Multi-station joint electromagnetic pulse event identification method, system and device
Through the multi-station joint electromagnetic pulse event recognition method, the multi-station joint convolutional neural network combines independent feature branches and global feature vectors, the problem of low single-station recognition accuracy is solved, and more accurate and diverse electromagnetic pulse event recognition is achieved. .
Patent Information
- Application Number
- CN202111487997.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-08
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2041-12-08
AI Technical Summary
The existing electromagnetic pulse event recognition technology only considers a single detection station, resulting in low recognition accuracy and positioning accuracy, and it is impossible to effectively distinguish between nuclear explosion and lightning electromagnetic pulse.
The multi-station joint electromagnetic pulse event recognition method is adopted to collect multi-station electromagnetic pulse signals through distributed detection sites, and use the trained multi-station joint convolutional neural network (M-CNN) for identification, combining independent feature branches and The global feature vector performs classification of electromagnetic pulse event types.
It improves the recognition accuracy and diversity of electromagnetic pulse events, can more accurately identify more electromagnetic pulse events, provide more comprehensive electromagnetic pulse information, and achieve more efficient dynamic change adaptability.
Smart Images

Figure CN114118168B_ABST
Abstract
Description
Background Art
[0002] Since the electromagnetic pulse of a nuclear explosion is very similar to the electromagnetic pulse of lightning in nature, lightning has become the main interference source for the monitoring of electromagnetic pulses of nuclear explosions, and the identification of electromagnetic pulses of nuclear explosions has become a key issue in nuclear explosion detection. Many experts at home and abroad have used normalized spectrum features, time-frequency domain signal processing, artificial neural networks and other methods to identify nuclear explosions and lightning electromagnetic pulses (Liu Zhigang's time-varying spectrum feature analysis of nuclear explosion electromagnetic pulse signals in 2002, Li Peng's nuclear explosion and lightning electromagnetic pulse signal identification based on AR model and neural network in 2010, Qi Shufeng's nuclear explosion and lightning electromagnetic pulse identification based on time-frequency image analysis in 2013). In 2019, Zhu Baoyou et al. applied for a patent "Intelligent Lightning Identification Method Based on Convolutional Neural Network" and proposed the use of convolutional neural network to classify lightning waveforms (patent number CN109829407). These methods all use a single electromagnetic pulse detection station to identify and classify electromagnetic pulse waveforms, and the positioning accuracy and event recognition accuracy are low. Summary of the invention
[0003] In order to solve the above-mentioned problems in the prior art, that is, the prior electromagnetic pulse event recognition technology only considers recognition from a single detection station, and the recognition accuracy and positioning accuracy are insufficient and incomplete, the present invention provides a multi-station joint electromagnetic pulse event recognition method, the method comprising:
[0004] Step S100, collecting multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area;
[0005] Step S200: Based on the multi-site electromagnetic pulse signal, the multi-site electromagnetic pulse signal is identified by a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type.
[0006] In some preferred embodiments, the multi-station joint convolutional neural network is constructed based on the M-CNN convolutional neural network, including independent feature branches with the same number as the detection sites, 1 aggregation layer and 1 fully connected classification layer.
[0007] In some preferred embodiments, the multi-station joint convolutional neural network, wherein each independent feature branch includes one input layer connected in sequence, m groups of alternating one-dimensional convolutional layers and average pooling layers.
[0008] In some preferred implementations, the step S200 specifically includes:
[0009] Step S210, normalizing and filtering the electromagnetic pulse signals of multiple sites and performing waveform numbering to obtain numbered electromagnetic wave data;
[0010] Step S220, inputting the numbered electromagnetic wave data into independent feature branches of a multi-station joint convolutional neural network respectively;
[0011] Step S230: The multi-station joint convolutional neural network obtains the feature map of the input image through a one-dimensional convolutional layer.
