A method, apparatus, and medium for optical fiber acoustic event recognition

Abstract

The application discloses a kind of optical fiber acoustic wave event identification method, equipment and medium, it is related to acoustic wave signal processing field, the method comprises: obtaining the time series signal of optical fiber acoustic wave;Wavelet denoising processing is carried out to time series signal;Adopt multimodal feature conversion method and convert the time series signal after denoising into multimodal time series image;The fusion processing is carried out to multimodal time series image, obtains RGB fusion image;Transform-CNN hybrid network is constructed;RGB fusion image is input into Transform-CNN hybrid network, obtains the event identification result of optical fiber acoustic wave.The application can significantly enhance the discriminant ability of neural network model to complex signal form, solve the problem that image single mode modeling ability is limited, single image representation mode cannot comprehensively express the multidimensional characteristics of event, improve identification precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of acoustic wave signal processing, in particular to a method and device for identifying acoustic wave events in optical fibers and a medium. BACKGROUND

[0002] Distributed Acoustic Sensing (DAS) technology is a new intelligent sensing method that has developed rapidly in recent years and is widely used in oil and gas pipeline protection, railway state monitoring, border intrusion detection and other scenarios. The DAS system can sense vibration events in the surrounding environment by detecting small changes in scattered signals along the optical fiber. However, due to the complexity of the signal source, strong environmental noise, and large differences in event patterns, the original vibration signals output by the DAS often have low signal-to-noise ratio, fuzzy event boundaries, and unstable spectral characteristics, which brings great challenges to signal recognition and classification.

[0003] Traditional methods mostly use shallow machine learning models to classify after manually extracting features such as spectral energy, signal envelope, and periodicity indicators, but this approach relies heavily on prior experience and has poor generalization ability, making it difficult to meet the actual needs in complex scenarios. In recent years, the development of deep learning has driven the intelligent transformation of DAS signal recognition, and some studies have converted signals into time-frequency graphs or phase space graphs, then used convolutional neural networks (CNN), Transformer networks, and other methods for image classification, achieving certain progress. However, existing methods still have the following problems:

[0004] (1) The image single-modal modeling capability is limited, and a single image representation method (such as only using short-time Fourier transform) cannot fully express the multi-dimensional features of the event, affecting the recognition accuracy;

[0005] (2) The network model structure is redundant or complex to deploy, and standard CNN or Transformer network has a large number of parameters, which is not suitable for real-time deployment in edge devices;

[0006] (3) The time sequence structure of the DAS signal is not modeled well, and its dynamic evolution, system regression, and trend information are not fully mined. SUMMARY

[0007] To solve the above problems existing in the prior art, the present application provides a method and device for identifying acoustic wave events in optical fibers and a medium.

[0008] To achieve the above purpose, the present application provides the following solutions:

[0009] In a first aspect, the present application provides a method for identifying acoustic wave events in optical fibers, comprising:

[0010] obtaining a time series signal of an optical fiber acoustic wave;

[0011] performing wavelet denoising processing on the time series signal;

[0012] converting the denoised time series signal into a multi-modal time series image by using a multi-modal feature conversion method;

[0013] performing fusion processing on the multi-modal time series image to obtain an RGB fusion image;

[0014] constructing a Transformer-CNN hybrid network;

[0015] inputting the RGB fusion image into the Transformer-CNN hybrid network to obtain an event recognition result of the optical fiber acoustic wave.

[0016] Optionally, the multi-modal feature conversion method includes short-time Fourier transform, recursive graph conversion, and Gram matrix analysis; and the multi-modal time series image includes a time-frequency distribution matrix image, a recursive graph, and a two-dimensional GAF image.

[0017] performing short-time Fourier transform on the denoised time series signal to obtain the time-frequency distribution matrix image;

[0018] converting the denoised time series signal into the recursive graph;

[0019] performing Gram matrix analysis on the denoised time series signal to obtain the two-dimensional GAF image.

