Wireless sensor network main feature extraction method, system, device and storage medium

Abstract

The present invention relates to the field of multimodal network technology, specifically a method, system, device and storage medium for extracting main features of wireless sensor networks, including: S1, receiving wireless sensor network data; S2, constructing a target loss function according to the data mode of the wireless sensor network, evaluating the main features of the wireless sensor network and the operation and optimization process of extracting the main features; S3, constructing an adjacency tensor describing the spatiotemporal relationship between nodes according to the wireless sensor network data, extracting the main features according to the wireless sensor network data and the adjacency tensor, and fusing them, extracting the fused main features that meet the preset accuracy requirements, and storing them. This scheme acts on wireless sensor network data and can extract main features that are in line with statistical laws and with high accuracy, so as to solve the problem that the existing technology cannot perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal networks, and in particular to a method, system, device and storage medium for extracting main features of a wireless sensor network. Background Art

[0002] With the in-depth development of a new round of scientific and technological revolution and industrial transformation, an era with computing power as the core productivity is accelerating. Computing power has become the cornerstone of the digital transformation of the whole society. As the main artery connecting users, data, and computing power, the integration and symbiosis of wireless sensor networks and computing power are deepening. The development and implementation of a large number of scientific and technological technologies are inseparable from wireless sensor networks. Data-driven wireless sensor network analysis has been valued and studied in the new generation of computing network service systems. One of the most core problems is the main feature extraction in wireless sensor networks. Among them, wireless sensor networks, as a natural non-Euclidean network structure, can use graph neural network technology to extract spatiotemporal patterns from time-varying wireless sensor network data, and deeply explore the potential laws and interaction patterns in wireless sensor networks, thereby providing important decision support for wireless sensor network analysis.

[0003] Wireless sensor networks cover multiple dimensions such as communication, computing, and storage, and integrate heterogeneous network features such as WIFI communication, cellular communication, and optical communication. In particular, the network topology changes over time, making it particularly difficult to characterize network service capabilities. Existing wireless sensor network representation methods focus on discovering spatial connectivity information or capturing continuous dynamic patterns on a single time slice, and are unable to uniformly model the spatiotemporal patterns in wireless sensor networks. Therefore, existing methods suffer from a large amount of information loss when processing wireless sensor networks, resulting in insufficient representation capabilities of the extracted main features.

[0004] In addition, existing graph neural network methods lack targeted consideration of neighborhood information of different orders, ignore the differences in neighborhood features at different levels, and further weaken the representation ability of the main features; and non-Euclidean topological structures and spatiotemporal patterns are widely present in wireless sensor networks. In most cases, extracting non-Euclidean structural features in wireless sensor networks and deeply mining their implicit spatiotemporal patterns are sufficient to express the main characteristics of wireless sensor networks.

[0005] Therefore, there is an urgent need for a method, system, device and storage medium for extracting the main features of wireless sensor networks, which can be applied to wireless sensor network data and can perform main feature extraction with high accuracy in accordance with statistical laws, so as to solve the problem that the existing technology is unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems. Summary of the invention

[0006] One of the purposes of the present invention is to provide a main feature extraction method for a wireless sensor network, which can extract main features that conform to statistical laws and have high accuracy when applied to wireless sensor network data, so as to solve the problem that the existing technology cannot perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0007] The basic solution 1 provided by the present invention: a method for extracting main features of a wireless sensor network, comprising the following contents:

[0008] S1, receiving wireless sensor network data;

[0009] S2. According to the data mode of the wireless sensor network, a target loss function is constructed to evaluate the operation and optimization process of the main features of the wireless sensor network and the main feature extraction;

[0010] S3. Constructing an adjacency tensor according to the wireless sensor network data, extracting main features according to the wireless sensor network data and the adjacency tensor, and fusing them, extracting the fused main features that meet the preset accuracy requirements, and storing them;

[0011] Among them, S3 includes:

[0012] S304, extracting the main feature spatiotemporal pattern, including:

[0013] According to the wireless sensor network data and the adjacency tensor, the parameters of the initialized main feature tensor are used to extract the spatiotemporal patterns contained in the time-varying wireless sensor network, wherein the spatiotemporal patterns are extracted by using a spatiotemporal graph convolutional network to perform hierarchical information mining on the tensors describing the time-varying wireless sensor network;

[0014] The tensor multiplication-driven spatiotemporal graph convolution operation is:

[0015] ;

[0016] in for The multiplication operator, The transformation matrix of a layer is directly defined as a band indicator matrix capturing the neighboring time slice information; For the The adjacency tensor used by the layer spatiotemporal graph convolution, , for Conversion operations, For the The transformation matrix for layer tensor multiplication;

[0017] S305, fusion main features:

[0018]

[0019] inE is the main feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors; , …, is the importance coefficient of the main features of different layers.

[0020] Further, the S2 includes:

[0021] S201, obtaining an instruction for constructing a target loss function;

[0022] S202: Read the target tensor of the wireless sensor network data according to the instruction of constructing the target loss function A known data set , which only includes non-zero wireless sensor network data;

[0023] Based on known data sets , construct the form of the target loss function:

[0024]

[0025] The target loss function is specifically:

[0026] .