[0012]
[0013] in, Indicates the connection to the previous feature map With the current feature map One-dimensional kernels, each kernel sweeps over the input feature map to generate a high-level representation, * represents the convolution operation, Represents the feature map of the previous layer With the current feature map The bias, M j Indicates the connection to the previous feature map and the output feature map of the previous layer, f(·) represents nonlinear activation, k represents the index of the feature branch, i represents the i-th feature point connected to the feature map of the previous layer, j represents the j-th feature point of the feature map of this layer, l represents the index value of the current layer, and the rectified linear unit function f(x) = max(0,x) is used to activate the deep learning structure. In particular, when i = 1, l = 1, k = 1, that is, Represents input data;
[0014] Step S240, the multi-station joint convolutional neural network downsamples the feature map through the shift pooling layer to generate a compressed feature map, and uses it as the feature map of the input image of the next convolution layer:
[0015]
[0016] Where R represents the maximum length of the pooling sequence, r represents the length of the pooling sequence of a single feature map, S represents the step size of the pooling, and n represents the index of the nth feature vector in the feature map;
[0017] Step S250, repeat the method of step S230-step S240 until the last pooling layer of the independent feature branch is completed, and the compressed feature map of the last layer of all independent feature branches is mapped into a one-dimensional feature vector V k :
[0018]
[0019] Among them, V k represents the feature vector of the kth feature branch, represents the last layer of the kth independent feature branch, L represents the last layer, and J represents the number of final output feature maps;
[0020] Step S260, the multi-station joint convolutional neural network organizes the one-dimensional feature vectors of all independent feature branches into a global one-dimensional feature vector V through a pooling layer:
[0021] V=[V1,V2,……,V k ]
[0022] Where V is the global eigenvector;
[0023] Step S270, performing softmax calculation on V through a fully connected layer, and outputting an element vector C, where the element vector C is the probability value of various preset electromagnetic pulse event types corresponding to the feature map of the input image.
[0024] In some preferred embodiments, the softmax calculation can be expressed as:
[0025]
[0026] Among them, W c It indicates that the weight vector is connected to the neurons of the output layer, Y represents the final output vector, each element of Y represents a positive number from 0 to 1, and the sum of all output elements is 1. C represents the maximum value of various preset electromagnetic pulse event types corresponding to the feature map of the input image, that is, how many types can be recognized at most, and c is the index value representing the cth event type.
[0027] In some preferred embodiments, the multi-station joint convolutional neural network is trained by:
[0028] Step A100, acquiring historical data of electromagnetic pulse signals from multiple sites, normalizing and filtering them, and numbering them to obtain numbered electromagnetic wave data as a training set;
[0029] Step A200, randomly selecting numbered electromagnetic wave data of the same electromagnetic pulse event from the training set, and obtaining the electromagnetic pulse event type of the training set by the method of steps S220 to S270;
[0030] Step A300, optimizing the matrix of trainable parameters by stochastic gradient descent method until the loss function is lower than a preset threshold, and obtaining a trained multi-station joint convolutional neural network; the loss function is:
[0031]
[0032] Where Q represents the minimum batch size, q represents the qth batch, C represents the maximum value of the feature map of the input image corresponding to various preset electromagnetic pulse event types, c represents the index value of the cth event type, and D q(c) represents the standard value corresponding to the cth output of the qth batch, Y q (c) represents the estimated value of the model output for the cth output quantity in the qth batch.
[0033] In some preferred embodiments, the matrix θ of the trainable parameters t+1 for:
[0034]
[0035] Among them, η represents the learning rate, represents the gradient operator and t represents the training round.
[0036] In another aspect of the present invention, a multi-station joint electromagnetic pulse event identification system is proposed, the system comprising a multi-station electromagnetic pulse signal collection module and an electromagnetic pulse event classification module;
[0037] The multi-site electromagnetic pulse signal collection module is configured to collect multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area;
[0038] The electromagnetic pulse event classification module is configured to identify the multi-site electromagnetic pulse signals based on the multi-site electromagnetic pulse signals through a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type.
[0039] The third aspect of the present invention proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned multi-station joint electromagnetic pulse event identification method.
[0040] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned multi-station joint electromagnetic pulse event identification method.
