A method and system for underwater unmanned submersible track fusion based on graph convolutional neural network
By converting track data into graph data and fusion using graph convolution neural network, the problem of the need for more prior knowledge and insufficient adaptability in complex environments in the existing technology is solved, and the problem of incomplex marine environments is realized. High accuracy and rapid track fusion.
Patent Information
- Application Number
- CN202410536580.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-30
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-04-30
AI Technical Summary
The existing track fusion method requires a lot of prior knowledge, poor adaptability, and insufficient fault tolerance in complex dynamic marine environments, making it difficult to achieve real-time and high-accurate track fusion in complex environments.
Using a graph convolution neural network-based method, the track data after spatiotemporal calibration is converted into graph data, nodes and edges are represented through undirected graphs, and a pre-trained graph convolution neural network model is input to achieve track fusion, using multiple The graph convolution network of the receptive field processes multi-sensor data, revealing the patterns and trends implicitly in data.
In complex dynamic marine environments, the accuracy and speed of track fusion are improved, complex track data can be processed more effectively, and has good engineering practicality and adaptability.
Smart Images

Figure CN118445753B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of track fusion, and more specifically, relates to a method and system for underwater unmanned submersible track fusion based on graph convolutional neural network. Background Art
[0002] The rapid development of science and technology in the modern military field has made the way of war no longer limited to land, sea and air, but covers a wider range of fields. In modern command automation technology systems, the information provided by a single sensor can no longer meet the needs of modern warfare, especially in complex electronic warfare environments. The combat performance of a single platform is severely restricted, resulting in a decline in positioning, tracking and navigation performance. In order to make up for this deficiency, multi-sensor information fusion technology has gradually become the focus of attention and is widely used in various weapon system platforms.
[0003] Traditional data fusion methods have many limitations. Algorithms based on Bayesian estimation require a priori measurement variance, which limits their scope of application. The Kalman filter method is suitable for multi-sensor redundant data fusion, but its recursive characteristics and storage space requirements limit real-time performance. The nearest neighbor data association algorithm performs poorly when the number of targets or interference density is large. The global nearest neighbor data association algorithm is based on the NN algorithm. Although it improves the fusion performance, it has a large amount of calculation. The probabilistic data association method is only applicable to single target tracking problems. The multiple hypothesis tracking method (MHT) solves multi-target tracking, but the number of hypotheses increases dramatically over time. The weighted fusion algorithm directly operates on the data source, but the weights are difficult to determine. Although the DS evidence theory does not require prior information, it has strict requirements on the conditions for evidence combination. These defects make traditional methods limited in dealing with complex environments, improving real-time performance, and determining weights. Summary of the invention
[0004] In view of the above defects or improvement needs of the prior art, the present invention provides an underwater unmanned submersible track fusion method and system based on graph convolutional neural network, which aims to solve the problems of existing track fusion methods in complex dynamic ocean environments, such as high requirements for prior knowledge, weak adaptability, and weak fault tolerance.
[0005] To achieve the above object, according to a first aspect of the present invention, a track fusion method based on a graph convolutional neural network is provided, comprising:
[0006] S1, obtain the track data Z observed by observation platforms A and B respectively A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `;
[0007] S2, respectively Z A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph;
[0008] The preprocessing includes: normalizing the data to be processed, dividing the data according to a preset length to obtain a plurality of subsamples, taking each subsample as a node, and determining the connected edges between the nodes according to the similarity between the nodes to construct an undirected graph;
[0009] S3, Z A `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
[0010] According to a second aspect of the present invention, there is provided an underwater unmanned submersible track fusion device based on a graph convolutional neural network, comprising:
[0011] The first processing module is used to obtain the track data Z observed by the observation platforms A and B respectively on the observation target A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `;
[0012] The second processing module is used to respectively A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph;
[0013] The preprocessing includes: normalizing the data to be processed, dividing the data according to a preset length to obtain a plurality of subsamples, taking each subsample as a node, and determining the connected edges between the nodes according to the similarity between the nodes to construct an undirected graph;
[0014] The third processing module is used to convert Z A `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
[0015] According to a third aspect of the present invention, there is provided an underwater unmanned submersible track fusion system based on a graph convolutional neural network, comprising: a computer-readable storage medium and a processor;
[0016] The computer-readable storage medium is used to store executable instructions;
[0017] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the method described in the first aspect.
