A perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios

Through the improved SE-DenseNet model and intensive linking technology, the problem of the decrease in the recognition rate of distributed fiber sensing after migration scenarios is solved, and a high-precision event classification network is quickly built on low-configuration devices, saving resources and improving identification accuracy is achieved. Rate.

CN118781723BActive Publication Date: 2024-11-22QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES)
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411273012.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-11-22
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

In the prior art, the recognition rate of distributed fiber sensing decreases after migrating the scene, and it is necessary to re-acquire a large amount of data and re-train the model, which consumes a lot of manpower and computing resources, and the low-latitude characteristics of the signal are in multiple convolution operations. The problem of disappearing afterwards.

Method used

Adopting the improved SE-DenseNet deep learning network model, weight parameters are obtained by training in experimental scenarios, and only the fully connected layer of the model is trained after the migration scenario, combining dense links and SENet layers to reduce gradient vanishing and feature redundancy. Problem, improve the generalization ability and recognition rate of the model.

Benefits of technology

A high-accurate event classification is achieved in the migration scenario, saving computing power and human resources, and can quickly build a high-precision event classification network on low-configuration machines, reducing the impact of noise on recognition results, and improving the migration of models and recognition accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118781723B_ABST
    Figure CN118781723B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of perimeter security intrusion event monitoring, and in particular to a perimeter security intrusion signal recognition method based on distributed optical fiber sensing in a migration scenario, comprising the following steps: constructing a distributed optical fiber vibration sensing system in fences of different scenarios; constructing a signal data set of different intrusion events in different scenarios; preprocessing the intrusion signals in the event signal data set to construct a spatiotemporal feature data set; using an improved SE‑DenseNet deep learning network model as a feature extractor, and training it using the spatiotemporal data in an experimental scenario to obtain weight parameters; after migrating the scenario, only the fully connected layer of the model is trained by using the trained weight parameters to complete the recognition and classification of intrusion events in the new scenario.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of perimeter security intrusion event monitoring, and in particular to a perimeter security intrusion signal recognition method based on distributed optical fiber sensing in a migration scenario. Background Art

[0002] With the development of society and the advancement of technology, perimeter security has become increasingly important for protecting property and personnel safety. Traditional perimeter security methods often rely on the deployment of sensors and cameras, but there are problems such as blind spots. By utilizing distributed fiber optic sensing technology, real-time monitoring of larger areas can be achieved with high accuracy and reliability.

[0003] In the early days, with the development and widespread application of artificial intelligence, more and more scholars have introduced it into the field of fiber optic sensing. After extracting the features of the intrusion signal, its features are selected and classified according to the extracted features. The multi-domain analysis method enriches the extraction information of the detection signal, so in the early days, there appeared an empirical learning method based on "fixed artificial feature extraction + specific recognition model". During this period, it mainly relied on expert knowledge and experience to select appropriate features, such as: after the wavelet transform improves the signal-to-noise ratio, the spectrum energy distribution, the time domain amplitude zero crossing rate and other features are artificially extracted, and then the intrusion signal is identified and classified by selecting traditional recognition models such as support vector machine (SVM), K-nearest neighbor algorithm (KNN), traditional neural network, etc. Later, with the significant breakthroughs made in the field of computer vision and signal processing by deep learning technology: it can automatically learn and extract features from a large amount of data to realize the recognition and classification of complex patterns. The introduction of deep learning methods provides new ideas and opportunities for the application of distributed fiber optic sensing in perimeter security, pipeline safety, highway safety monitoring, etc. The distributed fiber optic sensing security intrusion signal recognition system based on deep learning continuously monitors the surrounding environment for a long time through optical fibers deployed in hidden environments such as walls, underground or fences. When physical disturbances such as people walking or climbing occur near the optical fiber, the optical signal in the optical fiber will change. The signal is acquired through the distributed fiber optic vibration monitoring system (DVS), and then the deep learning algorithm is used to identify and classify the signal, forming an intrusion recognition system that combines deep learning and distributed fiber optic sensing.

