Heterogeneous cooperative path planning method for offshore floating object target search task

Abstract

The application provides a heterogeneous cooperative path planning method for a marine floating object target search task. A Parcel v2 particle simulator based on a Lagrange framework is used to predict the future trajectory and position of a plurality of particles according to real sea waves, ocean currents, wind fields and other information, and then a Parzen window method based on a Gaussian kernel function is used to obtain a target probability map. In the traditional GBNN method, the regional centroid and motion constraint are added to improve the external stimulation value of the neuron to plan the cooperative search path of the plurality of unmanned ships, so that the local optimum can be effectively avoided and the search efficiency is improved. A solar unmanned aerial vehicle is used as a communication relay of the plurality of unmanned ships, an optimal communication network connection is obtained by using a minimum spanning tree, and then the path of the unmanned aerial vehicle is planned. The application can improve the method for solving the marine search target of the plurality of unmanned ships, so as to improve the efficiency of the marine search target of the unmanned ships.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of autonomous navigation and control technology of unmanned vehicles, and particularly relates to a heterogeneous cooperative path planning method for unmanned ship / unmanned aerial vehicle in the target search task of sea floating objects. BACKGROUND

[0002] In recent years, with the deepening of human development and utilization of the ocean, various sudden accidents occur frequently on the sea. Among them, how to quickly search for sea floating object targets (such as aircraft wreckage, etc.) has always been a problem that people pay attention to and strive to solve, and at the present stage, the target search mainly relies on manned ships and manned aircraft, which has a high search cost. With the development of science and technology, unmanned systems represented by unmanned surface vehicles (USV) and unmanned aerial vehicles (UAV) have developed rapidly. Because unmanned systems have the advantages of low cost, high efficiency, high flexibility, and high safety factor, they have obvious advantages in the field of sea target search. In addition, compared with single unmanned ship or unmanned aircraft platforms, unmanned ship / UAV heterogeneous cooperation can more efficiently complete complex search tasks.

[0003] At the present stage, the model of the sea target search problem mostly does not combine the actual sea conditions, that is, it is assumed that the position of the searched target has been determined, and the unmanned system only needs to navigate to the target position, but the position of the sea floating object target will change with the sea conditions, so such a model has poor applicability to the actual situation. Unlike the above method, the sea target search based on the target probability map is closer to the fact and is convenient for path planning to obtain maximum benefit and efficiency. Therefore, how to predict the target probability map combined with sea condition information will directly determine the efficiency of the subsequent search task. This patent will use the Parcel v2 particle simulator based on the Lagrangian framework to predict the position of the virtual sea particles, and then use the Parzen window method based on the Gaussian kernel function to obtain the target probability map.

[0004] For the problem of searching for a target at sea, the present patent takes multiple unmanned ships as the main search platform, and how to plan the path of the unmanned ship to obtain the maximum benefit and efficiency is the key. There are various existing target search path planning methods, such as the exhaustive coverage method (parallel line method, spiral line method, random method, etc.), graph-based heuristic method, adaptive search strategy (such as near-sighted search, fruit fly-inspired search, and saccade search), and opportunity learning, but these methods may not effectively handle the problem in some complex situations, and the unmanned ship may fall into a local optimum, resulting in reduced search efficiency. Unlike the above methods, the neuron working process in the Glasius Bio-inspired Neural Network (GBNN) can express the dynamic changes in the environment, thereby more effectively solving the path planning problem. The present patent introduces regional centroid guidance and motion constraints in the GBNN algorithm to better avoid local optimum and improve search efficiency.

[0005] Because the area of the sea area is large and the communication distance between unmanned ships is limited, the search range of the unmanned ship is limited, or the unmanned ships cannot communicate with each other, i.e., cannot share the target probability map, resulting in repeated search. Both of the above two cases will affect the search efficiency. SUMMARY

[0006] The present patent aims to improve the method for searching for a target at sea by multiple unmanned ships to improve the efficiency of searching for a target at sea by unmanned ships, in view of the low communication efficiency and limited communication distance of unmanned ships in the prior art.

