Graph Convolution Clustering Method for Mesoscale Eddy Time Series in Towed Sensor Array

Abstract

The present invention discloses a graph convolution clustering method for mesoscale vortex time series of towed sensor array. The method comprises the following steps: building the overall link of the towed system; deploying and recovering the towed sensor array to obtain time series data; using the continuous bag-of-words model in the word vector to calculate the similarity of the time series segmented feature vectors and perform hierarchical clustering, using the cluster center to uniformly characterize the feature vectors within the category, dividing the segmented time series feature vector values ​​represented by this into training samples and test samples, selecting the key feature vector categories of the mesoscale vortex with stage representativeness as labels; establishing and training a graph convolution network model, establishing an affinity graph, inputting the test samples into the trained GCN, obtaining the representative clustering output of the mesoscale vortex stage; based on the graph convolution network secondary clustering, finally obtaining the clustering result of the time series. The present invention finally obtains a high-accuracy time series classification result through two clusterings.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean observation and deep learning, and more specifically to a convolution clustering method for a mesoscale eddy time series graph in a towed sensor array, and belongs to the technical field of time series classification. Background Art

[0002] As the world's largest resource, the ocean contains rich mineral resources and oil and gas resources. In the exploration of various marine resources, the ocean towing system plays an important role and function, not only involving the exploration of marine resources, but also in the fields of civil, fishery and military. Marine science is a knowledge system that studies the natural phenomena, properties and laws of change of the ocean, as well as the development and utilization of the ocean. Its research object is the ocean, which accounts for 71% of the earth's surface, including seawater, substances dissolved and suspended in seawater, organisms living in the ocean, seabed sediments and seabed lithosphere, as well as the atmospheric boundary layer and estuary coastal zone on the sea surface. Therefore, marine science is an important part of earth science. The research field of marine science is very broad, and its main contents include basic research on physical, chemical, biological and geological processes in the ocean, and applied research on the development and utilization of marine resources and maritime military activities. The towing system is a powerful tool that can currently meet the above requirements. The towing system integrates multidisciplinary sensors, and the observation elements include temperature, salinity, pressure and conductivity. To ensure observation accuracy and depth, the towing system needs to have good hydrodynamic performance, unobstructed water exchange channels, and the sensor needs to have fast response capabilities and high-precision measurement performance.

[0003] Graph Convolutional Network (GCN) is a method based on deep learning to process graph domain information. By combining graph broadcast operations with deep learning algorithms, it can allow both the structural information and vertex attribute information of the graph to participate in learning. It shows good results and interpretability in applications such as vertex classification, graph classification, and link prediction, and has become a widely used graph analysis method.

[0004] However, GCN semi-supervised learning also has certain defects: (1) If there are too few nodes with labels during semi-supervision, the performance of GCN will be seriously reduced. (2) Shallow GCN networks cannot propagate label information over a large range (the deeper the layer, the larger the receptive field of the node). (3) Deep GCN networks will cause the problem of over-smoothing. Summary of the invention

[0005] The purpose of the present invention is to propose a towed sensor array mesoscale eddy time series graph convolution clustering method to solve the above-mentioned problems.

[0006] In order to achieve the above object, the specific technical solution proposed by the present invention is:

[0007] A method for convolution clustering of mesoscale eddy time series graphs in a towed sensor array, the method comprising the following steps:

[0008] S1: Build the overall link of the towing system;

[0009] S2: Deployment and recovery of the towed sensor array to collect three-dimensional profile observation data of the towed sensor array and obtain time series data;

[0010] S3: The Continuous Bag-of-Words (CBOW) model in word vectors is used to calculate the similarity of the time series segment feature vectors and perform hierarchical clustering. The feature vectors within the category are uniformly represented by the cluster center. The segmented time series feature vector values ​​represented by this are divided into training samples and test samples. The key feature vector categories that are representative of the mesoscale vortex at each stage are selected as labels.

[0011] S4: Establish and train a graph convolutional network (GCN) model, input the feature vector representation of the original time series segmentation, the ideal output is the representative time series feature vector category of the stage of the mesoscale vortex life cycle, establish an affinity graph, and then input the sample to be tested into the trained GCN to obtain the possible representative clustering output of the mesoscale vortex stage.

