A DAS data enhancement method based on improved DRAGAN

Abstract

The present invention relates to the technical field of DAS data enhancement processing, and specifically designs a DAS data enhancement method based on an improved DRAGAN, comprising the following steps: S1, constructing a DAS perimeter security system, collecting and preprocessing data; S2, improving the DRAGAN data enhancement model, introducing an EnhancedMCAM module into a generator, and introducing a residual block and a CBAM module into a discriminator to form an MCDRAGAN data enhancement model; S3, training the data enhancement model, and adopting an adaptive learning rate adjustment strategy; S4, inputting the preprocessed data into the MCDRAGAN data enhancement model to obtain generated samples; S5, using an AlexNet classification model to evaluate the MCDRAGAN data enhancement model and verifying the effectiveness of the data enhancement model. The present invention generates new data samples by using the improved MCDRAGAN data enhancement model, which well simulates the distribution characteristics of the data, expands the diversity of the data set, and can obtain a better recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of DAS data enhancement processing, and specifically designs a DAS data enhancement method based on an improved DRAGAN. Background Art

[0002] Traditional perimeter security technology has long relied on technologies such as infrared detection alarms, vibration cables, and electronic fences, but these technologies have significant drawbacks. Infrared detection alarms are easily restricted by terrain, with high false alarm rates in complex terrain or harsh weather conditions, and are unable to achieve long-distance, accurate positioning. Vibration cable solutions rely on active sensing units, resulting in high system power consumption, short sensor lifespans, and high maintenance costs. Solutions such as electronic fences pose a risk to personnel and experience reduced sensitivity in electromagnetic interference environments. In addition, traditional systems generally lack positioning capabilities and are unable to respond to intrusions in a timely manner. Perimeter security technology based on distributed fiber acoustic sensing systems (DAS) is gaining increasing attention. Compared to other traditional perimeter security technologies, it offers advantages such as the ability to obtain continuously distributed information, long detection distances, high spatial resolution, low deployment costs, high anti-interference capabilities, and privacy protection.

[0003] In perimeter security scenarios, normal states (no intrusion) account for over 99.7%, while abnormal events follow a low-incidence Poisson distribution, resulting in an extremely low probability of obtaining positive samples and making it difficult to collect sufficient samples for model training. DAS raw data must be labeled by experts, which is not only costly, time-consuming, and prone to errors, severely limiting the scale of labeled data. In actual deployment environments, complex interference factors such as weather changes and human activity make data unusable for model training. Data scarcity can lead to unstable model training, making deep learning models prone to falling into local optimality, resulting in degraded model performance. Furthermore, data scarcity exacerbates the problem of insufficient data, affecting the model's adaptability to complex scenarios and limiting the intelligent recognition and early warning capabilities of perimeter security systems. Therefore, researching data augmentation methods for DAS to address data scarcity is of great significance.

[0004] Traditional data augmentation methods generate synthetic data by simply altering the original samples. However, this method produces low-quality data and can lead to severe homogeneity in the augmented data. Generative adversarial networks (GANs), proposed by Goodfellow et al. in 2014, leverage game theory and adversarial thinking to generate high-quality images in a short period of time. They have become a mainstream approach to addressing data scarcity. Through adversarial training of the generator and discriminator, GANs can generate high-quality data with a distribution similar to real data. However, existing GAN methods still have some problems when processing DAS data. For example, GANs are prone to mode collapse, resulting in insufficient diversity in the generated data, and the discriminator is susceptible to noise, resulting in poor quality of generated data. To address these issues, researchers have proposed various improved methods. LMA-DRAGAN is an improved model based on DRAGAN that introduces an improved LSKblock module and MobileViT module to enhance the stability of the generator and the diversity of the generated data. However, it still fails to fully capture the global characteristics of the data, and the discriminator is susceptible to noise when processing data, resulting in limited diversity in the generated data.

[0005] Therefore, it is necessary to propose a DAS data augmentation method based on improved DARGAN, which will focus on solving the shortcomings of LMA-DRAGAN to enhance the model's ability to capture global features and improve the quality and diversity of generated data. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the present invention proposes a DAS data enhancement method based on an improved DRAGAN.

[0007] The present invention is achieved through the following technical solutions:

[0008] A DAS data augmentation method based on an improved DRAGAN includes the following steps:

[0009] S1: Build a DAS perimeter security system, collect data and pre-process it;

[0010] S2: Improve the DRAGAN data enhancement model. The generator introduces the EnhancedMCAM module, and the discriminator introduces the residual block and CBAM module to form the MCDRAGAN data enhancement model.

[0011] S3: Data augmentation model training with adaptive learning rate adjustment strategy;

[0012] S4: Input the preprocessed data into the MCDRAGAN data augmentation model to obtain generated samples;

[0013] S5: Use the AlexNet classification model to evaluate the MCDRAGAN data augmentation model and verify the effectiveness of the data augmentation model.