[0041] Beneficial effects of the present invention:
[0042] (1) The M-CNN structure proposed in the present invention recognizes electromagnetic pulse events through a multi-station joint electromagnetic pulse convolutional neural network, learns different functions from different waveforms through independent functional feature branches, and at the same time meets the integrity and diversity of event recognition to provide more comprehensive electromagnetic pulse event information. It can automatically and dynamically learn efficient functions from electromagnetic pulse data to adapt to the dynamic changes caused by different electromagnetic pulse events, and provide a more accurate electromagnetic pulse recognition rate.
[0043] (2) The present invention uses the proposed M-CNN structure to identify electromagnetic pulse event information through multiple monitoring stations. Because of its comprehensiveness, it can identify more types of events. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:
[0045] Figure 1 It is a flowchart of a multi-station joint electromagnetic pulse event identification method in an embodiment of the present invention;
[0046] Figure 2 It is a schematic diagram of the arrangement of the electromagnetic pulse detection station and the electromagnetic pulse data processing central station in an embodiment of the present invention;
[0047] Figure 3 is a schematic diagram of differences in electromagnetic pulse signals received from different sites in an embodiment of the present invention;
[0048] Figure 4 It is a schematic diagram of a group of electromagnetic pulse waveform sliding mutual correlation coefficient calculation results in an embodiment of the present invention;
[0049] Figure 5 are different schematic diagrams of signals of the same electromagnetic pulse event from different detection stations in an embodiment of the present invention;
[0050] Figure 6 Schematic diagram of the structure of a multi-station joint convolutional neural network in an embodiment of the present invention;
[0051] Figure 7 1 is a schematic diagram of time domain waveforms of ten typical electromagnetic pulse signals in an embodiment of the present invention;
[0052] Figure 8 It is a schematic diagram of model training and testing results in an embodiment of the present invention. DETAILED DESCRIPTION
[0053] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.
[0054] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0055] The present invention provides a multi-station joint electromagnetic pulse event identification method, which satisfies both the integrity and diversity of event identification to provide more comprehensive electromagnetic pulse event information and more accurate electromagnetic pulse identification efficiency.
[0056] A multi-station joint electromagnetic pulse event identification method of the present invention comprises:
[0057] Step S100, collecting multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area;
[0058] Step S200: Based on the multi-site electromagnetic pulse signal, the multi-site electromagnetic pulse signal is identified by a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type.
[0059] In order to more clearly explain the multi-station joint electromagnetic pulse event identification method of the present invention, the following is combined with Figure 1 Each step in the embodiment of the present invention is described in detail.
[0060] The multi-station joint electromagnetic pulse event identification method of the first embodiment of the present invention includes steps S100 to S200, and each step is described in detail as follows:
[0061] Step S100, collecting multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area;
[0062] In this embodiment, more than two detection stations and a data processing center station are set up in each detection area. The detection station collects electromagnetic pulse signals and transmits them to the data processing center station, which screens and identifies the data. Figure 2 shown.
[0063] The electromagnetic pulse signals in this embodiment include four major categories: lightning electromagnetic pulse signals, radio navigation station signals, mains power frequency interference pulse signals, and nuclear explosion electromagnetic pulse signals. Among them, lightning electromagnetic pulse signals are subdivided into seven subcategories according to their discharge types: positive polarity cloud-to-ground flash, negative polarity cloud-to-ground flash, positive polarity ordinary cloud flash, negative polarity ordinary cloud flash, positive polarity pocket cloud flash, negative polarity pocket cloud flash, and far-field lightning electromagnetic pulse signals (sky waves) propagated via the ground-ionosphere waveguide; there are a total of ten categories, and the waveforms of various typical events are as follows: Figure 7 shown.
[0064] The waveform signals received by multiple electromagnetic pulse stations reflect the performance characteristics of the same electromagnetic pulse event in different regions. Compared with the waveform signals of a single station, they can provide more comprehensive information about the electromagnetic pulse event. After the electromagnetic pulse signal propagates over a long distance in the air medium, the signal strength will be significantly attenuated. At the same time, during the propagation process, affected by the multipath effect and the ground-ionosphere waveguide propagation, the signal time domain and frequency domain will undergo significant changes. These propagation factors ultimately lead to great differences in the same-source electromagnetic pulse signals received in different regions. For example Figure 3 As shown in the figure, different stations receive different waveforms for the same event. However, these differences are all caused by one event and are the manifestations of the same event in different regions. The multi-station joint recognition model fully utilizes these difference features to achieve more accurate recognition of electromagnetic pulse signals, which is impossible for a single-station recognition method.