[0020] Optionally, performing short-time Fourier transform on the denoised time series signal to obtain the time-frequency distribution matrix image includes:

[0021] decomposing the denoised time series signal into a plurality of short-time period signals of time windows;

[0022] determining Fourier transform of each short-time period signal to obtain a time-frequency distribution matrix and generate the time-frequency distribution matrix image.

[0023] Optionally, converting the denoised time series signal into the recursive graph includes:

[0024] determining a phase space vector of the denoised time series signal;

[0025] determining a distance matrix of each phase space vector and performing binary processing on the distance matrix to generate the recursive graph.

[0026] Optionally, performing Gram matrix analysis on the denoised time series signal to obtain the two-dimensional GAF image includes:

[0027] segmenting the de-noised time series signal into multiple subsequences with equal length;

[0028] determining an autocorrelation matrix of each subsequence;

[0029] converting diagonal elements in the autocorrelation matrix into angles to obtain an angle field;

[0030] converting non-diagonal elements in the autocorrelation matrix into radii to obtain a radius field;

[0031] fusing the angle field and the radius field to generate the two-dimensional GAF image.

[0032] Optionally, the constructed Transformer-CNN hybrid network includes five stages, a feature aggregation and dimension reduction module, and a classifier; the five stages are used for feature processing; wherein, stage 1 includes one convolutional layer and two reverse residual modules; stages 2 and 3 are the same in structure, and each includes two reverse residual modules, a global-local feature fusion module, and a C3k2 module; stages 4 and 5 are the same in structure, and each includes two reverse residual modules and a C3k2 module.

[0033] Optionally, the reverse residual module is used for local spatial feature extraction and channel dimension transformation.

[0034] Optionally, the global-local feature fusion module is composed of convolution and Transformer network, and is used for realizing local feature re-extraction and global feature correlation capture.

[0035] In a second aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the optical fiber acoustic wave event recognition method provided above.

[0036] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the optical fiber acoustic wave event recognition method provided above.

[0037] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0038] The application provides a kind of optical fiber acoustic wave event identification method, device and medium, by adopting multimodal feature conversion method, the time series signal after denoising is converted into multimodal time sequence image, to obtain RGB fusion image by fusion, obtains optical fiber acoustic wave event identification result based on RGB fusion image using Transform-CNN hybrid network, can significantly enhance the discriminant ability of neural network model to complex signal form, to solve the problems such as limited modeling ability of image single mode, single image representation mode cannot express the multidimensional characteristics of event comprehensively, improve recognition precision. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 A flowchart of an optical fiber acoustic wave event identification method provided by an embodiment of the present application is shown.

[0041] Figure 2 A schematic diagram of the structure of Transform-CNN hybrid network provided by an embodiment of the present application is shown.

[0042] Figure 3 A schematic diagram of the structure of reverse residual module provided by an embodiment of the present application is shown.

[0043] Figure 4 A schematic diagram of the structure of global-local feature fusion module provided by an embodiment of the present application is shown.

[0044] Figure 5 A schematic diagram of the structure of a computer device provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0046] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail with reference to the drawings and specific embodiments.

[0047] In one exemplary embodiment, the present application provides a fiber acoustic event recognition method suitable for event recognition of distributed acoustic sensing signals (DAS signals), which is executed by a computer device with data processing capability, specifically, can be executed by a terminal or a server alone or jointly by a terminal and a server. In the embodiment of the present application, the method is applied to a server as an example. The server is electrically connected to a DAS system. As shown in Figure 1 The method comprises the following steps:

[0048] Step 100: Obtain a time series signal of fiber acoustic waves. The time series signal is obtained by a DAS system. Based on the Rayleigh scattering principle, the DAS system converts the vibration / acoustic wave signals of each point along the fiber into a time series electrical signal by detecting the phase change of backscattered light in the fiber. In this step, the collected time series signal is a continuous signal in the time domain, denoted as wherein, represents a time variable (unit: second). Based on this, the purpose of this step is to obtain the original time series data of fiber acoustic events, thereby providing a data basis for subsequent processing.

[0049] Step 101: Perform wavelet denoising processing on the time series signal.

[0050] Step 102: Convert the denoised time series signal into a multi-modal time series image using a multi-modal feature conversion method.