[0004] Most of today's algorithm models have poor perception of small changes and details, which leads to the disappearance of low-dimensional features after multiple convolution operations. In addition, when the test data is quite different from the training data, especially when the model is migrated to different scenarios, the algorithm recognition rate decreases due to the influence of uncertainties on the signal in different scenarios. To address the above two issues, most practices today are to re-collect a large amount of data for the migrated scenarios and retrain the models. However, this approach consumes a lot of manpower and computing resources, and retraining has high requirements for the configuration of the host computer. Therefore, finding a method that can effectively avoid feature loss and quickly build a high-precision event classification network on a lower-configuration machine after migrating the scenario is the key to the current field. Summary of the invention

[0005] The present invention provides a perimeter security intrusion signal recognition method based on distributed fiber optic sensing in a migration scenario to solve the problems in the prior art that the low-latitude features of the signal disappear after multiple convolution operations, the recognition rate decreases after the migration scenario, a large amount of data needs to be re-collected, and the model needs to be retrained, which consumes a lot of manpower and computing resources.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A perimeter security intrusion signal recognition method based on distributed optical fiber sensing in a migration scenario includes the following steps:

[0008] S1: Scenario construction: Build a distributed optical fiber vibration sensing system on fences in different scenarios;

[0009] S2: Data preparation: Construct signal datasets of different intrusion events in different scenarios;

[0010] S3: Signal preprocessing: preprocess the intrusion signals in the event signal dataset and construct a spatiotemporal feature dataset;

[0011] S4: Model training: The improved SE-DenseNet deep learning network model is used as a feature extractor, and the spatiotemporal data in the experimental scene is used to train it to obtain weight parameters;

[0012] S5: Retrain after migrating the scene: After migrating the scene, only the fully connected layer of the model is trained using the trained weight parameters to complete the recognition and classification of intrusion events in the new scene.

[0013] Furthermore, the step S1 specifically includes the following steps:

[0014] S1.1: Distributed optical fibers are laid in three scenarios: training base, villa area, and laboratory, and connected to distributed optical fiber vibration sensing systems to collect signals of the same intrusion event in different scenarios;

[0015] S1.2: There are three ways to deploy distributed optical fibers in each scenario, namely, straight-line optical fiber, S-shaped optical fiber, and underground optical fiber, which are used to collect different intrusion events.

[0016] Furthermore, the step S2 specifically includes the following steps:

[0017] S2.1: Using the distributed optical fiber vibration sensing system built in three scenarios, collect the distributed optical fiber sensing intrusion signal near the fence in each complex background environment;

[0018] S2.2: There are six types of intrusion events: climbing, walking, digging, drones, rolling stones, and throwing objects. The sampling frequency of the distributed fiber optic vibration sensing system is 100-10000Hz, the spatial resolution is 1-10m, the sampling time of each type of time in each scene is 30-60min, and the event frequency is once every 2s.

[0019] Furthermore, the step S3 specifically includes the following steps:

[0020] S3.1: Based on the data signal transmitted from the optical fiber, high-pass filtering is used to reduce the noise of the signal, thereby reducing the impact of low frequencies on the intrusion signal. When high-frequency signals pass through the system, the attenuation is reduced;

[0021] S3.2: The distributed optical fiber vibration sensing system returns the original signal trajectory distributed along the space at each moment, continuously accumulates N original signal trajectories on the time axis, and constructs a space-time signal matrix with N time dimensions and S space dimensions;

[0022] S3.3: Through data processing, the obtained spatiotemporal signal matrix is ​​converted into a 224×224 spatiotemporal graph, and the spatiotemporal graphs of the six types of intrusion events in each scenario are obtained, where each type of event has 3000 samples in each scenario;

[0023] S3.4: For the data in the laboratory scene, it is divided into training set and test set in a ratio of 8:2. The training base and villa area scene data are used as the updated data set after the migration scene.

[0024] Furthermore, the step S4 specifically includes the following steps:

[0025] S4.1: The six types of intrusion event spatiotemporal graph samples are passed into the first 7×7 convolutional layer with a stride of 2 in the SE-DenseNet network to extract features from the input image, and then the extracted data is passed into the 3×3 maximum pooling layer with a stride of 2 to reduce the size of the feature map;

[0026] S4.2: Input the output of the maximum pooling layer into the first Dense Block module of SE-DenseNet. The first Dense Block module includes six Dense Layers. One Dense Layer consists of two BatchNormalization layers, two Relu activation functions, a 3×3 convolution, and a 1×1 convolution. The pre-activation arrangement is in the front and the convolution layer in the back. This method can alleviate the gradient vanishing problem and improve the training effect and generalization ability of the model. The arrangement order is: BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv;

[0027] Furthermore, the dense layers of each Dense Block are densely connected for data transmission to alleviate the problem of gradient disappearance, improve feature utilization, and reduce the recognition rate after model migration. The method is to use the output of all previous layers of the same Dense Block as the input of the current Dense Layer. The formula is as follows:

[0028] ,in, represents the input of the current layer, It represents a nonlinear transformation function, which is a combination of Bn, Relu, Pooling and Conv. Represents the concatenation of the i outputs of the previous layer.