[0012] Furthermore, in S1, the construction of the overall link of the towing system includes the following steps:

[0013] The towing system consists of a deck unit, a towing chain, a depth-fixing submersible, electrodes, etc. Assume that the total length of the towing cable is L, which contains N s sensor integrated module, assuming the distance from the i-th node to the previous node is L i .

[0014] In S2, the deployment and recovery of the towed sensor (towed optical, temperature, salinity, pressure sensor) array includes the following steps:

[0015] 1) The deployment and recovery process of the towing chain is implemented by the deck unit winch, A-frame, and module assembly and disassembly device. Assume that the ship speed is v, the towing time is T, and the cable length from the water surface to the i-th node is l i The depth-fixing submersible ensures that the bottom of the towing chain can carry out three-dimensional profile observation within a certain depth, and tests the tension F(t) in real time to ensure that it is within the winch bearing capacity limit F max Within.

[0016] 2) Based on the principle of electromagnetic coupling, the towing system realizes contactless power supply and data transmission in the same transmission link. The above-water electric control system converts DC power into high-frequency AC signals, and supplies power to the underwater sensor integrated module and performs data communication. Suppose the absorbance, fluorescence, temperature, conductivity, and pressure time series collected by the i-th sensor integrated module for the jth time are A i,j (t), F i,j (t), T i,j (t), C i,j (t), D i,j (t), t∈(0,T), i=1, 2,…,N s , j = 1, 2, ..., N s , the multi-dimensional stereo observation time series set is S i,j (t).

[0017] Furthermore, the S3 is specifically as follows:

[0018] 1) The time series data S i,j (t) Divide into m segments to obtain subsequences, let the subsequence S of the fth segment be i,j,f (t) The length is T′, t∈[0,T′], and the subsequence is also divided into training subsequence and test subsequence;

[0019] 2) Establish a continuous bag-of-words model (CBOW) in word vectors, calculate the similarity of the feature vectors of the training time series segments and perform hierarchical clustering, use the cluster center to uniformly represent the feature vectors within the category, and train the classifier. Input the subsequence to be tested into the trained classifier for classification, and use a hierarchical clustering algorithm to classify the subsequences.

[0020] Furthermore, the classification specifically includes: firstly, the time series data S i,j (t) is divided into m subsequences, and the subsequence S of the fth segment is i,j,f (t) is input into the continuous bag of words model CBOW, and S i,j,f (t) Perform feature extraction and obtain feature vectors C1, C2............C m , assuming that each feature vector is a cluster, use discrete Fréchet distance to calculate similarity, merge the two clusters with the highest similarity, and update the similarity matrix; use the hierarchical clustering algorithm to divide the trajectory set into a certain number of clusters, and finally obtain the new feature vectors K1, K2,............,K after clustering n, where (m ≥ n). The discrete Fréchet distance is used to evaluate the nearest neighbor connection because it takes into account both the position and time order and the spatial distance factor, and has the advantages of high diagnostic accuracy, fast speed and strong adaptability.

[0021] The discrete Fréchet distance evaluation of the nearest neighbor connection is specifically as follows:

[0022] Let the eigenvector C i It consists of p feature points, and the feature vector C j It consists of q trajectory points; using σ(C i ) and σ(C j ) represent the sequential sets of two feature points, then σ(C i )=(u1,…,u p ) and σ(C j )=(u1,…,u q ), we can get the following sequence point pair L as shown in the formula:

[0023] L=(u a1 ,v b1 ),(u a2 ,v b2 ),...,(u am ,v bm ) (1)

[0024] Among them, a1=1, b1=1, a m =p,b m =q; for any i = 1, 2, 3, ..., q has a i+1 =a i or a i+1 =a i+1 and b i+1 =b i ;

[0025] C j The length of the sequence between feature points ||L|| is expressed as follows:

[0026]

[0027] Then its discrete Fréchet distance is defined as follows:

[0028] δ F (C i ,C j )=min||L|| (3)

[0029] The silhouette distance is used to evaluate the quality of the clustering task. For the i-th (the distance from the i vector to other points in the cluster) data point, the silhouette distance is given by the following formula:

[0030]

[0031] a(i) is the average similarity and dissimilarity between the ith data point and all other points in the cluster to which the ith data point belongs.