[0007] To achieve the above-mentioned purpose, the present patent provides the following technical solutions:

[0008] A heterogeneous collaborative path planning method for searching for a floating target at sea is used for planning the path of an unmanned ship searching for a floating target at sea, and includes the following steps:

[0009] S1: Releasing virtual marine particles in a target area where a floating target is located, the virtual marine particles being used to represent the floating target; inputting the virtual marine particles into a Parcel v2 particle simulator in a Lagrangian framework to predict and track the position of the virtual marine particles, and using a Parzen window method based on a Gaussian kernel function to obtain a target probability map of the virtual marine particle distribution;

[0010] S2: Adding a regional centroid to the GBNN method to improve the external stimulation value of the neuron, taking the distance between the location of the unmanned ship and the regional centroid as a parameter to plan the collaborative search path of multiple unmanned ships;

[0011] S3: Utilize solar-powered drones as communication relays for multiple unmanned vessels, and use the minimum spanning tree method to obtain the optimal communication network connection between unmanned vessels and drones, thereby planning the drone paths.

[0012] In some embodiments of the present invention, step S2 further includes:

[0013] The target probability map of the virtual ocean particle distribution is rasterized, and each grid is regarded as a neuron;

[0014] Motion constraints are added to the GBNN method to modify the external stimulus values ​​of neurons and update the GBNN model. Based on the updated GBNN model, the activity dynamic value of neurons is calculated. The unmanned vessel selects the neuron with the largest activity dynamic value in its current neuron's receptive region as the next path point. This process is repeated iteratively to complete the unmanned vessel's path planning.

[0015] In some embodiments of the present invention, step S1 includes:

[0016] Solve the equation:

[0017]

[0018] Where: X is the three-dimensional position of the virtual ocean particle, t represents any time, Δt represents the sampling period, v(x,τ) is the three-dimensional velocity field of the ocean current at that position, and ΔX b (t) represents the positional change of the particle caused by its behavior at sea;

[0019] The Parcel v2 particle simulator was used to predict the positions of virtual ocean particles;

[0020] Determine the corresponding Gaussian function based on the predicted position of each particle:

[0021]

[0022] in: For the particle's geographical location, β i ∈[0,1] represents the particle's credibility. Let S represent the particle distribution range, where i represents the i-th particle, and the particle has the property S = {s} i ,β i ,λ i}, i = 1, ..., N, where N represents the total number of particles in the target area;

[0023] Sample K from the Gaussian function of each particle. i Samples: K i =β i K max ;

[0024] Where: K i It is β i Number of samples when = 1;

[0025] Expand the sample size to:

[0026] The target probability map of the virtual ocean particle distribution is represented as:

[0027]

[0028] in, It is the position of the kth particle, σ = (σ x ,σ y ) is the standard deviation;

[0029]

[0030]

[0031] in, and It is the standard deviation of the entire sample, IQR. x and IQR y This represents the difference between the 75th percentile and the 25th percentile of the sample.

[0032] In some embodiments of the present invention, the invention further includes:

[0033] The obtained target probability map is rasterized, and the probability of each grid m (m∈{1,2,…,M}) is represented by p(x). m p(x) ∈ [0,1] represents p(x) m The probability p(x) is obtained by integrating the probability of the target within the grid; the probability p(x) of all grids is obtained by... m The sum of ) is:

[0034]

[0035] Where: vector x m This indicates the position of each grid cell.

[0036] In some embodiments of the present invention, Bayesian rules are used to update the target probability map at time t:

[0037]

[0038] Where: λ is the normalization factor:

[0039]

[0040] Where t is any search time. The detection event of the mth grid at time t can be detected only when the unmanned ship is in the mth grid, that is, when the mth grid is detected is 1 when the mth grid is not detected is 0.

