Lightweight fiber optic vibration intrusion event identification method and system for perimeter security

Abstract

The application discloses a light-weighted optical fiber vibration intrusion event identification method and system for perimeter security, belongs to the field of optical fiber sensing technology and pattern recognition technology, and comprises the following steps: collecting original vibration signals of a perimeter monitoring area through a distributed optical fiber vibration sensing system; performing wavelet threshold denoising pretreatment on the original vibration signals to obtain denoised signals; performing feature extraction on the denoised signals to construct a high-dimensional feature vector; performing dimension reduction processing on the high-dimensional feature vector by using a linear discriminant analysis method to obtain a low-dimensional classification feature vector; inputting the low-dimensional classification feature vector into a pre-trained light-weighted convolutional neural network model for classification and recognition, and outputting corresponding intrusion event categories. The application cooperates the front-end LDA dimension reduction with the rear-end light-weighted network, greatly reduces the model parameter quantity and the calculation cost while ensuring high recognition accuracy, realizes efficient and real-time deployment on an edge device, and is suitable for intrusion detection in perimeter security.

Description

TECHNICAL FIELD

[0001] The present application relates to a lightweight optical fiber vibration intrusion event recognition method and system for perimeter security, belonging to the field of optical fiber sensing technology and pattern recognition technology. BACKGROUND

[0002] With the increasing demand for security in society, perimeter security systems play an increasingly important role in the protection of critical infrastructure. Distributed fiber optic vibration sensing system (DVS) has become an important technical means in the field of perimeter security due to its advantages such as long-distance monitoring, anti-electromagnetic interference, no power supply, and good concealment. The system detects the change of backscattering Rayleigh scattering light phase or intensity in the optical fiber caused by external intrusion behavior, and realizes the distributed sensing and positioning of vibration signals.

[0003] However, when applying the DVS system to actual perimeter security, it faces the challenge of insufficient adaptability to complex environments. Background noise, climate interference, animal activity, etc. will produce a large number of invalid alarms, while the characteristics of real intrusion signals (such as climbing, digging, walking) are different. The traditional threshold-based or simple signal processing method has high false alarm rate and low recognition rate. In recent years, deep learning models have been introduced to improve recognition accuracy, but they usually have high computational complexity and large number of parameters, requiring powerful computing platforms to support, resulting in high system power consumption, large delay, and high cost, making it difficult to deploy in resource-constrained edge or embedded devices, limiting the widespread application and productization of the technology.

[0004] Therefore, there is an urgent need for a lightweight optical fiber vibration intrusion event recognition scheme that can significantly reduce computational complexity and resource consumption while ensuring high recognition accuracy, suitable for edge deployment. SUMMARY

[0005] To solve the above problems, the present application provides a lightweight optical fiber vibration intrusion event recognition method and system for perimeter security.

[0006] The technical scheme adopted by the present application to solve its technical problems is:

[0007] In a first aspect, the present application provides a lightweight optical fiber vibration intrusion event recognition method for perimeter security, comprising the following steps:

[0008] Step S1, acquiring the original vibration signal of the perimeter monitoring area by the distributed fiber optic vibration sensing system;

[0009] Step S2, wavelet threshold denoising (WTD) preprocessing is performed on the original vibration signal to obtain a denoised signal;

[0010] Step S3, feature extraction is performed on the denoised signal to construct a high-dimensional feature vector;

[0011] Step S4, linear discriminant analysis (LDA) is used to reduce the dimension of the high-dimensional feature vector to obtain a low-dimensional classification feature vector;

[0012] Step S5, the low-dimensional classification feature vector is input into a pre-trained lightweight convolutional neural network model for classification and recognition, and the corresponding intrusion event category is output; wherein the lightweight convolutional neural network model is a MobileNet-V4 (MNV4) model.

[0013] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system detects vibration signals by monitoring the phase change of backscattered Rayleigh light in the sensing optical fiber based on the principle of phase-sensitive optical time domain reflectometry.

[0014] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system comprises:

[0015] A narrow linewidth laser for generating continuous laser light;

[0016] An acousto-optic modulator for modulating the continuous laser light into pulsed light;

[0017] An erbium-doped fiber amplifier for amplifying the pulsed light signal;

[0018] A circulator for injecting the amplified pulsed light into the sensing optical fiber and guiding the backscattered Rayleigh light to the detection end;

[0019] A photodetector for converting the backscattered Rayleigh light into an electrical signal;

[0020] A data acquisition card for converting the electrical signal into a digital signal to form the original vibration signal.

[0021] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system collects the original vibration signal at a sampling frequency of no less than 1 kHz, and divides the continuous signal stream into multiple signal segments according to a preset time window length for subsequent processing.

[0022] As a possible implementation manner of the embodiment, the perimeter monitoring area includes a fence, a pipeline, a bridge or a railway line, and the sensing optical fiber is laid along the perimeter.

[0023] As a possible implementation manner of the embodiment, the step S2 comprises the following steps:

[0024] Step S21, selecting a wavelet base function and a decomposition layer number, and performing multi-scale wavelet decomposition on the original vibration signal to obtain wavelet coefficients of each layer;

[0025] Step S22, performing threshold processing on the wavelet coefficients, and setting coefficients less than a set threshold to zero or shrinking to filter out coefficients corresponding to noise;

[0026] Step S23, performing wavelet reconstruction using the threshold-processed coefficients to obtain the denoised signal to retain the transient mutation characteristics of the signal.

[0027] As a possible implementation manner of the embodiment, the wavelet base function is one of Symlets, Daubechies or Coiflets series wavelets.

[0028] As a possible implementation manner of the embodiment, the threshold processing adopts a soft threshold function or a hard threshold function, and the threshold is adaptively determined according to the signal noise level by a general threshold method or a minimax criterion.

[0029] As a possible implementation manner of the embodiment, the step S3 comprises the following steps:

[0030] Step S31, converting the denoised one-dimensional vibration signal into a two-dimensional image tensor representation by a signal reconstruction and channel mapping method;

[0031] Step S32, flattening the two-dimensional image tensor to obtain a first feature vector;

[0032] Step S33, extracting time domain features, frequency domain features and time-frequency domain features from the denoised signal to obtain a second feature vector;

[0033] Step S34, fusing the first feature vector and the second feature vector to form a high-dimensional feature vector.

[0034] As a possible implementation manner of the embodiment, the time domain features include one or more of signal mean, variance, root mean square, peak factor, kurtosis and pulse factor; the frequency domain features include one or more of spectral centroid, spectral variance, spectral entropy and specific frequency band energy proportion; and the time-frequency domain features include sub-band energies obtained by wavelet threshold denoising or wavelet packet decomposition.