[0014] Furthermore, the specific steps of S1 are set as follows:

[0015] A network of optical cables was used, with horizontal and vertical cables laid across the fence in a staggered pattern, creating a three-bend loop within each section of the fence for complete coverage. After the fiber was laid, different types of DAS data were collected. IIR filters were used to denoise the raw signals, and differential averaging was used to compress the data. The preprocessed data was then segmented into 4-second time units in the time domain. The resulting data and its corresponding labels constituted the raw dataset.

[0016] Furthermore, the specific steps of S2 are set as follows:

[0017] S2.1: The EnhancedMCAM module is introduced into the generator. The EnhancedMCAM module is obtained by making the following improvements to the existing MCAM module: (1) replacing the single convolutional layer with convolution kernels of different sizes to extract multi-scale features; (2) replacing the single self-attention mechanism with a multi-head self-attention mechanism; (3) introducing a channel attention mechanism to dynamically adjust the channel weights through global average pooling and two fully connected layers; (4) introducing a learnable gating mechanism to control the residual connection strength; and inserting the EnhancedMCAM module at different levels of the generator.

[0018] By extracting multi-scale features using convolution kernels of different sizes, the EnhancedMCAM module is able to capture feature information of these different scales, thereby introducing more diversity in the generation process.

[0019] DAS data contains a large amount of noise and interference signals, and traditional single-head self-attention mechanisms cannot effectively distinguish important features from noise. The multi-head self-attention mechanism maps input features into different subspaces, obtains attention vectors from multiple subspaces, and then combines them to analyze the correlation between embedding methods from multiple perspectives. This enables the generator to more accurately focus on features related to intrusion behavior, improving the quality of generated data.

[0020] The channel attention mechanism dynamically adjusts channel weights through global average pooling and two fully connected layers to enhance the features of important channels.

[0021] To better control the strength of the residual connection, a learnable gating mechanism is introduced. This gating mechanism generates gating weights through a convolutional layer and a sigmoid activation function, dynamically adjusting the ratio of output features to input features. This allows the generator to flexibly adjust the strength of the residual connection based on different input features when generating data, further improving the quality of the generated data.

[0022] S2.2: The residual block and CBAM module are introduced into the discriminator. The residual block alleviates the gradient vanishing problem in deep networks by introducing residual connections, enabling the network to learn deeper features; the CBAM module enhances the feature expression ability through channel attention and spatial attention mechanisms.

[0023] In the discrimination process, the gradient vanishing problem of deep networks is a major factor affecting model performance. Residual blocks alleviate this problem by introducing residual connections, enabling the network to learn deeper features.

[0024] The CBAM module combines channel attention and spatial attention mechanisms to enhance the expressive power of features. In DAS data, features from different channels and spatial locations have varying importance for identifying intrusions. The channel attention mechanism dynamically adjusts channel weights through global average pooling and global maximum pooling to enhance the features of important channels. The spatial attention mechanism focuses on the importance of different spatial locations in the feature map, weighting the channel-weighted feature map in the spatial dimension. This dual attention mechanism enables the discriminator to more accurately distinguish between real and generated samples, reducing the possibility of the generator "deceiving" the discriminator and making adversarial training more stable.

[0025] S2.3: Construct the MCDRAGAN model. The network structure of its generator and discriminator is as follows:

[0026] Generator construction: Input noise is passed through a fully connected layer to generate preliminary features; the preliminary features are upsampled through a transposed convolution layer, and the Enhanced MCAM module is inserted for multi-scale feature extraction and multi-head self-attention calculation; the features are upsampled through a transposed convolution layer, and the Enhanced MCAM module is inserted for channel attention and gating mechanism calculation; the features are upsampled through a transposed convolution layer, and finally the image is generated;

[0027] Discriminator construction: The input image first passes through an initial convolutional layer to extract preliminary features; then, the features pass through three residual blocks, each of which is followed by a CBAM module; the residual blocks enhance the transmission of features through residual connections, and the CBAM module enhances the expressiveness of features through channel and spatial attention mechanisms; after the last residual block, the features pass through a multi-scale feature extraction module to extract feature information at different scales; finally, the features pass through a final convolutional layer to output a feature vector; this feature vector is concatenated with the feature vector output by the multi-scale feature extraction module, and then classified through a fully connected layer to output the final discrimination result.