[0041] Wherein: is the credibility of the unmanned ship to the grid m:

[0042] Each grid is regarded as a neuron, and the connection weight w between neurons is defined mk is:

[0043]

[0044] Wherein |x m -x k | is the Euclidean distance between neuron m and neuron k, and γ is a constant greater than 0.

[0045] The activity dynamic value of neuron m is defined as:

[0046]

[0047] Wherein: N(m) is the neuron in the receptive field of neuron m, is the external stimulus signal of neuron m;

[0048]

[0049] Wherein: E is the internal punishment value of the neuron, which will not be transmitted to other neurons.

[0050] The centroid of the search area is defined as:

[0051]

[0052]

[0053] V is the entire area of the target probability map, p(x m ) is the probability density function, A is the quality of the target probability map, and CM is the centroid of the target probability map.

[0054] The centroid is used as the external stimulus signal of the neuron, and the distance between the position of the unmanned ship and the centroid of the search area is used as the parameter to plan the cooperative search path of the multiple unmanned ships.

[0055] In some embodiments of the application, the method for adding motion constraints in the GBNN method comprises:

[0056] Define Δθ as the vector x k -xm with x m -x l between the angle, define the motion cost as

[0057]

[0058] x k Indicate the position of neuron k, x m Indicate the position of neuron m, x l Indicate the position of neuron l;

[0059] Improved part of the external stimulus value of the GBNN model:

[0060]

[0061] Wherein: q is a constant greater than 0, The Euclidean distance between the regional centroid and the next time position of the unmanned ship, alpha is a normal number less than 1;

[0062] The external stimulus value signal in the improved GBNN model is:

[0063]

[0064] In some embodiments of the application, the step S3 comprises:

[0065] Define the weight, set it as the negative correlation function of the transmission success probability:

[0066]

[0067] Wherein: The communication probability between the ith and the jth;

[0068] The optimal communication network connection of the unmanned aerial vehicle / unmanned ship heterogeneous system is determined through the MST, and the communication performance index is set as:

[0069]

[0070] Wherein: U1 is the number of solar unmanned aerial vehicles, U2 is the number of unmanned ships, R is the optimal adjacency matrix of the communication network, if the node i and the node j are directly connected, then R ij =1, otherwise R ij =0;

[0071] The communication performance index is set as:

[0072]

[0073] J sThe communication network connection performance index is an index, and the smaller the index is, the better the communication effect is.

[0074] Compared with the prior art, the application has the advantages and positive effects that:

[0075] (1) The application can realize high-precision prediction of the target probability map of the floating object on the sea by using the Parcel v2 particle simulator based on the Lagrange framework to predict the position of the virtual marine particle, and using the Parzen window method based on the Gaussian kernel function to obtain the target probability map.

[0076] (2) The improved GBNN algorithm introducing the regional centroid guidance and motion constraint is used to plan the path of the unmanned ship, which can better avoid local optimization and greatly improve the search efficiency.

[0077] (3) The application introduces a solar unmanned aerial vehicle as a communication relay between multiple unmanned ships, establishes an optimal communication network connection and plans the path of the unmanned aerial vehicle, so as to ensure that the unmanned ships can still share information and coordinate behavior under a large range of sea surface, and at the same time, the target probability map is fused and updated. The search range of the unmanned ship can be expanded, and the search efficiency can be effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1a is a simulated initial particle distribution map;

[0079] Figure 1b is a simulated particle distribution map after two days;

[0080] Figure 2 is a target probability map to be searched;

[0081] Figure 3 is a GBNN network model schematic diagram;

[0082] Figure 4a is an updated target probability grid map based on the improved GBNN;

[0083] Figure 4b is an unmanned ship path map based on the improved GBNN;

[0084] Figure 5 is a Kruskal algorithm flowchart;

[0085] Figure 6 is an unmanned aerial vehicle path planning diagram. DETAILED DESCRIPTION

[0086] Hereinafter, the specific embodiments of the application will be further described with reference to the accompanying drawings.