[0032] Further, the S4 includes the following:

[0033] Establish an affinity graph and calculate the corresponding sparse symmetric adjacency matrix: N is the total number of nodes in the low-dimensional mapping of the mesoscale eddy data in historical observations, and D is the characteristic dimension of the data detected within a given translation window;

[0034] GCN is used as the backbone network to further extract its significant connection relationship. The calculation process of its Lth layer is:

[0035]

[0036] Add a self-loop to the adjacency matrix A, that is I is the identity matrix, where Angle Matrix for F l represents the embedding feature of the lth layer, F0 is the input feature vector K1, K2, ............, K n ; is a learnable matrix that maps the embedded features to a new space, σ is a nonlinear activation function ReLU; F L represents the output feature map of layer L; F L The predicted affinity graph aggregates information from the neighborhood and encodes the graph structure. During training, a pair of features connected by the affinity graph edge is used as the classifier input. The supervised contrast loss between the predicted edge confidence and the true edge label is minimized.

[0037] The edges in the affinity graph consist of dense connections within each mesoscale eddy propagation process cluster and sparse connections between approximate clusters. Structure-preserving sampling reorganizes the training nodes in the cluster. Its goal is to sample the subgraph formed by the data feature vector obtained by mesoscale eddy detection. The representative retained samples have the ability to characterize the structural characteristics of the mesoscale eddy itself, that is, the edge connections within the cluster and the connections between near clusters. At the same time, the approximate clusters of the subgraph will also be sampled, and the edges between approximate clusters will be used as negative samples based on probability to improve performance gain. First, M clusters are randomly selected as samples and expanded to their N neighbor clusters to obtain a subgraph Q consisting of M×N clusters. n ; Introduce cluster randomness strategy, from Q nRandomly select K n clusters, and the sample randomness strategy, from Q n Randomly select K m nodes, and reconstruct the affinity graph Q based on the sampled nodes;

[0038] In clustering reasoning, the attributes of the mesoscale eddy samples to be classified are characterized in the same way and sent to GCN to predict the edge scores. Graph parsing directly prunes according to the predicted edge scores, and graph refinement calculates node intimacy to further delete irrelevant edges, thereby obtaining clustering reasoning results. Assume that two nodes N i and N j , respectively connect n i and n j edges, and have k common neighbor nodes, where C represents the clustering function, and its node intimacy I can be expressed as an aggregation operation:

[0039] I=C(k / n i ,k / n j ) (6)

[0040] Let the given adjacency matrix be The number of mutual neighbors of a node pair is Each element in N i and N j The mutual neighbor number, P represents the maximum value function, and v represents the vector function, then the node intimacy I is expressed as:

[0041]

[0042] After the above steps, the nodes can be classified, and then the representative time series of the mesoscale eddy stage can be classified.

[0043] Advantages and technical effects of the present invention:

[0044] The present invention uses a word vector model to calculate the similarity of time series data, performs hierarchical clustering, forms a new feature vector, constructs an affinity graph, and then performs secondary clustering based on a graph convolutional network (GCN), and finally obtains the clustering result of the time series. Through two clusterings, the present invention finally obtains a highly accurate time series classification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0046] Figure 2 It is a specific flow chart of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, implementation scheme and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and through specific examples.

[0048] Example:

[0049] A towed sensor array mesoscale eddy time series graph convolution clustering method, the method comprises the following steps (the specific process is as follows Figure 2 shown):

[0050] Step 1: Deploy and recover the towed optical, temperature, salinity, and pressure sensor arrays during the sea trial, and collect multi-dimensional stereoscopic observation time series data of the towed sensor array, such as Figure 1 as shown in .