[0027] Furthermore, the S3 further includes: S301, receiving an instruction for extracting main features of a wireless sensor network;

[0028] S302, initializing relevant parameters according to the instructions of the main characteristics of the wireless sensor network;

[0029] S303, constructing an adjacency tensor describing the spatiotemporal pattern of the wireless sensor network;

[0030] Also includes:

[0031] S306: Determine and extract iterative control variables Has the limit been reached? If yes, execute S308, if no, execute S307;

[0032] S307, determine whether the distance target loss function value is within If the upper convergence is yes, then execute S309, if not, then execute S308; The target tensor for wireless sensor network data A known data set ;

[0033] S308, using the gradient descent method, Perform iterative optimization and update ;

[0034] S309: extracting and storing the fused main feature results that meet the preset accuracy requirements.

[0035] Furthermore, the related parameters include: main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , main feature initial feature , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold .

[0036] Furthermore, the judgment distance target loss function value Is it in Upper convergence, including: before the start of this round of iteration Compared with the value before the last iteration , whether the absolute value of the difference between the two is less than the convergence judgment threshold If yes, it is considered converged, if not, it is considered not converged.

[0037] Beneficial effects of basic scheme 1: In this scheme, for the received wireless sensor network data, according to the data mode of the wireless sensor network, a distance measurement function is constructed, the main features of the wireless sensor network and the operation and optimization process of extracting the main features are evaluated, and then the adjacency tensor is constructed according to the wireless sensor network data, and the main features are extracted and fused according to the wireless sensor network data and the adjacency tensor, and the fused main features that meet the preset accuracy requirements are extracted and stored;

[0038] In the process of extracting the main feature spatiotemporal pattern, the tensor multiplication-driven spatiotemporal graph convolution operation extends the graph convolution operation on the static graph to the three-dimensional space, which can simultaneously capture the spatiotemporal pattern in the time-varying wireless sensor network. After multiple layers of graph convolution operations, each node can perceive the information of its multi-hop spatiotemporal neighbors. Compared with the traditional method that uses a simple combination of timing modules and static graph convolution networks, the spatiotemporal graph convolution operation of this scheme can capture richer structural features, thereby greatly reducing information loss and can extract the main features in accordance with statistical laws and with high accuracy.

[0039] In the process of fusing the main features, an adaptive layer aggregation mechanism is used to alleviate the over-smoothing problem. Compared with the traditional method that does not consider the fusion of main features or gives the same weight to the main features of different layers, the adaptive main feature fusion of this scheme has higher generalization ability and stronger adaptability.

[0040] To sum up, this solution acts on wireless sensor network data and can extract main features that conform to statistical laws and have high accuracy, so as to solve the problem that existing technologies are unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0041] The second purpose of the present invention is to provide a wireless sensor network main feature extraction system, which acts on wireless sensor network data and can perform main feature extraction that conforms to statistical laws and has high accuracy, so as to solve the problem that the existing technology cannot perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0042] The present invention provides a second basic solution: a wireless sensor network main feature extraction system, comprising: a wireless sensor network data receiving module, a target loss function building module, a wireless sensor network data storage module, and a main feature extraction module;

[0043] The wireless sensor network data receiving module is used to import and manage the wireless sensor network data, and control the wireless sensor network data storage module to store the imported wireless sensor network data;

[0044] The target loss function construction module is used to construct the target loss function according to the data mode of the wireless sensor network, which is used to evaluate the operation and optimization process of the main features of the wireless sensor network and the main feature extraction module;

[0045] A wireless sensor network data storage module is used to store the wireless sensor network data received by the receiving module, the adjacency tensor of the spatial mode and the main features of the extracted wireless sensor network;

[0046] The main feature extraction module is used to construct an adjacency tensor according to the wireless sensor network data, extract the main features according to the wireless sensor network data and the adjacency tensor, and fuse them, extract the fused main features that meet the preset accuracy requirements, and store them in the main feature storage unit in the wireless sensor network data storage module;

[0047] The main feature extraction module includes: a main feature spatiotemporal pattern extraction unit, which is used to extract the main features of the high-order spatiotemporal pattern of each sensor node contained in the wireless sensor network according to the wireless sensor network data stored in the wireless sensor network data storage unit and the adjacent tensor stored in the adjacent tensor storage unit, combined with the main features initialized by the initialization unit and other related parameters; wherein the spatiotemporal pattern is extracted by using a spatiotemporal graph convolutional network to perform hierarchical information mining on the tensor describing the time-varying wireless sensor network;

[0048] The tensor multiplication-driven spatiotemporal graph convolution operation is:

[0049] ;

[0050] in for The multiplication operator, The transformation matrix of a layer is directly defined as a band indicator matrix capturing the neighboring time slice information; For the The adjacency tensor used by the layer spatiotemporal graph convolution, , for Conversion operations, For the The transformation matrix for layer tensor multiplication;

[0051] The main feature fusion unit is used to fuse the main features of different orders extracted by the main feature spatiotemporal pattern extraction unit and the initialized structural main features to obtain the overall main features:

[0052]

[0053] in E is the main feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors; , …, is the importance coefficient of the main features of different layers.