[0051] Step 103: Perform fusion processing on the multi-modal time series image to obtain an RGB fusion image.

[0052] Step 104: Construct a Transformer-CNN hybrid network.

[0053] Step 105: Input the RGB fusion image into the Transformer-CNN hybrid network to obtain the event recognition result of the fiber acoustic waves.

[0054] By implementing the above steps 100 to 105, the present application converts the original distributed fiber acoustic sensing signals (i.e., the obtained time series signal of fiber acoustic waves) into a plurality of two-dimensional image representations, and then performs RGB image fusion. Using a Transformer-CNN hybrid network, the RGB fusion image is subjected to feature extraction and classification recognition, which can effectively improve the recognition accuracy of complex events in the DAS signal and has good generalization ability and edge deployment performance.

[0055] In another exemplary embodiment of the present application, the time series signal Generally contains environmental noise (such as electromagnetic interference, fiber inherent noise), need to be denoised by wavelet transform, to eliminate the noise interference in the original signal, improve the accuracy of subsequent feature conversion and signal to noise ratio. Based on this, the implementation process of the above step 101 includes:

[0056] Step 11: Select wavelet basis function , such as Db4 wavelet, has good time-frequency localization characteristics.

[0057] Step 12: Set wavelet transform parameters. Wherein, the scale factor (control the frequency band width of signal decomposition, > 0) and the translation factor (control the time position of signal decomposition, ∈R).

[0058] Step 13: Calculate the wavelet transform coefficient. The calculation formula is:

[0059] .

[0060] In the formula, denotes the conjugate complex of the wavelet basis function . is the wavelet transform coefficient (i.e. the wavelet transform result), which is a complex number, and the modulus reflects the energy of the signal at the scale-time point.

[0061] Step 14: Threshold processing (such as soft threshold method) is performed on the wavelet transform coefficient: the wavelet transform coefficient greater than the threshold value (corresponding to the effective signal) is retained, and the wavelet transform coefficient less than the threshold value (corresponding to the noise) is suppressed.

[0062] The denoised time series signal is reconstructed by wavelet inverse transform, denoted as .

[0063] In another exemplary embodiment of the present application, in order to solve the problem that the existing time sequence structure modeling of DAS signal is insufficient and fails to fully mine its dynamic evolution, system regression and trend information, the denoised time series signal is respectively subjected to three kinds of modal feature conversion methods, time-frequency domain joint features are extracted, and three two-dimensional matrix images are generated. Based on this, in the above step 102, the multi-modal feature conversion method adopted includes: short-time Fourier transform, recursive graph conversion and gram matrix analysis. The multi-modal time sequence image includes: time-frequency distribution matrix image, recursive graph and two-dimensional GAF image. Among them:

[0064] Step 21: Perform short-time Fourier transform on the denoised time series signal to obtain a time-frequency distribution matrix image.

[0065] Step 21-1: decompose the denoised time series signal into short-term signals .

[0066] Step 21-2: calculate the Fourier transform of each short-term signal to obtain a time-frequency distribution matrix, and generate a time-frequency distribution matrix image. The calculation formula is:

[0067] .

[0068] wherein, represents a window function, such as a Hamming window, used to limit the window length and balance the time-frequency resolution. represents a time shift parameter, the starting time of the window, with the unit of seconds. represents an angular frequency, =2πf, f is the frequency, with the unit of Hz. represents a time-frequency distribution matrix, with the dimension of time window number x frequency point number, reflecting the energy distribution of the signal in time and frequency. represents an imaginary unit, represents a natural logarithm.

[0069] Step 22: convert the denoised time series signal into a recurrence plot. The recurrence plot visualizes the repeating patterns of the signal by analyzing the similarity of the signal at different times, and is suitable for detecting nonlinear dynamic behavior.

[0070] Step 22-1: determine the phase space vectors of the denoised time series signal. Select the embedding dimension (e.g. =3) and the time delay (e.g. =1 sampling interval) to construct the phase space vector , , , , where N is the signal length.