[0035] As a possible implementation manner of the embodiment, after constructing the high-dimensional feature vector, the method further comprises a step of normalizing or normalizing the high-dimensional feature vector.

[0036] As a possible implementation manner of the embodiment, the step S4 comprises the following steps:

[0037] A step S41 is included, in which the within-class scatter matrix and the between-class scatter matrix of the high-dimensional feature vectors of all training samples are calculated.

[0038] A step S42 is included, in which a generalized eigenvalue equation is solved to obtain a projection vector that maximizes the between-class difference and minimizes the within-class difference.

[0039] A step S43 is included, in which the eigenvectors corresponding to the first k-1 largest generalized eigenvalues are selected to form a projection matrix, where k is the total number of intrusion event categories.

[0040] A step S44 is included, in which the high-dimensional feature vectors are linearly transformed by using the projection matrix to obtain low-dimensional classification feature vectors with a dimension of k-1.

[0041] As a possible implementation manner of the embodiment, before the step S41 is performed, a principal component analysis method is used to pre-dimensionally reduce the high-dimensional feature vectors to reduce the feature dimension and ensure the invertibility of the within-class scatter matrix.

[0042] As a possible implementation manner of the embodiment, the intrusion event categories include background noise (no intrusion event), high-frequency strong vibration signals caused by mechanical operation, and low-frequency weak vibration signals caused by personnel walking; the total number of intrusion event categories is 3, and the optimal dimension reduction dimension is 2.

[0043] As a possible implementation manner of the embodiment, the objective function for dimensionally reducing the high-dimensional feature vectors is to maximize the ratio of the trace of the between-class scatter matrix to the trace of the within-class scatter matrix.

[0044] As a possible implementation manner of the embodiment, the lightweight convolutional neural network model is a MobileNet-V4 model, and the step S5 comprises the following steps:

[0045] A step S51 is included, in which the low-dimensional classification feature vectors are converted into a multi-dimensional tensor format suitable for input of the MobileNet-V4 model through dimension expansion and reshaping processing;

[0046] A step S52 is included, in which the converted tensor is input into the MobileNet-V4 model for feature extraction and classification.

[0047] A step S53 is included, in which the corresponding intrusion event category is determined according to the output of the MobileNet-V4 model.

[0048] As a possible implementation manner of the embodiment, the MobileNet-V4 model comprises a general inverted bottleneck module and a mobile multi-query attention module, which are used for efficient calculation and key information focusing on the input features.

[0049] As a possible implementation manner of the embodiment, before the low-dimensional classification feature vector is input into the MobileNet-V4 model, a feature remodeling step is further included: the one-dimensional low-dimensional classification feature vector is reconstructed into a multi-dimensional tensor to adapt to the input structure of the MobileNet-V4 model.

[0050] As a possible implementation manner of the embodiment, the lightweight convolutional neural network model is a model pre-trained by the following steps:

[0051] A training sample set containing labeled categories is obtained;

[0052] The training samples are sequentially processed through steps S1 to S4 to obtain corresponding low-dimensional classification feature vectors and labels;

[0053] The MobileNet-V4 network is supervised trained with the low-dimensional classification feature vectors as input and the corresponding labels as supervision signals until the model converges.

[0054] As a possible implementation manner of the embodiment, before step S5, a model training step is further included: the overall model comprising the linear discriminant analysis dimension reduction layer and the MobileNet-V4 classification network is supervised trained by using the training sample set containing labeled categories to obtain the trained lightweight convolutional neural network model.

[0055] In a second aspect, the embodiment of the present application provides a lightweight optical fiber vibration intrusion event recognition system for perimeter security, which is used to implement the method described above, and the system comprises:

[0056] A signal acquisition module is configured to acquire original vibration signals by using a distributed optical fiber vibration sensing system.

[0057] A signal preprocessing module is configured to perform wavelet threshold denoising on the original vibration signals.

[0058] A feature extraction and dimension reduction module is configured to extract signal features and perform dimension reduction by using a linear discriminant analysis method.

[0059] A classification and recognition module is configured to have a trained lightweight convolutional neural network model built-in, and is configured to perform classification and recognition on the features after dimension reduction.

[0060] An output module is configured to output an intrusion event category and alarm information.

[0061] One of the technical solutions has the following advantages or beneficial effects:

[0062] 1. The signal quality is enhanced by wavelet threshold denoising, supervised dimension reduction is performed by linear discriminant analysis, low-dimensional features most conducive to classification are extracted, and a high-performance lightweight network MNV4 is combined to jointly ensure high-accuracy recognition of various intrusion events in a high-noise complex environment.

[0063] 2. The LDA front-end dimension reduction compresses the feature dimension from hundreds of thousands to a single digit (such as 2 dimensions), greatly reducing the computational pressure of the rear-end network. Combined with the efficient architecture of MNV4 itself, the overall model parameter quantity can be reduced by more than 99%, and the inference time on a general CPU is reduced to the millisecond level, meeting the real-time requirement.

[0064] 3. The extremely low computational overhead and memory occupation enable the technical solution to be easily deployed on resource-constrained edge devices or embedded platforms, realize data processing in place, reduce the bandwidth dependence on the central server and the overall power consumption of the system, and improve the practicability and economy of the system.

[0065] 4. The combination of WTD in the signal processing field, LDA in the pattern recognition field, and lightweight CNN in the deep learning field forms an end-to-end processing framework of "feature enhancement-supervised dimension reduction-efficient classification", providing a new technical path for optical fiber vibration signal processing. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a flowchart of a lightweight optical fiber vibration intrusion event recognition method for perimeter security according to an example embodiment;

[0067] Figure 2 is a schematic diagram of a lightweight optical fiber vibration intrusion event recognition system for perimeter security according to an example embodiment;

[0068] Figure 3 is a flowchart of an LDA-MNV4 algorithm according to an example embodiment;

[0069] Figure 4 is a data dimension reduction flowchart according to an example embodiment;

[0070] Figure 5 is a schematic diagram of a projection vector solution according to an example embodiment;

[0071] Figure 6 is a feature extraction and classification flowchart according to an example embodiment;

[0072] Figure 7A module structure diagram of an initial convolution block and a feature extraction layer is shown according to an example embodiment;

[0073] Figure 8 A DVS system structure schematic diagram is shown according to an example embodiment;

[0074] Figure 9 A schematic diagram of three collected signals is shown according to an example embodiment, Figure 9 Among them, (a) is a background noise (no intrusion event) signal schematic diagram, (b) is a mechanical operation high-frequency strong vibration signal schematic diagram, and (c) is a personnel walking low-frequency weak vibration signal schematic diagram);