[0051] 1. The towing system consists of a deck unit, a towing chain, a depth-fixing submersible, electrodes, etc. Assume that the total length of the towing cable is L, which contains N s sensor integrated module, assuming the distance from the i-th node to the previous node is L i The deployment and recovery process of the towing chain is implemented by the deck unit winch, A-frame, and module assembly and disassembly device. Assume that the ship speed is v, the towing time is T, and the cable length from the water surface to the i-th node is l i The depth-fixing submersible ensures that the bottom of the towing chain can carry out three-dimensional profile observation within a certain depth, and tests the tension F(t) in real time to ensure that it is within the winch bearing capacity limit F max Within.

[0052] 2. Assume that the absorbance, fluorescence, temperature, conductivity, and pressure time series collected by the i-th sensor integrated module for the jth time are A i,j (t), F i,j (t), T i,j (t), C i,j (t), D i,j (t), t∈(0,T), i=1, 2,…,N s , j = 1, 2, ..., N s , the multi-dimensional stereo observation time series set is S i,j (t).

[0053] Step 2: Use the Continuous Bag-of-Words (CBOW) model in the word vector to calculate the similarity of the time series segment feature vectors and perform hierarchical clustering. Use the cluster center to uniformly represent the feature vectors within the category. The segmented time series feature vector values ​​represented by this are divided into training samples and test samples. The key feature vector categories that are representative of the stage of the mesoscale vortex are selected as labels.

[0054] 1. Multi-dimensional stereoscopic observation time series S of sensor integration module i,j(t) Divide into m segments to obtain subsequences, let the subsequence S of the fth segment be i,j,f (t) has a length of T′, t∈[0,T′].

[0055] 2. Establish the continuous bag-of-words model CBOW in the word vector, calculate the similarity of the feature vectors of the training time series segment and perform hierarchical clustering, use the cluster center to uniformly represent the feature vectors within the category, and perform classifier training. Input the subsequence to be tested into the trained classifier for classification, and use the hierarchical clustering algorithm to classify the subsequence. First, the time series data S i,j (t) is divided into m subsequences, and the subsequence S of the fth segment is i,j,f (t) is input into the continuous bag of words model CBOW, and S i,j,f (t) Perform feature extraction and obtain feature vectors C1, C2............C m , assuming that each feature vector is a cluster, use the discrete Fréchet distance to calculate the similarity, merge the two clusters with the highest similarity, and update the similarity matrix; use the hierarchical clustering algorithm to divide the trajectory set into a certain number of clusters, and finally obtain the new feature vectors K1, K2,............,K after clustering n , where (m≥n). The discrete Fréchet distance is used to evaluate the nearest neighbor connection because it takes into account both the order of position and time, and the factor of spatial distance, and has the advantages of high diagnostic accuracy, fast speed, and strong adaptability. Let the feature vector C i It consists of p feature points, and the feature vector C j It consists of q trajectory points; using σ(C i ) and σ(C j ) represent the sequential sets of two feature points, then σ(C i )=(u1,…,u p ) and σ(C j )=(u1,...,u q ), we can get the following sequence point pair L as shown in the formula:

[0056] L=(u a1 ,v b1 ),(u a2 ,v b2 ),...,(u am ,v bm ) (1)

[0057] Among them, a1=1, b1=1, a m =p,bm =q. For any i=1,2,3,........,q has a i+1 =a i or a i+1 =a i+1 and b i+1 =b i .

[0058] C j The length of the sequence between feature points ||L|| is expressed as follows:

[0059]

[0060] Then its discrete Fréchet distance is defined as follows:

[0061] δ F (C i ,C j )=min||L|| (3)

[0062] The silhouette distance is used to evaluate the quality of the clustering task. For the i-th (the distance from the i vector to other points in the cluster) data point, the silhouette distance is given by the following formula

[0063]

[0064] a(i) is the average similarity and dissimilarity between the ith data point and all other points in the cluster to which the ith data point belongs.

[0065] Step 3: Establish and train a graph convolutional network (GCN) model, input the feature vector representation of the original time series segmentation, and the ideal output is the representative time series feature vector category of the stage of the mesoscale vortex life cycle. Establish an affinity graph, and then input the sample to be tested into the trained GCN to obtain the possible representative clustering output of the mesoscale vortex stage.