[0029] Furthermore, a Dense Block with i layers has connections, each Dense Layer outputs k unique feature maps, that is, the number of channels of the obtained feature map is k, then the number of channels of the i-th layer input is:

[0030]

[0031] Among them, S i is the number of channels input to the i-th layer, k represents the growth rate of the network, which is used to extract different features, and k0 represents the number of input channels of the current Dense Block.

[0032] Furthermore, due to feature reuse, the input of the subsequent layers will be very large. Therefore, a bottleneck structure is used inside each Dense Block to reduce the amount of calculation. That is, a 1×1 convolution is introduced as a bottleneck layer before each 3×3 convolution to reduce the number of input feature maps, thereby improving calculation efficiency.

[0033] S4.3: For transfer learning, the data of different scenarios will be more complicated, and DenseNet still has limitations such as feature redundancy and overfitting. Therefore, the present invention adds a SENet layer after each Dense Block module to solve this problem. The SENet layer first performs global average pooling on the Dense Block input feature map to obtain global information on each channel.

[0034] The obtained global description is denoted as , the input feature map is X, and its size is C×H×W, where C represents the number of channels, H represents the height, and W represents the width; i Represents the average value on the i-th channel:

[0035]

[0036] Apply a nonlinear activation function Relu to the result of global average pooling. This process is used to learn the relationship between channels and focus on the channels that contribute to network performance. The result is called compressed representation , where s i is the compressed representation on the i-th channel, and the formula is as follows:

[0037] ,in is the activation function, is a fully connected layer. Multiply the compressed representation s by the original input feature map X to obtain the excited feature map. The formula is as follows:

[0038] ,

[0039] The overall formula is expressed as:

[0040]

[0041] in is the activation function and FC is the fully connected layer.

[0042] The SENet layer dynamically adjusts the importance of each channel by learning a weight on each channel, increasing the focus on useful features while suppressing irrelevant features, thereby enhancing the generalization ability of DenseNet.

[0043] The SENet layer then passes the data to the Transition layer, which has the following structure: BN+ReLU+1x1 Conv+2x2AvgPooling. It is used to compress the incoming data again as the input of the next Dense Block. The Transition layer generates features, among which is the compression factor, and m is the number of input channels.

[0044] S4.4: The present invention includes four Dense Block modules, each of which contains 6, 12, 24 and 16 Dense Layers respectively, with a growth rate k of 32. The features of the events after the migration scene may be obliterated by noise. The DenseBlock module uses features at different levels so that the pre-trained weights still have high adaptability after migration. Except for the last layer, a SENet layer and a Transition layer are added after each Dense Block module to enhance the model's ability to extract global information and reduce feature redundancy caused by Dense Block. Finally, a global average pooling layer and a linear layer are used for classification. The present invention uses the above network model as a feature extractor and uses the spatiotemporal data in the experimental scene to train it to obtain weight parameters.

[0045] Furthermore, the step S5 specifically includes the following steps:

[0046] S5.1: After changing the scene, load the weights trained in S4, use a small amount of data in the new scene, and retrain only the fully connected layer of the model. In this stage, the cross entropy loss function is used to estimate the difference between the model's predicted value and the true value. The formula is as follows:

[0047]

[0048] where C is the number of categories, is the jth element of the true label of the i-th sample, indicating whether the i-th sample belongs to category j. is the model's predicted probability that the i-th sample belongs to category j;

[0049] S5.2: Use the Adam optimizer to update the network parameters based on the gradient information to minimize the loss function, and use ReduceLROnPlateau to dynamically adjust the learning rate and use the linear layer to classify events in new scenes.