[0054] Furthermore, the objective function building module is based on the known data set , construct the form of distance target loss function:

[0055]

[0056] The distance target loss function is specifically:

[0057] .

[0058] Furthermore, the wireless sensor network data storage module includes:

[0059] Wireless sensor network data storage unit, adjacency tensor storage unit, main feature storage unit;

[0060] A wireless sensor network data storage unit, used for storing the wireless sensor network data received by the receiving module;

[0061] An adjacency tensor storage unit, used for storing an adjacency tensor describing a spatiotemporal topological structure of a wireless sensor network;

[0062] A main feature storage unit, used to store the main features of the wireless sensor network extracted by the main feature extraction module;

[0063] The main feature extraction module further includes: an initialization unit, a spatiotemporal adjacency tensor construction unit, and an output unit;

[0064] Initialization unit, used to initialize the relevant parameters involved in the main feature extraction process of the wireless sensor network based on tensor decomposition; the relevant parameters include: main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , main feature initial feature , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold ;

[0065] A spatiotemporal adjacency tensor construction unit, used to construct an adjacency tensor describing the spatiotemporal pattern of wireless sensor networks;

[0066] The output unit is used to extract the fused main feature results that meet the preset accuracy requirements and store them in the main feature storage unit in the wireless sensor network data storage module.

[0067] Beneficial effects of basic scheme II: This scheme includes a receiving module, an objective function building module, a wireless sensor network data storage module, and a main feature extraction module working together. For the received wireless sensor network data, a distance measurement function is constructed according to the data mode of the wireless sensor network, and the main features of the wireless sensor network and the operation and optimization process of the main feature extraction are evaluated. Then, an adjacency tensor is constructed according to the wireless sensor network data. The main features are extracted and fused according to the wireless sensor network data and the adjacency tensor. The fused main features that meet the preset accuracy requirements are extracted and stored.

[0068] In the process of extracting the main feature spatiotemporal pattern, the tensor multiplication-driven spatiotemporal graph convolution operation extends the graph convolution operation on the static graph to the three-dimensional space, which can simultaneously capture the spatiotemporal pattern in the time-varying wireless sensor network. After multiple layers of graph convolution operations, each node can perceive the information of its multi-hop spatiotemporal neighbors. Compared with the traditional method that uses a simple combination of timing modules and static graph convolution networks, the spatiotemporal graph convolution operation of this scheme can capture richer structural features, thereby greatly reducing information loss and can extract the main features in accordance with statistical laws and with high accuracy.

[0069] In the process of fusing the main features, an adaptive layer aggregation mechanism is used to alleviate the over-smoothing problem. Compared with the traditional method that does not consider the fusion of main features or gives the same weight to the main features of different layers, the adaptive main feature fusion of this scheme has higher generalization ability and stronger adaptability.

[0070] To sum up, this solution acts on wireless sensor network data and can extract main features that conform to statistical laws and have high accuracy, so as to solve the problem that existing technologies are unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0071] The third purpose of the present invention is to provide a device that acts on wireless sensor network data and can extract main features that conform to statistical laws and have high accuracy, so as to solve the problem that the existing technology is unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0072] The present invention provides basic solution three: a device that adopts the above-mentioned wireless sensor network main feature extraction method.

[0073] Beneficial effects of Basic Solution Three: This device acts on wireless sensor network data and can extract main features that conform to statistical laws and have high accuracy, so as to solve the problem that existing technologies are unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0074] The fourth purpose of the present invention is to provide a storage medium that acts on wireless sensor network data and can extract main features that conform to statistical laws and have high accuracy, so as to solve the problem that the existing technology is unable to perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems.

[0075] The present invention provides basic solution 4: a storage medium having a computer program stored thereon, and the computer program, when executed by a processor, implements the steps of the main feature extraction method of the wireless sensor network.

[0076] Beneficial effects of Basic Scheme 4: This scheme provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the wireless sensor network main feature extraction method are implemented. It acts on wireless sensor network data and can perform main feature extraction that complies with statistical laws and has high accuracy, so as to solve the problem that the existing technology cannot perform high-precision main feature extraction containing spatiotemporal patterns for wireless sensor networks in modern large-scale computing network systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A schematic diagram of a flow chart of an embodiment of a method for extracting main features of a wireless sensor network according to the present invention;

[0078] Figure 2 A schematic diagram of the structure of an embodiment of a device of the present invention;

[0079] Figure 3 The figure is a logic block diagram of an embodiment of a main feature extraction system for a wireless sensor network according to the present invention. DETAILED DESCRIPTION

[0080] The following is further described in detail through specific implementation methods:

[0081] Embodiment 1

[0082] The embodiment is basically as shown in the attached Figure 1 As shown: The main feature extraction method of wireless sensor network includes the following contents:

[0083] S1, receiving wireless sensor network data; that is, collecting and receiving wireless sensor network data;