[0028] Furthermore, S3 includes the following steps:

[0029] S3.1: The model ensures gradient flow by constraining the Lipschitz continuity of the discriminator; the discriminator must satisfy the 1-Lipschitz condition, which requires that for any point in the input space x , the gradient of the discriminator function at this point must be less than or equal to 1; when the data distribution generated by the generator p g With the real data distribution p data The optimal discriminator can be obtained theoretically. In the Wasserestein framework, the optimal discriminator D The gradient norm of * approaches 1 at the equilibrium point; the optimization goal of the generator is to minimize the Wasserstein distance as much as possible. By introducing the gradient penalty term and imposing constraints, the update direction of the generator can be made more stable;

[0030] S3.2: During model training, the generator and discriminator are trained alternately to update their respective parameters. First, the generator is fixed to train the discriminator. At this stage, the parameters of the generator are fixed, and the discriminator receives real data. x and generate data G ( z ) for training; then fix the discriminator to train the generator. At this stage, the parameters of the discriminator are fixed, and the generator receives random noise. z Repeat the above steps to train the MCDRAGAN data enhancement model until the Nash equilibrium is reached. At this time, the data generated by the generator is G ( z ) and real data x Indistinguishable from the perspective of the discriminator;

[0031] S3.3: In the initial stage of training, the CosineAnnealingLR scheduler is used to dynamically adjust the generator's learning rate based on the shape of the cosine function. This allows the generator to quickly adapt to the discriminator's capabilities in the early stages of training. At the same time, a low learning rate is used for fine-tuning in the later stages of training to avoid overfitting.

[0032] At the same time, the model is trained using the RMSprop optimizer, whose idea is to use the moving average of the squared gradient to dynamically adjust the learning step size of each parameter. The update process can be described as: first, calculate the current gradient; then, calculate an exponentially decaying moving average of the squared gradient; next, divide the current gradient by the square root of the moving average of the squared gradient plus a very small constant; finally, multiply this scaled gradient by the learning rate to update the model parameters; among them, the decay rate is taken as 0.99, and the very small constant is added to prevent division by zero errors during the calculation process.

[0033] Furthermore, S3.1: The model ensures gradient flow by constraining the Lipschitz continuity of the discriminator; its objective function is:

[0034] ;

[0035] In the formula, the third term is the gradient penalty term, l is the gradient penalty coefficient, d is Gaussian noise perturbation, c To control the disturbance intensity, cI is the covariance matrix; after experiments, the hyperparameters are tuned and finally set l =10, c =8;

[0036] Among them, the 1-Lipschitz condition requires: .

[0037] For any point x , its gradient must satisfy: .

[0038] When generating distribution p g =p data When , the optimal discriminator satisfies:

[0039] .

[0040] In the Wasserestein framework, D The gradient norm of * approaches 1 at the equilibrium point.

[0041] The optimization goal of the generator is to minimize the Wasserstein distance: .

[0042] The gradient penalty term is constrained , making the update direction of the generator more stable.

[0043] S3.2: During model training, the generator and discriminator are trained alternately to update their respective parameters. In order to make the training more stable, the generator is first fixed to train the discriminator. At this stage, the parameters of the generator are fixed, and the discriminator receives real data. x and generate data G ( z ) for training, the loss function of the discriminator can be expressed as:

[0044] .

[0045] Then fix the discriminator to train the generator. At this stage, the parameters of the discriminator are fixed and the generator receives random noise. z For training, the loss function of the generator can be expressed as:

[0046] .

[0047] Repeat the above steps to train the MCDRAGAN data augmentation model until the Nash equilibrium is reached. G ( z ) and real data x Indistinguishable from the perspective of the discriminator.

[0048] S3.3: In the initial stage of training, the generator may not be able to effectively deceive the discriminator due to its weak generation ability, resulting in too small a gradient update of the generator and a mismatch in model learning difficulty. The CosineAnnealingLR scheduler is used to dynamically adjust the learning rate of the generator according to the shape of the cosine function, so that it can quickly adapt to the ability of the discriminator in the early stage of training. At the same time, a low learning rate is used for fine-tuning in the later stage of training to avoid overfitting. Its expression is:

[0049] ;

[0050] in, t Indicates the current iteration number, or t Indicates the t The learning rate at the iteration, or min represents the minimum value of the learning rate, or maxrepresents the maximum value of the learning rate, T max Indicates the number of iterations in a cycle.

[0051] At the same time, in order to overcome the problem of excessive swing in the loss function during the update process, the model is trained using the RMSprop optimizer, which is expressed as:

[0052] , ;

[0053] in, v t represents the moving average of the square of the gradient; Represents the current gradient; Indicates the attenuation rate, with a value of 0.99; represents the learning rate; is a small constant used to avoid the denominator being zero.

[0054] Furthermore, the specific steps of S4 are set as follows:

[0055] The preprocessed data is input into the MCDRAGAN data enhancement model, where the generator generates data similar to the real data. x Generated data G ( z ) to deceive the discriminator until the discriminator cannot distinguish between real samples and generated samples, thereby generating more realistic and diverse samples; the mean absolute error MAE, root mean square error RMSE and initial score IS indicators are used to evaluate the quality of the generated samples, and the formula is:

[0056] , , ;

[0057] in, N is the sample size, y i It is i The true value of the sample, It is i The predicted value of the sample, p ( y | G i ) is the i Zhang generates the probability distribution of the categories of samples, p ( y ) is the average category probability distribution of all generated samples.

