Optical fiber anomaly detection method, device and equipment and storage medium

Abstract

The invention discloses an optical fiber anomaly detection method, device and equipment and a storage medium, and relates to the technical field of target detection, a target two-dimensional time-frequency atlas is input into a target detection model, an optical fiber anomaly detection result is obtained, the target detection model is an improved yolov8 model, the target detection model comprises a RepVGG convolution block, and the RepVGG convolution block comprises a RepVGG convolution block; the RepVGG convolution block comprises a main branch and at least one auxiliary branch in the training stage; the RepVGG convolution block fuses a plurality of branches in the training stage into an equivalent 3 * 3 convolution layer through structural re-parameterization in the reasoning stage. Therefore, the lightweight design of multiple branches during training and single branch during reasoning is realized, and the reasoning speed is improved while the feature extraction capability is ensured.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to a method, apparatus, device and storage medium for optical fiber anomaly detection. Background Technology

[0002] With the rapid development of information technology, fiber optic sensing technology has been widely used in many fields due to its advantages such as high sensitivity, resistance to electromagnetic interference, and remote monitoring capabilities. DAS (Distributed Acoustic Sensing), as one of the key technologies, monitors the impact of external physical quantities on optical fibers by analyzing changes in backscattered light transmitted in the optical fiber, thereby realizing the perception and identification of environmental changes.

[0003] In related technologies, data is collected from sensing optical fibers deployed in the monitoring area using DAS (Digital Animation Sensor), resulting in DAS data. A two-dimensional time-frequency spectrum is then generated based on the DAS data. This spectrum is input into a YOLOv8 model, which outputs bounding boxes and category labels to identify the time period and spatial location of abnormal activity, thereby triggering alarms or further analysis. However, existing YOLOv8 models often suffer from high false negative rates and slow inference speeds when detecting anomalies. Summary of the Invention

[0004] In view of this, this application provides a fiber optic anomaly detection method, apparatus, device, and storage medium to solve the problems of high false negative rate and slow inference speed in the prior art.

[0005] The objective of this application can be achieved through the following technical solutions: The first aspect of this application is to provide a method for detecting optical fiber anomalies, including: Data is collected from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data; Generate a two-dimensional time-frequency spectrum of the target based on the target DAS data; The target's two-dimensional time-frequency spectrum is input into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block.

[0006] In an optional embodiment, the RepVGG convolutional block further includes a residual branch and a first feature fusion unit; The residual branch is configured as an identity mapping, through which the sample input feature map is passed; The first feature fusion unit adds the output features of each branch element by element to generate a sample output feature map.

[0007] In an optional embodiment, Neck includes an improved SPPF module, which includes a first 1×1 convolutional layer, a parallel-connected 3×3 max-pooling layer, a 5×5 max-pooling layer and a 7×7 max-pooling layer, a second feature fusion unit, and an LSKA attention module. The first target feature map is input into the first 1×1 convolutional layer to reduce the dimensionality of the first target feature map, resulting in the dimensionality-reduced first target feature map. The first target feature map after dimensionality reduction is input into a 3×3 max pooling layer, a 5×5 max pooling layer, and a 7×7 max pooling layer, respectively, to obtain the output of each pooling layer. The resolution of the output of each pooling layer is consistent with the resolution of the input of each pooling layer, and the number of channels of the output of each pooling layer is consistent with the number of channels of the input of each pooling layer. The outputs of each pooling layer are spliced ​​along the channel dimension by the second feature fusion unit to obtain a multi-scale fused feature map; The multi-scale fused feature map is input into the LSKA attention module, and after channel splitting, depthwise separable convolution processing and concatenation, the intermediate feature map is output. Apply the Sigmoid function to the intermediate feature map to generate an attention weight map.

[0008] In one optional embodiment, the LSKA attention module includes a channel splitting unit, a second 1×1 convolutional layer, a 7×7 depth-separable convolutional layer, a third 1×1 convolutional layer, and a third feature fusion unit. The multi-scale fusion feature map is split by channel splitting unit to obtain multiple branches of multi-scale fusion feature map; The multi-scale fusion feature map branch is reduced in dimensionality by using the second 1×1 convolutional layer along the channel dimension to obtain the dimensionality-reduced multi-scale fusion feature map branch. Long-distance spatial modeling is performed on the multi-scale fused feature map branch after dimensionality reduction by using a 7×7 depth separable convolutional layer to obtain the enhanced feature map branch; The enhanced feature map branch is increased in dimension along the channel dimension by the third 1×1 convolutional layer to obtain the enhanced feature map branch in dimension. The enhanced feature map branches after dimensionality increase are spliced ​​together by the third feature fusion unit to obtain an intermediate feature map. The number of channels in the intermediate feature map is consistent with the number of channels in the multi-scale fusion feature map.