[0018] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method described in the first invention.
[0019] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:
[0020] 1. The method provided by the present invention is based on the idea of deep learning, which converts the track data after time and space calibration into graph data, so as to reveal the relationship between different nodes; by representing data points as nodes of an undirected graph and the relationship between nodes as edges, it is possible to model and analyze data more intuitively, revealing the patterns and trends hidden in the data; the graph data is input into a pre-trained graph convolutional neural network model to obtain the track fusion result, which can realize the track fusion of underwater unmanned submersibles in a complex dynamic ocean environment, and effectively solve the problems of accuracy and speed in the track fusion process. Taking the track planning simulation of the unmanned submersible in the dynamic ocean simulation environment as an experimental example, in the actual application process, it can be extended to unmanned equipment such as surface unmanned boats and underwater robots, and has good engineering practicality.
[0021] 3. The method provided by the present invention adopts a graph convolutional network with multiple receptive fields. This network processes multiple graph inputs and considers the specific receptive fields of different nodes, so that the network can understand and analyze the information in the track data more comprehensively and flexibly. This is very useful for processing complex track data with diverse connection relationships. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 A flow chart of an underwater unmanned submersible vehicle track fusion method based on a graph convolutional neural network provided in an embodiment of the present invention;
[0023] Figure 2 A schematic diagram of normalizing and segmenting track data provided by an embodiment of the present invention;
[0024] Figure 3 A schematic diagram of a process for constructing graph nodes from underwater unmanned submersible track sample data provided by an embodiment of the present invention;
[0025] Figure 4 A schematic diagram of a multi-receptive field graph convolution layer processing used in an underwater unmanned submersible graph convolution network provided in an embodiment of the present invention;
[0026] Figure 5 A schematic diagram of a simulated unmanned underwater vehicle provided by an embodiment of the present invention;
[0027] Figure 6 (a) and (b) are respectively the first and second comparison diagrams of the results of track fusion using the method provided by the embodiment of the present invention and using other neural networks. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0029] The embodiment of the present invention provides an underwater unmanned submersible track fusion based on a graph convolutional neural network, such as Figure 1 As shown, including:
[0030] S1, obtain the track data Z observed by observation platforms A and B respectively A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `;.
[0031] The observation target is an underwater unmanned submersible, a surface unmanned boat, an underwater robot or an aircraft, etc.; the track data includes information such as the coordinates and time of the observation target.
[0032] Taking the underwater unmanned submersible as an example, the track data of the underwater unmanned submersible in normal motion state is obtained from multiple sensors (respectively set on the observation platform); after each single observation platform obtains the position estimate of the target, it uses the tracking processor owned by each observation platform to process the feature information into underwater acoustic local track data and transmit it to the fusion center.
[0033] It can be understood that the method provided by the present invention is used to fuse the track data of the target unmanned underwater vehicle obtained based on the observation data of the same target unmanned underwater vehicle by any two observation platforms.
[0034] When performing time registration, any existing time registration algorithm can be used, and the present invention does not limit this to a unique algorithm. For example, a time registration algorithm based on the least squares criterion is used to perform time registration on the sensor data of an underwater unmanned submersible:
[0035] Assume that there are sensor a (installed on observation platform A) and sensor b (installed on observation platform B), with sampling periods τ and T respectively, and the relationship between the two is τ:T=m. The last observation of sensor a in the target state occurred at time point (k-1)τ, k is the number of sampling cycle times, and the time of the next observation and update of the target state is kτ=(k-1)τ+mT.
[0036] Then the measured value z of sensor b is i It can be expressed by the following formula:
[0037]
[0038] ν i To measure the noise, rewrite it into vector form:
[0039] Z m =W m U+V m
[0040] According to the least squares criterion, the optimization target can be obtained:
[0041]
[0042] According to this optimization goal, the partial derivative is calculated to make it equal to 0, and the final measurement value is:
[0043]
[0044] Among them, z i is the measurement value of sensor b, c1 and c2 are constants, and the noise variance of this measurement result is:
[0045]
[0046] Among them, σ r 2 is the position measurement noise variance before fusion.