[0065] For event identification within a region, the dimensions of diversity and completeness need to be considered:
[0066] Diversity: Each station's waveform signal provides a unique signature corresponding to the specific distance and angle of the EMP event, so each single station's waveform contains single station identification capabilities.
[0067] Integrity: Waveforms from all detection sites reflect the same EMP event and should be treated as a whole.
[0068] Different from identifying electromagnetic pulse signals only through a single detection station, the problem of homology needs to be solved. After the data processing center receives the electromagnetic pulse signal, that is, after step S100, it will first determine whether multiple groups of electromagnetic pulse signals belong to the same event source, and then jointly identify the electromagnetic pulse signals of the same event source collected by multiple detection stations;
[0069] Specifically include:
[0070] Step B100, encapsulating the multi-site electromagnetic pulse signal and sending it to the data processing center station in the form of UDP or TCP;
[0071] Step B200, the data processing center station performs homologous pulse matching on the electromagnetic pulse signals of multiple sites based on the time and waveform correlation of the electromagnetic pulse signals of multiple sites, and screens out homologous electromagnetic pulse waveforms; after this step, most electromagnetic pulse interference data, such as mains power frequency interference signals and radio navigation signals, can be removed. In this embodiment, step B200 includes:
[0072] Step B210, for the electromagnetic pulse signals from multiple sites received by the data processing center, query the maximum distance D between the electromagnetic pulse detection stations max , calculate the maximum time difference W of the event source arriving at each detection stationt As shown in formula (1), the time difference is used to preliminarily determine the homologous pulse events:
[0073]
[0074] Among them, c 传 Indicates the propagation speed of electromagnetic pulse in air medium, usually set to c 传 =3×10 8 m / s;
[0075] If the error of the maximum time difference is less than the time window, the electromagnetic pulse event is preliminarily determined to be a homologous event;
[0076] Step B220, calculating the correlation coefficient r of the pulse signals between the detection stations by sliding the time window method t As shown in formula (2), the homologous pulse events can be further accurately determined based on the mutual correlation coefficient:
[0077]
[0078] Among them, X i and Y i Represents the electromagnetic pulse sequence collected by different detection stations, and represents, n represents the sequence length;
[0079] When the mutual correlation coefficient is higher than a preset mutual correlation threshold, it is identified as a homologous event, and the mutual correlation threshold is preferably 0.6.
[0080] By calculating the mutual correlation coefficient through the sliding window, an accurate correlation can be obtained to prevent the influence of the acquisition signal deviation on the result. The calculation results of the sliding mutual correlation coefficient of a set of electromagnetic pulse waveforms are as follows: Figure 4 shown.
[0081] That is, the electromagnetic pulse event information collected from multiple angles is input into the multi-dimensional convolutional neural network; in this embodiment, the length of the electromagnetic pulse time domain waveform data sequence is set to 1000, that is, 1000 discrete data points are used to represent an electromagnetic pulse signal. The sampling rate of the electromagnetic pulse detection station channel analog / digital conversion circuit is 1MSPS, that is, 1 data point is sampled per microsecond, so the original signal time length corresponding to the collection of 1000 data points is 1 millisecond.
[0082] Step S200: Based on the multi-site electromagnetic pulse signal, the multi-site electromagnetic pulse signal is identified by a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type.
[0083] In this embodiment, the multi-station joint convolutional neural network, such as Figure 6As shown in the figure, it is constructed based on the M-CNN convolutional neural network, including independent feature branches with the same number of detection sites, 1 pooling layer and a fully connected classification layer. Each station waveform corresponds to an independent functional branch in the multi-station joint recognition convolutional neural network, which can utilize the diversity of the electromagnetic pulse reception waveforms of multiple stations.