[0087] S1: release virtual ocean particles in the area where the floating target is located, input the virtual ocean particles into the Parcel v2 particle simulator based on the Lagrangian framework, predict and track the position of the virtual ocean particles, and use the Parzen window method based on the Gaussian kernel function to obtain the target probability map of the virtual ocean particle distribution.

[0088] Due to the existence of factors such as sea waves, ocean currents and sea winds, the floating target to be searched will move, so the high-value area where it is located will change over time. The target probability of the floating target is used to show the probability of the existence of the floating target in each area of the sea, which is used to assist the guidance of the unmanned aerial vehicle search. In order to predict the target probability map of the floating target, virtual ocean particles are released in the area determined according to prior information (for example, in an air crash, the last position information sent by the aircraft is used as prior information), the virtual ocean particles are used to represent the floating target to be searched, and the particle information is input into the Parcel v2 particle simulator based on the Lagrangian framework. The relevant data of the actual sea state are used for Lagrangian simulation.

[0089] Solve the equation:

[0090]

[0091] Where: X is the three-dimensional position of the virtual ocean particle, t represents any time, Δt represents the sampling period, v(x,τ) is the three-dimensional velocity field of the ocean current at that position, ΔX b (t) is the change in the position of the particle due to the behavior of the sea;

[0092] The Parcel v2 particle simulator is used to predict the position of the virtual ocean particle; as shown in FIGS. Figure 1a and Figure 1b show the particle position prediction results of the current and two days later, respectively. Each particle is associated with three parameters: s i , β i and λ i .

[0093] According to the predicted position of each particle, the corresponding Gaussian function is determined:

[0094]

[0095] Where: is the geographical position of the particle, β i ∈[0,1] is the credibility of the particle, is the distribution range of the particle, i represents the i-th particle, and the particle has attributes S={s i , β i , λ i}, i = 1, …, N, N represents the total number of particles in the target area; according to prior information, the N particles released in the area are each associated with a two-dimensional standard Gaussian distribution function G(s i , λ i ), and there is no correlation between the two dimensions.

[0096] The target probability map can be obtained by adding these Gaussian functions, but the result is rough and does not take into account the entire area, so the present application uses the Parzen window method with Gaussian kernel to obtain more accurate data.

[0097] K i samples are sampled from the Gaussian function of each particle: K i = β i K max ;

[0098] Where: K i is the number of samples when β i = 1.

[0099] The number of samples is expanded to:

[0100] The target probability map of the virtual marine particle distribution is represented as:

[0101]

[0102] Where, is the position of the kth particle, σ = (σ x , σ y ) is the standard deviation;

[0103]

[0104]

[0105] Where, and are the standard deviations of the entire sample, IQR x and IQR y represent the difference between the 75th percentile and the 25th percentile of the sample.

[0106] Further: the obtained target probability map is rasterized, and the probability of each grid m (m ∈ {1, 2, …, M}) is represented by p(x m ) ∈ [0, 1], p(x m ) is obtained by integrating the target probability in the grid; the sum of the probabilities p(x m ) of all grids is:

[0107]

[0108] where vector x m represents the position of each grid.

[0109] As Figure 2 shown, the generated target probability map can well fit the particle distribution.

[0110] S2: In the GBNN method, the regional centroid is added to improve the external stimulation value of the neuron, taking the distance between the location of the unmanned ship and the regional centroid as the parameter, and planning the cooperative search path of multiple unmanned ships.

[0111] Further, step S2 further comprises:

[0112] The target probability map of the virtual marine particle distribution is rasterized, and each grid is regarded as a neuron;

[0113] In the GBNN method, the motion constraint is added to improve the external stimulation value of the neuron, the GBNN model is updated, and based on the updated GBNN model, the activity dynamic value of the neuron is calculated, the unmanned ship selects the neuron with the maximum activity dynamic value in its current neuron receptive field as the next path point, and the iteration is run, and the path planning of the unmanned ship is completed.