[0075] Figure 10 A comparison diagram of time domain waveform diagrams before and after wavelet denoising of three events of background noise, mechanical excavation and personnel walking is shown according to an example embodiment, Figure 10 Among them, (a) is a wavelet denoising before waveform diagram of background noise, (b) is a wavelet denoising after waveform diagram of background noise, (c) is a wavelet denoising before waveform diagram of mechanical excavation, (d) is a wavelet denoising after waveform diagram of mechanical excavation, (e) is a wavelet denoising before waveform diagram of personnel walking, and (f) is a wavelet denoising after waveform diagram of personnel walking);

[0076] Figure 11 A data dimension reduction UMAP result diagram is shown according to an example embodiment, Figure 11 Among them, (a) is an original data distribution diagram, and (b) is a data distribution diagram after LDA projection);

[0077] Figure 12 A comparison diagram of inter-class distances before and after dimension reduction is shown according to an example embodiment;

[0078] Figure 13 An LDA-MNV4 recognition result schematic diagram is shown according to an example embodiment;

[0079] Figure 14 An MNV4 recognition result schematic diagram is shown according to an example embodiment. DETAILED DESCRIPTION

[0080] In order to more clearly illustrate the technical features of the scheme of the present application, the present application will be described in detail below with reference to specific embodiments and accompanying drawings.

[0081] As shown in Figure 1 The light-weight optical fiber vibration intrusion event recognition method for perimeter security provided by the embodiment of the present application comprises the following steps:

[0082] Step S1, collecting original vibration signals of a perimeter monitoring area through a distributed optical fiber vibration sensing system;

[0083] Step S2, wavelet threshold denoising preprocessing is performed on the original vibration signal to obtain a denoised signal;

[0084] Step S3, feature extraction is performed on the denoised signal to construct a high-dimensional feature vector;

[0085] Step S4, linear discriminant analysis is used to perform dimension reduction processing on the high-dimensional feature vector to obtain a low-dimensional classification feature vector;

[0086] Step S5, the low-dimensional classification feature vector is input into a pre-trained lightweight convolutional neural network model for classification and recognition, and the corresponding intrusion event category is output; wherein the lightweight convolutional neural network model is a MobileNet-V4 model.

[0087] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system detects the vibration signal by monitoring the phase change of the backscattered Rayleigh light in the sensing optical fiber based on the principle of phase-sensitive optical time domain reflectometry.

[0088] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system comprises:

[0089] a narrow linewidth laser for generating continuous laser;

[0090] an acousto-optic modulator for modulating the continuous laser into pulsed light;

[0091] an erbium-doped fiber amplifier for amplifying the pulsed light signal;

[0092] a circulator for injecting the amplified pulsed light into the sensing optical fiber and guiding the backscattered Rayleigh light to the detection end;

[0093] a photodetector for converting the backscattered Rayleigh light into an electrical signal;

[0094] a data acquisition card for converting the electrical signal into a digital signal to form the original vibration signal.

[0095] As a possible implementation manner of the embodiment, the distributed optical fiber vibration sensing system collects the original vibration signal at a sampling frequency of not less than 1 kHz, and divides the continuous signal stream into multiple signal segments according to a preset time window length for subsequent processing.

[0096] As a possible implementation manner of the embodiment, the original vibration signal is a one-dimensional time sequence signal containing time domain waveform and position information of vibration event occurrence. The collected original vibration signal includes background noise (no intrusion event) signal, high frequency strong vibration signal generated by mechanical operation, and low frequency weak vibration signal generated by personnel walking. The continuously collected original vibration signal is segmented according to a preset time window length, and each segment of signal is taken as an independent sample for subsequent processing.

[0097] As a possible implementation manner of the embodiment, the perimeter monitoring area includes a railway, a highway, a bridge, a pipeline or a fence along a line, and the sensing optical fiber is laid along the perimeter.

[0098] As a possible implementation manner of the embodiment, the step S2 includes the following steps.

[0099] Step S21, selecting a wavelet base function and a decomposition layer number, and performing multi-scale wavelet decomposition on the original vibration signal to obtain wavelet coefficients of each layer.

[0100] Step S22, performing threshold processing on the wavelet coefficients, and setting or shrinking the coefficients less than a set threshold to zero to filter out the coefficients corresponding to noise.

[0101] Step S23, performing wavelet reconstruction by using the threshold-processed coefficients to obtain the denoised signal to retain the transient mutation characteristics of the signal.

[0102] As a possible implementation manner of the embodiment, the wavelet base function is one of Symlets, Daubechies or Coiflets series wavelets.

[0103] As a possible implementation manner of the embodiment, the threshold processing adopts a soft threshold function or a hard threshold function, and the threshold is adaptively determined according to the signal noise level by a universal threshold method or a minimax criterion.

[0104] The soft threshold function is:

[0105] ,

[0106] ,

[0107] wherein, is a wavelet coefficient, is a threshold, wherein is a noise standard deviation estimate, is a signal length.

[0108] The wavelet threshold denoising preprocessing is suitable for analyzing non-stationary vibration signals, and can effectively filter out noise while retaining the transient mutation information of the signal, in view of the characteristics of the original vibration signal being a non-stationary noisy signal. The wavelet threshold denoising preprocessing as a feature enhancement means improves the accuracy and robustness of the subsequent feature extraction and classification steps on the premise of retaining the essential features of the signal.

[0109] The denoised signal obtained in step S23 can effectively highlight the core features of different types of events:

[0110] For background noise (no intrusion event), the denoised signal presents a smooth and low-amplitude waveform.

[0111] For high-frequency strong vibration events of mechanical operation, the denoised signal clearly presents the characteristics of "pre-stable and post-burst strong vibration";

[0112] For personnel walking low-frequency weak vibration events, the denoised signal presents a periodic and intermittent fluctuation pattern.

[0113] As a possible implementation manner of the embodiment, the step S3 includes the following steps:

[0114] Step S31: converting the denoised one-dimensional vibration signal into a two-dimensional image tensor representation by a signal reconstruction and channel mapping method, the two-dimensional image tensor having a dimension of 224x224 and conforming to an RGB image format standard to be compatible with the standard input structure of the subsequent MobileNet-V4 model;

[0115] Step S32: flattening the two-dimensional image tensor to obtain a first feature vector;

[0116] Step S33: extracting time domain features, frequency domain features and time-frequency domain features from the denoised signal to obtain a second feature vector;

[0117] Step S34: fusing the first feature vector and the second feature vector to form a high-dimensional feature vector, the fusion being vector splicing; the high-dimensional feature vector having a dimension of 224x224x3=150528, and the high-dimensional feature vector being used as an input of a subsequent linear discriminant analysis module for supervised dimension reduction processing. The high-dimensional feature vector directly reflects the time-frequency distribution characteristics of the original vibration signal, and provides rich distinguishing information for subsequent supervised dimension reduction.