[0084] In this embodiment, the fully connected classification layer can be implemented by a fully connected softmax layer, which can be expressed as shown in formula (3):
[0085]
[0086] Among them, W c It means that the weight vector is connected to the neurons of the output layer, Y represents the final output vector, each element of Y represents a positive number from 0 to 1, and the sum of all output elements is 1.
[0087] In this embodiment, the multi-station joint convolutional neural network, wherein each independent feature branch includes one input layer connected in sequence, m groups of alternating one-dimensional convolutional layers and average pooling layers. Preferably, m can be set to 3. In order to maintain the balance between multiple electromagnetic pulse waveforms, the deployment of each functional branch follows a unified structure; the fully connected classification layer is used to summarize information from all element branches. In addition to the trainable parameters of the model, there are some parameters determined by repeated experiments, such as kernel size and merging stride. The one-dimensional convolutional layer can automatically learn and extract local effective features of the pulse signal through a suitable convolution kernel, and the number of features is unlimited. The more layers of the one-dimensional convolutional layer, the larger the number of parameters of the constructed neural network model, and the more features are learned.
[0088] In this embodiment, in step S210, the electromagnetic pulse signals of multiple sites are normalized and filtered and waveforms are numbered to obtain numbered electromagnetic wave data; Figure 5 As shown, the neurons in the two stations sense the same electromagnetic pulse sample at different angles, and the sensing field of each neuron is different, which can improve the feature extraction ability of the model while avoiding the risk of overfitting; preferably, in order to achieve standard accuracy, the number of stations needs to be greater than or equal to 4;
[0089] Step S220, inputting the numbered electromagnetic wave data into independent feature branches of a multi-station joint convolutional neural network respectively;
[0090] Step S230: The multi-station joint convolutional neural network obtains the feature map of the input image through a one-dimensional convolutional layer. As shown in formula (4):
[0091]
[0092] in, Indicates the connection to the previous feature map With the current feature map One-dimensional kernels, each kernel sweeps over the input feature map to generate a high-level representation, * represents the convolution operation, Represents the feature map of the previous layer With the current feature map The bias, M j Indicates the connection to the previous feature map and the output feature map of the previous layer, f(·) represents nonlinear activation, k represents the index of the feature branch, i represents the i-th feature point connected to the feature map of the previous layer, j represents the j-th feature point of the feature map of this layer, l represents the index value of the current layer, and the rectified linear unit function f(x) = max(0,x) is used to activate the deep learning structure. In particular, when i = 1, l = 1, k = 1, that is, Represents input data;
[0093] In step S240, the multi-station joint convolutional neural network downsamples the feature map through the shift pooling layer to generate a compressed feature map, which is used as the feature map of the input image of the next convolution layer as shown in formula (5):
[0094]
[0095] Where R represents the maximum length of the entire pooling sequence, r represents the length of the pooling sequence of a single feature map, S represents the step size of the pooling, and n represents the index of the nth feature vector in the feature map. The pooling layer enhances the invariance of the electromagnetic pulse waveform by compressing features, while reducing the computational burden of the model and alleviating the overfitting problem.
[0096] Step S250, repeat the method of step S230-step S240 until the last pooling layer of the independent feature branch is completed, and the compressed feature map of the last layer of all independent feature branches is mapped into a one-dimensional feature vector V k As shown in formula (6):
[0097]
[0098] Among them, V k represents the feature vector of the kth feature branch, represents the last layer of the kth independent feature branch, L represents the last layer, and J represents the number of final output feature maps;
[0099] Step S260, the multi-station joint convolutional neural network organizes the one-dimensional feature vectors of all independent feature branches into a global one-dimensional feature vector V through a pooling layer:
[0100] V=[V1,V2,......,Vk ]
[0101] Wherein, V is the global feature vector; the one-dimensional convergence layer is used to splice the feature parameters calculated by each branch to form a complete multi-dimensional feature that fully reflects the feature information of all detection stations;
[0102] Step S270, perform vector-matrix multiplication on V through the fully connected layer, and output the element vector C, where the element vector C is the probability value of various preset electromagnetic pulse event types corresponding to the feature map of the input image. The one-dimensional fully connected layer is used to fit and map the relationship between the multi-dimensional features and types of electromagnetic pulses, while continuously reducing the dimension size of the network layer, and the output result is a 1×10 one-dimensional vector. The output of the one-dimensional fully connected layer uses the Softmax logistic regression activation function to convert the 1×10 one-dimensional vector into the regression distribution of each type of probability, and present the multi-classification results in the form of probability.