[0084] S2. According to the data mode of the wireless sensor network, a target loss function is constructed to evaluate the operation and optimization process of the main features of the wireless sensor network and the main feature extraction;

[0085] Specifically include:

[0086] S201, obtaining an instruction for constructing a target loss function; in this embodiment, if an instruction for constructing a target loss function is received from a server, S202 is executed;

[0087] S202: Read the target weighted tensor of the wireless sensor network data according to the instruction of constructing the target loss function. A known data set , which only includes non-zero wireless sensor network data. Due to the large number of nodes in the wireless sensor network and the large amount of data accumulated over time, the target weighted tensor of the wireless sensor network data is XIt shows high-dimensional incomplete characteristics, that is, there are many 0 values ​​in it. Therefore, in order to improve the calculation efficiency and reduce the estimation error, only consider establishing the objective function on the known data; the node in the wireless sensor network is the sensor in the wireless sensor network;

[0088] Based on known data sets , construct the target loss function as the distance target loss function, which is in the form of:

[0089]

[0090] In this embodiment, the distance target loss function is specifically:

[0091]

[0092] Establish the Euclidean distance of the corresponding approximation value as the optimization target; use L2 regularization to constrain the optimization process; the distance target loss function is the target loss function;

[0093] S3. According to the wireless sensor network data, an adjacency tensor describing the spatiotemporal relationship between nodes is constructed. According to the wireless sensor network data and the adjacency tensor, main features are extracted and fused. The fused main features meeting the preset accuracy requirements are extracted and stored.

[0094] Specifically include:

[0095] S301, receiving an instruction to extract the main features of a wireless sensor network; that is, receiving an instruction from a server to extract the main features of a wireless sensor network, executing S302;

[0096] S302, initializing relevant parameters according to the instructions of the main characteristics of the wireless sensor network;

[0097] The relevant parameters include: main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , the main feature of initialization , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold ;

[0098] in , and The initial value of is a positive integer;

[0099] The number of rows is , the number of columns is , the number of time slices is Tensor of ;

[0100] , …, is the importance parameter of adaptive learning, which controls the degree of extraction of neighborhood information of different orders;

[0101] Initialized to a very small number, it controls the initial step size of the Adam optimizer gradient descent;

[0102] Control the degree of regularization to prevent overfitting of the model and set it to a very small positive number; the very small positive number is a positive number close to zero, which is set to 10 in this embodiment -4 ;

[0103] It is a variable that controls the iteration process. The maximum upper limit of the iteration allowed is R, and the parameter is initialized to a positive integer;

[0104] Initialized to integer 1, indicating that the current model is in During the training iteration;

[0105] Indicates the lower limit of the error between two rounds of iterations;

[0106] S303, constructing an adjacency tensor, that is, constructing an adjacency tensor that describes the spatiotemporal pattern of the wireless sensor network;

[0107] The time-varying wireless sensor network has a different topological structure in each time slice. The adjacency matrices on all time slices are connected in series according to the time dimension to obtain the adjacency tensor that describes the time-varying wireless sensor network as a whole.

[0108] In order to avoid the problems of numerical instability and over-smoothing that may occur when using the adjacency tensor to extract the topological structure information of the entire graph, the adjacency tensor is symmetrically normalized to obtain the normalized adjacency tensor:

[0109]

[0110] in is the symmetric normalized adjacency tensor, is the degree tensor describing the importance of the node at each time slice, represents the area operation, which is a matrix product along the third dimension.

[0111] S304, extracting main feature spatiotemporal patterns;

[0112] In this step, based on the wireless sensor network data and adjacency tensors, the important spatiotemporal patterns contained in the time-varying wireless sensor network are extracted using the initialized main feature tensor and other related parameters. The spatiotemporal patterns in the network are extracted mainly by using a spatiotemporal graph convolutional network to perform hierarchical information mining on the tensors describing the time-varying wireless sensor network.

[0113] Specifically, we use tensor multiplication to obtain the rich and sparse adjacency tensor so that each time slice can obtain the spatial correlation pattern between nodes in adjacent time slices, thereby obtaining the next layer of adjacency tensor:

[0114]

[0115] in For the The adjacency tensor used by the layer spatiotemporal graph convolution, , for" Convert For the The transformation matrix for layer tensor multiplication;

[0116] Specifically, the above formula is calculated as:

[0117]

[0118] in and They are and The above operation connects the topological structures on different time slices in series, so that each node on any time slice can perceive the topological structures on other adjacent time slices;

[0119] Similarly, to capture the main features on other time slices, the main feature tensor also performs the following tensor multiplication transformation:

[0120]

[0121] Here, each element in the tensor is calculated as:

[0122]

[0123] After obtaining the adjacency tensor and main feature tensor after tensor multiplication transformation, the graph convolution operation is performed to capture the node dependencies in space, specifically:

[0124]

[0125] Using the above formula, each node can capture the main features of nodes that are close in space, specifically:

[0126]