[0071] Step 22-2: determine the distance matrix of each phase space vector, and perform binaryzation processing on the distance matrix to generate a recurrence plot. The determination formula of the distance matrix is:

[0072] .

[0073] wherein, represents the element of the distance matrix D, , represents the phase space vector at time t, represents the denoised time series signal at time t.

[0074] ​​Set the threshold to Typically, 10% to 30% of the average distance is taken, and the distance matrix is ​​binarized and represented as follows:

[0075] .

[0076] In the formula, This represents the elements of the distance matrix after binarization.

[0077] The final generated recursion graph is represented as follows Among them, white pixels ( ) indicates that the signal is in and The states at different times are similar, black pixels ( ) indicates that the signal is in and They are not similar at different times. This represents the number field. Based on this, by realizing the visualization of the repeatability of signal states, nonlinear time-series features (such as periodic oscillations and abrupt events) can be extracted.

[0078] Step 23: Perform Gram Matrix Analysis (GAF) on the denoised time series signal to obtain a two-dimensional GAF image. Specifically, the Gram angle field transforms the time series signal into a two-dimensional angle / radius field image by constructing a Gram matrix for the signal segment, which is suitable for capturing the local similarity and dynamic changes of the signal.

[0079] Step 23-1: Divide the denoised time series signal into multiple subsequences of equal length. For example, divide the denoised time series signal... Equal-length segments are Subsequences , , The length of the subsequence. The sampling interval is denoted as .

[0080] Step 23-2: Determine the autocorrelation matrix for each subsequence. Specifically, for each subsequence... Calculate its autocorrelation matrix (Dimension:) ), where the autocorrelation matrix elements , indicating the first subsequence in the subsequence and the The product of the sampling points, Indicates the first Subsequence of sampling points, Indicates the first Subsequence of sampling points, This indicates the matrix transpose.

[0081] Step 23-3: Convert the diagonal elements of the autocorrelation matrix into angles to obtain the angle field. Specifically, [the following steps are involved]. diagonal elements Convert to angle (i.e., normalization to) ).

[0082] Step 23-4: Convert the off-diagonal elements in the autocorrelation matrix into radii to obtain the radius field. Specifically, [the following steps are involved]. Convert off-diagonal elements to radius (i.e., normalization to) ).

[0083] Step 23-5: Fuse the angle field and radius field to generate a two-dimensional GAF image. Specifically, the angle field and radius field are superimposed pixel-wise to generate a two-dimensional GAF image, reflecting the similarity distribution within the signal segment (angle represents phase relationship, radius represents amplitude correlation). Based on this, through local similarity analysis of signal segments, the dynamic evolution characteristics of the time-series signal (such as vibration mode switching, energy mutation, etc.) can be extracted.

[0084] In another exemplary embodiment of this application, in order to integrate the feature information of three modalities (time-frequency distribution, repetition pattern, and local similarity) to form a fused image containing richer discriminative information, thereby solving the problem of the one-sidedness of single-modal features, the time-frequency distribution matrix image generated in steps 21 to 23 (denoted as...) can be used. ), recursive graph (denoted as ) and two-dimensional GAF images (denoted as The RGB fused image in step 103 is constructed by fusing along the channel dimension. The specific steps are as follows:

[0085] Step 31: Normalization. Normalize the pixel values ​​of the three matrices to the interval [0,255] or [0,1] respectively (the specific interval can be adjusted according to the input requirements of the Transformer-CNN hybrid network) to ensure that the feature intensity of each modality is consistent.

[0086] Step 32: Channel stitching. The time-frequency distribution matrix is ​​used as the red channel (R), the recursive graph as the green channel (G), and the 2D GAF image as the blue channel (B), and stitched together to form a single RGB image. .in, Image height, This represents the image width.

[0087] In another exemplary embodiment of this application, such as Figure 2As shown, the constructed Transformer-CNN hybrid network includes five stages, a feature aggregation and dimension reduction (C2PSA) module, and a classifier. The five stages are used for feature processing. Among them, stage 1 includes one convolutional layer and two reverse residual (MBConv) modules. Stages 2 and 3 are the same structure, each including two reverse residual (MBConv) modules, a global-local feature fusion (MobileViT) module, and a C3k2 module. Stages 4 and 5 are the same structure, each including two reverse residual modules and a C3k2 module.