[0118] As one possible implementation of this embodiment, the time-domain features include one or more of the following: signal mean, variance, root mean square, peak factor, kurtosis, and impulse factor; the frequency-domain features include one or more of the following: spectral centroid, spectral variance, spectral entropy, and energy percentage of a specific frequency band; and the time-frequency domain features include the energy of each sub-band obtained by wavelet threshold denoising or wavelet packet decomposition.

[0119] As one possible implementation of this embodiment, after constructing the high-dimensional feature vector, the method further includes a step of standardizing or normalizing the high-dimensional feature vector.

[0120] As one possible implementation of this embodiment, step S4 includes the following steps:

[0121] Step S41: Calculate the intra-class scatter matrix and inter-class scatter matrix of the high-dimensional feature vectors of all training samples;

[0122] Step S42: Solve the generalized eigenvalue equation to obtain the projection vector that maximizes the inter-class difference and minimizes the intra-class difference;

[0123] Step S43: Select the eigenvectors corresponding to the first k-1 largest generalized eigenvalues ​​to form a projection matrix, where k is the total number of intrusion event categories;

[0124] Step S44: Use the projection matrix to perform a linear transformation on the high-dimensional feature vector to obtain the low-dimensional classification feature vector with dimension k-1.

[0125] The intra-class scatter matrix is ​​used to measure the degree of dispersion of data within each class, and its calculation formula is as follows:

[0126] ,

[0127] ,

[0128] in, It is the within-class scatter matrix. It is the first The scatter matrix of the class, The total number of categories, Indicates the first A sample set of classes It is a sample vector in this set. Indicates the first The sample mean of the class of samples.

[0129] The inter-class scatter matrix is ​​used to measure the degree of dispersion of data between different classes, and its calculation formula is as follows:

[0130] ,

[0131] wherein, is the between-class scatter matrix, is the number of samples in the i-th class, represents the total mean vector of all samples.

[0132] The generalized eigenvalue equation is:

[0133] ,

[0134] wherein, is the eigenvalue, is the projection matrix.

[0135] The optimal projection matrix is composed of eigenvectors corresponding to the first Kmaxmax eigenvalues, wherein Kmaxis the number of selected maximum eigenvalues, and K is the target dimension after dimension reduction.

[0136] As a possible implementation manner of the embodiment, before step S41 is performed, the high-dimensional feature vector is pre-reduced by using a principal component analysis method, so as to reduce the feature dimension and ensure the invertibility of the within-class scatter matrix.

[0137] The dimension of the low-dimensional classification feature vector satisfies wherein K is the total number of intrusion event classes.

[0138] As a possible implementation manner of the embodiment, the intrusion event classes include background noise (no intrusion event), high-frequency strong vibration signals caused by mechanical operation, and low-frequency weak vibration signals caused by personnel walking; the total number of intrusion event classes is 3, and the optimal dimension reduction dimension of the low-dimensional classification feature vector is 2.

[0139] As a possible implementation manner of the embodiment, the target function for performing dimension reduction processing on the high-dimensional feature vector is to maximize the ratio of the trace of the between-class scatter matrix and the trace of the within-class scatter matrix:

[0140] ,

[0141] wherein, is the target function, is the within-class scatter matrix, is the between-class scatter matrix, is the projection matrix.

[0142] ​​​The linear discriminant analysis method is a supervised linear dimension reduction method, and the core thereof is to find an optimal projection direction by calculating an intra-class scatter matrix and an inter-class scatter matrix of data, and map high-dimensional data to a low-dimensional space, so that the mapped data satisfies the maximum inter-class difference and the minimum intra-class difference.

[0143] As a possible implementation manner of the embodiment, the lightweight convolutional neural network model is a MobileNet-V4 model, and the step S5 includes the following steps.

[0144] In step S51, the low-dimensional classification feature vector is converted into a multi-dimensional tensor format suitable for input of the MobileNet-V4 model through dimension expansion and reshaping processing.

[0145] In step S52, the converted tensor is input into the MobileNet-V4 model for feature extraction and classification.

[0146] In step S53, the corresponding intrusion event category is determined according to the output of the MobileNet-V4 model.

[0147] The dimension expansion and reshaping processing includes: expanding the low-dimensional classification feature vector into a one-dimensional vector of a predetermined length through interpolation, repetition or splicing with zero values, then reshaping the one-dimensional vector into a two-dimensional matrix, and copying the two-dimensional matrix to three channels to form a three-dimensional tensor. Since the lightweight convolutional neural network model generally accepts a two-dimensional or three-dimensional tensor as input (for example, an image format), and the low-dimensional classification feature vector obtained by LDA dimension reduction is a one-dimensional vector, dimension conversion is required. In an embodiment, the dimension conversion includes: expanding the low-dimensional classification feature vector (for example, 2 dimensions) into a one-dimensional vector (for example, 224*224=50176 dimensions) of a predetermined length through interpolation, repetition or splicing with zero values, then reshaping the one-dimensional vector into a two-dimensional matrix (for example, 224*224), and copying the two-dimensional matrix to three channels to form a three-dimensional tensor (for example, 3*224*224). The three-dimensional tensor can be used as input of the lightweight convolutional neural network model.

[0148] As a possible implementation manner of the embodiment, the MobileNet-V4 model comprises a general inverted bottleneck module and a mobile multi-query attention module, and through the cooperative design of depth separable convolution, inverse residual connection and attention mechanism, the calculation complexity and parameter quantity are significantly reduced while ensuring the feature extraction capability. The general inverted bottleneck module balances the feature extraction performance and calculation efficiency according to the hardware resources and task requirements by dynamically selecting the convolution kernel size and the number of layers; the mobile multi-query attention module maintains the high resolution of the query while reducing the memory bandwidth requirement by sharing the key and value and spatially downsampling the key and value, thereby realizing efficient key feature focusing. The MobileNet-V4 model is an MNV4ConvSmall architecture, the input size is 3x224x224, and the parameter quantity is 2.38M.

[0149] The feature extraction layer comprises a combination of multiple general inverted bottleneck modules and mobile multi-query attention modules, which are used for deep feature extraction and semantic mapping of the low-dimensional classification feature vector.

[0150] As a possible implementation manner of the embodiment, before the low-dimensional classification feature vector is input into the MobileNet-V4 model, a feature remodeling step is further included, in which the one-dimensional low-dimensional classification feature vector is reconstructed into a multi-dimensional tensor to adapt to the input structure of the MobileNet-V4 model.

[0151] As a possible implementation manner of the embodiment, before step S5, a model training step is further included, in which the overall model comprising the linear discriminant analysis dimension reduction layer and the MobileNet-V4 classification network is supervised trained by using a training sample set with labeled categories, thereby obtaining the trained lightweight convolutional neural network model.

