. The method for recognizing and capturing the space debris based on the cranial neural network according to claim 1, wherein in S1, one or more feature vectors are extracted from the image of the space debris by using an image feature extraction tool, and are merged into an overall vector x.
3. The brain-like neural network-based space debris identification capture method according to claim 1, wherein each network node in S2 is modeled as follows:
wherein V represents a set of network nodes; s i Represents each node v i The strategy to be adopted; p is a radical of 1 Representing the probability of each network node adopting a "growth" policy; p is a radical of 2 Representing the probability of each network node adopting a "fallback" policy; s denotes a mixing strategy.
4. The method for recognizing and capturing the space debris based on the cranial nerve network according to claim 3, wherein in S2, the network cost and the network error of the subject network are calculated according to the strategy; obtaining an optimization objective function by using the network cost and the network error, and adjusting the connection structure and the network precision of the main network by using the optimization objective function;
the calculation formula of the network cost is as follows:
in the above formula, M i Representing a weight matrix, which is a variable quantity; t represents matrix transposition; a is ij Representing the connection condition of the ith node and the jth node in the connection matrix A; theta j Representing the weight of node i to node j; u shape i Represents each node v i Network cost under current policy; a' represents a connection matrix of the modeled subject network; l is 2，t Representing a total network cost of the subject network; the lower subscript t denotes time t;
the calculation formula of the network error is as follows:
L 1，t ＝||X t -Y t || 2 ；
in the above formula, X t Image features representing input space debris, Y t Is the output of the host network; l is 1，t Representing the accuracy of the subject network;
the calculation formula of the optimization objective function is as follows:
in the above formula, λ g Representing a weight coefficient.
5. The method for recognizing and capturing the space debris based on the brain-like neural network as claimed in claim 2, wherein the calculation formula of the distribution complexity of the space debris in the step S3 is as follows:
H u ＝f(x)；
wherein f (-) represents the ambient entropy calculation network, H u Representing the ambient entropy.
6. The method for recognizing and capturing the space debris based on the cranial neural network as claimed in claim 5, wherein the calculation formula of the network complexity in S4 is as follows:
H n ＝g(H graph )；
wherein, regarding the target identification network as a graph G ═ (V, E), the network computing entropy network node set V ═ { V ═ V } i I 1, 2.. k, each node v i And d i The nodes are connected with each other with the connection strength of
H graph Representing graph entropy for calculating network connection complexity under graph theory representation; g (-) represents a multilayer perceptron, entropy H of the graph graph Mapping to a high-dimensional feature space to obtain a network entropy H n 。
7. The method for recognizing and capturing the space debris based on the brain-like neural network as claimed in claim 6, wherein the calculation formula of the driving factor using entropy balance in S5 is as follows:
Q＝-λ(H u -H n )；
wherein Q is an entropy balance driving factor, represents the information difference between the environment and the network, and is used for adjusting the complexity of the network; λ is a proportionality coefficient.
8. The method for spatial debris recognition and capture based on the cranial neural network according to claim 1, wherein the entropy balance driving factor is used to adjust the network cost of the target recognition network in S6, and the strategy of "increase" and "decrease" of the target recognition network is changed to change the network structure of the target recognition network.
9. The brain-like neural network-based space debris identification capture method according to claim 1, wherein in S7, the space-based probe comprises a controller and a capture device; the target identification network catalogs and positions the space debris and outputs the space debris to the controller in real time; the controller controls the capturer to capture the corresponding space debris according to the position information of the space debris; and the space-based detector freezes the target identification network in the moving process until the acquisition task is completed, and then continues to adjust the structure of the target identification network.