[0058] Furthermore, the specific steps of S5 are set as follows:

[0059] In order to verify the effectiveness of the data enhancement model, the AlexNet model is used and the five-fold cross-validation method is used. The specific steps are as follows:

[0060] The original training set is divided into five equal parts, each of which is used as a validation set, and the remaining four parts are used as training sets. For each set of training sets and validation sets, the following steps are repeated: use the training set to train the MCDRAGAN model; use the generated image data to enhance the training set; use the enhanced training set to train the AlexNet classification model; evaluate the performance of the AlexNet classification model on the validation set and record the classification accuracy; calculate the average classification accuracy of the five validations as the final evaluation indicator.

[0061] At the same time, the data is filled at 20%, 40%, 60%, 80%, and 100% of the original dataset size to verify the necessity of data enhancement.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] First, the present invention improves the DRAGAN data augmentation model and introduces the EnhancedMCAM module into the generator. The EnhancedMCAM module is obtained by making the following improvements to the existing MCAM module: (1) replacing the single convolution layer with convolution kernels of different sizes to extract multi-scale features; (2) replacing the single self-attention mechanism with a multi-head self-attention mechanism; (3) introducing a channel attention mechanism to dynamically adjust the channel weights through global average pooling and two fully connected layers; (4) introducing a learnable gating mechanism to control the residual connection strength; and inserting the EnhancedMCAM module at different levels of the generator. Through multi-scale feature extraction and multi-head self-attention mechanism, EnhancedMCAM can capture richer feature information, thereby introducing more diversity in the generation process; the introduction of channel attention and gating mechanism enables EnhancedMCAM to more effectively utilize feature information, enhance the representation ability of important features, and thus improve the quality of generated images. The discriminator incorporates residual blocks and CBAM modules. The residual blocks, by introducing residual connections, alleviate the vanishing gradient problem in deep networks, enabling the network to learn deeper features. The CBAM module enhances the expressiveness of features through channel-wise and spatial-attention mechanisms. By focusing attention, it more accurately distinguishes between real and generated samples, reducing the likelihood of the generator "deceiving" the discriminator and making adversarial training more stable. An adaptive learning rate adjustment strategy is employed to address mode collapse caused by a mismatch in model learning difficulty. Ultimately, the MCDRAGAN data augmentation model is constructed to generate new data samples, expand the diversity of the dataset, and thus reduce the cost of acquiring real data. This research provides a reliable augmentation strategy for DAS perimeter security under small sample conditions, facilitating its wider engineering application. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is the overall network structure diagram of the present invention;

[0065] Figure 2 This is a structural diagram of the DAS system in the present invention;

[0066] Figure 3 This is a network structure diagram of the model generator in the present invention;

[0067] Figure 4 This is the model discriminator network structure diagram in the present invention. DETAILED DESCRIPTION

[0068] This paper provides a DAS data enhancement method based on improved DRAGAN (see Figure 1 ), the specific steps are as follows:

[0069] S1: Build a DAS perimeter security system, collect data and pre-process it;

[0070] Specifically, a DAS system is used for data collection and fiber optic laying, such as Figure 2 As shown, it includes a narrow linewidth laser, an acousto-optic modulator, an erbium-doped amplifier I, a circulator, an erbium-doped amplifier II, a grating filter, a coupler I and a coupler II, which are connected in sequence through a sensing optical fiber. The circulator is connected to a signal acquisition optical fiber laid in an application scenario. In this embodiment, the application scenario is perimeter security, that is, the circulator is connected to a signal acquisition optical fiber laid in a fence. Coupler II is connected to photodetector I, photodetector II, and photodetector III respectively. Photodetector I, photodetector II, and photodetector III are all connected to a data acquisition card through a data cable, and the data acquisition card is connected to a computer through a data cable. The DAS system also includes a pulse generator, which is connected to the acousto-optic modulator and the data acquisition card respectively.

[0071] Specifically, the optical fiber is laid horizontally on one section of the fence and vertically on another section of the fence, forming three curved loops on each section of the fence, which are laid horizontally and vertically in an alternating manner. After the optical fiber is laid, data is collected through the DAS system to collect intrusion events at the fence. The data collected in this application includes five typical perimeter security data types: climbing, digging, throwing objects, rolling stones, and people walking, as shown in Table 1.

[0072] Table 1. Dataset

[0073]

[0074] Specifically, the data collected by the DAS system contains a large amount of environmental low-frequency noise. The IIR filter is used to denoise the collected original signal. Its output signal is a series of addition and multiplication operations of the current input signal and the filter's historical output and historical input signals. The IIR filter is represented by a difference equation as shown below:

[0075] ;

[0076] in, x ( n ) is the input sequence, y ( n ) is the output sequence, a k 、 b k is the filter coefficient, N is the filter order; the sampling frequency is determined to be 2000Hz, the filtering frequency is 5Hz, and the filter order is fixed to fourth order.