[0009] In an optional embodiment, the LSKA attention module further includes an attention application unit; The attention application unit performs element-wise multiplication of the attention weight map and the dimensionality-reduced first target feature map to output the second target feature map.

[0010] In one optional embodiment, the Head includes a main detection head, an auxiliary detection head, and a detection result fusion unit; The first target feature map is input into the main detection head to obtain the first detection result; The second target feature map is input into the auxiliary detection head to obtain the second detection result; The first and second detection results are fused by the detection result fusion unit to obtain the fiber optic anomaly detection result.

[0011] In an optional embodiment, it further includes: Determine the center grid of the target bounding box; When the intersection-union ratio of the adjacent grids of the target bounding box center grid to the target bounding box center grid is greater than a preset threshold, the adjacent grids are identified as positive samples.

[0012] A second aspect of this application is to provide an optical fiber anomaly detection device, comprising: The acquisition module is used to acquire data from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data. The generation module is used to generate a target two-dimensional time-frequency spectrum based on the target DAS data; The input module is used to input the target's two-dimensional time-frequency spectrum into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block.

[0013] A third aspect of this application is to provide an electronic device, including 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 perform the method as described in the first aspect.

[0014] A fourth aspect of this application is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method as described in the first aspect.

[0015] Compared with existing technologies, the fiber optic anomaly detection method provided in this application inputs the two-dimensional time-frequency spectrum of the target into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. During the inference phase, the RepVGG convolutional block merges multiple branches from the training phase into an equivalent 3×3 convolutional layer through structural reparameterization. In this way, a lightweight design of "multiple branches during training and single branch during inference" is achieved, which improves the inference speed while ensuring feature extraction capabilities. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of an optical fiber anomaly detection method provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the process of obtaining sample output feature maps during the training phase of a RepVGG convolutional block provided in this application embodiment; Figure 3 A schematic diagram illustrating the process of obtaining the target output feature map during the inference stage of a RepVGG convolutional block provided in this application embodiment; Figure 4 A flowchart illustrating the process of obtaining a second target feature map by improving the SPPF module, as provided in this application embodiment; Figure 5 is a schematic diagram of a method for determining positive samples provided by related technologies; Figure 6 is a schematic diagram of determining a positive sample according to an embodiment of this application; Figure 7 A schematic diagram illustrating the training effect of the target detection model provided for related technologies; Figure 8 A schematic diagram illustrating the training effect of the target detection model provided in the embodiments of this application; Figure 9 Another schematic flowchart of the fiber optic anomaly detection method provided in the embodiments of this application; Figure 10 A structural block diagram of the fiber optic anomaly detection device provided in the embodiments of this application; Figure 11 This is a structural block diagram of an electronic device for implementing an optical fiber anomaly detection method, provided in an embodiment of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0020] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0021] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0022] To address the technical problems existing in related technologies, embodiments of this application provide a method, apparatus, electronic device, and storage medium for detecting optical fiber anomalies.

[0023] The fiber optic anomaly detection method provided in this application can be executed by an electronic device, such as a terminal or a server. The terminal can be a smartphone, tablet, laptop, or other similar device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. It is understood that this application does not limit the specific entity executing the fiber optic anomaly detection method.

[0024] The technical solution of this application will be described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments described below are used to explain the technical solution of this application and are not intended to limit actual use.

[0025] To address the technical problems existing in related technologies, embodiments of this application provide a method for detecting optical fiber anomalies, such as... Figure 1 As shown, Figure 1 This is a flowchart of a fiber optic anomaly detection method provided in an embodiment of this application. It should be noted that the steps shown may be executed in a different logical order than those shown in the flowchart. The method may include the following steps S101 to S103.

[0026] Step S101: Use DAS to collect data from the sensing optical fibers deployed in the monitoring area to obtain target DAS data.

[0027] It should be noted that the DAS data is a spatiotemporal phase map obtained using Rayleigh scattering. The data format is as follows: the phase signal of a single pulse of light is subtracted from the phase signal of the light coming back and the adjacent phases to obtain the phase signal, and then the audio signals at different distances are obtained by unwinding.

[0028] In one optional embodiment, assuming a sampling rate of 10kHz, a sampling length of 40km, and a demodulation interval of 4 meters per point, the data format is {Data(0,1), Data(0,2), Data(0,3), Data(0,4), Data(0,5)...Data(0,N)}. Data(0,1) represents the maximum audio signal value from 0 to 4 meters in the first second, Data(0,2) represents the maximum audio signal value from 4 to 8 meters in the first second, and so on. The audio data is then superimposed for 2 minutes, resulting in 120 signal sets. N is the number of sampling points.