[0047] Between these two observations, sensor b obtained m measurements. By using the least squares method, the m measurements of sensor b are fused to form a measurement that is aligned in time with sensor a. This process eliminates the asynchrony of the two sensors' observation track points of the target state, effectively eliminating the negative impact on multi-sensor information fusion.
[0048] When performing spatial registration, any existing spatial registration algorithm may be used, and the present invention does not limit this to a unique algorithm. For example, a spatial registration algorithm based on Gauss-Krüger projection is used to perform spatial registration on sensor data of an underwater unmanned submersible.
[0049] The Gauss-Krüger projection is a plane obtained by projecting the area within a certain diameter difference range on both sides of the central meridian onto an elliptical cylinder, and then unfolding the cylinder. It is mainly used to map the three-dimensional geographic coordinates (longitude, latitude, elevation) on the earth's surface to a two-dimensional coordinate system on the plane. The origin O of the Gauss plane rectangular coordinate system is the projection of the intersection of the central meridian and the equator, the horizontal axis is the projection of the equator, and the vertical axis is the projection of the central meridian. The relationship between the Gauss plane coordinates (x, y) and the geodetic coordinates (L, B) can be expressed by the forward and inverse calculation formulas of the Gauss projection:
[0050] Among them, B is the geodetic latitude, l is the longitude difference from the central meridian, L0 is the longitude of the central meridian. X represents the arc length from the equator to the meridian, C is the polar radius of the reference ellipsoid, e is the second eccentricity of the reference ellipsoid, N represents the radius of curvature of the meridian circle, M represents the radius of curvature of the meridian circle, and t represents tanB.
[0051] This method uses a time registration algorithm based on the least squares criterion and a space registration algorithm based on Gauss-Krüger projection to align data in time and space, which helps to eliminate time and space errors caused by different sensor sampling rates, position errors, or system clock differences. By accurately matching time and position, data fusion and analysis can be performed more accurately, improving the overall performance of the system.
[0052] S2, respectively Z A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph;
[0053] The preprocessing includes: normalizing the data to be processed, dividing it according to a preset length to obtain a plurality of sub-samples, taking each sub-sample as a node, and determining the connecting edges between the nodes according to the similarity between the nodes to construct an undirected graph.
[0054] Z A `、Z B `Normalize and segment the data to obtain a series of nodes, so that the local trajectory data is converted into graph data, and the corresponding edges are connected between the graph nodes through a specific similarity calculation to respectively convert Z A `、Z B `Convert to graph data.
[0055] For the original time series X with a signal length of L, data normalization is usually performed on it, and the normalized data can be expressed as:
[0056] X nol =normalize(X)
[0057] Among them, Xnol is a normalized time series, and this method uses maximum and minimum normalization. After data normalization, the normalized data can be divided into a series of data samples with a predefined sub-sample length or dimension d.
[0058] Normalized data sub-sample division is as follows Figure 2 shown.
[0059] Therefore, the constructed dataset is:
[0060]
[0061] Where M is the constructed sample set, represents the sub-sample, which represents each node data group after segmentation, L is the total length of the data to be processed, n is the total number of sub-samples, floor() means rounding to the lower limit, and d is the length of the sub-sample. In order to reduce the influence of noise, each sub-sample is subjected to a fast Fourier transform (FFT), and the obtained data is considered to be a new node. This process is expressed as:
[0062]
[0063] Each group of sub-data in the normalized and Fast Fourier Transformed (FFT) data is mapped to a low-dimensional space as a node of an undirected graph.
[0064] After data segmentation, it is necessary to calculate the similarity between nodes to determine whether they should be connected by edges. Usually, similarity calculation can use metrics such as cosine similarity and Euclidean distance. In this method, the ε radius is mainly used to find the edges that should be connected:
[0065]
[0066] Where ψ(x i ) represents the node x i The collection of nodes whose similarity is greater than or equal to ε, For output and node The nodes whose similarity is greater than the threshold ε, y i For The set of all nodes except , ε is the similarity threshold which can be set according to the actual situation.
[0067] The weights of the edges of the generated graph are:
[0068]
[0069] The process of constructing graph nodes from samples is as follows: Figure 3 shown.