[0077] Specifically, since the sampling frequency is 2000Hz, each sampling point collects 2000 data points per second, resulting in a very large amount of data. A differential averaging method is used to compress the data in the time domain. The collected data is grouped into groups of 256 in the time dimension, and a differential sequence is calculated. The differences between adjacent data points are calculated through a simple loop, and the difference between two adjacent data points in the time series is calculated and stored in a new sequence. The differential sequence is then summed and divided by the length of the differential sequence to obtain the average of the differences. This operation compresses every 256 data points in the time domain into a single data point. This average value contains the information of all 256 sampled data points, achieving data compression.

[0078] Specifically, the system sampling point count is 8192, the sampling frequency is 2000 Hz, and the data is compressed 256 times after differential averaging. Therefore, the 32 data in the time domain are equivalent to the 8192 data before the differential averaging operation, which is also equivalent to the data collected within a time range of 4.096 seconds. For ease of processing, the data obtained after preprocessing are selected as a group of 32 data in the time domain, that is, they are divided into 4s as a time unit. The obtained data and their corresponding labels constitute the original data set. In order to eliminate the influence of the spatial position of the event on the classification, we shuffle the spatial position of the data set to ensure that the spatial distribution of different events is more random.

[0079] S2: Improve the DRAGAN data enhancement model. The generator introduces the EnhancedMCAM module, and the discriminator introduces the residual block and CBAM module to form the MCDRAGAN data enhancement model.

[0080] Specifically, an EnhancedMCAM module is introduced into the generator. The Enhanced MCAM module is obtained by making the following improvements to the existing MCAM module: (1) replacing the single convolutional layer and using convolution kernels of different sizes to extract multi-scale features; (2) replacing the single self-attention mechanism with a multi-head self-attention mechanism; (3) introducing a channel attention mechanism to dynamically adjust the channel weights through global average pooling and two fully connected layers; (4) introducing a learnable gating mechanism to control the residual connection strength; and inserting the EnhancedMCAM module at different levels of the generator.

[0081] The generator network structure is as follows Figure 3As shown in the figure, the input noise is used to generate preliminary features through the fully connected layer; the preliminary features are upsampled through the transposed convolution layer and inserted into the EnhancedMCAM module for multi-scale feature extraction and multi-head self-attention calculation; the features are upsampled through the transposed convolution layer and inserted into the EnhancedMCAM module for channel attention and gating mechanism operations; the features are upsampled through the transposed convolution layer and finally generate the image;

[0082] Specifically, a multi-scale convolutional layer is defined. For each scale s, the feature extraction process is as follows:

[0083] ;

[0084] in, x is the input feature, Indicates that the convolution kernel size is k s The convolution operation, BatchNorm is the batch normalization operation, ReLU is the activation function;

[0085] Using the multi-head self-attention mechanism, assuming that the feature representation of a sample after network learning is:

[0086] ;

[0087] Query Matrix Q , keyword matrix K Sum Matrix V Both H The linear projection of n is the key number of subspaces; scaled dot product attention is an important component of the multi-head self-attention mechanism, which is defined as follows:

[0088] ;

[0089] Where: Q , K , V Represents a linear mapping format, defined as Q = W q e i , K = W k e i , V = W v e i ; d k Indicates the size of the projection;

[0090] Through the multiple self-attention operations in the above formula, the input features are mapped to different subspaces, and the attention vectors of multiple subspaces are obtained. Then, they are combined to analyze the correlation between embedding methods from multiple perspectives and improve model performance. The definition of multi-head attention is as follows:

[0091] , ;

[0092] in, , , Respectively Q , K , V In the i The weight matrix of the attention heads is used to concatenate the outputs of multiple attention heads using the fusion layer;

[0093] For the channel self-attention mechanism (SE module), the spatial feature compression is first performed on the feature map, and global average pooling is implemented in the spatial dimension to obtain a 1×1×C feature map. After learning through the FC fully connected layer, a feature map with channel attention is obtained, and its dimension is still 1×1×C. The feature map of the channel attention 1×1×C and the feature map of the original input H×W×C are multiplied by the weight coefficient channel by channel, and then processed by the Sigmoid function to finally output the feature map with channel attention.

[0094] The EnhancedMCAM module introduces a learnable gating mechanism that generates gating weights through a convolutional layer and a Sigmoid activation function, dynamically adjusting the fusion ratio of output features to input features. The gate weight calculation formula is defined as:

[0095] ;

[0096] The final output is: ;

[0097] in, Output It is a feature processed by multi-scale feature extraction and multi-head self-attention mechanism.