[0029] Overlay time: {Data(0,1),Data(0,2),Data(0,3),Data(0,4),Data(0,5)...Data(0,N}; {Data(1,1),Data(1,2),Data(1,3),Data(1,4),Data(1,5)...Data(1,N}; {Data(2,1),Data(2,2),Data(2,3),Data(2,4),Data(2,5)...Data(2,N}; ... {Data(M,1), Data(M,2), Data(M,3), Data(M,4), Data(M,5)...Data(M,N)}. M is the Mth data set.

[0030] Assign amplitude values ​​to these data P using a color table. For example, a value equal to 10 is blue, the middle color is yellow, and a value greater than 20 is red. This will generate a color heatmap. The data P is the corresponding DAS data mentioned below.

[0031] In one alternative embodiment, the target DAS data can be DAS data from different scenarios. DAS utilizes changes in optical signals within an optical fiber to sense external vibrations, thereby enabling the monitoring of various physical phenomena.

[0032] In one specific embodiment, physical phenomena can be classified according to their application scenarios as: vibrations caused by human activities, vibrations caused by the natural environment, vibrations caused by industrial equipment, and safety and disaster events.

[0033] In a more specific embodiment, vibrations caused by human activity include walking / running, such as low-frequency vibrations generated by people walking near optical fibers, which can be used for perimeter security. Vibrations caused by human activity include vehicle traffic, such as ground vibrations caused by vehicles such as cars, which can be used for traffic flow monitoring. Vibrations caused by human activity include digging / drilling / piling, such as high-frequency, high-intensity vibrations generated by construction machinery operations, which are the primary monitoring targets for third-party construction damage. Vibrations caused by human activity include climbing / cutting fences, such as perimeter intrusion behavior, which has obvious vibration characteristics and can be used for intelligent security alarms.

[0034] In another, more specific embodiment, vibrations related to the natural environment include wind vibration, rainfall, rockfalls, earthquakes, and water flow / water hammer. Vibrations related to industrial equipment include pipeline leaks, machinery operation, train operation, and oil and gas well fracturing. Vibrations related to safety and disasters include explosions / shock waves, landslides / mudslides, and fires.

[0035] It should be noted that the above is merely an exemplary generalization of physical phenomena and does not limit this application.

[0036] In another optional embodiment, DAS data from different scenarios are collected as training sample data when training the target detection model. For example, data including 8 types of construction data, 3 types of natural environmental disturbance data, and 4 types of typical interference are collected. The annotation adopts a three-dimensional annotation method of "event category + location coordinates + confidence level" to ensure the generalization ability of the training sample dataset, thereby ensuring that the accuracy of the target detection model in cross-scenario testing remains stable at over 97%.

[0037] It should be noted that event categories can include things like excavation, drilling, and wind vibration, used for classifying tasks; location coordinates are usually two-dimensional coordinates that include time and fiber optic distance, representing the spatiotemporal location where the event occurred; confidence level indicates the degree of credibility of the annotation results.

[0038] Step S102: Generate a two-dimensional time-frequency spectrum of the target based on the target DAS data.

[0039] It should be noted that the target DAS data is usually a one-dimensional DAS vibration signal. In the target two-dimensional time-frequency spectrum, the horizontal axis represents time, the vertical axis represents frequency, and the pixel color represents the signal energy intensity, which is used for subsequent vibration event detection based on deep learning.

[0040] In one optional embodiment, generating a target two-dimensional time-frequency spectrum based on target DAS data specifically includes the following steps: Set the window length and overlap rate of the STFT (Short-Time Fourier Transform); convert the target DAS data within the sliding window into a target two-dimensional time-frequency spectrum.

[0041] In one specific embodiment, the window length can be set to a range of 256 to 512 sampling points, and the overlap rate can be set to a range of 50% to 75%. By setting the window length and overlap rate, the characteristics of the DAS vibration signal are ensured to be clearly presented.

[0042] It should be noted that the settings for window length and overlap rate can be adjusted based on actual circumstances, and this application does not impose any limitations on them.

[0043] Step S103: Input the two-dimensional time-frequency spectrum of the target into the target detection model to obtain the fiber optic anomaly detection result.

[0044] In one optional embodiment, the fiber optic anomaly detection results include: the category of the anomaly, the spatiotemporal location of the anomaly, and the confidence level of the anomaly.

[0045] In another alternative embodiment, the object detection model is an improved YOLOv8 model, which includes a Backbone, a Neck, and a Head. It's important to note that the Backbone typically refers to the main part of the deep neural network, responsible for extracting features from the input data. The Neck, located between the Backbone and the Head, is responsible for further processing and integrating the features extracted by the Backbone. The Neck acts as a bridge connecting the Backbone and the Head, helping to better pass features to the Head for the final prediction or classification task. The Head is the top part of the entire network structure, responsible for performing specific tasks such as classification, object detection, or semantic segmentation. The Head typically consists of fully connected layers or convolutional layers, used to map the features extracted by the Neck to the final output space, generating the network's final prediction result. The structure of the Head may vary in different tasks to adapt to different task requirements.