[0070] S3, ZA `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
[0071] Preferably, the graph convolutional neural network model is a graph neural network model with multiple receptive fields, wherein the graph convolutional neural network is a graph convolutional network with multiple receptive fields. This network has a relatively powerful node feature extraction capability. This network includes at least one multi-receptive field graph convolution layer for aggregating information from multiple different receptive fields. This process can aggregate information from different numbers of nodes, and then fuse the aggregated features into a vector representation of the aggregated information.
[0072] Build a graph neural network model with multiple receptive fields, send the original data samples into the graph neural network as training data, train the model with the difference between the actual fusion result and the model output result as the loss function, and obtain a graph neural network model with spatiotemporal capture characteristics.
[0073] Preferably, the graph convolutional neural network model includes a first, a second, and a third graph convolutional layer and a fully connected layer connected in sequence.
[0074] Preferably, the graph convolutional neural network model is a multi-receptive field graph convolutional neural network model; at least one of the first, second or third graph convolutional layers is a multi-receptive field graph convolutional layer.
[0075] This network has three layers: the first layer performs four convolution operations on the graph data, the number of graph nodes after convolution remains unchanged, and the number of features of a single node is doubled; the second layer performs a second convolution operation on the graph data, the number of graph nodes after convolution remains unchanged, and the number of feature maps of a single node is doubled; the third layer performs a convolution operation on the image, the size of the graph data after convolution is reduced to 1 / 4 of the size of the previous feature map, and the number of feature maps is increased to 4 times the number of the previous feature map. Finally, an identity mapping is introduced between feature maps with the same size and number of node features, which is implemented through a fully connected layer. The entire network structure can be expressed as:
[0076]
[0077] f(X,A)=MP3(MP2(MP1(X,A),A),A).
[0078] MP i Represents each convolutional layer, W i represents the weights of different layers, (X, A) represents graph data, Degree matrix representing the graph data.
[0079] Preferably, this model uses the cross entropy cost function as the loss function.
[0080] The graph data obtained in step S2 can be defined as an undirected graph G = (V, ξ, A), where V = n represents a finite number of nodes, ξ is the edge set of graph G, and A∈R n*n Represents the adjacency matrix of the graph G. For two nodes in the graph <ν i ,ν j >, A ij The value of can be expressed as 0 or 1, which is 1 when the two nodes are connected by an edge and 0 when they are not connected by an edge.
[0081] As a type of data on a non-Euclidean domain, graphs are usually represented by adjacency matrices. The Laplacian matrix of a graph is defined as: L = DA where D is the diagonal matrix of A. Usually, the symmetric normalized graph Laplacian operator is used, which is defined as:
[0082] L=I N -D -1 / 2 AD -1 / 2
[0083] Among them I N is the identity matrix:
[0084] The eigenvalue decomposition of the symmetric normalized graph Laplacian operator is:
[0085]
[0086] where Λ=diag(λ1,λ2,...,λ n ) is a diagonal matrix consisting of the eigenvalues of L, and U is an orthogonal matrix consisting of the eigenvectors of L, for example, U -1 =U T The spectral graph convolution of node v and node features can be defined as:
[0087] h=(x*f) θ =U(U T xU T )f
[0088] Where x is the node feature, h is the feature map after graph convolution. f is the eigenfunction of Λ, and θ is the learnable parameter. θ =U T f is used as a learnable graph convolution filter, then the above formula is simplified to:
[0089] h=Uf θ U T x
[0090] After approximating the filter with Chebyshev polynomials, the graph convolution of node features and the filter in the spectral domain can be mathematically defined as follows:
[0091]
[0092] In order to increase the nonlinearity of graph convolution, the graph convolution result is nonlinearly activated by the activation function, which is expressed as:
[0093]
[0094] ReLU is usually used as an activation function. In this method, a three-layer graph convolutional network is used as the model, and ReLU function is used for activation. This structure can be expressed as:
[0095]
[0096] f(X,A)=MP3(MP2(MP1(X,A),A),A).