[0088] Based on the structure given above, the pre-trained hybrid neural network model is called in step 104. During model training, image enhancement such as rotation, cropping, random masking, etc. is used, referring to the training process of yolo11 network. Based on this, the implementation process of step 104 is mainly to input the RGB fusion image (size ) generated above into the Transformer-CNN hybrid network, ensure that the pixel value is normalized (such as [0, 1] or [0, 255]) and aligned with the model training, and provide uniform format data for feature extraction.

[0089] The Transformer-CNN hybrid network takes the path of basic extraction→global-local fusion→deep reinforcement, and gradually improves the discriminability of features through module reuse of stage 1→stage 2&3→stage 4&5. Among them:

[0090] Stage 1: Basic feature bottoming (dimension increasing, down sampling).

[0091] The input image (i.e. the RGB fusion image) is first processed by a convolutional layer (Conv) to extract preliminary local features, such as frequency band edge structures in time-frequency distribution matrix, regression contours in recursive graph, etc. Then, two MBConv modules are used for channel dimension increasing and spatial feature enhancement. In the structure of the MBConv module, 1x1 convolution is used for channel adjustment, BatchNorm normalization processing, SiLU activation to enhance expression, and then 3x3 deep convolution is used to extract spatial details.

[0092] Based on this, stage 1 can provide stable bottom-level semantic information and complete a down sampling.

[0093] Stages 2 and 3: Reuse and reinforcement of global and local fusion.

[0094] In stage 2 and stage 3, two consecutive MBConv modules can further refine local features, improve feature clarity and stability, and achieve local enhancement. MobileViT modules build long-range dependencies through the unfolding-Transformer-folding process, strengthen inter-modal connections, and achieve global modeling. C3k2 modules are used for channel and spatial information integration, enhancing network perception range and achieving feature fusion.

[0095] Based on this, stage 2 and stage 3 can achieve more sufficient local-global feature fusion through two times of reuse, and complete one down-sampling in each stage.

[0096] Stage 4 and stage 5: deep feature enhancement and attention focusing.

[0097] In stage 4 and stage 5, two MBConv modules continue to enhance local features, making repetitive patterns in the recursion graph more obvious and achieving detail mining. C3k2 modules automatically filter the feature regions with the strongest event relevance by introducing attention mechanisms, suppress background noise, and achieve discriminative feature purification.

[0098] Based on this, stage 4 and stage 5 achieve primary purification and deep focusing in combination, and each stage down-samples once to output high-recognizability features for the classifier.

[0099] Make classification decisions: map features to events.

[0100] The feature map output by stage 5 is transmitted to the C2PSA module for further purification of discriminative features and compression to a dimension suitable for the input of the classifier.

[0101] Classifier operation: the classifier constructed by the fully connected layer (FC) and the Softmax function maps the high-dimensional feature vector to the probability distribution of event types. For example:

[0102] .

[0103] where, is the predicted score of the i-th event class, is the total number of event classes, is the predicted probability of the i-th event class, is the predicted score of the i-th event class, is the predicted probability of the i-th event class, is the predicted score of the i-th event class, represents the input.

[0104] Based on the above description, the event type corresponding to the maximum probability can be selected as the event recognition result of the fiber acoustic wave, completing the mapping from multi-modal images to signal events.

[0105] ​Further, the Transformer-CNN hybrid network provided in the application can use a cross-entropy loss function for the optimization objective in the training stage, and the specific expression is as follows:

[0106] .

[0107] In the formula, is the cross-entropy loss value, is the one-hot encoding of the true label, is the predicted probability.

[0108] In another exemplary embodiment of the application, the structure of the MBConv module is as shown in Figure 3 It is the core of lightweight convolution, responsible for local spatial feature extraction + channel dimension transformation, and the process is as follows:

[0109] (1) Residual branch judgment.