[0114] Define the search time of the unmanned ship as t (t∈{1,2,…,T}), and set the detection event of the unmanned ship as When the unmanned ship is in grid m, the grid is detected, that is, when the mth grid is detected, 1, otherwise 0.

[0115] All historical observation events are independent of each other. The initial time is equal to the prior probability p(x m ). Based on the above assumptions, the target probability map at time t can be updated by Bayes rule:

[0116]

[0117] where λ is the normalization factor:

[0118]

[0119] Assume that the observation system equipped by the unmanned ship can cover all the grids, and the observation probability is g s ∈[0,1], which quantifies the credibility and accuracy of the observation system and reduces instability.

[0120] where: is the credibility of the unmanned ship to detect grid m:

[0121]

[0122] Define the conditional probability that the target is not detected at time t as: This value is calculated from the updated target probability map and the detection confidence level:

[0123]

[0124] From the above analysis, the joint probability that the target has not been detected by time t is:

[0125]

[0126] Therefore, the probability of the target being detected, i.e., the cumulative detection reward, up to time t is:

[0127]

[0128] The unmanned system's search gradually covers all high-value areas over time, so the cumulative detection reward D 1:t It gradually approaches 1. The target search problem can be viewed as a path planning problem for an unmanned system with the goal of maximizing the cumulative detection reward.

[0129] This invention employs the GBNN algorithm to solve the target search problem. The principle diagram of GBNN is shown below. Figure 3 As shown, after obtaining the rasterized target probability map, each grid cell serves as a neuron in the GBNN, and each neuron can transmit activity values ​​to eight surrounding neurons. The receptive region of each neuron is defined as a circular region with radius R. The connection weights between each neuron are defined as w. mk :

[0130]

[0131] Where: |x m -x k | is the Euclidean distance between neurons m and k, and γ is a constant greater than 0.

[0132] Taking neuron m as an example, the dynamic activity value of the neuron It is iteratively updated, determined by the dynamic activity values ​​transmitted from its neighboring neurons and the external stimulation values ​​of neuron m itself.

[0133]

[0134] Where N(m) is the number of neurons within the receptive region of neuron m, i.e., the connection weight w. mk Neurons with a non-zero value, is the external stimulus signal of neuron m. The external stimulus signal of neuron m in the standard GBNN model is composed of the reward value of single detection (a positive constant) and the penalty value of obstacle avoidance (a negative constant). Since there are some obstacles such as reefs in some sea areas, the obstacle avoidance constraint S F is set. When the obstacle avoidance constraint is not satisfied, i.e., x m ∈ S F , the external stimulus signal of neuron m is defined as the penalty value, i.e., a negative constant, regardless of the reward value of single detection. If the obstacle avoidance constraint is satisfied, i.e., x , the external stimulus signal of neuron m only considers the reward value of single detection, denoted as , which can be obtained from the above formula.

[0135]

[0136] Therefore, the external stimulus signal is defined as:

[0137]

[0138] -E is used as the penalty value of neuron m and is not transmitted to other neurons. The function f(·) is used to normalize the activity output value of the entire neural network. Based on the above analysis, the activity dynamic value of neuron m can be redefined as:

[0139]

[0140] where N(m) is the neuron in the receptive field of neuron m, and is the external stimulus signal of neuron m.

[0141] Based on the above analysis, each neuron can transmit the activity value to the adjacent neurons, and therefore the detection reward of each grid is gradually transmitted to the entire neural network. It can be seen that the GBNN can guide the unmanned ship to search the target from the perspective of global optimization to a certain extent.