[0050] Compared with the prior art, the effective results of the present invention are:

[0051] (1) The present invention is suitable for extracting the spatiotemporal structural information of distributed optical fiber vibration and acoustic wave sensor signals, has strong feature extraction capabilities, and can obtain classification results with high accuracy through simple training in complex migration scenarios. Compared with the current method of collecting a large amount of data to retrain the model after the migration scenario, the present invention can save a lot of computing power and human resources, and can be trained directly on a low-configuration on-site host computer.

[0052] (2) The present invention introduces DenseNet into the distributed fiber optic sensor signal recognition. In view of the fact that there may be uncertain noise in the scene that may cause the original signal to be submerged, each layer of DenseNet can directly participate in the propagation of the error signal, thus realizing implicit "deep supervision". This means that each layer has the opportunity to "know" information about the final classification result and make adjustments based on this information, thereby reducing the impact of noise on the recognition result and increasing the generalization ability of the model.

[0053] (3) The present invention proposes an SE-Densenet model, which uses SENet to learn a weight on each channel of the output of DenseBlock to dynamically adjust the importance of each channel, increase the focus on useful features, and suppress irrelevant features, thereby increasing the recognition rate of DenseNet.

[0054] (4) In different scenarios, the present invention can achieve a higher recognition accuracy and better transferability by using the weight file pre-trained in the original scenario and adding a new fully connected layer to simply retrain the model after migration.

[0055] (5) When there is a lack of a large amount of training data, the present invention can obtain a model with a higher recognition rate by relying on a small amount of training samples, and effectively avoid the common gradient vanishing problem in the CNN model.

[0056] (6) The present invention aims to solve the problem that as the number of layers in a common network increases, low-dimensional features are lost, resulting in a decrease in recognition rate. The present invention adopts a dense linking method to preserve low-dimensional features and fully utilize the characteristic information of the signal to improve the recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of perimeter security intrusion signal recognition in the migration scenario of the present invention;

[0058] Figure 2 This is a structural diagram of the pre-training algorithm in the present invention;

[0059] Figure 3 This is a schematic diagram of retraining after migrating scenes in the present invention;

[0060] Figure 4 It is a schematic diagram of the distributed vibration sensing system (DVS) in the present invention;

[0061] Figure 5 This is a plan diagram for laying optical fiber for perimeter security in three scenarios: training base, laboratory and villa area;

[0062] Figure 6 The spatiotemporal diagrams of six randomly selected intrusion events in the present invention, where (a) is stone rolling, (b) is climbing, (c) is throwing objects, (d) is walking, (e) is digging, and (f) is drone. DETAILED DESCRIPTION

[0063] In order to clearly illustrate the technical features of this solution, the solution is described below through specific implementation methods.

[0064] A perimeter security intrusion signal recognition method based on distributed optical fiber sensing in a migration scenario includes the following steps:

[0065] Taking the fences in training bases, villa areas, and laboratory scenarios as specific examples, the flow chart of the perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios is as follows: Figure 1 shown.

[0066] S1. Scenario construction: Distributed optical fiber vibration sensing system is constructed on fences in different scenarios; the construction method is as follows Figure 5 As shown, the single-mode optical fiber is laid in an "S" shape and a "line" shape for 100 meters each, and the single-mode optical fiber is buried in the open space near the fence for 60 meters. Several tens of meters of coils are retained at the end and beginning of each section. One end of the optical fiber is connected to the distributed fiber vibration system (DVS) and the other end is coiled.

[0067] S2. Data preparation: Data signals are collected through a distributed fiber vibration sensing system (DVS) in three scenarios. The principle is as follows: Figure 4The main equipment used in the system includes ultra-narrow line laser, acousto-optic modulator, erbium-doped fiber amplifier, circulator, photoelectric detector, data acquisition card, and host computer. The system of this embodiment adopts a narrow line laser with a line width of 5kHz as the light source. The ultra-narrow line width laser generates a continuous coherent optical signal, which is modulated into an optical pulse signal by the acousto-optic modulator to generate an optical pulse sequence with a frequency shift of 60mhz. The optical pulse signal is centrally amplified by the erbium-doped fiber amplifier. The amplified optical pulse signal is injected into the detection optical cable through port I and port II of the circulator. The optical pulse signal generates Rayleigh scattering during the transmission process along the optical cable, and then the Rayleigh scattered optical signal returns along the optical cable, is received by port II and port III of the circulator, and enters the photoelectric detector 7. The signal is collected by a 100MHZ data acquisition card, and the phase change information introduced is used to obtain the effect signal of the sound wave and vibration on the optical fiber through phase demodulation, and the event signal is stored through a personal computer or a storage device. The sampling frequency of the distributed fiber optic vibration sensing system is 3000 Hz, the spatial resolution is 3m, the sampling time for each type of event in each scenario is 30-60min, and the event frequency is once every 2s.