[0098] Specifically, the residual block and CBAM module are introduced into the discriminator. The residual block alleviates the gradient vanishing problem in deep networks by introducing residual connections, enabling the network to learn deeper features; the CBAM module enhances the expressiveness of features through channel attention and spatial attention mechanisms;

[0099] The discriminator network structure is as follows Figure 4As shown in the figure, the input image first passes through an initial convolutional layer to extract preliminary features; then, the features pass through three residual blocks, each of which is followed by a CBAM module; the residual blocks enhance the transmission of features through residual connections, and the CBAM module enhances the expressiveness of features through channel and spatial attention mechanisms; after the last residual block, the features pass through a multi-scale feature extraction module to extract feature information at different scales; finally, the features pass through a final convolutional layer to output a feature vector; this feature vector is concatenated with the feature vector output by the multi-scale feature extraction module, and then classified through a fully connected layer to output the final discrimination result;

[0100] Specifically, the residual block introduces skip connections to output the input features to the output layer, alleviating the vanishing gradient problem in the network. Specifically, a 3×3 convolution kernel is used instead of the original 5×5 convolution to reduce the amount of computation while maintaining the receptive field. Instance normalization (InstanceNorm) replaces the original non-normalization design to stabilize the training process. The learnable downsampling operation maintains the consistency of feature map size changes.

[0101] CBAM is a combination of the channel attention mechanism and the spatial attention mechanism. It first uses the channel attention module to focus on the importance of different channels, weights the channels of the feature map, and automatically enhances the feature channels related to the discrimination task. Then, it uses the spatial attention module to focus on the importance of different spatial positions in the feature map, and weights the channel-weighted feature map in the spatial dimension, thereby achieving more refined adjustment of the feature map.

[0102] The channel attention mechanism dynamically adjusts the channel weights through global average pooling and global maximum pooling to enhance the features of important channels, and then generates channel attention weights through a shared multi-layer perceptron (MLP). The formula is as follows:

[0103] ;

[0104] The spatial attention mechanism performs global average pooling and global maximum pooling on the feature map after the channel attention mechanism to generate spatial attention weights:

[0105] ;

[0106] Finally, the channel and spatial attention weights are combined by element-wise multiplication to enhance the representation of key features:

[0107] ;in, F is the input feature map, is the feature map weighted by the channel attention mechanism, Represents the sigmoid function;

[0108] S3: Data augmentation model training with adaptive learning rate adjustment strategy;

[0109] Specifically, the model ensures gradient flow by constraining the Lipschitz continuity of the discriminator; its objective function is:

[0110] ;

[0111] In the formula, the third term is the gradient penalty term, l is the gradient penalty coefficient, d is Gaussian noise perturbation, c To control the disturbance intensity, cI is the covariance matrix; after experiments, the hyperparameters are tuned and finally set l =10, c =8.

[0112] Among them, the 1-Lipschitz condition requires: .

[0113] For any point x , its gradient must satisfy: .

[0114] When generating distribution p g =p data When , the optimal discriminator satisfies: .

[0115] In the Wasserestein framework, D The gradient norm of * approaches 1 at the equilibrium point;

[0116] The optimization goal of the generator is to minimize the Wasserstein distance: .

[0117] The gradient penalty term is constrained , making the update direction of the generator more stable;

[0118] Specifically, during model training, the generator and discriminator are trained alternately to update parameters. To make the training more stable, the generator is first fixed to train the discriminator, and then the discriminator is fixed to train the generator. The training process is shown in Algorithm 1.

[0119] Algorithm 1: The training strategy of the model. In one iteration, the model parameters of the generator and the discriminator are updated once respectively.

[0120]

[0121] Specifically, in the initial stage of training, the generator may not be able to effectively deceive the discriminator due to its weak generation ability, resulting in too small a gradient update of the generator and a mismatch in model learning difficulty. The CosineAnnealingLR scheduler is used to dynamically adjust the learning rate of the generator according to the shape of the cosine function, so that it can quickly adapt to the ability of the discriminator in the early stage of training. At the same time, fine-tuning is performed through a low learning rate in the later stage of training to avoid overfitting. Its expression is:

[0122] ;

[0123] in, t Indicates the current iteration number, or t Indicates the t The learning rate at the iteration, or min represents the minimum value of the learning rate, or max represents the maximum value of the learning rate, T max Indicates the number of iterations in a cycle; after experiments, set T max is 200.

[0124] At the same time, in order to overcome the problem of excessive swing in the loss function during the update process, the model is trained using the RMSprop optimizer, which is expressed as:

[0125] , ;

[0126] in, v t represents the moving average of the square of the gradient; Represents the current gradient; Indicates the attenuation rate, with a value of 0.99; represents the learning rate; is a small constant used to avoid the denominator being zero.