[0142] Although the standard GBNN model has certain global optimization characteristics, the transmission of neuron information requires time, and the transmission of activity value will gradually decrease in the process (w mk <1), so the unmanned ship system still has the possibility of falling into local optimization in some complex situations.

[0143] Therefore, the regional centroid weight stimulus can be introduced into the GBNN model as part of the external stimulus signal of the neuron, and the distance between the unmanned ship and the regional centroid is used as a parameter to assist in guiding the search process and ensure that the search path is in a more efficient direction. The regional centroid is defined as follows:

[0144]

[0145] V is the entire area of the target probability map, p(x m ) is the probability density function, A is the quality of the target probability map, and CM is the area centroid of the target probability map.

[0146] In addition, in some embodiments, the motion cost of the unmanned ship's travel direction is also taken into account. The conversion cost of the motion direction has a great influence on the energy consumption and search efficiency of the unmanned ship. Assuming that the position of the unmanned ship at a moment is neuron l, the current position is neuron m, and the next step moves to neuron k, define Δθ as the angle between vectors x k -x m and x m -x l , and the motion cost is defined as

[0147]

[0148] wherein x k represents the position of neuron k, x m represents the position of neuron m, and x l represents the position of neuron l.

[0149] Based on the above analysis, the improvement of the external stimulation value is as follows:

[0150]

[0151] q is a constant greater than 0, and selecting a suitable q can make the additional stimulation value have a suitable weight in the entire external stimulation signal. is the Euclidean distance between the area centroid and the position of the unmanned ship at the next moment. α is a normal number less than 1.

[0152] In the improved GBNN model, the external stimulation signal is as follows:

[0153]

[0154] Based on the updated GBNN model, the activity dynamic value of the neuron is calculated, and the unmanned ship selects the neuron with the largest activity dynamic value in its current neuron receptive field as the next path point. Such iterative operation can complete the path planning of the unmanned ship. As shown in FIG. 4, four unmanned ships can cover most of the high-value areas, and have a higher search efficiency.

[0155] S3: Use solar unmanned aerial vehicles as communication relays for multiple unmanned ships, use the minimum spanning tree method to obtain the optimal communication network connection between the unmanned ships and the unmanned aerial vehicles, set the communication performance indicators, and use the greedy strategy to plan the path of the unmanned aerial vehicles.

[0156] In the process of multi-agent unmanned aerial vehicle and unmanned ship based on solar energy to complete the target search task, keeping the communication of each unit of the system is the key to the efficient operation of the system, but the actual execution search task, unmanned ship will be affected by many factors between the communication terminal. The present application researches the network topology connectivity maintenance problem, and establishes a communication model. The communication model in the real environment is used, and the successful transmission probability between each node (i.e. agent) is used to quantify the communication strength between agents.

[0157] Each unmanned aerial vehicle and unmanned ship is regarded as a node, and it is assumed that node i sends a signal to node j, and the transmission power is P i , and the average power of the signal transmission process is noisy. Therefore, the received SNR (signal-to-noise ratio) of the signal sent by node i to node j is defined as Γ ij :

[0158]

[0159] Where: h ij represents the factor of multipath fading.

[0160] Where the channel gain is G ij :

[0161]

[0162] C ij is a constant related to antenna gain and shadowing, h ij is the factor of multipath fading, D ij is the distance between two unmanned aerial vehicle nodes, and alpha is the propagation loss factor. For successful transmission with acceptable small packet loss, the SNR needs to be high enough, or higher than the required minimum level of link quality gamma. Therefore, if h ij is a Gaussian distribution with zero mean and unit variance, the probability of successful transmission between two nodes i and j can be expressed as:

[0163]

[0164] If is less than a certain threshold, the two nodes are considered disconnected. For communication relay, it is assumed that the network is connected to each other with sufficient transmission power, and the probability of such connectivity is used to optimize the network.