[0110] If the condition is met (i.e., the input and output channels / size match), the residual connection (i.e., the addition operation) is enabled, allowing the original features to be fused with the processed features to avoid information loss.

[0111] If the condition is not met, the residual is skipped and the features are directly processed.

[0112] (2) Channel transformation and activation.

[0113] 1). 1×1 convolution: dimension increase / dimension decrease (adjust the number of channels to adapt to the subsequent deep convolution).

[0114] 2). Batch normalization (BatchNorm): normalization, accelerate training convergence.

[0115] 3). SiLU activation function: activation function, introduce nonlinearity, enhance feature expression ability.

[0116] (3) Deep convolution to extract spatial details.

[0117] Depthwise separable convolution (Dwise3×3): use a single-channel convolution kernel to traverse the feature map to extract local spatial details (such as frequency blocks of time-frequency graphs, angle patterns of GAF graphs), which greatly reduces the computational complexity.

[0118] (4) Output local features: batch normalization → SiLU activation function → 1×1 convolution (adjust channels) → batch normalization, output enhanced local feature maps to provide spatial detail materials for the MobileViT module. Figure 3 False indicates an error, and True indicates correctness.

[0119] In another exemplary embodiment of the present application, the structure of the MobileViT module is as shown in Figure 4 It combines convolution and Transformer network to realize local feature re-extraction + global correlation capture, and the process is as follows:

[0120] (1) Local feature re-extraction (convolution part):

[0121] The local feature is first subjected to convolution-3x3 (spatial convolution) and convolution-1x1 (channel adjustment) to extract fine-grained local features again (to supplement the details not covered by MBConv).

[0122] (2) Global correlation construction (Transformer part):

[0123] 1. Expansion (Flatten): flatten the feature map (dimension CxHxW) into a sequence (dimension d x N x P, d is the dimension, N is the sequence length, and P is the number of blocks) to adapt the input of the Transformer network.

[0124] 2. Transformer processing: through multi-head self-attention (Multi-Head Self-Attention), the global dependence relationship of the feature sequence is excavated (such as the correlation of cross-region time-frequency patterns, long-distance interaction of different modal features).

[0125] 3. Folding (Unflatten): convert the sequence output by the Transformer network back to a feature map (dimension restored) to realize the fusion of local details + global correlation.

[0126] (3) Output fusion feature: through convolution-3x3 (integrated fusion feature) and convolution-1x1 (channel adjustment), the global-local fusion feature map is output, which not only retains the spatial details of MBConv, but also integrates the global correlation of the Transformer network.

[0127] The C3k2 module and the C2PSA module adopt the existing structure, so they will not be described here.

[0128] Based on the above description, the method provided by the application is a fiber acoustic event recognition method based on multi-modal time sequence image fusion and Transformer-CNN structure. First, the original time domain DAS signal is denoised by wavelet, and then converted into a three-channel image through short-time Fourier transform (STFT), recurrence plot (RP) and Gramian angular field (GAF) and the like, to capture the time-frequency features, dynamic behavior and overall trend of the signal. Subsequently, a recognition model combining Transformer network and convolution structure is designed for image processing; among them, the Transformer part adopts a lightweight MobileViT structure to enhance the global modeling capability, and the convolution module borrows the improved structure in the YOLO11 network to strengthen the local perception and multi-scale feature extraction capability while keeping the model lightweight.

[0129] In an exemplary embodiment, the advantages of the method provided by the application are illustrated by comparison with single-modal recognition, pure Transformer network structure recognition, and end-to-end original signal recognition methods.

[0130] Single-modal recognition method: directly using STFT image or original waveform as input, using traditional CNN (such as ResNet, VGG, etc.) for training and recognition. Although this method is simple in structure, it is difficult to distinguish similar events in a high interference environment due to the lack of inter-modal information collaborative expression.

[0131] Pure Transformer network structure recognition method: using pure ViT (Vision Transformer) architecture to model the time-frequency image, relying on large-scale pre-training to learn global features. Although the global modeling capability is strong, the parameter quantity is large, the training cost is high, and the dependence on data quantity is significant, which is not suitable for resource-limited engineering deployment scenarios.