[0127] in For the Layers and Nodes In the The number of directly connected nodes on a time slice, summarizing the above process, obtains the complete tensor multiplication driven spatiotemporal graph convolution operation as:

[0128]

[0129] in, for The multiplication operator, The transformation matrix of the layer can be directly defined as a band indicator matrix to capture the adjacent time slice information, and can also be adaptively trained. In this embodiment, it is chosen to be initialized as a band matrix. The above-mentioned spatiotemporal graph convolution process extends the graph convolution operation on the static graph to three-dimensional space, and can simultaneously capture the spatiotemporal patterns in the time-varying wireless sensor network. After multiple layers of graph convolution operations, each node can perceive the information of its multi-hop spatiotemporal neighbors. Compared with the traditional method that uses a simple combination of timing modules and static graph convolution networks, the spatiotemporal graph convolution operation of this scheme can capture richer structural features, thereby greatly reducing information loss.

[0130] S305, merging main features;

[0131] In this step, the main features containing spatiotemporal information extracted in S304 are E 1~K and each node's own main feature E 0 , adaptive fusion obtains the overall main features, namely the structural main features; the spatiotemporal information is the non-Euclidean topological structure in each time slice and the correlation information between time slices; the node is the sensor;

[0132] The layer-by-layer graph convolution operation follows the message propagation framework. A common problem encountered is over-smoothing, which makes it difficult to capture higher-order neighborhood information, thereby damaging the representation ability of the captured wireless sensor network main features. Therefore, this solution uses an adaptive layer aggregation mechanism to alleviate the over-smoothing problem. The fusion method is:

[0133]

[0134] in E It is the main structural feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors; , …, is the importance coefficient of the main features of different layers, which is adaptively trained here.

[0135] Compared with the traditional method that does not consider the fusion of main features or assigns the same weight to the main features of different layers, the adaptive main feature fusion of this scheme has higher generalization ability and stronger adaptability.

[0136] S306: Determine and extract iterative control variables Has the limit been reached? If yes, execute S308, if no, execute S307;

[0137] S307, determine whether the distance target loss function value is within If the upper convergence is yes, then execute S309; ​​if not, then execute S308;

[0138] The distance target loss function value is judged Is it in Upper convergence, including: before the start of this round of iteration Compared with the value before the last iteration , whether the absolute value of the difference between the two is less than the convergence judgment threshold If yes, it is considered converged, if no, it is considered not converged;

[0139] S308, using gradient descent method, such as Adam, Perform iterative optimization and then execute S303, and update after each round of training. ;

[0140] S309: Extracting the main feature results of the fusion that meet the preset accuracy requirements , and store; extract the main structural features that meet the preset accuracy requirements Storage: In this embodiment, the main feature result with the highest accuracy is taken for storage.

[0141] like Figure 2 As shown, this embodiment also provides a device, which includes: at least one processor, at least one storage medium and at least one bus. In this embodiment, it also includes at least one communication interface. The bus is used to realize the direct connection and communication of these components, the communication interface is used to communicate signaling or data with other node devices, and the storage medium stores machine-readable instructions executable by the processor. When the device is running, the processor communicates with the storage medium through the bus, and the machine-readable instructions are executed by the processor to implement the steps of the above-mentioned wireless sensor network main feature extraction method.

[0142] This embodiment also provides a storage medium storing a computer program, which implements the steps of the above-mentioned method for extracting main features of a wireless sensor network when executed by a processor.

[0143] If the above-mentioned introduction text method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiment can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0144] Embodiment 2

[0145] This embodiment is basically as shown in the attached Figure 3 As shown, a wireless sensor network main feature extraction system is provided, including: a wireless sensor network data receiving module, a target loss function building module, a wireless sensor network data storage module, and a main feature extraction module;

[0146] The wireless sensor network data receiving module is used to import and manage the wireless sensor network data, and control the wireless sensor network data storage module to store the imported wireless sensor network data;

[0147] A target loss function construction module is used to construct a target loss function according to the data mode of the wireless sensor network, which is used to evaluate the operation and optimization process of the main feature and the main feature extraction module of the wireless sensor network; specifically, in this embodiment, it is a distance target loss function;

[0148] Specifically, according to the target weighted tensor of wireless sensor network data , the tensor size is OK, List, time slices, the elements in the target tensor Represents the matrix Row, No. Column, No. The data of time slices, the target is a known data set , whose value is greater than or equal to 0. In order to optimize the error between the predicted value and the true value, the distance target loss function adopts the Euclidean distance function, which is in the form of:

[0149] ;

[0150] In this embodiment, the distance target loss function is specifically:

[0151]

[0152] Establish the Euclidean distance of the corresponding approximation value as the optimization target; use L2 regularization to constrain the optimization process; the distance target loss function is the target loss function; in addition, the target function construction module can also obtain the instruction to construct the target loss function; specifically, receive the instruction to construct the target loss function from the server, and read the target weighted tensor of the wireless sensor network data in the wireless sensor network data storage module according to the instruction to construct the target loss function ;

[0153] A wireless sensor network data storage module is used to store the wireless sensor network data received by the receiving module and capture the adjacency tensor of the spatial pattern and the main features of the extracted wireless sensor network;