[0097] The main idea of the graph convolutional network (GCN) is to use the neighbor information in the graph structure to learn the representation of the node in order to better capture the relationship between multiple track data and the topological structure of the graph. The graph convolutional network updates the representation of the node to the weighted average of its neighbor nodes by aggregating the neighbor information of each track data node. Such an operation enables the node to use the information of its surrounding nodes to more comprehensively express the characteristics of the node. The graph convolutional network introduces a graph convolution layer on the graph structure, which is similar to the convolution layer in the traditional convolutional neural network. The graph convolutional layer considers the neighbors of the node in the node representation update and keeps the parameter size small by weight sharing. The graph convolution operation can be regarded as a process of filtering the input track data features. This filtering operation adapts to different track data structures by learning the weight matrix, making the model more universal. The network parameters of the graph convolutional network are shown in Table 1:
[0098] Table 1
[0099]
[0100] The multi-receptive field graph neural network model processes graph data as follows Figure 4 shown.
[0101] The graph data obtained in step S2 is used as the input of the trained graph neural network to obtain the result after target track fusion.
[0102] In order to verify the effectiveness of the underwater unmanned submersible track fusion method and system proposed in the present invention, model training and performance verification are carried out in combination with actual application scenarios.
[0103] The specific steps are as follows:
[0104] (1) Scenario Construction
[0105] Build the main scene of track planning including unmanned submersible, starting point and destination. To ensure the generalization and practicality of the trained underwater equipment track fusion algorithm. In the scene, the starting point and destination positions of the unmanned submersible are random values within a certain range, and the unmanned submersible randomly selects one of the designed routes for movement. Set up three observation stations to observe multiple targets. Each observation station observes the target. Through the observation matrix and observation noise variance, the distance and azimuth information with errors can be obtained respectively, thereby obtaining the local track information, where the noise variance changes with the distance between the target and the observation station.
[0106] Specifically, in order to simulate the actual application tasks of the target location of the unmanned underwater vehicle, combined with the actual marine environment, a main scene containing the track data fusion of the unmanned underwater vehicle and the destination is built. To ensure the generalization and practicality of the training, the starting point and the end point, the starting point and the destination position of the unmanned underwater vehicle in the scene are all random values within a certain range. In order to reduce unnecessary collisions, the pool has been pre-processed with debris. Figure 5 As shown in the figure, the unmanned underwater vehicle is set as a small ball with a radius of 0.5m and a maximum speed of 1m / s. The initial horizontal coordinate of the center of the circle is randomly selected in the range of [-2m, 2m], and the initial vertical coordinate is -5m. The destination is set as a square range with a length of 3m and a width of 3m. The initial horizontal coordinate of the center of the square is randomly selected in the range of [-2m, 2m], and the initial vertical coordinate is 0m. The sampling interval of the data is 1s, and the main scene is set as a square area with a side length of 15m and a center of [0m, 0m]. The horizontal and vertical coordinates of the three observation stations are [0m, 5m], [5m, 0m], and [7.5m, 7.5m], respectively. The noise variance of the target initial state process is 0.012. Three observation stations are set to observe four targets. Each observation station observes the target. Through the observation matrix and the observation noise variance, the distance and azimuth information with errors can be obtained respectively, thereby obtaining the local track information, where the noise variance changes with the distance between the target and the observation station. The observation station transmits local track information to the fusion center every 2 seconds, and each observation platform transmits local track data 1000 times.
[0107] The target's motion model is modeled and analyzed through the track data of multiple sensors, and the interference caused by the complexity of the underwater environment is fully considered to adjust the observation noise model.
[0108] The motion models include uniform linear motion (CV), uniform accelerated motion (CA) or uniform circular motion (CT). This example generates tracks of local hydroacoustic positioning data based on the three motion models. By adopting local tracks that are more consistent with actual hydroacoustic positioning data, the features of the hydroacoustic positioning data can be more fully extracted in the track association and fusion steps, thereby ensuring the reliability of the method.
[0109] In the uniform linear motion model, the state space variable model of the system is:
[0110] Z(k)=HX(k)+ν(k)
[0111]
[0112] In the uniform acceleration motion model, the state space variable model of the system is:
[0113]
[0114] In the uniform circular motion model, the state space variable model of the system is:
[0115]
[0116] Among them, v(k) is the noise, Z(k) is the observed state of the target, and X(k) is the true state of the target.