10. A brain-like neural network-based space debris identification and capture system, which is suitable for the brain-like neural network-based space debris identification and capture method according to any one of claims 1-9, and comprises:
an input module for obtaining image information of space debris observed by the space-based detector
The network optimization module is used for constructing a main network, analyzing a strategy of connection and disconnection of each network node in the main network, modeling each network node according to the strategy and optimizing the main network into a target identification network;
the environment entropy calculation module is used for calculating the distribution complexity of the network description space debris by utilizing the environment entropy;
a network entropy calculation module for describing the network complexity of the target identification network by using the network entropy calculation network
The entropy balance driving factor calculation module is used for obtaining an entropy balance driving factor according to the network complexity of the target identification network and the distribution complexity of the space debris;
the self-adaptive adjusting module is used for self-adaptively adjusting the network structure of the target identification network by utilizing the entropy balance driving factor under the guidance of the game theory framework so as to enable the network complexity of the target identification network to be matched with the distribution complexity of the space debris; the game theory framework guides that:
adjusting the connection strategy of the target identification network by using the entropy balance driving factor so as to change the network structure of the target identification network; through the change of the network structure of the target identification network, the network entropy calculation network recalculates the network complexity of the target identification network, matches the network complexity of the space debris again, and repeats the steps until the network complexity of the target identification network matches the distribution complexity of the space debris;
and
and the capturing module is used for establishing an information closed loop between the space debris and the space-based detector, enabling the target recognition network to continuously learn, and outputting the relative positions of the space-based detector and the space debris in real time until the space debris is captured.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011104427.8A CN112232484B (en) | 2020-10-15 | 2020-10-15 | Space debris identification and capture method and system based on brain-like neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011104427.8A CN112232484B (en) | 2020-10-15 | 2020-10-15 | Space debris identification and capture method and system based on brain-like neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112232484A CN112232484A (en) | 2021-01-15 |
| CN112232484B true CN112232484B (en) | 2022-08-05 |
Family
ID=74118380
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011104427.8A Active CN112232484B (en) | 2020-10-15 | 2020-10-15 | Space debris identification and capture method and system based on brain-like neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112232484B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116578109B (en) * | 2023-06-05 | 2023-11-21 | 北京控制工程研究所 | Emergency evasion maneuver control method and device for spacecraft |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6169981B1 (en) * | 1996-06-04 | 2001-01-02 | Paul J. Werbos | 3-brain architecture for an intelligent decision and control system |
| WO2020071618A1 (en) * | 2018-10-02 | 2020-04-09 | 인하대학교 산학협력단 | Method and system for entropy-based neural network partial learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018058426A1 (en) * | 2016-09-29 | 2018-04-05 | 清华大学 | Hardware neural network conversion method, computing device, compiling method and neural network software and hardware collaboration system |
-
2020
- 2020-10-15 CN CN202011104427.8A patent/CN112232484B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6169981B1 (en) * | 1996-06-04 | 2001-01-02 | Paul J. Werbos | 3-brain architecture for an intelligent decision and control system |
| WO2020071618A1 (en) * | 2018-10-02 | 2020-04-09 | 인하대학교 산학협력단 | Method and system for entropy-based neural network partial learning |
Non-Patent Citations (2)
| Title |
|---|
| Deep Convolutional Neural Network Based Small Space Debris Saliency Detection;Jiang Tao, 等;《Proceedings of the 25th International Conference on Automation & Computing》;20190907;全文 * |
| Sapienza Space debris Observatory Network (SSON): A high coverage infrastructure for space debris monitoring;Shariar Hadji Hossein 等;《Journal of Space Safety Engineering》;20191210;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112232484A (en) | 2021-01-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP6514166B2 (en) | Machine learning apparatus, robot system and machine learning method for learning robot operation program | |
| CN108921893B (en) | Image cloud computing method and system based on online deep learning SLAM | |
| Zhao et al. | A spatial-temporal attention model for human trajectory prediction. | |
| US7672920B2 (en) | Apparatus and method for embedding recurrent neural networks into the nodes of a self-organizing map | |
| Rashid et al. | Interspecies knowledge transfer for facial keypoint detection | |
| Shiba et al. | Secrets of event-based optical flow | |
| Passalis et al. | Deep reinforcement learning for controlling frontal person close-up shooting | |
| CN110463376B (en) | Machine plugging method and machine plugging equipment | |
| CN110941272A (en) | Automatic driving control method and device | |
| Wang et al. | Research on robot path planning based on fuzzy neural network algorithm | |
| Baldini et al. | Learning pose estimation for UAV autonomous navigation and landing using visual-inertial sensor data | |
| CN112232484B (en) | Space debris identification and capture method and system based on brain-like neural network | |
| Sinha et al. | Convolutional neural network based sensors for mobile robot relocalization | |
| CN109764876B (en) | Multi-mode fusion positioning method of unmanned platform | |
| CN113910218B (en) | Robot calibration method and device based on kinematic and deep neural network fusion | |
| Prasetyo et al. | Spatial Based Deep Learning Autonomous Wheel Robot Using CNN | |
| Praczyk | Using evolutionary neural networks to predict spatial orientation of a ship | |
| Liu et al. | $\alpha $-MDF: An Attention-based Multimodal Differentiable Filter for Robot State Estimation | |
| CN111539364B (en) | Multi-somatosensory human behavior recognition algorithm based on feature fusion and multi-classifier voting | |
| CN112377332A (en) | Rocket engine polarity testing method and system based on computer vision | |
| Chen et al. | Learning trajectories for visual-inertial system calibration via model-based heuristic deep reinforcement learning | |
| Crupi et al. | Sim-to-Real Vision-Depth Fusion CNNs for Robust Pose Estimation Aboard Autonomous Nano-quadcopters | |
| Zeng et al. | Tutor-guided interior navigation with deep reinforcement learning | |
| Baldini et al. | Learning pose estimation for uav autonomous navigation andlanding using visual-inertial sensor data | |
| CN114266824A (en) | Non-cooperative target relative pose measurement method and system based on deep learning |
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 |