[0103] In this embodiment, the multi-station joint convolutional neural network is trained by:
[0104] Step A100, acquiring historical data of electromagnetic pulse signals from multiple sites, normalizing and filtering them, and numbering them to obtain numbered electromagnetic wave data as a training set;
[0105] Step A200, randomly selecting numbered electromagnetic wave data of the same electromagnetic pulse event from the training set, and obtaining the electromagnetic pulse event type of the training set by the method of steps S220 to S270;
[0106] Step A300, optimizing the matrix of trainable parameters by stochastic gradient descent method until the loss function is lower than a preset threshold, and obtaining a trained multi-station joint convolutional neural network; the loss function is:
[0107]
[0108] Where Q represents the minimum batch size, q represents the qth batch, C represents the maximum value of the feature map of the input image corresponding to various preset electromagnetic pulse event types, c represents the index value of the cth event type, and D q (c) represents the standard value corresponding to the cth output of the qth batch, Y q(c) represents the estimated value of the model output for the cth output in the qth batch. Neurons in the fully connected layer are connected to all neurons in the adjacent layers; these dense connections generate most of the trainable weights for multi-station joint neural network recognition. The model is trained using tensorflow, which is Google's open source machine learning framework based on data flow graphs. Due to the large amount of data, the training speed is slow on the CPU, so GPU computing is used here. The GPU model is NVIDIA Quadro RTX 8000.
[0109] In this embodiment, the matrix θ of the trainable parameters t+1 for:
[0110]
[0111] Among them, η represents the learning rate, represents the gradient operator and t represents the training round.
[0112] The present invention can be directly applied to multiple mobile detection arrays, and the relative position information of each detection station is jointly input into the multi-station joint convolutional neural network.
[0113] In this embodiment, electromagnetic pulse data collected by a total of 80 detection stations distributed in various places are used for homology matching and signal type labeling, and an electromagnetic pulse data set is formed for verification.
[0114] The model was trained with 40 sample data sets, and the results of each training and verification are as follows: Figure 8 As shown in the figure, it can be seen that the recognition accuracy of the multi-branch deep convolutional neural network recognition model on the training set reaches 99.38%, and the recognition accuracy on the test set reaches 98.65%.
[0115] The multi-station joint electromagnetic pulse event identification system of the second embodiment of the present invention comprises: a multi-station electromagnetic pulse signal collection module and an electromagnetic pulse event classification module;
[0116] The multi-site electromagnetic pulse signal collection module is configured to collect multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area;
[0117] The electromagnetic pulse event classification module is configured to identify the multi-site electromagnetic pulse signals based on the multi-site electromagnetic pulse signals through a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type.
[0118] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0119] It should be noted that the multi-station joint electromagnetic pulse event identification system provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiment can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not regarded as improper limitations of the present invention.
[0120] An electronic device according to the third embodiment of the present invention comprises: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned multi-station joint electromagnetic pulse event identification method.
[0121] A fourth embodiment of the present invention is a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned multi-station joint electromagnetic pulse event identification method.
[0122] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the storage device and processing device described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.
[0123] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.
[0124] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article, or apparatus / device.
[0125] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.
