Distributed optical fiber sensing signal processing method and system

Abstract

The invention relates to the technical field of optical fiber sensing systems, and discloses a distributed optical fiber sensing signal processing method and system. The method comprises: collecting sensing signals of different events through distributed optical fibers and preprocessing the sensing signals; improving a YOLOv8 model to obtain a target detection model; training the target detection model through normal spatiotemporal data images and abnormal spatiotemporal data images to obtain a trained target detection model; applying the model; replacing a Bottleneck module in a C2F module of the YOLOv8 model with a combination of a fastnet module and an MRFA module, and introducing a deep separable convolution and a multi-receptive field attention mechanism into the improved YOLOv8 model backbone, so as to better utilize the computing power of a device, greatly reduce the computing complexity of network learning, reduce the number of parameters, and improve network efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber sensing systems, and in particular to a distributed optical fiber sensing signal processing method and system. Background Art

[0002] Traditional distributed fiber optic sensing systems (DFOS) mainly focus on perception and lack efficient processing of perception information. DFOS usually includes distributed fiber acoustic sensors (DAS) and distributed fiber vibration sensing systems (DVS). DFOS uses the backward Rayleigh scattered light beam in the optical fiber to locate and restore the changes in environmental physical quantities at any position on the optical fiber link. It has the advantages of strong environmental adaptability, large sensing range, and high spatiotemporal resolution accuracy. The distributed fiber optic event monitoring system not only focuses on perception, but also pays attention to the classification and identification of real-time dynamic perception information. It expresses the collected perception information into a spatiotemporal graph, then finds abnormal event points from the massive noisy spatiotemporal data graph, and conducts real-time monitoring and early warning.

[0003] With the continuous development of artificial intelligence technology and computer vision, image target detection technology based on deep learning has also emerged. A large number of studies have shown that models built based on deep learning methods can be applied to distributed fiber event monitoring systems, and can achieve good performance for target detection problems in spatiotemporal data images. However, due to the characteristics of distributed fiber optic sensor spatiotemporal data such as high noise, mixed data, and huge data volume, the existing network cannot meet its real-time monitoring needs. For example, the speed of network forward inference is not fast enough, it is difficult to fully mine effective information, and it is difficult to detect small targets. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a distributed optical fiber sensing signal processing method and system.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A distributed optical fiber sensing signal processing method, characterized in that it comprises the following steps:

[0007] Step 1: collect sensor signals of different events through distributed optical fibers, pre-process the sensor signals, and obtain multiple spatiotemporal data images X;

[0008] Step 2: Improve the YOLOv8 model to obtain a target detection model; the improvements to the YOLOv8 model include replacing the C2F module in the YOLOv8 model with a custom C2F_FB_MRFA module, and replacing the convolutional layer of the YOLOv8 model backbone network with a custom DwConv_MRFA layer;

[0009] The Bottleneck module in the C2F module of the YOLOv8 model is replaced with a combination of the FastNet module and the Multi-Receptive Field Attention module to obtain the C2F_FB_MRFA module; the Multi-Receptive Field Attention module is used to weight the input feature map according to specific rules and output the weighted feature map;

[0010] The standard convolution operation in the convolution layer is replaced by a depth-wise separable convolution operation, and the result of the convolution operation is input into the multi-receptive field attention module before being output to the next layer, thereby obtaining the DwConv_MRFA layer;

[0011] Step 3: divide the spatiotemporal data image X into a normal spatiotemporal data image and an abnormal spatiotemporal data image containing events that affect the cable by manual comparison, and then select and mark the events in the abnormal spatiotemporal data image according to the types of different events; train the target detection model through the normal spatiotemporal data image and the abnormal spatiotemporal data image to obtain a trained target detection model;

[0012] Step 4: Application of target detection model: Input the spatiotemporal data image X0 of the sensor signal collected in real time by the distributed optical fiber sensor into the trained target detection model to obtain the spatiotemporal data image with the event location marked, and then convert the event location in the spatiotemporal data image into the actual cable location.