[0132] End-to-end original signal recognition method: for example, inputting the original DAS time sequence signal into 1D-CNN, RNN or LSTM network for feature learning and event recognition. This method skips the image construction process and simplifies the structure, but the time sequence modeling capability is limited, it is difficult to extract multi-scale spatial semantic features, and the performance is limited.

[0133] The identification method provided by the application: by converting the original optical fiber distributed acoustic sensing (DAS) signal into a plurality of two-dimensional image representations (such as STFT, RP, GAF) respectively, and then performing RGB image fusion. A hybrid neural network model constructed by integrating a Transformer structure (such as MobileViT) and a lightweight convolution structure derived from YOLO11 is used to extract features and classify images, which can effectively improve the recognition accuracy of complex events in the DAS signal and has good generalization ability and edge deployment performance.

[0134] Based on the above description, although there are some alternative solutions that can realize optical fiber acoustic event recognition in the prior art, the identification method based on multi-modal image fusion and Transformer-CNN hybrid structure adopted by the application still has obvious advantages in recognition accuracy, computational efficiency and deployment adaptability. The application realizes the expression of signals in multiple dimensions (frequency domain, time sequence, angle) through multi-modal image coding, solving the problem of insufficient perception of traditional single-modal models. The structure of the lightweight Transformer-CNN hybrid network controls the computational complexity while taking into account local feature extraction and global semantic modeling. The Transformer-CNN hybrid network structure integrates the efficient convolution modules of YOLO11 (such as MBConv modules, C3k2 modules, etc.), which is significantly better than traditional backbone networks such as ResNet. The model as a whole has a modular design, supports different depth pruning and deployment strategies, and is highly practical. Although there are other implementation paths that can complete the basic function of optical fiber acoustic event recognition, the application has comprehensive technical advantages in feature extraction dimension, model structure design, system lightweight and recognition robustness, and is a better solution.

[0135] In summary, compared with existing optical fiber distributed acoustic signal recognition technology, the application has the following significant technical advantages:

[0136] (1) High recognition accuracy, strong multi-modal information collaborative modeling capability.

[0137] By introducing STFT, RP, and GAF three kinds of time series two-dimensional image coding methods, the application effectively preserves the frequency domain changes, time dependence structure and nonlinear dynamic characteristics of the original signal, and then inputs the RGB fusion into the Transformer-CNN hybrid network, which significantly enhances the model's ability to distinguish complex signal patterns. This advantage mainly comes from the design of image conversion and image fusion.

[0138] In the experiment, the fusion image can better distinguish high similarity events (such as construction interference and vibration interference) than a single modality.

[0139] (2) Lightweight design, low computational resource occupation.

[0140] The application comprehensively adopts lightweight convolution modules and compact structure design in the neural network architecture, mainly including:

[0141] The MBConv module in YOLO11 replaces the standard convolution with 1x1 convolution and depth separable convolution, greatly reducing the number of parameters and calculation amount.

[0142] The C3k2 module is efficient in channel and spatial feature fusion and supports fast forward propagation.

[0143] The MobileViT module uses a low-complexity Transformer network to replace the traditional ViT model, combines local convolution and small-size attention window, and balances global modeling capability and inference speed.

[0144] The above design makes the entire Transformer-CNN hybrid network side-deployable, which can run on embedded devices, industrial edge servers or on-site diagnosis terminals, meeting the engineering requirements for energy consumption, storage, response speed and other aspects. The lightweight advantage mainly comes from the neural network structure and the relay design of the modules inside.

[0145] Compared with traditional ResNet, ViT and other networks, the inference speed of the Transformer-CNN hybrid network provided by the application is improved by about 30-50%, and the number of parameters is reduced by more than 50%.

[0146] (3) Robustness, adapt to complex background changes.

[0147] By embedding the MobileViT module into the backbone network, the Transformer mechanism can model non-local dependent relationships in different modal images, effectively adapting to environmental changes, signal occlusion and noise disturbance. The attention mechanism in the C3k2 module further improves the positioning accuracy of abnormal events such as sudden leaks and abnormal vibrations. This effect depends on the module stacking and reuse design from stage 2 to stage 5, which embodies the progressive strategy of local detail enhancement, global semantic modeling and attention focusing.