[0068] S3. Signal preprocessing: Preprocess the intrusion signals in the event signal data set to construct a spatiotemporal feature data set. First, use high-pass filtering to reduce the noise of the signal to reduce the impact of lower frequencies on the intrusion signal. Then, arrange each denoised data in position order into a spatiotemporal matrix, and convert the matrix into a 224*224 spatiotemporal graph.

[0069] S4. Model training: The improved SE-DenseNet deep learning network model is used as a feature extractor, and the spatiotemporal data in the experimental scene is used to train it to obtain weight parameters. The improved SE-DenseNet model is as follows: Figure 2 The present invention uses 18,000 spatiotemporal graphs of six types of events in laboratory scenes in a ratio of 8:2 as the validation set and test set in the training phase.

[0070] The 224*224 spatiotemporal graph is passed into the first 7*7 convolutional layer with a step size of 1 of the network to obtain the convolution result; then the output result is passed into the first 3*3 maximum pooling layer with a step size of 2 to obtain the result.

[0071] The result of the maximum pooling layer is passed to the first Dense Block module. The first Dense Block includes 6 DenseLayers. Each Dense Layer uses dense links to transmit data to alleviate the problem of gradient disappearance, improve feature utilization, and reduce the recognition rate after model migration. The method is to use the output of all previous layers of the same Dense Block as the input of the current Dense Layer. The formula is as follows:

[0072]

[0073] in, represents the input of the current layer, It represents a nonlinear transformation function, which is a combination of Bn, Relu, Pooling and Conv. Represents the concatenation of the i outputs of the previous layer.

[0074] The densely connected features are input into the Dense Layer. Due to feature reuse, the input of the subsequent layers will be very large. Therefore, a bottleneck structure is used inside each Dense Block to reduce the amount of calculation, that is, each DenseLayer consists of a structure of BN+ReLU+CONV (1*1) + ReLU+CONV (3*3); a 1×1 convolution is introduced as a bottleneck layer before each 3×3 convolution to reduce the number of input feature maps, thereby improving computational efficiency.

[0075] After 6 Dense Layers are repeated, the features flow out of the first Dense Block and reach SENet. The SENet layer first performs global average pooling on the Dense Block input feature map to obtain global information on each channel. The obtained global description is recorded as , where z i Represents the average value on the i-th channel:

[0076]

[0077] Apply a nonlinear activation function Relu to the result of global average pooling. This process is used to learn the relationship between channels and focus on the channels that contribute to network performance. The result is called compressed representation , where s i is the compressed representation on the i-th channel, and the formula is as follows:

[0078]

[0079] in is the activation function, is a fully connected layer. Multiply the compressed representation s by the original input feature map X to obtain the excited feature map. The formula is as follows:

[0080] ,

[0081] The overall formula is expressed as:

[0082] ,in is the activation function and FC is the fully connected layer.

[0083] The SENet layer dynamically adjusts the importance of each channel by learning a weight on each channel, increasing the focus on useful features while suppressing irrelevant features, thereby enhancing the generalization ability of DenseNet.

[0084] The SENet layer data is passed to the Transition layer. The structure of this layer is: BN+ReLU+1x1 Conv+2x2AvgPooling. The function is to compress the transmitted data again as the input of the next Dense Block. The Transition layer generates features, among which is the compression factor, and m is the number of input channels.

[0085] The data is repeatedly passed through Dense Block containing 6, 12, 24, and 16 Dense Layers, as well as 3 Transitions and 4 SENets.

[0086] The output of the last SENet is input into a global average pooling layer to obtain the pooled result, which is then input into the linear layer to classify the result. After 300 rounds of training, the training weights of the model are obtained.

[0087] S5. Retrain after migrating the scene: After migrating the scene, add a new fully connected layer for training by loading the trained weight parameters to complete the recognition and classification of intrusion events in the new scene. The process is as follows: Figure 3 shown.