[0066] Compute the corresponding sparse symmetric adjacency matrix as N is the total number of nodes in the low-dimensional mapping of historical mesoscale eddy data, and D is the characteristic dimension of the data detected within a given translation window.

[0067] GCN is used as the backbone network to further extract its significant connection relationship. The calculation process of its Lth layer is:

[0068]

[0069] Add a self-loop to the adjacency matrix A, that is I is the identity matrix, where Angle Matrix for F l represents the embedding feature of the lth layer, F0 is the input feature vector K1, K2, ............, K n ; is a learnable matrix that maps the embedded features to a new space, σ is a nonlinear activation function ReLU; F L represents the output feature map of layer L; F L The predicted affinity graph aggregates information from the neighborhood and encodes the graph structure. During training, a pair of features connected by the affinity graph edge is used as the classifier input. The supervised contrast loss between the predicted edge confidence and the true edge label is minimized.

[0070] The edges in the affinity graph consist of dense connections within each mesoscale eddy propagation process cluster and sparse connections between approximate clusters. Structure-preserving sampling reorganizes the training nodes in the cluster. Its goal is to sample the subgraph formed by the data feature vector obtained by mesoscale eddy detection. The representative retained samples have the ability to characterize the structural characteristics of the mesoscale eddy itself, that is, the edge connections within the cluster and the connections between near clusters. At the same time, the approximate clusters of the subgraph will also be sampled, and the edges between approximate clusters will be used as negative samples based on probability to improve performance gain. First, M clusters are randomly selected as samples and expanded to their N neighbor clusters to obtain a subgraph Q consisting of M×N clusters. n ; Introduce cluster randomness strategy, from Q n Randomly select K n clusters, and the sample randomness strategy, from Q n Randomly select K m nodes, and reconstruct the affinity graph Q based on the sampled nodes;

[0071] In clustering reasoning, the attributes of the mesoscale eddy samples to be classified are characterized in the same way and sent to GCN to predict the edge scores. Graph parsing directly prunes according to the predicted edge scores, and graph refinement calculates node intimacy to further delete irrelevant edges, thereby obtaining clustering reasoning results. Assume that two nodes N i and N j , respectively connect n i and n j edges, and have k common neighbor nodes, where C represents the clustering function, and its node intimacy I can be expressed as an aggregation operation:

[0072] I=C(k / n i ,k / n j ) (6)

[0073] Let the given adjacency matrix be The number of mutual neighbors of a node pair is Each element in N i and N j The mutual neighbor number, P represents the maximum value function, and v represents the vector function, then the node intimacy I is expressed as:

[0074]

[0075]

[0076] On the basis of the above embodiments, the present invention continues to describe in detail the technical features involved therein and the functions and roles played by the technical features in the present invention, so as to help technicians in this field fully understand the technical solution of the present invention and reproduce it.