[0152] As a possible implementation manner of the embodiment, the model training step specifically comprises:

[0153] A training sample set comprising labeled categories is acquired, and the training set comprises multiple types of intrusion event signals;

[0154] The training set signals are sequentially subjected to wavelet transform denoising, feature extraction and linear discriminant analysis dimension reduction, thereby obtaining low-dimensional classification feature vector training samples;

[0155] The low-dimensional classification feature vector is taken as the input, and the corresponding event category label is taken as the supervision signal, and the AdamW optimizer and the CosineAnnealingLR learning rate scheduler are used for training;

[0156] The cross-entropy loss function is used in the training process, and the model is iteratively trained until convergence, and the best model is saved, and the initial learning rate is set to 0.001, the weight decay coefficient is set to 1e-4, and the training iteration number is 100.

[0157] As shown in Figure 2 The embodiment of the application provides a light-weight optical fiber vibration intrusion event identification system for perimeter security, which comprises:

[0158] A signal acquisition module is configured to acquire original vibration signals through a distributed optical fiber vibration sensing system.

[0159] A signal preprocessing module is configured to perform wavelet threshold denoising on the original vibration signals.

[0160] A feature extraction and dimension reduction module is configured to extract signal features and perform dimension reduction by using a linear discriminant analysis method.

[0161] A classification and identification module is internally provided with a trained MobileNet-V4 model and is configured to classify and identify the features after dimension reduction.

[0162] An output module is configured to output an intrusion event category and alarm information.

[0163] The application provides a WTD-LDA-MNV4 method combining wavelet transform denoising (WTD), linear discriminant analysis (LDA) and MobileNet-V4 model, which is applied to DVS for perimeter security intrusion detection. The high-dimensional feature dimension reduction capability of LDA and the lightweight and efficient feature extraction capability of MNV4 are organically combined to form an end-to-end processing framework of "feature enhancement-dimension reduction-classification", which significantly reduces the calculation overhead while ensuring the recognition accuracy, and adapts to the hardware constraints of edge devices. The method first acquires optical fiber vibration original signals by DVS, extracts essential features after preprocessing, converts high-dimensional data into low-dimensional feature vectors, and finally completes pattern classification through a classifier, thereby providing a new effective solution for perimeter security.

[0164] The application mainly relates to the following key technologies.

[0165] I. LDA-MNV4 innovative algorithm design

[0166] The core logic of the LDA-MNV4 algorithm is to realize lightweight compression of high-dimensional data in the feature processing front end through LDA, and to retain key classification information. The efficient network structure of MNV4 is used to perform deep extraction and pattern classification on the low-dimensional features after dimension reduction, and the two are optimized in coordination to balance "feature retention" and "computing efficiency".

[0167] As shown in Figure 3As shown, the original vibration signal collected by the DVS system is first input, and after preprocessing, it is converted into a high-dimensional feature vector; then LDA is used to reduce the dimension of the high-dimensional features, and through maximizing the inter-class difference and minimizing the intra-class difference, the high-dimensional data is mapped to a low-dimensional space to retain the core classification features; the MNV4 network processes the low-dimensional features after dimension reduction, and through the feature enhancement of the UIB and the key feature focusing of the Mobile MQA, the final output of the classification layer is the intrusion event category.

[0168] II. Front-end LDA data dimension reduction:

[0169] LDA is a linear dimension reduction model based on statistical methods. By calculating the inter-class scatter and intra-class scatter of the data, the optimal projection direction is found to map high-dimensional data to low-dimensional space, so that the mapped data maximizes the inter-class difference and minimizes the intra-class difference, thereby enhancing the inter-class discrimination. As the front-end feature processing module of the algorithm, it is responsible for solving the problem of high computational burden caused by high dimensionality of the original vibration signal features. The core is to achieve "dimension compression while preserving classification information" through linear transformation.

[0170] As shown in Figure 4 , the input two-dimensional tensor is flattened into a one-dimensional vector through preprocessing. Since the feature dimension of the flattened original image is extremely high, directly using LDA may lead to computational complexity or intra-class scatter matrix singularity, so PCA is used to pre-reduce to 10 to reduce the computational load and avoid numerical stability problems when LDA processes high-dimensional data. Then LDA core dimension reduction is performed, i.e., the optimal projection matrix is solved to project high-dimensional data to low-dimensional space. Through this dimension reduction process, the high-dimensional image features are gradually reduced from 150528 dimensions to 2 dimensions, the purpose being to simplify the data while preserving the class distinction information, thereby enhancing the discrimination of the subsequent classification task.

[0171] Intra-class scatter matrix is used to measure the dispersion of data within each class. For a data set with classes, assume that the sample size of the th class data is , the sample mean is , and the total mean of all samples is . The calculation formula of the intra-class scatter matrix is:

[0172] ,

[0173] ,

[0174] where is the scatter matrix of the th class, represents the A sample set of classes It is a sample vector in this set. Indicates the first i The sample mean of the class of samples. Essentially, it is the sum of the scatter matrices of each category, reflecting the degree of dispersion of data within each category. The smaller the value, the more clustered the data of the same type.

[0175] Inter-class scatter matrix The formula used to measure the dispersion of data between different categories is as follows:

[0176] ,

[0177] This reflects the degree of difference between the mean of each category and the overall mean. The larger the value, the more dispersed the different categories of data.

[0178] The goal of LDA is to find the projection matrix. , high-dimensional data Projected into a lower-dimensional space ,Right now ,like Figure 5 As shown.

[0179] To maximize inter-class differences and minimize intra-class differences, a generalized eigenvalue problem needs to be solved. The objective function is defined as follows:

[0180] ,

[0181] Through the Taking the derivative and setting it to zero, we obtain the generalized eigenvalue equations:

[0182] ,

[0183] here It is an eigenvalue. These are the corresponding feature vectors. Usually, the largest one is chosen. The eigenvectors corresponding to each eigenvalue form the projection matrix. ( The dimension after dimensionality reduction, and In actual calculations, first... Perform the inverse (provided that) If it is reversible (or if it is not reversible, some improvement methods can be used, such as regularization), to transform the generalized eigenvalue problem into a standard eigenvalue problem:

[0184] ,

[0185] Obtain the projection matrix Then, dimensionality reduction can be performed on the original data. Assume the original data matrix... The dimension is ( For the sample size, (The original feature dimension) is projected onto a lower-dimensional space:

[0186] ,

[0187] It is the data matrix after dimensionality reduction, with dimensions of . ( (This refers to the dimension after dimensionality reduction). Thus, by choosing a suitable projection matrix... LDA projects high-dimensional data into a low-dimensional space, preserving the classification information of the data in the process, so that different categories of data can be separated as much as possible in the low-dimensional space.