[0165] Due to the large number of solar-powered unmanned aerial vehicles and unmanned ships, the connection mode is also complex and diverse, in this case, the minimum spanning tree (MST) based on graph theory is used to obtain the minimum cost link to obtain the highest success transmission probability. The distance of each communication node of the unmanned system is defined as the weight of the tree branch. The connection network with the minimum cost sum is obtained by comparing the weight sum of all possibilities. The communication network is connected by using the MST, and the optimal adjacency matrix is solved.

[0166] To solve the MST problem, Kruskal algorithm with high efficiency is used, which will always select the smallest available edge and judge whether the endpoints are connected, if not, connect the two endpoints, until each point is connected. The algorithm block diagram is shown in Figure 5 .

[0167] The weight W ij is defined as the negative correlation function of the transmission success probability:

[0168]

[0169] The transmission probability is higher, the weight W ij is smaller, and the possibility of message transmission to each node of the network is greater, so the network with the highest transmission success probability can be obtained by MST.

[0170] Taking the maximum network communication performance as the index, the solar-powered unmanned aerial vehicle path is optimized. Assuming that there are U1 solar-powered unmanned aerial vehicles and U2 unmanned ships in the communication network, the optimal adjacency matrix R is solved by connecting the communication network by MST, if node i and node j are directly connected, R ij = 1, otherwise R ij = 0. According to the above analysis, the communication performance index is set as:

[0171]

[0172] J s is the communication network connection performance index, the smaller the index, the better the communication effect, and the next path point of the unmanned aerial vehicle is directly determined by using the greedy strategy. Under the premise of the unmanned ship path shown in Fig. 4, the planned path of the solar-powered unmanned aerial vehicle is shown in Figure 6 .

[0173] The application can realize high-precision prediction of a sea floating object probability graph, and conforms to the real situation; the improved GBNN algorithm of regional centroid guidance and motion constraint is introduced to plan the unmanned ship path, which can better avoid local optimum and greatly improve the search efficiency; the solar unmanned plane is introduced as the communication relay between the multiple unmanned ships, the optimal communication network connection is established and the unmanned plane path is planned, so that the unmanned ships can still share information and coordinate behaviors under a large range of sea surface.

[0174] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in other forms. Any person skilled in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution content of the present application, in accordance with the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. A heterogeneous cooperative path planning method for searching floating objects at sea, used for planning the search path of unmanned surface vessels for floating objects at sea, characterized in that, Includes the following steps: S1: Release virtual ocean particles in the target area where the floating target is located at sea. The virtual ocean particles are used to represent the floating target at sea. Input the virtual ocean particles into the Parcelv2 particle simulator based on the Lagrange framework, predict and track the position of the virtual ocean particles, and use the Parzen window method based on the Gaussian kernel function to obtain the target probability map of the virtual ocean particle distribution. S2: In the GBNN method, the region centroid is added to improve the external stimulus value of the neuron. The distance between the location of the unmanned ship and the region centroid is used as a parameter to plan the cooperative search path of multiple unmanned ships. S3: Utilize solar-powered UAVs as communication relays for multiple unmanned surface vessels (USVs), use the minimum spanning tree method to obtain the optimal communication network connection for the heterogeneous USV / UAV swarm system, set communication performance indicators, and employ a greedy strategy to plan UAV paths.

2. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 1, characterized in that, Step S2 further includes: The target probability map of the virtual ocean particle distribution is rasterized, and each grid is regarded as a neuron; Motion constraints are added to the GBNN method to improve the external stimulus values ​​of neurons, and the GBNN model is updated. Based on the updated GBNN model, the activity dynamic value of neurons is calculated. The unmanned vessel selects the neuron with the largest activity dynamic value in its current neuron's receptive region as the next path point. This process is repeated iteratively to complete the unmanned vessel's path planning.

3. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 1 or 2, characterized in that, Step S1 includes: Solve the equation: Where: X is the three-dimensional position of the virtual ocean particle, t represents any time, Δt represents the sampling period, v(x,τ) is the three-dimensional velocity field of the ocean current at that position, and ΔX b (t) represents the positional change of the particle caused by its behavior at sea; The Parcel v2 particle simulator was used to predict the positions of virtual ocean particles; Determine the corresponding Gaussian function based on the predicted position of each particle: in: For the particle's geographical location, β i ∈[0,1] represents the particle's credibility. Let S represent the particle distribution range, where i represents the i-th particle, and the particle has the property S = {s} i ,β i ,λ i }, i = 1, ..., N, where N represents the total number of particles in the target area; Sample K from the Gaussian function of each particle. i Samples: K i =β i K max ; Where: K i It is β i Number of samples when = 1; Expand the sample size to: The target probability map of the virtual ocean particle distribution is represented as: in, It is the sampling position of the k-th particle, σ = (σ x ,σ y ) is the standard deviation; in, and It is the standard deviation of the entire sample, IQR. x and IQR y This represents the difference between the 75th percentile and the 25th percentile of the sample.

4. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 3, characterized in that, Further includes: The obtained target probability map is rasterized, and the probability of each grid m (m∈{1,2,…,M}) is represented by p(x). m p(x) ∈ [0,1] represents p(x) m The probability p(x) is obtained by integrating the probability of the target within the grid; the probability p(x) of all grids is obtained by... m The sum of ) is: Where: vector x m This indicates the position of each grid cell.

5. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 4, characterized in that, Update the target probability map at time t using Bayesian rules: Where: λ is the normalization factor. Where: t is any search time. Let t be the detection event where the m-th grid is detectable. The unmanned surface vessel is only detected when the m-th grid is detected. The value is 1 when the m-th cell is not detected. =0; in: The reliability of unmanned surface vessel detection of grid m: Treat each grid cell as a neuron and define the connection weights w between neurons. mk for: Where |x m -x k | is the Euclidean distance between neurons m and k, and γ is a constant greater than 0; Define the dynamic value of neuron m's activity: Where: N(m) is the number of neurons in the receptive region of neuron m. It is the external stimulation signal of neuron m; Where: -E is the penalty value inside the neuron, which will not be passed on to other neurons; Define the centroid of the search region: V represents the entire region of the target probability map, p(x) m ) is the probability density function, A is the quality of the target probability map, and CM is the centroid of the region of the target probability map; Using the centroid as the external stimulus signal for neurons, and taking the location of the unmanned vessel and the distance to the centroid of the region as parameters, a collaborative search path for multiple unmanned vessels is planned.

6. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 5, characterized in that, Methods for incorporating motion constraints into the GBNN method include: Define Δθ as a vector x k -x m With x m -x l The angle between them is defined as the motion cost. x k Indicates the location of neuron k, x m Indicates the location of neuron m, x l Indicates the location of neuron l; Improvements to the external stimulus values ​​of the GBNN model: Where: q is a constant coefficient greater than 0, Let α be the Euclidean distance between the region's centroid and the unmanned vessel's position at the next moment, where α is a positive constant less than 1; In the improved GBNN model, the external stimulus signal is:

7. The heterogeneous cooperative path planning method for searching floating objects at sea as described in claim 1 or 2, characterized in that, Step S3 includes: Define a weight, setting it as a negative correlation function of the probability of successful transmission: in: Let be the communication probability between the i-th and j-th nodes; The optimal communication network connection for the heterogeneous UAV / unmanned surface vessel system is determined using MST, and the communication performance indicators are set as follows: Where: U1 is the number of solar-powered drones, U2 is the number of unmanned boats, and R is the optimal adjacency matrix of the communication network. If node i and node j are directly connected, then R... ij =1, otherwise R ij =0; The communication performance indicators are set as follows: J s This is a performance indicator for communication network connectivity. The smaller the indicator, the better the communication effect. A greedy strategy is directly used to determine the next waypoint of the UAV.

Images