Claims

1. A method for convolution clustering of time series graphs of mesoscale eddies in towed sensor arrays, characterized in that: The method comprises the following steps: S1: Build the overall link of the towing system; S2: Deployment and recovery of the towed sensor array to collect three-dimensional profile observation data of the towed sensor array and obtain time series data; S3: Use the continuous bag-of-words model in the word vector to calculate the similarity of the time series segment feature vectors and perform hierarchical clustering. Use the cluster center to uniformly represent the feature vectors within the category. Divide the segmented time series feature vector values ​​represented by this into training samples and test samples, and select the key feature vector categories that are representative of the stage of the mesoscale vortex as labels; S4: Establish and train a graph convolutional network model, input the feature vector representation of the original time series segmentation, and the ideal output is the representative time series feature vector category of the stage of the mesoscale vortex life cycle, establish an affinity graph, and then input the sample to be tested into the trained GCN to obtain the possible representative clustering output of the mesoscale vortex stage; S4 includes the following: 1) Establish an affinity graph and calculate the corresponding sparse symmetric adjacency matrix A∈R N×D , N is the total number of nodes of the low-dimensional mapping of the mesoscale eddy data in historical observations, and D is the characteristic dimension of the data detected within a given translation window; GCN is used as the backbone network to further extract its significant connection relationship. The calculation process of its Lth layer is: Add a self-loop edge to the adjacency matrix A, that is I is the identity matrix, where Angle Matrix for F l represents the embedding feature of the lth layer, F0 is the input feature vector K1, K2, ............, K n ; is a learnable matrix that maps the embedded features to a new space, σ is a nonlinear activation function ReLU; F L represents the output feature map of layer L; F L The predicted affinity graph aggregates information from the neighborhood and encodes the graph structure. During training, a pair of features connected by the affinity graph edge are used as classifier input. The supervised comparison loss between the predicted edge confidence and the true edge label is minimized. 2) First, randomly select M clusters as samples and expand them to their N neighbor clusters to obtain a subgraph Q consisting of M×N clusters. n ; Introduce cluster randomness strategy, from Q n Randomly select K n clusters, and the sample randomness strategy, from Q n Randomly select K m nodes, and reconstruct the affinity graph Q based on the sampled nodes; 3) During clustering reasoning, the attributes of the mesoscale eddy samples to be classified are characterized in the same way and sent to GCN to predict the edge scores. The graph parsing is directly pruned according to the predicted edge scores, and the graph refinement calculates the node intimacy to further delete the irrelevant edges, and then the clustering reasoning results are obtained; assuming that two nodes N i and N j , respectively connect n i and n j edges, and have k common neighbor nodes, where C represents the clustering function and its node intimacy I is represented by the aggregation operation: I=C(k / n i ,k / n j )(6) Let the given adjacency matrix A∈R N×N , the number of mutual neighbors of a node pair is Each element in N i and N j The mutual neighbor number of , P represents the maximum value function, v represents the vector function, then the node intimacy I is expressed as: After the above steps, the node classification is realized, and then the representative time series of the mesoscale eddy stage are classified.

2. The clustering method according to claim 1, characterized in that: The S3 is as follows: 1) The time series data S i,j (t) Divide into m segments to obtain subsequences, let the subsequence S of the fth segment be i,j,f (t) The length is T′, t∈[0,T′], and the subsequence is also divided into training subsequence and test subsequence; 2) Establish the continuous bag-of-words model CBOW in the word vector, calculate the similarity of the segmented feature vectors of the training time series and perform hierarchical clustering, use the cluster center to uniformly represent the feature vectors within the category, and perform classifier training. Input the subsequence to be tested into the trained classifier for classification, and use the hierarchical clustering algorithm to classify the subsequences.

3. The clustering method according to claim 2, characterized in that: The classification specifically includes: firstly, the time series data S i,j (t) is divided into m subsequences, and the subsequence S of the fth segment is i,j,f (t) is input into the continuous bag of words model CBOW, and S i,j,f (t) Perform feature extraction and obtain feature vectors C1, C2............C m , assuming that each feature vector is a cluster, use discrete Fréchet distance to calculate similarity, merge the two clusters with the highest similarity, and update the similarity matrix; use the hierarchical clustering algorithm to divide the trajectory set into a certain number of clusters, and finally obtain the new feature vectors K1, K2,............,K after clustering n , where m≥n.

4. The clustering method according to claim 3, characterized in that: The discrete Fréchet distance evaluates the nearest neighbor connection as follows: Let the eigenvector C i It consists of p feature points, and the feature vector C j It consists of q trajectory points; using σ(C i ) and σ(C j ) represent the sequential sets of two feature points, then σ(C i )=(u1,...,u p ) and σ(C j )=(u1,...,u q ), we get the following sequence of point pairs L as shown in the formula: L=(u a1 ,v b1 ),(u a2 ,v b2 ),...,(u am ,v bm ) (1) Among them, a1=1, b1=1, a m =p,b m =q; for any i = 1, 2, 3, ..., q has a i+1 =a i or a i+1 =a i+1 and b i+1 =b i ; C j The length of the sequence between feature points ||L|| is expressed as follows: Then its discrete Fréchet distance is defined as follows: d F (C i ,C j )=min||L|| (3) The silhouette distance is used to evaluate the quality of the clustering task. For the i-th data point: the distance from the i vector to other points in the cluster, the silhouette distance is given by the following formula: a(i) is the average similarity and dissimilarity between the ith data point and all other points in the cluster to which the ith data point belongs.

Images