Claims
1. A multi-station joint electromagnetic pulse event identification method, characterized in that: The method comprises: Step S100, collecting multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area; Step S200, based on the multi-site electromagnetic pulse signal, identifying the multi-site electromagnetic pulse signal through a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type, includes: Step S210, normalizing and filtering the electromagnetic pulse signals of multiple sites and performing waveform numbering to obtain numbered electromagnetic wave data; Step S220, inputting the numbered electromagnetic wave data into independent feature branches of a multi-station joint convolutional neural network respectively; Step S230: The multi-station joint convolutional neural network obtains the feature map of the input image through a one-dimensional convolutional layer. : ; in, Indicates the connection to the previous feature map With the current feature map One-dimensional kernels, each kernel sweeps over the input feature map to generate a high-level representation, * represents the convolution operation, Represents the feature map of the previous layer With the current feature map The bias, Indicates the connection to the previous feature map and the output feature map of the previous layer, represents nonlinear activation, k represents the index of the feature branch, i represents the i-th feature point connected to the feature map of the previous layer, j represents the j-th feature point of the feature map of this layer, l represents the index value of the current layer, and the rectified linear unit function f(x)=max(0,x) is used to activate the deep learning structure. In particular, when i =1, l =1, k =1, that is Represents input data; Step S240, the multi-station joint convolutional neural network downsamples the feature map through the shift pooling layer to generate a compressed feature map, and uses it as the feature map of the input image of the next convolution layer: ; Where R represents the maximum length of the pooling sequence, r represents the length of the pooling sequence of a single feature map, S represents the step size of the pooling, and n represents the index of the nth feature vector in the feature map; Step S250, repeat the method of step S230-step S240 until the last pooling layer of the independent feature branch is completed, and the compressed feature map of the last layer of all independent feature branches is mapped into a one-dimensional feature vector : ; in, represents the feature vector of the kth feature branch, represents the last layer of the kth independent feature branch, L represents the last layer, and J represents the number of final output feature maps; Step S260, the multi-station joint convolutional neural network organizes the one-dimensional feature vectors of all independent feature branches into a global one-dimensional feature vector V through a pooling layer: ; Where V is the global eigenvector; Step S270, performing vector-matrix multiplication on V through a fully connected layer, and outputting an element vector C, where the element vector C is the electromagnetic pulse event type with the highest probability; After step S100, it is first determined whether the multiple groups of electromagnetic pulse signals belong to the same event source, and then the electromagnetic pulse signals of the same event source collected by multiple detection stations are jointly identified, including: Step B100, encapsulating the multi-site electromagnetic pulse signal and sending it to the data processing center station in the form of UDP or TCP; Step B200, the data processing central station performs homologous pulse matching on the electromagnetic pulse signals of multiple sites based on the time and waveform correlation of the electromagnetic pulse signals of multiple sites, and screens out homologous electromagnetic pulse waveforms.
2. The multi-station joint electromagnetic pulse event identification method according to claim 1, characterized in that: The multi-station joint convolutional neural network is constructed based on the M-CNN convolutional neural network, including independent feature branches with the same number as the detection sites, 1 aggregation layer and 1 fully connected classification layer.
3. The multi-station joint electromagnetic pulse event identification method according to claim 2, characterized in that: The multi-station joint convolutional neural network, wherein each independent feature branch includes one input layer connected in sequence, m groups of alternating one-dimensional convolutional layers and average pooling layers.
4. The multi-station joint electromagnetic pulse event identification method according to claim 3 is characterized in that: The multi-station joint convolutional neural network is trained by: Step A100, acquiring historical data of electromagnetic pulse signals from multiple sites, normalizing and filtering them, and numbering them to obtain numbered electromagnetic wave data as a training set; Step A200, randomly selecting numbered electromagnetic wave data of the same electromagnetic pulse event from the training set, and obtaining the electromagnetic pulse event type of the training set by the method of steps S220 to S270; Step A300, optimizing the matrix of trainable parameters by stochastic gradient descent method until the loss function is lower than a preset threshold, and obtaining a trained multi-station joint convolutional neural network; the loss function is: ; Where Q represents the minimum batch size, q represents the qth batch, C represents the maximum value of the feature map of the input image corresponding to various preset electromagnetic pulse event types, and c represents the index value of the cth event type. Indicates the standard value corresponding to the cth output of the qth batch, Represents the estimated value of the model output for the cth output quantity in the qth batch.
5. The multi-station joint electromagnetic pulse event identification method according to claim 4, characterized in that: The matrix of trainable parameters for: ; in, represents the learning rate, represents the gradient operator and t represents the training round.