[0127] S4: Input the preprocessed data into the MCDRAGAN data augmentation model to obtain generated samples;

[0128] Specifically, it is compared with 5 data enhancement models:

[0129] (1) GAN: Generate realistic data through adversarial training of generator and discriminator;

[0130] (2) DCGAN: Use convolutional layers instead of fully connected layers to improve the quality of generated images;

[0131] (3) LSGAN: uses the least squares loss function to alleviate the vanishing gradient problem and generate clearer images;

[0132] (4) WGAN-GP: Wasserstein GAN with gradient penalty, solving the gradient problem;

[0133] (5) DRAGAN: Enhances training stability and generation diversity through local region adjustment optimization;

[0134] (6) LMA-DRAGAN: An improved model based on DRAGAN.

[0135] Specifically, we define the evaluation indicators and use MAE, RMSE and IS indicators to evaluate the quality of the generated samples. The formula is:

[0136] , , ;

[0137] in, N is the sample size, y i It is i The true value of the sample, It is i The predicted value of the sample, p ( y | G i ) is the i Zhang generates the probability distribution of the categories of samples, p ( y ) is the average category probability distribution of all generated samples.

[0138] Specifically, the analysis and comparison results are shown in Table 2:

[0139] Table 2. Quality evaluation of generated samples

[0140]

[0141] As shown in the table, the MCDRAGAN data augmentation model achieves the highest MAE value, reaching 0.1660; the highest RMSE value, reaching 0.2271; and the highest IS value, reaching 2.7756. Compared with other models, the samples generated by this model have higher diversity.

[0142] S5: Use the AlexNet classification model to evaluate the MCDRAGAN data augmentation model and verify the effectiveness of the data augmentation model;

[0143] Specifically, to further illustrate the effectiveness of the MCDRAGAN method proposed in this paper, the AlexNet model was trained using the original dataset and the enhanced dataset respectively, and the five-fold cross-validation method was used for verification. The specific steps are as follows:

[0144] The original training set was divided into five equal parts, each of which served as a validation set, and the remaining four as training sets. For each set of training and validation sets, the following steps were repeated: using the training set to train the MCDRAGAN model; using the generated image data to augment the training set; using the augmented training set to train the AlexNet classification model; evaluating the performance of the AlexNet classification model on the validation set, recording the classification accuracy; and calculating the average classification accuracy of the five validation runs as the final evaluation metric.

[0145] The data was padded at 20%, 40%, 60%, 80%, and 100% of the original dataset size. The experimental results are shown in Table 3:

[0146] Table 3. MCDRAGAN augmented experimental data table

[0147]

[0148] Analyzing the data in the table, we can see that MCDRAGAN is effective on small-scale datasets, and the proposed model achieves excellent performance in all evaluation metrics. Accuracy reaches 98.23%, indicating that the model has high reliability in overall sample classification. Precision reaches a maximum of 97.57%, indicating that the samples obtained by this model retain the information of the original samples, making it easier for the model to distinguish between positive and negative samples. Recall is 97.10%, indicating that the generated samples cover a wider range of sample features, enabling the model to better capture positive samples. The F1-score is 97.33%, further demonstrating the model's ability to strike a balance between Precision and Recall, reflecting the stability and robustness of its classification results. At the same time, the recognition accuracy of various evaluation metrics improves with increasing augmentation ratio. When the augmentation ratio increases from 0 to 20%, the Accuracy value increases the most. Compared with the augmentation ratio of 0, when the augmentation ratio reaches 100%, the growth rate of Accuracy and other metrics is around 20%. Therefore, it is very necessary to address the problem of data scarcity through data augmentation models.

Claims

1. A DAS data augmentation method based on an improved DRAGAN, characterized by: The following steps are involved: S1: Build a DAS perimeter security system, collect data and pre-process it; S2: Improve the DRAGAN data enhancement model. The generator introduces the EnhancedMCAM module, and the discriminator introduces the residual block and CBAM module to form the MCDRAGAN data enhancement model. S3: Data augmentation model training with adaptive learning rate adjustment strategy; S4: Input the preprocessed data into the MCDRAGAN data augmentation model to obtain generated samples; S5: Use the AlexNet classification model to evaluate the MCDRAGAN data augmentation model and verify the effectiveness of the data augmentation model; In S2, the generator introduces the EnhancedMCAM module, which is obtained by making the following improvements to the existing MCAM module: (1) replacing a single convolutional layer and using convolution kernels of different sizes to extract multi-scale features; (2) Replace the single self-attention mechanism with a multi-head self-attention mechanism; (3) Introduce a channel attention mechanism to dynamically adjust channel weights through global average pooling and two fully connected layers; (4) Introduce a learnable gating mechanism to control the strength of residual connections; and insert EnhancedMCAM modules at different levels of the generator; In S2, the residual block and CBAM module are introduced into the discriminator. Specifically, the residual block alleviates the gradient vanishing problem in deep networks by introducing residual connections, enabling the network to learn deeper features. The CBAM module enhances the expressiveness of features through channel attention and spatial attention mechanisms; In S2, the network structures of the generator and discriminator of the constructed MCDRAGAN data enhancement model are as follows: Generator construction: Input noise generates preliminary features through a fully connected layer; preliminary features are upsampled through a transposed convolution layer and inserted into an EnhancedMCAM module for multi-scale feature extraction and multi-head self-attention calculation; features are upsampled through a transposed convolution layer and inserted into an EnhancedMCAM module for channel attention and gating mechanism calculation; features are upsampled through a transposed convolution layer to finally generate an image; Discriminator construction: The input image first passes through an initial convolutional layer to extract preliminary features. Next, the features pass through three residual blocks, each followed by a CBAM module. The residual blocks enhance feature transfer through residual connections, and the CBAM module enhances the expressiveness of features through channel and spatial attention mechanisms. After the last residual block, the features pass through a multi-scale feature extraction module to extract feature information at different scales. Finally, the features pass through a final convolutional layer, outputting a feature vector. This feature vector is concatenated with the feature vector output by the multi-scale feature extraction module, and then classified through a fully connected layer to output the final discrimination result.