[0154] Specifically, it includes: a wireless sensor network data storage unit, an adjacency tensor storage unit, and a main feature storage unit;

[0155] A wireless sensor network data storage unit is used to store the wireless sensor network data received by the receiving module; wherein the wireless sensor network data is collected by the receiving module from the computing network system;

[0156] An adjacency tensor storage unit, used for storing an adjacency tensor describing a spatiotemporal topological structure of a wireless sensor network;

[0157] A main feature storage unit, used to store the main features of the wireless sensor network extracted by the main feature extraction module;

[0158] The main feature extraction module is used to construct an adjacency tensor describing the spatiotemporal relationship between nodes based on the wireless sensor network data, extract the main features based on the wireless sensor network data and the adjacency tensor, and fuse them to extract the fused main features that meet the preset accuracy requirements and store them in the main feature storage unit in the wireless sensor network data storage module; wherein the node is a sensor in the wireless sensor network;

[0159] Specifically, it includes: an initialization unit, a spatiotemporal adjacency tensor construction unit, a main feature spatiotemporal pattern extraction unit, a main feature fusion unit, and an output unit;

[0160] An initialization unit is used to initialize relevant parameters involved in the main feature extraction process of the wireless sensor network based on tensor decomposition; the initialization unit will receive an instruction to extract the main feature of the wireless sensor network, that is, receive an instruction from the server to extract the main feature of the wireless sensor network, and initialize relevant parameters according to the instruction;

[0161] The relevant parameters include: main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , the main feature of initialization , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold ;

[0162] in , and The initial value of is a positive integer;

[0163] The number of rows is , the number of columns is , the number of time slices is Tensor of ;

[0164] , …, is the importance parameter of adaptive learning, which controls the degree of extraction of neighborhood information of different orders;

[0165] Initialized to a very small number, it controls the initial step size of the Adam optimizer gradient descent;

[0166] Control the degree of regularization to prevent overfitting of the model and set it to a very small positive number; the very small positive number is a positive number close to zero, which is set to 10 in this embodiment -4 ;

[0167] It is a variable that controls the iteration process. The maximum upper limit of the iteration allowed is R, and the parameter is initialized to a positive integer;

[0168] Initialized to integer 1, indicating that the current model is in During the training iteration;

[0169] Indicates the lower limit of the error between two iterations.

[0170] A spatiotemporal adjacency tensor construction unit, used to construct an adjacency tensor describing the spatiotemporal pattern of wireless sensor networks;

[0171] Specifically, the time-varying wireless sensor network has a different topological structure in each time slice. The adjacency matrices on all time slices are connected in series according to the time dimension to obtain the adjacency tensor that describes the time-varying wireless sensor network as a whole.

[0172] In order to avoid the problems of numerical instability and over-smoothing that may occur when using the adjacency tensor to extract the topological structure information of the entire image, the adjacency tensor is symmetrically normalized to obtain the normalized adjacency tensor as the input of the main feature extraction module:

[0173]

[0174] in is the symmetric normalized adjacency tensor, is the degree tensor describing the importance of the node at each time slice, represents the area operation, which is a matrix product along the third dimension.

[0175] The main feature spatiotemporal pattern extraction unit is used to extract the main features of the high-order spatiotemporal pattern of each sensor node contained in the wireless sensor network based on the wireless sensor network data stored in the wireless sensor network data storage unit and the adjacency tensor stored in the adjacency tensor storage unit, combined with the main features initialized by the initialization unit and other related parameters;

[0176] The classic graph convolutional neural network structure mainly includes input features, adjacency matrix, feature transformation and nonlinear activation. Existing studies have shown that nonlinear activation and feature transformation are redundant in latent feature analysis. Therefore, this scheme follows the lightweight graph convolution structure and is applied to wireless sensor network data. In addition, the existing graph convolution paradigm is only applied to static graphs. Although recurrent neural networks and graph convolutional networks can be combined for dynamic networks, this separate modeling method loses a lot of useful information, such as high-order spatiotemporal connectivity.

[0177] Therefore, this scheme proposes a novel spatiotemporal graph convolution module based on the general tensor multiplication framework as the main feature spatiotemporal pattern extraction unit to extract the main features of the high-order spatiotemporal patterns of each sensor node contained in the wireless sensor network;

[0178] The specific processing process is as follows:

[0179] By multiplying the rich and sparse adjacency tensor by tensors, each time slice can obtain the spatial correlation pattern between nodes in adjacent time slices, thereby obtaining the next layer of adjacency tensors:

[0180]

[0181] in For the The adjacency tensor used by the spatiotemporal graph convolution. , for" Convert For the The transformation matrix for layer tensor multiplication;

[0182] Specifically, the above formula is calculated as:

[0183]

[0184] in and They are and The above operation connects the topological structures on different time slices in series, so that each node on any time slice can perceive the topological structures on other adjacent time slices;

[0185] Similarly, to capture the main features on other time slices, the main feature tensor also performs the following tensor multiplication transformation:

[0186]

[0187] Here, each element in the tensor is calculated as:

[0188]

[0189] After obtaining the adjacency tensor and main feature tensor after tensor multiplication transformation, the graph convolution operation is performed to capture the node dependencies in space, specifically:

[0190]

[0191] Using the above formula, each node can capture the main features of nodes that are close in space, specifically:

[0192]

[0193] in For the Layers and Nodes In the The number of directly connected nodes on a time slice, summarizing the above process, obtains the complete tensor multiplication driven spatiotemporal graph convolution operation as:

[0194]

[0195] in, for The multiplication operator, The transformation matrix of the layer can be directly defined as a band indicator matrix to capture the adjacent time slice information, and can also be adaptively trained. In this embodiment, it is chosen to be initialized as a band matrix. The above-mentioned spatiotemporal graph convolution process extends the graph convolution operation on the static graph to three-dimensional space, and can simultaneously capture the spatiotemporal patterns in the time-varying wireless sensor network. After multiple layers of graph convolution operations, each node can perceive the information of its multi-hop spatiotemporal neighbors. Compared with the traditional method that uses a simple combination of timing modules and static graph convolution networks, the spatiotemporal graph convolution operation of this scheme can capture richer structural features, thereby greatly reducing information loss.

[0196] The main feature fusion unit is used to combine the main features of different orders extracted by the main feature spatiotemporal pattern extraction unit E 1~K and the main features of initialization E 0 The main feature of the whole is obtained by fusion, that is, the main feature of the structure; the main feature extracted by the main feature spatiotemporal pattern extraction unit contains spatiotemporal information, that is, the non-Euclidean topological structure in each time slice and the correlation information between time slices;

[0197] Specifically, the layer-by-layer graph convolution operation follows the message propagation framework. A common problem encountered is over-smoothing, which makes it difficult to capture higher-order neighborhood information, thereby damaging the representation ability of the captured main features of the wireless sensor network. Therefore, this scheme uses an adaptive layer aggregation mechanism to alleviate the over-smoothing problem. The fusion method is:

[0198]

[0199] in E It is the main structural feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors, , …, is the importance coefficient of the main features of different layers, which is adaptively trained here.

[0200] Compared with the traditional method that does not consider the fusion of main features or gives the same weight to the main features of different layers, the adaptive main feature fusion of this solution has higher generalization ability and stronger adaptability. This solution uses the inner product to predict the interaction value of two nodes at different times, receives the optimization target of the objective function construction module, and uses the Adam optimizer with weight decay to optimize all parameters to reduce the difference between the predicted value and the observed true value. Repeat the above spatiotemporal graph convolution and main feature fusion process until the objective function is Upper convergence: the convergence condition is that the absolute value of the difference between the value of the objective function before the start of this iteration and the value before the start of the previous iteration is less than the convergence termination threshold , or iterative control variables Reached limit.

[0201] The output unit is used to extract the fused main feature results that meet the preset accuracy requirements and store them in the main feature storage unit in the wireless sensor network data storage module; in this embodiment, the main feature results with the highest accuracy are stored.

[0202] Specifically, extract the main structural features that meet the preset accuracy requirements The main feature storage unit is stored in the wireless sensor network data storage module; wherein the preset accuracy requirement is the main feature of the structure with the highest accuracy.

[0203] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for extracting main features of wireless sensor networks, characterized in that: It includes the following: S1, receiving wireless sensor network data; S2. According to the data mode of the wireless sensor network, a target loss function is constructed to evaluate the operation and optimization process of the main features of the wireless sensor network and the main feature extraction; S3. Constructing an adjacency tensor according to the wireless sensor network data, extracting main features according to the wireless sensor network data and the adjacency tensor, and fusing them, extracting the fused main features that meet the preset accuracy requirements, and storing them; Among them, S3 includes: S304, extracting the main feature spatiotemporal pattern, including: According to the wireless sensor network data and the adjacency tensor, the space-time pattern contained in the time-varying wireless sensor network is extracted by using the relevant parameters of the initialized main feature tensor, wherein the space-time pattern is extracted by using a space-time graph convolutional network to perform hierarchical information mining on the tensor describing the time-varying wireless sensor network; The tensor multiplication-driven spatiotemporal graph convolution operation is: ； in for The multiplication operator, The transformation matrix of a layer is directly defined as a band indicator matrix capturing the adjacent time slice information; For the The adjacency tensor used by the spatiotemporal graph convolution. , for" Convert" operation, For the The transformation matrix for layer tensor multiplication; S305, fusion main features: in E is the main feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors; , …, is the importance coefficient of the main features of different layers.