6. The multi-station joint electromagnetic pulse event identification method according to claim 2, characterized in that: The fully connected classification layer can be implemented by a fully connected softmax layer, which can be expressed as: ; in, It means that the weight vector is connected to the neurons of the output layer, Y represents the final output vector, each element of Y represents a positive number from 0 to 1, and the sum of all output elements is 1.
7. A multi-station joint electromagnetic pulse event identification system, characterized in that: The system comprises: a multi-site electromagnetic pulse signal collection module and an electromagnetic pulse event classification module; The multi-site electromagnetic pulse signal collection module is configured to collect multi-site electromagnetic pulse signals through all detection sites dispersedly arranged in the detection area; The electromagnetic pulse event classification module is configured to identify the multi-site electromagnetic pulse signals based on the multi-site electromagnetic pulse signals through a trained multi-site joint convolutional neural network to obtain the electromagnetic pulse event type, specifically including: The electromagnetic pulse signals of multiple sites are normalized and filtered and waveforms are numbered to obtain numbered electromagnetic wave data; Inputting the numbered electromagnetic wave data into independent feature branches of a multi-station joint convolutional neural network respectively; The multi-station joint convolutional neural network obtains the feature map of the input image through a one-dimensional convolutional layer : ; in, Indicates the connection to the previous feature map With the current feature map One-dimensional kernels, each kernel sweeps over the input feature map to generate a high-level representation, * represents the convolution operation, Represents the feature map of the previous layer With the current feature map The bias, Indicates the connection to the previous feature map and the output feature map of the previous layer, represents nonlinear activation, k represents the index of the feature branch, i represents the i-th feature point connected to the feature map of the previous layer, j represents the j-th feature point of the feature map of this layer, l represents the index value of the current layer, and the rectified linear unit function f(x)=max(0,x) is used to activate the deep learning structure. In particular, when i =1, l =1, k =1, that is Represents input data; The multi-station joint convolutional neural network downsamples the feature map through the shift pooling layer to generate a compressed feature map, which is used as the feature map of the input image of the next convolution layer: ; Where R represents the maximum length of the pooling sequence, r represents the length of the pooling sequence of a single feature map, S represents the step size of the pooling, and n represents the index of the nth feature vector in the feature map; Repeat the process of iterating the multi-station joint convolutional neural network to obtain the feature map of the input image through the one-dimensional convolutional layer and the multi-station joint convolutional neural network to downsample the feature map through the shift pooling layer to generate a compressed feature map, and use it as the feature map of the input image of the next convolution layer until the last pooling layer of the independent feature branch is completed, and the compressed feature map of the last layer of all independent feature branches is mapped to a one-dimensional feature vector : ; in, represents the feature vector of the kth feature branch, represents the last layer of the kth independent feature branch, L represents the last layer, and J represents the number of final output feature maps; The multi-station joint convolutional neural network organizes the one-dimensional feature vectors of all independent feature branches into a global one-dimensional feature vector V through a pooling layer: ; Where V is the global eigenvector; Perform vector-matrix multiplication on V through the fully connected layer and output element vector C, which is the type of electromagnetic pulse event with the highest probability; After the multi-site electromagnetic pulse signal collection module collects the electromagnetic pulse signals of the multi-sites, it first determines whether the multiple groups of electromagnetic pulse signals belong to the same event source, and then jointly identifies the electromagnetic pulse signals of the same event source collected by multiple detection stations, including: Encapsulating the multi-site electromagnetic pulse signal and sending it to the data processing center station in the form of UDP or TCP; The data processing central station performs homologous pulse matching on the electromagnetic pulse signals of multiple sites based on the time and waveform correlation of the electromagnetic pulse signals of multiple sites, and screens out homologous electromagnetic pulse waveforms.
8. An electronic device comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being used to be executed by the processor to implement the multi-station joint electromagnetic pulse event identification method as described in any one of claims 1-6.
9. A computer-readable storage medium storing computer instructions, wherein the computer instructions are used to be executed by the computer to implement the multi-station joint electromagnetic pulse event identification method according to any one of claims 1 to 6.