[0046] In one specific embodiment, the backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract spatial local features from the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input to the RepVGG convolutional block. The 1×1 convolutional layer includes convolution operations, batch normalization, and an activation function, and the 3×3 convolutional layer includes convolution operations, batch normalization, and an activation function.

[0047] In one specific embodiment, the activation function is the ELU (Exponential Linear Unit) activation function. ELU has non-zero outputs in the negative region, which helps maintain neuron activity, improves sensitivity to weak vibration signals, and thus enhances the nonlinear expression ability of the target detection model for weak DAS signals (low amplitude, low frequency). Batch normalization layers are used to eliminate gradient offset.

[0048] In another specific embodiment, the RepVGG convolutional block merges multiple branches from the training phase into an equivalent 3×3 convolutional layer through structural reparameterization during the inference phase. The equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map and outputs a first target feature map. The target input feature map is the feature representation of the target two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. The equivalent 3×3 convolutional layer includes convolution operations, batch normalization, and activation functions.

[0049] In this embodiment, a lightweight design of "multi-branch training and single-branch inference" is implemented, which reduces the number of model parameters by 20% while ensuring feature extraction capabilities and improving inference speed.

[0050] In another alternative embodiment, the RepVGG convolutional block further includes a residual branch and a first feature fusion unit; The residual branch is configured as an identity mapping, through which the sample input feature map is passed; The first feature fusion unit adds the output features of each branch element by element to generate a sample output feature map, which is then input into the subsequent modules.

[0051] In one specific embodiment, such as Figure 2 and Figure 3 As shown, Figure 2 A flowchart illustrating the process of obtaining sample output feature maps during the training phase of RepVGG convolutional blocks is provided. Figure 3 This document provides a flowchart illustrating the process of obtaining the target output feature map during the inference phase of a RepVGG convolutional block, specifically including the following steps: During the training phase, the sample input feature maps are fed into a 3×3 convolutional layer, a 1×1 convolutional layer, and a residual branch, respectively, to obtain the sample output feature maps. The 1×1 convolutional layer includes a convolutional operation, a batch normalization layer, and an activation function layer connected in sequence, and the 3×3 convolutional layer includes a convolutional operation, a batch normalization layer, and an activation function layer connected in sequence.

[0052] During the inference phase, the target input feature map is fed into an equivalent 3×3 convolutional layer and a residual branch to obtain the target output feature map.

[0053] In a more specific embodiment, both the input feature map and the output feature map are three-dimensional tensors with dimensions of H×W×C.

[0054] In another alternative embodiment, Neck includes an improved SPPF (Spatial Pyramid Pooling) module, which includes a first 1×1 convolutional layer, a parallel 3×3 max pooling layer, a 5×5 max pooling layer and a 7×7 max pooling layer, a second feature fusion unit and an LSKA attention module. The first target feature map is input into the first 1×1 convolutional layer to reduce its dimensionality, resulting in a dimensionality-reduced first target feature map. For example, if the first target feature map has dimensions H / 8×W / 8×2C, the dimensionality-reduced first target feature map can be H / 8×W / 8×C. The dimensionality-reduced first target feature map is then input into a 3×3 max pooling layer, a 5×5 max pooling layer, and a 7×7 max pooling layer, respectively, to obtain the outputs of each pooling layer. The resolution of each pooling layer output is consistent with the resolution of its input, and the number of channels in each pooling layer output is consistent with the number of channels in its input. The second feature fusion unit concatenates the outputs of each pooling layer along the channel dimension to obtain a multi-scale fused feature map. The multi-scale fused feature map is then input into the LSKA (Large Separable Kernel Attention) attention module. After channel splitting, depthwise separable convolution processing, and concatenation, an intermediate feature map is output. A Sigmoid function is applied to the intermediate feature map to generate an attention weight map.

[0055] It's worth noting that the sigmoid function is one of the commonly used activation functions in deep learning, especially in early neural network models. It maps the input to the (0, 1) interval, making it suitable for probability estimation in binary classification problems. The weights in the attention weight map range from 0 to 1.

[0056] In a more specific embodiment, the LSKA attention module includes a channel splitting unit, a second 1×1 convolutional layer, a 7×7 depth-separable convolutional layer, a third 1×1 convolutional layer, and a channel splicing unit. The multi-scale fusion feature map is split into multiple branches by a channel splitting unit. For example, the dimension of each branch can be H / 8×W / 8×C / 4. A second 1×1 convolutional layer reduces the dimensionality of each branch along the channel dimension, resulting in a dimensionality-reduced branch with C / 8 channels. A 7×7 depthwise separable convolutional layer performs long-range spatial modeling on the dimensionality-reduced branch, resulting in an enhanced feature map. A third 1×1 convolutional layer increases the dimensionality of the enhanced branch along the channel dimension, resulting in an enhanced branch with H / 8×W / 8×C / 4 dimensions. Finally, a channel concatenation unit concatenates the enhanced branch to obtain an intermediate feature map. The number of channels in the intermediate feature map is the same as that in the multi-scale fusion feature map. For example, the dimension of the intermediate feature map can be H / 8×W / 8×C.