[0188] LDA's advantage in dimensionality reduction lies in its full utilization of the data's categorical information. Compared to some unsupervised dimensionality reduction methods (such as Principal Component Analysis (PCA), it is more suitable for data processing scenarios with classification requirements. In fields such as image recognition and bioinformatics, LDA is often used to extract key features, reduce data dimensionality, and improve the performance and efficiency of classification models.

[0189] III. Backend MNV4 Feature Extraction and Classification:

[0190] The low-dimensional features after dimensionality reduction by LDA are input into the MNV4 network for deep feature extraction and classification. As the latest iteration of lightweight convolutional neural networks, MNV4 achieves efficient feature learning on edge devices through the innovative design of UIB and Mobile MQA, forming a synergistic effect of "dimensionality reduction-enhancement" with the front-end LDA.

[0191] In the process of deep feature extraction and classification, such as Figure 6 and Figure 7As shown, MNV4 takes "lightweight efficient feature extraction + accurate semantic mapping" as the core: first, a pre-processed 3x224x224 RGB image is received, and after the initial convolution block (conv0), it is down-sampled to 112x112 by 3x3 convolution, the channel is expanded to 32, the low-level features are extracted and the calculation amount is reduced; the feature extraction layer (layer1~layer5) relies on the UIB / Fused_IB block to deepen the features with the inverse residual logic of "1x1 convolution channel expansion → depth separable convolution spatial feature extraction → 1x1 convolution channel compression", and the residual connection is used to prevent gradient disappearance; then the 14x14x1280 features are compressed into 1x1x1280 global features by global average pooling (GAP), and then mapped to the category space (such as 3 classification output 3-dimensional scores) by the fully connected layer. The overall relies on depth separable convolution, inverse residual, and attention to cooperatively extract low-to-high features while reducing the parameter and calculation amount, providing strong discriminative support for classification and balancing performance and efficiency.

[0192] UIB is an improved extension of MobileNetV2 inverted bottleneck (IB), which adds two configurable DW layers to flexibly adjust the structure compared to the standard IB. It supports NAS to dynamically select large core to enhance feature extraction or skip part of the layer to optimize efficiency, and balances performance and resource consumption through modular design.

[0193] Mobile MQA is an attention module optimized for mobile accelerators in MNV4, which improves inference efficiency by improving multi-query attention (MQA). MQA reduces memory bandwidth requirements by sharing keys (Key) and values (Value), and is suitable for the later stage of mobile visual models. At the same time, spatial reduction attention (SRA) strategy is further introduced, using 3x3 depth separable convolution with a step of 2 to down-sample the keys and values, while maintaining the high resolution of the query (Query), reducing the calculation amount while enhancing the model's expression ability.

[0194] MNV4 cooperates with UIB and Mobile MQA to achieve significant efficiency improvement in processing LDA reduced dimensionality features compared to previous generation models: on CPU, the inference speed is 2 times faster than MobileNet V3, on EdgeTPUs, the speed is doubled and the accuracy is equivalent, perfectly matching the resource constraints of edge devices, and forming a "dimensionality reduction load reduction-network speedup" double optimization with the front-end LDA.

[0195] The input layer of the present application: LDA provides "low dimensionality, high discriminability" feature input for MNV4, reducing the network's computational burden (such as parameters and Mult-Adds);

[0196] The characteristic level of the application: the UIB of MNV4 optimizes the feature characteristics (low dimension, strong inter-class difference) of Mobile MQA after LDA dimension reduction, the former strengthens the space / channel correlation of the feature, and the latter focuses on the key vibration feature, making up for the possible loss of details caused by dimension reduction;

[0197] The output level of the application: the recognition results of three types of events are output through the classification layer, and the overall process realizes end-to-end efficient processing of "dimension reduction-extraction-classification" on the edge device, with reasoning time as low as 127ms / sample, while maintaining an identification rate of more than 98%.

[0198] This fusion design of "front-end dimension reduction and lightweight-rear-end network efficiency" enables the algorithm to break through both "power consumption limit" and "precision bottleneck" on the edge device, providing an innovative solution for DVS perimeter security intrusion detection.

[0199] Specific examples:

[0200] The DVS used in the application is a technology based on Phase-Sensitive Optical Time Domain Reflectometry (Φ-OTDR). This technology realizes distributed detection of vibration signals along the line by monitoring the phase change of backscattered Rayleigh light in the optical fiber. Its core principle is: when external intrusion acts on the optical cable, it will produce vibration disturbance, which in turn modulates the phase of the light wave in the optical fiber, and the system can accurately locate the event position by analyzing these modulation signals, such as Figure 8 As shown in the figure, the DVS system structure includes: narrow line width laser NLL, acousto-optic modulator AOM, erbium-doped fiber amplifier EDFA, circulator CIR, measured optical fiber FUT, photodetector PD, data acquisition card DAQ. This configuration ensures that the system has high sensitivity and long distance monitoring capability, making it have outstanding advantages in perimeter security and other fields.

[0201] The system uses a narrow line width laser NLL as the light source, which is modulated into pulsed light by an acousto-optic modulator AOM, and then amplified by an erbium-doped fiber amplifier EDFA. The signal is transmitted to the sensing optical fiber by the circulator CIR to excite backscattered Rayleigh light. After processing, the electrical signal is converted into a digital signal by the data acquisition card DAQ and transmitted to the computer for analysis. After the collected signal is extracted, it is converted into a two-dimensional image tensor, and the recognition result is output after LDA data dimension reduction and lightweight network MNV4 recognition and classification.

[0202] In order to obtain representative monitoring data, experiments were carried out with DVS as the core experimental equipment, and three types of typical event signals were collected in real time in the test site: background noise (no intrusion event), mechanical operation high frequency strong vibration, and personnel walking low frequency weak vibration, as shown in Figure 9 the figure,Figure 9 In the figure, (a) is the schematic diagram of background noise (no intrusion event) signal, (b) is the schematic diagram of mechanical operation high-frequency strong vibration signal, (c) is the schematic diagram of personnel walking low-frequency weak vibration signal. The sampling frequency is set to 2 kHz, each signal segment contains 8192 sampling points, 2000 groups of each event, a total of 6000 groups constitute the database. This database is divided into training set and validation set according to the proportion of 7:3.

[0203] After data collection, the data needs to be preprocessed first, and here the denoising operation is carried out to enhance the signal. Analysis shows that this kind of signal belongs to non-stationary noise signal, and according to the characteristics of significant noise interference, large amplitude span and containing transient mutation information, wavelet denoising relies on multi-resolution analysis, strong adaptability and retention of transient characteristics, and is especially suitable for processing such non-stationary noise signals.