2. The DAS data augmentation method based on the improved DRAGAN according to claim 1, characterized in that: The specific steps of S1 are as follows: A meshed optical cable method is used, that is, horizontal and vertical cables are laid alternately on the fence, forming a three-bend loop on each section of the fence for full coverage. After the optical fiber is laid, different types of DAS data are collected. An IIR filter is used to denoise the collected raw signals, and the differential averaging method is used to compress the data. After that, the data obtained after preprocessing is selected in the time domain and segmented into 4-second time units. The obtained data and its corresponding labels constitute the original data set.

3. The DAS data augmentation method based on the improved DRAGAN according to claim 1, characterized in that: The specific steps of S3 are as follows: S3.1: The model ensures gradient flow by constraining the Lipschitz continuity of the discriminator; the discriminator must satisfy the 1-Lipschitz condition, requiring that for any point in the input space x , the gradient of the discriminator function at this point must be less than or equal to 1; when the data distribution generated by the generator p g With the real data distribution p data The optimal discriminator is obtained in theory. Under the Wasserestein framework, the optimal discriminator D The gradient norm of * approaches 1 at the equilibrium point; the optimization goal of the generator is to minimize the Wasserstein distance as much as possible. By introducing a gradient penalty term and imposing constraints, the update direction of the generator is made more stable. S3.2: During model training, the generator and discriminator are trained alternately to update their respective parameters. First, the generator is fixed to train the discriminator. At this stage, the parameters of the generator are fixed, and the discriminator receives real data. x and generate data G ( z ) for training; then fix the discriminator to train the generator. At this stage, the parameters of the discriminator are fixed, and the generator receives random noise. z Conduct training; Repeat the above steps to train the MCDRAGAN data enhancement model until the Nash equilibrium is reached. G ( z ) and real data x Indistinguishable from the perspective of the discriminator; S3.3: In the initial stage of training, the CosineAnnealingLR scheduler is used to dynamically adjust the generator's learning rate based on the shape of the cosine function. This allows the generator to quickly adapt to the discriminator's capabilities in the early stages of training. At the same time, a low learning rate is used for fine-tuning in the later stages of training to avoid overfitting. At the same time, the model is trained using the RMSprop optimizer. The idea is to use the moving average of the squared gradient to dynamically adjust the learning step size of each parameter. The update process is described as follows: first, calculate the current gradient; then, calculate an exponentially decaying moving average of the squared gradient; next, divide the current gradient by the square root of the moving average of the squared gradient plus a very small constant; finally, update the model parameters by multiplying the scaled gradient by the learning rate; among them, the decay rate is taken as 0.99, and the very small constant is added to prevent division by zero errors during the calculation process.

4. The DAS data augmentation method based on the improved DRAGAN according to claim 1, characterized in that: The specific steps of S4 are as follows: The preprocessed data is input into the MCDRAGAN data enhancement model, where the generator generates data similar to the real data. x Generated data G ( z ) to deceive the discriminator until the discriminator cannot distinguish between real samples and generated samples, thereby generating more realistic and diverse samples; the mean absolute error MAE, root mean square error RMSE and initial score IS indicators are used to evaluate the quality of the generated samples.

5. The DAS data augmentation method based on the improved DRAGAN according to claim 1, characterized in that: The specific steps of S5 are as follows: The original training set is divided into five equal parts, each of which is used as a validation set, and the remaining four parts are used as training sets. For each set of training sets and validation sets, the following steps are repeated: use the training set to train the MCDRAGAN model; use the generated image data to enhance the training set; Use the enhanced training set to train the AlexNet classification model; The performance of the AlexNet classification model was evaluated on the validation set, and the classification accuracy was recorded. The average classification accuracy of the five validations was calculated as the final evaluation indicator. At the same time, the data was padded at 20%, 40%, 60%, 80%, and 100% of the original dataset size to verify the necessity of data augmentation.

Images