[0057] In another alternative embodiment, the LSKA attention module further includes an attention application unit; The attention application unit performs element-wise multiplication of the attention weight map and the dimensionality-reduced first target feature map to output the second target feature map.

[0058] By using 7×7 separable convolution to capture long-range dependent features (suitable for high-amplitude vibration signals to prevent external damage), and by allocating channel attention weights, vibration signal features (such as vibration peaks in specific frequency bands) are enhanced, while interference signal features (such as high-frequency noise from vehicle vibrations) are suppressed, resulting in a 12% improvement in multi-scale feature extraction accuracy.

[0059] It should be noted that the specific structure of the feature fusion unit is detailed in the relevant technologies and will not be elaborated here.

[0060] In a more specific embodiment, such as Figure 4 As shown, Figure 4 This is a flowchart illustrating a process for obtaining a second target feature map by improving the SPPF module, specifically including the following steps: The first target feature map is input into a 1×1 convolutional layer to reduce its dimensionality, resulting in a dimensionality-reduced first target feature map. The dimensionality-reduced first target feature map is then input into an improved SPPF module to obtain an attention weight map. The attention application unit performs element-wise multiplication (feature concatenation) between the attention weight map and the dimensionality-reduced first target feature map to output the second target feature map.

[0061] In another optional embodiment, the Head includes a main detection head, an auxiliary detection head, and a detection result fusion unit; a first target feature map is input into the main detection head to obtain a first detection result; a second target feature map is input into the auxiliary detection head to obtain a second detection result; the detection result fusion unit fuses the first detection result and the second detection result to obtain an optical fiber anomaly detection result.

[0062] In this embodiment, an auxiliary training head is added on the basis of the original detection head (responsible for final prediction). The auxiliary training head predicts the low-to-mid-level feature maps output by the neck network during the training process. Through the collaborative optimization of the "main-auxiliary" dual detection heads, the gradient vanishing problem during model training is alleviated. In particular, the false negative rate is reduced to below 0.5% for the recognition of micro-vibration signals.

[0063] In another alternative embodiment, considering that related technologies typically only consider the grid containing the center of the target bounding box as a positive sample candidate, the positive sample coverage is small, resulting in a high false negative rate. To address this issue, embodiments of this application provide a novel method for determining positive sample candidates, reducing the false negative rate and improving the positive sample recall rate.

[0064] The fiber optic anomaly detection method provided in this application embodiment also includes: Determine the center grid of the target bounding box; when the intersection-union ratio of the adjacent grids of the center grid of the target bounding box with the center grid of the target bounding box is greater than a preset threshold, the adjacent grids are determined as positive samples.

[0065] In one specific embodiment, the two adjacent grids (vertical or horizontal, depending on the pipeline direction) of the center grid of the target bounding box are included in the positive sample candidate range. When the IoU (Intersection over Union) between the adjacent grid and the target bounding box is greater than or equal to a preset threshold, it is marked as a positive sample. This strategy expands the coverage of positive samples, especially for the capture capability of edge leakage signals (target bounding boxes spanning multiple grids), improving the positive sample recall rate by 18%.

[0066] In a more specific embodiment, Figure 5 A schematic diagram for identifying positive samples for related technologies, such as Figure 5 As shown, in related technologies, the original YOLOv8 uses a positive sample matching method of "target box center grid + IoU threshold", only considering the grid where the target box center is located as a positive sample candidate. This application includes the two adjacent grids (up and down or left and right, determined according to the pipeline direction) of the target box center grid in the positive sample candidate range. When the IoU between the adjacent grid and the target box is ≥0.3, it is marked as a positive sample. This strategy expands the positive sample coverage, especially for the capture capability of edge leakage signals (target boxes spanning multiple grids), and improves the positive sample recall rate by 18%.

[0067] In related technologies, the center of the target box falls within "grid (3, 4)". When the IoU between the grid and the target box is ≥0.5, it is marked as a positive sample. The adjacent grids (2, 4) and (4, 4) are marked as "negative samples". The false negative rate is 4.2%.

[0068] In a more specific embodiment, Figure 6 This is a schematic diagram illustrating the determination of positive samples provided in an embodiment of this application. In this application, the target box is centered at grid (4, 4). Besides grid (4, 4) (IoU ≥ 0.5) being marked as a positive sample, its adjacent grids (3, 4) and (5, 4), if IoU ≥ 0.3, are also marked as positive samples. The false negative rate is 0.4%, and the positive sample recall rate is improved by 18%. It should be noted that... Figure 5 and Figure 6 In the diagram, the dashed box represents the target bounding box, and the solid box / solid box filled with black represents the positive sample grid.