[0204] Figure 10 In the figure, the waveforms before and after wavelet denoising of background noise, mechanical excavation and personnel walking are shown, Figure 10 In the figure, (a) is the waveform before wavelet denoising of background noise, (b) is the waveform after wavelet denoising of background noise, (c) is the waveform before wavelet denoising of mechanical excavation, (d) is the waveform after wavelet denoising of mechanical excavation, (e) is the waveform before wavelet denoising of personnel walking, (f) is the waveform after wavelet denoising of personnel walking. The original waveform of the background event is dense and irregular fluctuation with small amplitude, and the noise interference is significant; the denoised waveform is smoother, maintaining a low amplitude as a whole, which embodies the characteristics of "continuous weak and no obvious burst" of the background signal. The original waveform of the mechanical excavation event is relatively stable in the first half and suddenly appears a large number of dense and large amplitude oscillations in the second half; the waveform profile of the oscillation part after denoising is clearer, highlighting the characteristics of "calm in the early stage and sudden strong vibration in the late stage" of mechanical excavation. The original waveform of the personnel walking event has continuous and interval fluctuations; after denoising, the periodicity and peak-valley profile of the fluctuations are more distinct, showing a "stage fluctuation" pattern, which is consistent with the characteristics of "intermittent and regular generation" of the vibration signal when walking. In general, after wavelet denoising preprocessing, the redundant noise in the original waveform of each event is effectively filtered out, while the core characteristics of different event signals are retained and highlighted.

[0205] The experimental data set contains 3 different categories. According to the principle of LDA, the maximum possible dimension of reduction is the number of categories minus 1, so the maximum possible dimension of LDA for this data is 2, and the dimension of reduction is [1, 2]. The data is processed for 1-dimensional and 2-dimensional reduction respectively, and the reduction results of each dimension are evaluated.

[0206] Under different dimension reduction dimensions, multiple evaluation indicators are calculated, including inter-class dispersion trace, intra-class dispersion trace, variance sum, cross-validation accuracy and F1 score, and the specific results are shown in Table 1.

[0207] Table 1 Comparison of dimension reduction dimension indicators

[0208]

[0209] The inter-class dispersion trace measures the separation degree between different classes, and the larger the value, the more obvious the class distinction. The inter-class dispersion trace of 2-dimensional dimension reduction is significantly higher than that of 1-dimensional, indicating that 2-dimensional dimension reduction can better separate the distance between classes; the intra-class dispersion trace reflects the compactness of data within the same class, and the smaller the value, the more concentrated the data. The intra-class dispersion trace of 1-dimensional dimension reduction is smaller than that of 2-dimensional, indicating that 1-dimensional dimension reduction has better aggregation effect on intra-class data; the variance sum reflects the total amount of information retained after dimension reduction, and the variance sum of 2-dimensional dimension reduction is higher than that of 1-dimensional, indicating that 2-dimensional can retain more characteristic information of original data; cross-validation accuracy and F1 score are used to evaluate the performance of data after dimension reduction in classification task, and the accuracy and F1 score of 2-dimensional dimension reduction are significantly higher than those of 1-dimensional, indicating that 2-dimensional dimension reduction can more effectively distinguish different classes in classification task.

[0210] In this experiment, considering the above evaluation indicators, although 1-dimensional dimension reduction performs better in intra-class compactness, 2-dimensional dimension reduction has obvious advantages in inter-class distinction, information retention and classification performance. Especially, the cross-validation accuracy and F1 score increase by more than 40%, indicating that 2-dimensional dimension reduction can retain more characteristic information while significantly improving the recognition ability of the model for different classes. Therefore, the optimal dimension reduction dimension is determined as 2. The UMAP map of data after dimension reduction is shown in Figure 11 Figure 11 , where (a) is the original data distribution graph, and (b) is the data distribution graph after LDA projection.

[0211] To further verify the effectiveness of LDA dimension reduction, this experiment quantitatively compares and analyzes the original features and the features after LDA dimension reduction from the inter-class distance. The experimental data includes three types of signals: no event (class 0), mechanical operation (class 1) and personnel walking (class 2), and the original features are one-dimensional vectors after image flattening (dimension 224x224x3=150528). After LDA dimension reduction to the optimal dimension of 2, the feature dimension is compressed to 2 dimensions.

[0212] The inter-class distance uses Euclidean distance to measure the distance between the mean values of different classes, and the formula is:

[0213] ,

[0214] wherein, , are the mean values of classes​ with characteristic mean vector.

[0215] The inter-class distance of original features is close to zero (10 - ¹ 7 order of magnitude), and it is almost impossible to distinguish between classes; after dimensionality reduction, the inter-class distance is significantly improved to between 2.8 and 4.2, indicating that LDA effectively expands the class distinction. After LDA dimensionality reduction, the inter-class distance of all class pairs is significantly increased, indicating that dimensionality reduction effectively expands the distinction between classes, from the nearly zero inter-class distance before dimensionality reduction to the effective distinction after dimensionality reduction, as shown in Figure 12

[0216] In this experiment, an image classification model based on PyTorch was used for training and evaluation. In the selection of model architecture, various parameters such as model size, inference speed, and memory usage were considered, and finally the MNV4ConvSmall architecture was selected. The specific parameters are shown in Table 2.

[0217] Table 2 Model parameter settings

[0218]

[0219] During training, the AdamW optimizer and CosineAnnealingLR learning rate scheduler were used to ensure that the model could converge stably and efficiently. During the entire training process, the loss value and accuracy were recorded in detail to analyze the training effect of the model. After training, the best-performing model was saved, and the loss and accuracy curves were drawn to visually display the performance changes of the model during training. In addition, a confusion matrix was generated to calculate the recognition accuracy of each class and evaluate the classification performance of the model comprehensively.

[0220] To optimize the training effect, LDA was applied to the training set for dimensionality reduction during training, and the parameter amount before and after dimensionality reduction was carefully calculated to explore the changes in data features, and the model inference speed was tested on general hardware to explore the speed improvement of dimensionality reduction, and the results are shown in Table 3.

[0221] Table 3 Inference speed comparison of different hardware platforms

[0222]

[0223] ​After LDA dimensionality reduction, the parameter quantity is reduced from 632.22M to 2.38M, reducing 99.6% of the parameter quantity, saving a lot of computing resources, and making a great contribution to the real-time aspect of the experiment. At the same time, LDA-MNV4 consumes 127.0ms / sample on CPU, which is 3 times faster than MobileNetV4, thanks to the optimization of Mobile MQA module for CPU calculation and the dynamic selection of convolution kernel size by UIB block.