[0088] In the training base and villa area scenes, 1,800 spatiotemporal data maps (six types of events) were collected respectively. Figure 6 .

[0089] The data collected from the two scenarios are used to add a new fully connected layer to the weights obtained in S4 for retraining: the Adam optimizer is used to update the network parameters according to the gradient information to minimize the loss function, and ReduceLROnPlateau is used to dynamically adjust the learning rate adaptively, and a linear layer is added to classify the events in the new scenario.

[0090] The cross entropy loss function used is used to estimate the difference between the model's predicted value and the true value. The formula is as follows:

[0091] , where C is the number of categories, is the jth element of the true label of the i-th sample, indicating whether the i-th sample belongs to category j. is the model's predicted probability that the i-th sample belongs to category j;

[0092] In the experimental stage, the update training was carried out on a DVS without GPU, with 16GB RAM and Intel Core i7-12700 CPU. The transfer learning methods using different data amounts for updating and freezing different network layers were compared in terms of recognition accuracy and training time. The results after 30 rounds of training showed that the SE-DenseNet-based feature extractor transfer learning method with higher recognition rate in the post-migration scene achieved a test accuracy of 97.53% after 19.81 minutes of retraining using data from 600 new scenes. Compared with the other two transfer learning methods of freezing the first 1 / 4 and freezing the first 1 / 2, the time was shortened by 20.87 minutes and 9.34 minutes respectively, and the test accuracy was improved by 13.76% and 3.83% respectively. The comparison of the experimental transfer learning schemes is shown in Table 1.

[0093] Table 1. Comparison results of experimental transfer learning schemes

[0094]

[0095] The present invention provides a new, convenient and fast method for perimeter security intrusion signal recognition, which can quickly adapt to a new scenario and be applied to common engineering host computers.

Claims

1. A perimeter security intrusion signal recognition method based on distributed optical fiber sensing in a migration scenario, characterized in that: The following steps are involved: S1. Scenario construction: construct a distributed optical fiber vibration sensing system on fences in different scenarios; S2, data preparation: construct signal datasets of different intrusion events in different scenarios; S3, signal preprocessing: preprocess the intrusion signal in the event signal dataset and construct a spatiotemporal feature dataset; S4, model training: The improved SE-DenseNet deep learning network model is used as a feature extractor, and the spatiotemporal data in the experimental scene is used to train it to obtain weight parameters; S5. Retrain after migrating the scene: After migrating the scene, only the fully connected layer of the model is trained using the trained weight parameters to complete the recognition and classification of intrusion events in the new scene; The step S4 specifically comprises the following steps: S4.

1. The samples of various intrusion event spatiotemporal graphs are passed into the first 7×7 convolutional layer with a stride of 2 in the SE-DenseNet network to extract features from the input image. The extracted data is then passed into the 3×3 maximum pooling layer with a stride of 2 to reduce the size of the feature graph. S4.

2. Input the output of the maximum pooling layer into the first Dense Block module of SE-DenseNet. The first Dense Block module includes six Dense Layers. Each Dense Layer consists of two BatchNormalization layers, two Relu activation functions, a 3×3 convolution and a 1×1 convolution. The activation function is in front and the convolution layer is in the back. The pre-activation arrangement is: BN+ReLU+1x1 Conv+BN+ReLU+3x3 Conv. S4.

3. The dense layers of each Dense Block are densely connected for data transmission. The method is to use the output of all previous layers of the same Dense Block as the input of the current Dense Layer. The formula is as follows: ,in, represents the input of the current layer, It represents a nonlinear transformation function, which is a combination of Bn, Relu, Pooling and Conv. Represents the connection of i outputs of the previous layer; S4.

4. A Dense Block with i layers has connections, each Dense Layer outputs k unique feature maps, that is, the number of channels of the obtained feature map is k, then the number of channels of the i-th layer input is: , Among them, S i is the number of channels input to the i-th layer, k represents the growth rate of the network, which is used to extract different features, and k0 represents the number of input channels of the current Dense Block; A bottleneck structure is used inside each Dense Block to reduce the amount of calculation, that is, a 1×1 convolution is introduced as a bottleneck layer before each 3×3 convolution to reduce the number of input feature maps, thereby improving calculation efficiency; S4.