[0069] In another alternative embodiment, Figure 7This is a schematic diagram showing the changes in multiple loss functions and evaluation metrics of object detection models during the training process in related technologies. Figure 8 This is a schematic diagram showing the changes in multiple loss functions and evaluation metrics during the training process of the target detection model provided in this application embodiment.

[0070] exist Figure 7 and Figure 8 In the diagram, the first image in the first row is a schematic diagram of train / box_loss (boundary box regression loss on the training set), the second image in the first row is a schematic diagram of train / cls_loss (classification loss on the training set), the third image in the first row is a schematic diagram of train / dfl_loss (distribution focus loss on the training set), the fourth image in the first row is a schematic diagram of metrics / precision(B) (precision on the training set), and the fifth image in the first row is a schematic diagram of metrics / recall(B) (recall on the training set). The first image in the second row is a schematic diagram of val / box_loss (boundary box regression loss on the validation set), the second image in the second row is a schematic diagram of val / cls_loss (classification loss on the validation set), the third image in the second row is a schematic diagram of val / dfl_loss (distribution focus loss on the validation set), the fourth image in the second row is a schematic diagram of metrics / mAP50(B) (mAP@0.5 on the validation set), and the fifth image in the second row is a schematic diagram of metrics / mAP50-95(B) (mAP@0.5:0.95 on the validation set).

[0071] It should be noted that, Figure 7 The x-axis of all graphs in the diagram represents the training epoch (round number), ranging from 0 to 100. Figure 7The training / box_loss decreased from 4.5 to 1.5, showing a stable downward trend, indicating that the model's learning effect on target location is good; the training / cls_loss decreased from 5.0 to 2.0, also showing a downward trend, indicating that the classification ability is gradually improving; the training / dfl_loss decreased from 3.0 to 1.5, with a stable decrease, indicating that the model is effective in optimizing coordinate distribution; the val / box_loss decreased from 5.0 to 2.0, slightly higher than the training loss, but with the same trend, indicating good generalization ability; the val / cls_loss fluctuated sharply from 10.0 and then stabilized at around 1.0, the high initial value may be due to differences in data distribution or initialization issues; the val / dfl_loss decreased from 2.5 to 1.7, with small fluctuations, consistent with the training trend, indicating that the model can also effectively optimize on the validation set; metrics / precision(B) increased from 0.2 to 0.6 with slight fluctuations, showing a clear overall upward trend; metrics / recall(B) increased from 0.2... The increase was relatively slow but continued to improve; metrics / mAP50(B) rose from 0.0 to 0.5, steadily improving and reaching an above-average level (generally >0.5 is considered good); metrics / mAP50-95(B) rose from 0.0 to 0.25, which, although lower than mAP50, also continued to grow, indicating that the model has a certain ability even at high IoU.

[0072] It should be noted that, Figure 8The horizontal axis of all graphs represents the training epochs, ranging from 0 to over 200. The train / box_loss decreased steadily from 1.6 to 1.0, showing a gradual decline followed by a plateau, indicating that the model's ability to locate objects has largely converged. The train / cls_loss decreased significantly from 2.0 to 1.0, with slight fluctuations later but overall stability, indicating a good improvement in classification ability. The train / dfl_loss decreased from 1.4 to 1.0, showing a continuous decrease followed by stabilization, indicating good performance in learning coordinate distributions. The val / box_loss decreased from 2.0 to 1.2, slightly higher than the training loss, but with a consistent trend, indicating strong generalization ability. The val / cls_loss decreased rapidly from 2.5 to 1.0 initially, with slight fluctuations later, showing an overall downward trend, indicating that classification ability has also improved on the validation set. The val / dfl_loss decreased smoothly from 1.6 to 1.2, consistent with the training trend, showing no obvious signs of overfitting. The metrics / precision(B) increased from 0.3 to 0.7 with slight fluctuations, showing a steady upward trend overall, indicating a reduction in false positives and improved prediction quality. The metrics / recall(B) increased from 0.5 to 0.8, a rapid and stable increase, indicating a reduction in false negatives and an expansion of detection coverage. The metrics / mAP50(B) increased from 0.3 to 0.7, showing continuous growth and eventually reaching a high level (generally >0.5 is considered excellent), indicating good model detection performance. The metrics / mAP50-95(B) increased from 0.2 to 0.4, although lower than mAP50, it still showed a significant improvement, indicating strong capabilities even under strict matching conditions.

[0073] contrast Figure 7 and Figure 8 It is known that the related technologies have high starting values ​​and slow convergence speed, while the present application has low starting values ​​and fast convergence speed.