[0224] On the 6000-group data set (including no event, mechanical operation, and personnel walking signals), the test set accuracy of the LDA-MNV4 model reaches 98.28%, which is 4.67% higher than that of MobileNetV4 (93.61%). The performance of each event classification is as follows:

[0225] Background: accuracy 95.77%, low false alarm rate to ensure the stability of the system in the normal environment;

[0226] Mechanical operation signal: accuracy 99.30%, thanks to the high-resolution feature extraction capability of MobileNetV4 and the enhancement of LDA for high-frequency vibration;

[0227] Personnel walking signal: accuracy 99.21%, which proves that LDA has good discrimination ability for low-frequency weak signals.

[0228] Figure 13 The recognition result of the combination of LDA and lightweight CNN is shown in the confusion matrix of the recognition result of the present method; Figure 14 The recognition result of the lightweight CNN alone is shown in the confusion matrix of the recognition result of the lightweight CNN alone. Compared with the MNV4 shown in Figure 14 The average precision of the present LDA-MNV4 is improved from 93.61% to 98.28%, and the inference time is reduced from 450.3ms / sample to 127.0ms / sample, realizing the breakthrough of precision and efficiency.

[0229] The present application combines LDA with MNV4, and uses WTD for signal denoising, and proposes a WTD-LDA-MNV4 method for perimeter security intrusion detection. WTD effectively filters out redundant noise, retains and highlights the core features of different event signals, LDA dimensionality reduction significantly optimizes the feature space, and improves the performance of the classification model; the innovative design of MNV4 makes it perform well on mobile terminals. The experimental results show that the parameter quantity of the method is reduced by 99.6%, the sample inference time is 127ms, the inference speed is faster, the average recognition rate of the three classifications in the experiment reaches 98.28%, which is obviously superior to the traditional method and the baseline model, effectively solving the problem of embedded deployment of security recognition devices, and providing an efficient and practical new scheme for perimeter security.

[0230] The application cooperates with the front-end LDA dimension reduction and the rear-end lightweight network, greatly reduces the model parameter quantity and the calculation cost while ensuring high recognition accuracy, realizes efficient real-time deployment on the edge device, and is suitable for intrusion detection in perimeter security.

[0231] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced by the equivalent, without departing from the spirit and scope of the present application. Any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A lightweight fiber optic vibration intrusion event identification method for perimeter security, characterized in that, Includes the following steps: Step S1: Collect the original vibration signal of the perimeter monitoring area through a distributed fiber optic vibration sensing system; Step S2: Perform wavelet threshold denoising preprocessing on the original vibration signal to obtain the denoised signal; Step S3: Extract features from the denoised signal and construct a high-dimensional feature vector; Step S4: Use linear discriminant analysis to reduce the dimensionality of the high-dimensional feature vector to obtain a low-dimensional classification feature vector; Step S5: Input the low-dimensional classification feature vector into a pre-trained lightweight convolutional neural network model for classification and recognition, and output the corresponding intrusion event category; Step S3 includes the following steps: Step S31: The denoised one-dimensional vibration signal is converted into a two-dimensional image tensor representation by signal reconstruction and channel mapping methods. Step S32: Flatten the two-dimensional image tensor to obtain the first feature vector; Step S33: Extract time-domain features, frequency-domain features, and time-frequency-domain features from the denoised signal to obtain a second feature vector; Step S34: The first feature vector and the second feature vector are fused to form a high-dimensional feature vector; Step S4 includes the following steps: Step S41: Calculate the intra-class scatter matrix and inter-class scatter matrix of the high-dimensional feature vectors of all training samples; Step S42: Solve the generalized eigenvalue equation to obtain the projection vector that maximizes the inter-class difference and minimizes the intra-class difference; Step S43: Select the eigenvectors corresponding to the first k-1 largest generalized eigenvalues ​​to form a projection matrix, where k is the total number of intrusion event categories; Step S44: Use the projection matrix to perform a linear transformation on the high-dimensional feature vector to obtain the low-dimensional classification feature vector with dimension k-1.

2. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to claim 1, characterized in that, The distributed optical fiber vibration sensing system is based on the principle of a phase-sensitive optical time-domain reflectometer, and detects vibration signals by monitoring the phase change of backscattered Rayleigh light in the sensing fiber.

3. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to claim 2, characterized in that, The perimeter monitoring area includes fences, pipelines, bridges, or railway lines, and the sensing optical fiber is laid along the perimeter.

4. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Select a wavelet basis function and a decomposition level, and perform multi-scale wavelet decomposition on the original vibration signal to obtain the wavelet coefficients of each level; Step S22: Threshold processing is performed on the wavelet coefficients, and coefficients smaller than the set threshold are set to zero or shrunk to filter out coefficients corresponding to noise. Step S23: Wavelet reconstruction is performed using the coefficients after threshold processing to obtain the denoised signal.

5. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to claim 1, characterized in that, The time-domain features include one or more of the following: signal mean, variance, root mean square, peak factor, kurtosis, and impulse factor; the frequency-domain features include one or more of the following: spectral centroid, spectral variance, spectral entropy, and energy percentage of a specific frequency band; the time-frequency domain features include the energy of each sub-band obtained by wavelet threshold denoising or wavelet packet decomposition.

6. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to any one of claims 1-5, characterized in that, The lightweight convolutional neural network model is the MobileNet-V4 model, and step S5 includes the following steps: Step S51: The low-dimensional classification feature vector is converted into a multi-dimensional tensor format suitable for input to the MobileNet-V4 model by dimensional expansion and reshaping. Step S52: Input the transformed tensor into the MobileNet-V4 model for feature extraction and classification; Step S53: Determine the corresponding intrusion event category based on the output of the MobileNet-V4 model.

7. The lightweight fiber optic vibration intrusion event identification method for perimeter security according to claim 6, characterized in that, The lightweight convolutional neural network model is a MobileNet-V4 model pre-trained through the following steps: Obtain a training sample set containing labeled categories; The training samples are processed sequentially through steps S1 to S4 to obtain the corresponding low-dimensional classification feature vectors and labels. Using the low-dimensional classification feature vector as input and the corresponding label as supervision signal, the MobileNet-V4 network is trained in a supervised manner until the model converges.

8. A lightweight fiber optic vibration intrusion event recognition system for perimeter security, characterized in that, The system for implementing the method of any one of claims 1-7 comprises: The signal acquisition module is used to acquire raw vibration signals through a distributed fiber optic vibration sensing system. The signal preprocessing module is used to perform wavelet threshold denoising on the original vibration signal; The feature extraction and dimensionality reduction module is used to extract signal features and perform dimensionality reduction using linear discriminant analysis. The classification and recognition module has a built-in lightweight convolutional neural network model, which is used to classify and recognize the dimensionality-reduced features. The output module is used to output the intrusion event category and alarm information.

Images