[0148] (4) Flexible deployment, adapt to multiple scene applications.

[0149] Since the Transformer-CNN hybrid network is side-deployable, and the structure is modular and highly customizable, the depth and width of the feature extraction stage can be adjusted according to the actual scene, supporting multiple fiber acoustic application scenarios such as pipeline leaks, railway intrusions and perimeter protection. The flexible deployment comes from the decoupled design of the CSP structure and the Transformer module, supporting flexible loading and transfer learning on different computing platforms.

[0150] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores fiber optic acoustic event identification data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a fiber optic acoustic event identification method.

[0151] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0152] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0153] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0154] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0155] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0156] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0157] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0158] The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0159] The principles and implementation manners of the present application are described herein by using specific examples, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will have changes. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method of optical fiber acoustic event recognition, the method comprising: The method comprises the following steps: obtaining a time series signal of an optical fiber acoustic wave; wavelet denoising the time series signal; converting the denoised time series signal into a multi-modal time series image by using a multi-modal feature conversion method; the multi-modal feature conversion method comprises short-time Fourier transform, recurrence plot conversion and Gram matrix analysis; the multi-modal time series image comprises a time-frequency distribution matrix image, a recurrence plot and a two-dimensional GAF image; wherein the time-frequency distribution matrix image is obtained by performing short-time Fourier transform on the denoised time series signal; the recurrence plot is obtained by converting the denoised time series signal; the two-dimensional GAF image is obtained by performing Gram matrix analysis on the denoised time series signal; performing fusion processing on the multi-modal time series image to obtain an RGB fusion image; constructing a Transformer-CNN hybrid network; the constructed Transformer-CNN hybrid network comprises five stages, a feature aggregation and dimension reduction module and a classifier; the five stages are used for feature processing; wherein stage 1 comprises one convolutional layer and two reverse residual modules; stages 2 and 3 have the same structure and each comprises two reverse residual modules, a global-local feature fusion module and a C3k2 module; stages 4 and 5 have the same structure and each comprises two reverse residual modules and a C3k2 module; inputting the RGB fusion image into the Transformer-CNN hybrid network to obtain an event recognition result of the optical fiber acoustic wave.

2. The optical fiber acoustic event recognition method of claim 1, wherein, The time-frequency distribution matrix image is obtained by performing short-time Fourier transform on the denoised time series signal, comprising: decomposing the denoised time series signal into a plurality of short-time period signals of time windows; determining the Fourier transform of each short-time period signal to obtain a time-frequency distribution matrix and generate the time-frequency distribution matrix image.

3. The optical fiber acoustic event recognition method of claim 1, wherein, The recurrence plot is obtained by converting the denoised time series signal, comprising: determining the phase space vector of the denoised time series signal; determining the distance matrix of each phase space vector and performing binary processing on the distance matrix to generate the recurrence plot.

4. The optical fiber acoustic event recognition method of claim 1, wherein, The two-dimensional GAF image is obtained by performing Gram matrix analysis on the denoised time series signal, comprising: segmenting the denoised time series signal into a plurality of sub-sequences with equal length; determining the autocorrelation matrix of each sub-sequence; converting the diagonal elements in the autocorrelation matrix into angles to obtain an angle field; converting the non-diagonal elements in the autocorrelation matrix into radii to obtain a radius field; fusing the angle field and the radius field to generate the two-dimensional GAF image.

5. The optical fiber acoustic event recognition method of claim 1, wherein, The reverse residual module is used for local spatial feature extraction and channel dimension transformation.

6. The optical fiber acoustic event recognition method of claim 5, wherein, The global-local feature fusion module is composed of convolution and Transformer network and is used for realizing local feature re-extraction and global feature correlation capture.

7. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the optical fiber acoustic wave event recognition method of any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the optical fiber acoustic wave event recognition method of any one of claims 1-6.

Images