[0074] In this embodiment, through a dual-detector design and an improved positive sample matching strategy, the false negative rate is reduced from 3%-5% to below 0.5%, preventing micro-vibration events from escalating into major accidents due to unidentified events. The SPPF_LSKA module can simultaneously capture the features of micro-vibrations (small scale) and destructive signals (large scale), achieving an identification accuracy of ≥96% under different construction vibrations. The lightweight design of the RepVGG convolutional block reduces the number of model parameters by 20%, shortening the processing time of a single time-frequency spectrum from 0.5s to 0.25s, meeting the real-time monitoring requirements for long-distance data inference. The application of diverse datasets and the ELU activation function enables the model to work stably under different pipe diameters, media, and interference environments, demonstrating better adaptability than existing methods.

[0075] In one specific embodiment, Figure 9 This is a schematic flowchart of an optical fiber anomaly detection method provided in an embodiment of this application, such as... Figure 9 As shown, the specific steps include: Step 1: Input the target's two-dimensional time-frequency spectrum into Backbone to obtain the first target feature map; Step 2: Input the first target feature map into Neck to obtain the second target feature map; Step 3: Input the first target feature map and the second target feature map into the main detection head and the auxiliary detection head respectively to obtain the fiber optic anomaly detection results. The fiber optic anomaly detection results include event category, location coordinates and confidence level. Step 4: Visualize alarms based on fiber optic anomaly detection results.

[0076] Corresponding to the fiber optic anomaly detection method provided in the embodiments of this application, the embodiments of this application also provide a fiber optic anomaly detection device, such as... Figure 10 As shown, the fiber optic anomaly detection device includes: The acquisition module 1001 is used to acquire data from the sensing optical fiber deployed in the monitoring area using DAS to obtain target DAS data. Generation module 1002 is used to generate a target two-dimensional time-frequency spectrum based on the target DAS data; Input module 1003 is used to input the target's two-dimensional time-frequency spectrum into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. Feature representation during convolutional blocks.

[0077] Corresponding to the fiber optic anomaly detection method provided in the embodiments of this application, the embodiments of this application also provide an electronic device for performing the fiber optic anomaly detection method, such as... Figure 11As shown, the electronic device includes: a processor 1101; and a memory 1102 for storing a program for the fiber optic anomaly detection method. After the device is powered on and the program for the fiber optic anomaly detection method is run by the processor, the following steps are performed: Data is collected from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data; Generate a two-dimensional time-frequency spectrum of the target based on the target DAS data; The target's two-dimensional time-frequency spectrum is input into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block.

[0078] Corresponding to the fiber optic anomaly detection method provided in the embodiments of this application, the embodiments of this application also provide a computer-readable storage medium storing a program for the fiber optic anomaly detection method, which is executed by a processor to perform the following steps: Data is collected from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data; Generate a two-dimensional time-frequency spectrum of the target based on the target DAS data; The target's two-dimensional time-frequency spectrum is input into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block.

[0079] Corresponding to the fiber optic anomaly detection method provided in the embodiments of this application, the embodiments of this application also provide a computer program containing instructions, which, when executed by a computer, cause the computer to perform the following steps: Data is collected from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data; Generate a two-dimensional time-frequency spectrum of the target based on the target DAS data; The target's two-dimensional time-frequency spectrum is input into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract the spatial local features of the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block uses structural reparameterization to fuse multiple branches from the training phase into an equivalent 3×3 convolutional layer. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting the first target feature map. The target input feature map is the feature representation of the target's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block.

[0080] It should be noted that for a detailed description of the fiber optic anomaly detection device, electronic device, computer-readable storage medium, and computer program provided in the embodiments of this application, please refer to the relevant description of the fiber optic anomaly detection method embodiments provided in the embodiments of this application, which will not be repeated here.

[0081] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

[0082] In a typical configuration, an electronic device includes one or more processors (Central Processing Units), input / output interfaces, network interfaces, and memory.

[0083] Memory may include non-persistent storage in computer-readable media, such as random access memory and / or non-volatile memory, like read-only memory or flash memory. Memory is an example of computer-readable media.

[0084] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable operations, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PCM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital video disc (DMCD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0085] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, compact disc read-only memory, optical storage, etc.) containing computer-usable program code.

[0086] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for detecting optical fiber anomalies, characterized in that, include: Data is collected from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data; Generate a target two-dimensional time-frequency spectrum based on the target DAS data; The target two-dimensional time-frequency spectrum is input into the target detection model to obtain the fiber optic anomaly detection result. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During the training phase, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract spatial local features from the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the RepVGG convolutional block fuses multiple branches from the training phase into an equivalent 3×3 convolutional layer through structural reparameterization. This equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map, outputting a first target feature map. The target input feature map is the feature representation of the target two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. Feature representation during convolutional blocks.