5. Add a SENet layer after each Dense Block module. The SENet layer first performs global average pooling on the Dense Block input feature map to obtain the global information on each channel. The obtained global information description is recorded as , the input feature map is X, the size is C×H×W, where C represents the number of channels, H represents the height, and W represents the width; Perform global average pooling on the input feature map X to obtain the average value of each channel, where Represents the average value on the i-th channel: , a nonlinear activation function Relu is applied to the result of global average pooling, and the result is called compressed representation , where s i is the compressed representation on the i-th channel, and the formula is as follows: ,in is the activation function, is a fully connected layer; Multiply the compressed representation s by the original input feature map X to obtain the excited feature map. The formula is as follows: , , the overall formula is expressed as: ,in is the activation function, FC is the fully connected layer; The SENet layer then passes the data to the Transition layer, which has the structure of BN+ReLU+1x1 Conv+2x2AvgPooling. The transmitted data is compressed again as the input of the next Dense Block. The Transition layer generates features, among which is the compression factor, m is the number of input channels; S4.

6. The SE-DenseNet contains four Dense Block modules in total. Each Dense Block contains 6, 12, 24 and 16 Dense Layers respectively. The growth rate k is 32. The Dense Block modules use features at different levels. Except for the last layer, a SENet layer and a Transition layer are added after each Dense Block module. Due to the feature redundancy brought by the Dense Block, it is finally classified by a global average pooling layer and a linear layer. The above network model is used as a feature extractor, and the spatiotemporal data in the experimental scenario is used to train it to obtain the weight parameters.

2. The perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios according to claim 1 is characterized in that: The S1 specifically includes the following steps: S1.

1. Distributed optical fibers are set up in different scenarios and connected to distributed optical fiber vibration sensing systems to collect signals of the same intrusion event in different scenarios; S1.

2. There are three ways to deploy distributed optical fibers in each scenario, namely, straight-line optical fiber, S-shaped optical fiber, and buried optical fiber, which are used to collect different intrusion events.

3. The perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios according to claim 1 is characterized in that: The S2 specifically includes the following steps: The distributed fiber optic vibration sensing system built in different scenarios collects distributed fiber optic sensing intrusion signals near the fence in each complex background environment; the sampling frequency of the distributed fiber optic vibration sensing system is 100-10000Hz, the spatial resolution is 1-10m, the sampling time of each type of time in each scenario is 30-60min, and the event frequency is once every 2s.

4. The perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios according to claim 1 is characterized in that: The step S3 specifically comprises the following steps: S3.

1. According to the data signal transmitted from the optical fiber, the signal is subjected to noise reduction processing by high-pass filtering to reduce the influence of low frequency on the intrusion signal. When the high frequency signal passes through the system, the attenuation is reduced; S3.2, the distributed optical fiber vibration sensing system returns the original signal trajectory distributed along the space at each moment, continuously accumulates N original signal trajectories on the time axis, and constructs a space-time signal matrix with N time dimensions and S space dimensions; S3.3, through data processing means, the obtained space-time signal matrix is ​​converted into a 224×224 space-time graph, and the space-time graph of various intrusion events in each scenario is obtained, where each type of event has 3000 samples in each scenario; S3.

4. Divide the data in the experimental scene into a training set and a test set in a ratio of 8:2, and use the remaining scene data as the updated data set after migrating the scene.

5. The perimeter security intrusion signal recognition method based on distributed optical fiber sensing in migration scenarios according to claim 1 is characterized in that: The step S5 specifically comprises the following steps: S5.

1. After changing the scene, load the weights trained in S4, use a small amount of data in the new scene, and retrain only the fully connected layer of the model. At this stage, the cross entropy loss function is used to estimate the difference between the model's predicted value and the true value. The formula is as follows: , where C is the number of categories, is the jth element of the true label of the i-th sample, Indicates whether the i-th sample belongs to category j, which is the model's predicted probability that the i-th sample belongs to category j; S5.

2. Use the Adam optimizer to update network parameters based on gradient information to minimize the loss function, use ReduceLROnPlateau to dynamically adjust the learning rate, and use the linear layer to classify events in new scenes.

Citation Information

Cited By

  • Distributed vibration and temperature simultaneous measurement device

    CN119880037A