2. The main feature extraction method of wireless sensor network according to claim 1 is characterized in that: The S2 comprises: S201, obtaining an instruction for constructing a target loss function; S202: Read the target weighted tensor of the wireless sensor network data according to the instruction of constructing the target loss function. A known data set , which only includes non-zero wireless sensor network data; Based on known data sets , the objective function is constructed as the distance target loss function, which is in the form of: The objective loss function is: 。 3. The main feature extraction method of wireless sensor network according to claim 2 is characterized in that: The S3 further includes: S301, receiving an instruction to extract main features of a wireless sensor network; S302, initializing relevant parameters according to the instructions of the main characteristics of the wireless sensor network; S303, constructing an adjacency tensor describing the spatiotemporal pattern of the wireless sensor network; Also includes: S306: Determine and extract iterative control variables Has the limit been reached? If yes, execute S308, if no, execute S307; S307, determine whether the distance target loss function value is within If the upper convergence is yes, then execute S309, if not, then execute S308; The target tensor for wireless sensor network data A known data set ; S308, using the gradient descent method, Perform iterative optimization and update ; S309: extracting and storing the fused main feature results that meet the preset accuracy requirements.

4. The main feature extraction method of wireless sensor network according to claim 3 is characterized in that: The relevant parameters include: main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , the main feature of initialization , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold .

5. The main feature extraction method of wireless sensor network according to claim 4 is characterized in that: The judgment distance target loss function value Is it in Upper convergence, including: before the start of this round of iteration Compared with the value before the last iteration , whether the absolute value of the difference between the two is less than the convergence judgment threshold If yes, it is considered converged, if not, it is considered not converged.

6. A wireless sensor network main feature extraction system, characterized in that: include: Wireless sensor network data receiving module, target loss function building module, wireless sensor network data storage module, main feature extraction module; The wireless sensor network data receiving module is used to import and manage the wireless sensor network data, and control the wireless sensor network data storage module to store the imported wireless sensor network data; The target loss function construction module is used to construct the target loss function according to the data mode of the wireless sensor network, which is used to evaluate the operation and optimization process of the main features of the wireless sensor network and the main feature extraction module; A wireless sensor network data storage module is used to store the wireless sensor network data received by the receiving module, the adjacency tensor of the spatial mode and the main features of the extracted wireless sensor network; The main feature extraction module is used to construct an adjacency tensor according to the wireless sensor network data, extract the main features according to the wireless sensor network data and the adjacency tensor, and fuse them, extract the fused main features that meet the preset accuracy requirements, and store them in the main feature storage unit in the wireless sensor network data storage module; The main feature extraction module includes: a main feature spatiotemporal pattern extraction unit, which is used to extract the main features of the high-order spatiotemporal pattern of each sensor node contained in the wireless sensor network according to the wireless sensor network data stored in the wireless sensor network data storage unit and the adjacent tensor stored in the adjacent tensor storage unit, combined with the relevant parameters of the main features initialized by the initialization unit; wherein the spatiotemporal pattern extraction is to use the spatiotemporal graph convolution network to perform hierarchical information mining on the tensor describing the time-varying wireless sensor network; The tensor multiplication-driven spatiotemporal graph convolution operation is: ； in for The multiplication operator, The transformation matrix of a layer is directly defined as a band indicator matrix capturing the adjacent time slice information; For the The adjacency tensor used by the spatiotemporal graph convolution. , for" Convert" operation, For the The transformation matrix for layer tensor multiplication; The main feature fusion unit is used to fuse the main features of different orders extracted by the main feature spatiotemporal pattern extraction unit and the initialized structural main features to obtain the overall main features: in E is the main feature of the final output, which includes the main feature of each node itself and all its K The main feature information of the hopping time and space neighbors; , …, is the importance coefficient of the main features of different layers.

7. The wireless sensor network main feature extraction system according to claim 6, characterized in that: The target loss function construction module is based on the known data set , construct the target loss function as the distance target loss function, which is in the form of: The target loss function is specifically: 。 8. The wireless sensor network main feature extraction system according to claim 7, characterized in that: The wireless sensor network data storage module comprises: Wireless sensor network data storage unit, adjacency tensor storage unit, main feature storage unit; A wireless sensor network data storage unit, used for storing the wireless sensor network data received by the receiving module; An adjacency tensor storage unit, used for storing an adjacency tensor describing a spatiotemporal topological structure of a wireless sensor network; A main feature storage unit, used to store the main features of the wireless sensor network extracted by the main feature extraction module; The main feature extraction module further includes: an initialization unit, a spatiotemporal adjacency tensor construction unit, and an output unit; An initialization unit is used to initialize the relevant parameters involved in the main feature extraction process of the wireless sensor network based on tensor decomposition; the relevant parameters include: the main feature space dimension , The perception window size of the matrix , the number of spatiotemporal graph convolution layers , the main feature of initialization , fusion ratio , …, , learning rate , the regularization factor , the maximum number of iterations , the current iteration number , termination convergence threshold ; A spatiotemporal adjacency tensor construction unit, used to construct an adjacency tensor describing the spatiotemporal pattern of wireless sensor networks; The output unit is used to extract the fused main feature results that meet the preset accuracy requirements and store them in the main feature storage unit in the wireless sensor network data storage module.

9. A device, characterized in that: The main feature extraction method of the wireless sensor network as described in any one of claims 1 to 5 is adopted.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the main feature extraction method for a wireless sensor network as claimed in any one of claims 1 to 5 are implemented.

Images