2. The fiber optic anomaly detection method according to claim 1, characterized in that, The RepVGG convolutional block also includes a residual branch and a first feature fusion unit; The residual branch is configured as an identity mapping, through which the sample input feature map is passed; The first feature fusion unit adds the output features of each branch element by element to generate a sample output feature map.

3. The fiber optic anomaly detection method according to claim 1, characterized in that, The Neck includes an improved SPPF module, which includes a first 1×1 convolutional layer, a parallel 3×3 max pooling layer, a 5×5 max pooling layer and a 7×7 max pooling layer, a second feature fusion unit and an LSKA attention module; The first target feature map is input into the first 1×1 convolutional layer to reduce the dimensionality of the first target feature map, resulting in the dimensionality-reduced first target feature map. The first target feature map after dimensionality reduction is input into the 3×3 max pooling layer, the 5×5 max pooling layer and the 7×7 max pooling layer respectively to obtain the output of each pooling layer. The resolution of the output of each pooling layer is consistent with the input resolution of each pooling layer, and the number of channels of the output of each pooling layer is consistent with the number of channels of the input of each pooling layer. The outputs of each pooling layer are spliced ​​along the channel dimension by the second feature fusion unit to obtain a multi-scale fused feature map. The multi-scale fused feature map is input into the LSKA attention module, and after channel splitting, depthwise separable convolution processing and concatenation, an intermediate feature map is output. A sigmoid function is applied to the intermediate feature map to generate an attention weight map.

4. The fiber optic anomaly detection method according to claim 3, characterized in that, The LSKA attention module includes a channel splitting unit, a second 1×1 convolutional layer, a 7×7 depth-separable convolutional layer, a third 1×1 convolutional layer, and a third feature fusion unit. The multi-scale fusion feature map is split by the channel splitting unit to obtain multiple multi-scale fusion feature map branches; The multi-scale fusion feature map branch is reduced in dimensionality along the channel dimension by the second 1×1 convolutional layer to obtain the dimensionality-reduced multi-scale fusion feature map branch. The 7×7 depth-separable convolutional layer is used to perform long-distance spatial modeling on the dimensionality-reduced multi-scale fused feature map branch to obtain the enhanced feature map branch. The enhanced feature map branch is increased in dimension along the channel dimension by the third 1×1 convolutional layer to obtain the enhanced feature map branch with increased dimension. The enhanced feature map branches after dimensionality increase are spliced ​​together by the third feature fusion unit to obtain an intermediate feature map, the number of channels of which is consistent with the number of channels of the multi-scale fusion feature map.

5. The fiber optic anomaly detection method according to claim 4, characterized in that, The LSKA attention module also includes an attention application unit; The attention application unit performs element-wise multiplication of the attention weight map and the dimensionality-reduced first target feature map to output the second target feature map.

6. The fiber optic anomaly detection method according to claim 5, characterized in that, The Head includes a main detection head, an auxiliary detection head, and a detection result fusion unit; The first target feature map is input into the main detection head to obtain the first detection result; The second target feature map is input into the auxiliary detection head to obtain the second detection result; The detection result fusion unit fuses the first detection result and the second detection result to obtain the fiber optic anomaly detection result.

7. The fiber optic anomaly detection method according to claim 1, characterized in that, Also includes: Determine the center grid of the target bounding box; When the intersection-union ratio of the adjacent grids of the target bounding box center grid to the target bounding box center grid is greater than a preset threshold, the adjacent grids are determined as positive samples.

8. A fiber optic anomaly detection device, characterized in that, include: The acquisition module is used to acquire data from the sensing optical fibers deployed in the monitoring area using DAS to obtain target DAS data. The generation module is used to generate a target two-dimensional time-frequency spectrum based on the target DAS data; An input module is used to input the target's two-dimensional time-frequency spectrum into a target detection model to obtain fiber optic anomaly detection results. The target detection model is an improved YOLOv8 model, which includes a Backbone, Neck, and Head. The Backbone includes a RepVGG convolutional block. During training, the RepVGG convolutional block includes a main branch and at least one auxiliary branch. The main branch uses a 3×3 convolutional layer to extract spatial local features from the sample input feature map, and the auxiliary branch uses a 1×1 convolutional layer to perform channel transformation on the sample input feature map. The sample input feature map is the feature representation of the sample's two-dimensional time-frequency spectrum when input into the RepVGG convolutional block. During the inference phase, the convolutional block merges multiple branches from the training phase into an equivalent 3×3 convolutional layer through structural reparameterization. The equivalent 3×3 convolutional layer performs a linear transformation on the target input feature map and outputs a first target feature map. The target input feature map is the feature representation of the target two-dimensional time-frequency spectrum when it is input into the RepVGG convolutional block.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the fiber optic anomaly detection method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the fiber optic anomaly detection method according to any one of claims 1-7.

Images