[0117] The background interference of underwater acoustic equipment includes reverberation and noise. Generally speaking, they exist at the same time, but the impact is different in different situations. Under certain conditions, the intensity difference between reverberation and noise is so large that the smaller one can be ignored. If the noise masking level is noise masking level 1, R r <R s , at a distance R>R r When the sound signal detected by the sonar device is masked by the reverberation, it means that the effective range of the sonar is limited by the reverberation. If the noise masking level is level 2, then due to the high noise masking level, the limited distance is determined by R. s becomes R' s , which is better than R r If the range is small, it will be limited by noise. s is the noise masking level when reverberation is the main background interference, R' s is the noise masking level when the echo is the main background interference, R r It is the noise masking level when the reverberation masking curve intersects the echo curve.
[0118] In underwater noise research, in order to simplify the processing, a certain type of interference noise is usually assumed to be Gaussian white noise. Therefore, the observation noise ν(k) is assumed to be Gaussian white noise with zero mean and variance 2σv. From the above analysis, it can be seen that the size of the variance 2σv can be set by the distance between the observation station and the target. At present, the positioning accuracy of the domestic ultra-short baseline is about 2%, and the high accuracy can reach 0.2%. Therefore, the noise variance can be expressed as follows:
[0119]
[0120] represents the noise variance.
[0121] In view of the complexity of the underwater environment, the observation noise model was adjusted, and the key parameters that need to be retained were carefully selected to promote the full use of the local underwater acoustic track data by the track association and fusion model. This optimization strategy helps to improve the accuracy of the underwater target position and better adapt to the complexity of the actual underwater acoustic positioning scene; this example uses local track data modeling, which can make the model more adaptable to the characteristics and changes of the local environment and improve the computational efficiency of the model, improve the generalization ability of the model, and make it more suitable for application in a variety of underwater acoustic scenes, which is very important for tasks that require rapid decision-making and real-time application, and improve the adaptability and real-time performance of the method.
[0122] (2) Model construction and training
[0123] The actual multi-sensor data of the underwater unmanned submersible under different routes are obtained as the input of the graph convolutional neural network model. The parameters of the graph convolutional neural network model under different scenarios are set. The obtained observation data is preprocessed accordingly to obtain the graph structure. During the training process of the reinforcement learning algorithm, the learning rate is set to 0.0005, the batch_size is set to 128, the discount factor is set to 0.95, and the maximum iteration round is 300 steps. The features of the nodes of the graph data are used as input. This can be a vector of information such as the position and speed of the target. After that, through the graph convolution layer, the model can aggregate the neighbor information of each node and update the representation of the node. This helps to consider the relationship between the target and its neighboring targets. Stack multiple layers of graph convolution to gradually extract higher-level and more abstract features. This helps to capture more complex relationships between targets. The loss function used is the cross entropy loss function.
[0124] The final output of the graph neural network (i.e., the track fusion result) can be used as a new representation of each data node. This representation can include the learned information of neighboring data nodes. And a fusion strategy can be used to combine the graph neural network output with the results of traditional fusion methods. This can be simple linear fusion, weighted average, etc.
[0125] (3) Method comparison
[0126] Build an aircraft fusion architecture based on graph convolutional networks. Set up several underwater unmanned submersibles for training in the main scene, and compare the results of the fusion of the returned sensor data with the actual track to obtain the reliability of the experimental model. The track types used in this experiment are shown in Table 2:
[0127] Table 2
[0128]
[0129]
[0130] Four deep learning methods were used to demonstrate the effectiveness of this achievement, including convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM) and Chebyshev graph network (ChebyNet). The statistical results of 100 random experiments are compared and shown in Table 3:
[0131] Table 3
[0132] method Track fusion accuracy Training time CNN 76% 2h15m LSTM 86% 3h45m RNN 83% 3h10m ChebyNet 90% 4h21m Method of the present invention 94% 3h32m
[0133] Figure 6 The comparison chart of the results of track fusion using different neural networks is shown in the attached figure. Figure 6 As shown in (a) of the figure, the training comparison of different neural networks for track fusion is shown in the attached figure. Figure 6 As shown in (b) in .
[0134] Experimental results show that the proposed method further improves the accuracy of track fusion results while maintaining a low running time, and has the highest success rate in the entire track fusion task.
[0135] In summary, the method provided by the present invention includes: data acquisition, obtaining the track data of the underwater unmanned submersible through multiple sensors; feature calculation, normalizing and low-dimensional space mapping the track data of the underwater unmanned submersible, and then embedding the graph nodes so that each observation point is a graph node; edge connection strategy design, constructing the data through data processing and low-dimensional space mapping into the form of edge connection, and the process includes: graph structure establishment, familiarity calculation, selection of appropriate thresholds, and edge connection; graph neural network model construction, sending the original data samples as training data into the graph neural network, and training the model with the difference between the actual fusion result and the model output result as the loss function to obtain a graph neural network model with spatiotemporal capture characteristics.
[0136] The method provided by the present invention adopts local track data modeling, which can make the model more adaptable to the characteristics and changes of the local environment and improve the calculation efficiency of the model; at the same time, a time registration algorithm based on the least squares criterion and a space registration algorithm based on Gauss-Kruger projection are used to align the data in time and space, which helps to eliminate the time and space errors caused by factors such as different sensor sampling rates, position errors or system clock differences; the track data after time and space calibration is converted into graph data, so that the relationship between different nodes can be revealed; a graph convolution network with multiple receptive fields is used, so that the network can understand and analyze the information in the track data more comprehensively and flexibly. This is very useful for processing complex track data with diverse connection relationships.
[0137] An embodiment of the present invention provides an underwater unmanned submersible track fusion device based on a graph convolutional neural network, comprising:
[0138] The first processing module is used to obtain the track data Z observed by the observation platforms A and B respectively on the observation target A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `;
[0139] The second processing module is used to respectively A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph;
[0140] The preprocessing includes: normalizing the data to be processed, dividing the data according to a preset length to obtain a plurality of subsamples, taking each subsample as a node, and determining the connected edges between the nodes according to the similarity between the nodes to construct an undirected graph;
[0141] The third processing module is used to convert Z A `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
[0142] An embodiment of the present invention provides an underwater unmanned submersible track fusion system based on a graph convolutional neural network, comprising: a computer-readable storage medium and a processor;
[0143] The computer-readable storage medium is used to store executable instructions;
[0144] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method described in any of the above embodiments.
[0145] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method described in any of the above embodiments.
[0146] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A track fusion method based on graph convolutional neural network, characterized in that: include: S1, obtain the track data Z observed by observation platforms A and B respectively A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `; S2, respectively Z A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph; The preprocessing includes: normalizing the data to be processed, dividing the data according to a preset length to obtain a plurality of subsamples, taking each subsample as a node, and determining the connected edges between the nodes according to the similarity between the nodes to construct an undirected graph; S3, Z A `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
2. The method according to claim 1, characterized in that The graph convolutional neural network model includes a first, a second, and a third graph convolutional layer and a fully connected layer connected in sequence.
3. The method according to claim 1 or 2, characterized in that The graph convolutional neural network model is a multi-receptive field graph convolutional neural network model; at least one of the first, second or third graph convolutional layers is a multi-receptive field graph convolutional layer.
4. The method according to claim 1, characterized in that In step S2, after the data to be processed is divided into a plurality of sub-samples according to a preset length, the method further includes: performing a fast Fourier transform on each sub-sample.
5. The method according to claim 1, characterized in that The weights of the edges of the undirected graph are: Where i = 1, 2, ..., n, j = 1, 2, ..., n, n is the number of nodes, For Node similarity, β is the variance of all node data, ε is the similarity threshold, For output and node The nodes whose similarity is greater than the threshold ε, y i For The set of all nodes except .
6. The method according to claim 1 or 5, characterized in that The similarity is cosine similarity or Euclidean distance.
7. An underwater unmanned submersible track fusion device based on graph convolutional neural network, characterized in that: include: The first processing module is used to obtain the track data Z observed by the observation platforms A and B respectively on the observation target A , Z B , for Z A , Z B Perform temporal and spatial registration to obtain Z A `、Z B `; The second processing module is used to respectively A `、Z B `Preprocess the data to be processed to obtain the corresponding undirected graph; The preprocessing includes: normalizing the data to be processed, dividing the data according to a preset length to obtain a plurality of subsamples, taking each subsample as a node, and determining the connected edges between the nodes according to the similarity between the nodes to construct an undirected graph; The third processing module is used to convert Z A `、Z B `The corresponding undirected graph is input into the pre-trained graph convolutional neural network model to obtain the track fusion result.
8. An underwater unmanned submersible track fusion system based on graph convolutional neural network, characterized in that: include: A computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute the method according to any one of claims 1 to 6.