[0013] Furthermore, the C2F_FB_MRFA module includes two convolutional layers Conv1 and Conv2, a feature segmentation layer, a plurality of FB_MRFA modules connected in series, and a feature concatenation layer; wherein the convolutional layer Conv1 represents a set of operations, so that the feature map input to the C2F_FB_MRFA module undergoes a standard convolution operation, a batch normalization operation, and a SiLU activation function in sequence; the feature segmentation layer divides the feature map output by the convolutional layer Conv1 into two feature maps according to the number of channels, and the number of channels of each feature map is half of the feature map output by the convolutional layer Conv1, and the two feature maps output by the feature segmentation layer Figure 1 One is input to the first FB_MRFA module, and the other is input to the feature concatenation layer; the FB_MRFA module is a fast feature fusion module, which is used to adjust the weight of the feature map; the feature concatenation layer concatenates a feature map separated by the feature segmentation layer and the feature maps output by each FB_MRFA module to obtain the feature map output by the C2F_FB_MRFA module.

[0014] Furthermore, the FB_MRFA module is obtained by combining a fast net module and a multi-receptive field attention module; specifically, the FB_MRFA module includes a partial convolution operator, a convolution layer, a batch normalization operation, a ReLU activation function and a multi-receptive field attention module, wherein the partial convolution operator, the convolution layer, the batch normalization operation, and the ReLU activation function constitute the fast net module; the feature map input to the FB_MRFA module passes through the partial convolution operator, the convolution layer, the batch normalization operation, the ReLU activation function and the multi-receptive field attention module in sequence, and then the obtained feature map tensor is added to the feature map tensor input to the FB_MRFA module to obtain the feature map output by the FB_MRFA module.

[0015] Furthermore, it is characterized in that the multi-receptive field attention module includes two 1x1 standard convolution branches and 3x3, 5x5, and 7x7 depth-separable convolution branches, and the depth-separable convolution branches are placed in parallel with the standard convolution branches;

[0016] For any given input to the feature map of the multi-receptive field attention module represents the real number domain, W, H, C represent the width, height and number of channels of the feature map x, and divide the feature map x into G sub-feature map groups x0, x1, ..., x along the channel dimension. g ,…,x G-1 , to learn different semantics,

[0017] The multi-receptive field attention module encodes channels along two spatial directions in two 1x1 standard convolution branches, adopts two 1D global average pooling operations, and stacks 3x3, 5x5, and 7x7 depth-wise separable convolutions in the depth-wise separable convolution branch to capture multi-scale feature representations;

[0018] In the two standard convolution branches, after average pooling in two spatial directions, they are respectively subjected to the Sigmoid activation function, and the obtained results are weighted to re-assign the sub-feature map groups. The obtained feature maps are then group normalized to obtain the feature map x 11 ; Then perform average pooling and pass through the Softmax activation function to obtain the feature map x 12 ;

[0019] In the depthwise separable convolution branch, the feature map after a 3x3 depthwise separable convolution is denoted as x 21 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 22 ; The feature map that has undergone 3x3 and 5x5 depth-separable convolutions is denoted as x 31, then perform average pooling and pass through the Softmax activation function to obtain the feature map x 32 ; The feature map that has been sequentially processed by 3x3, 5x5, and 7x7 depth-separable convolutions is denoted as x 41 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 42 ;

[0020] Sequentially change x 11 Multiply by x 22 、x 11 Multiply by x 32 、x 11 Multiply by x 42 , x 12 Multiply by x 21 、x 12 Multiply by x 31 、x 12 Multiply by x 41 ; Then add the above results in sequence, and then pass through the Sigmoid activation function to finally form a weight parameter to reassign the sub-feature map and output a weighted feature map

[0021] A distributed optical fiber sensing signal processing system, comprising:

[0022] The acquisition module collects the sensor signals of different events through distributed optical fibers, pre-processes the sensor signals, and obtains multiple spatiotemporal data images X;

[0023] The model building module improves the YOLOv8 model to obtain the target detection model; the improvements to the YOLOv8 model include replacing the C2F module in the YOLOv8 model with a custom C2F_FB_MRFA module, and replacing the convolutional layer of the YOLOv8 model backbone network with a custom DwConv_MRFA layer;

[0024] The Bottleneck module in the C2F module of the YOLOv8 model is replaced with a combination of the FastNet module and the Multi-Receptive Field Attention module to obtain the C2F_FB_MRFA module; the Multi-Receptive Field Attention module is used to weight the input feature map according to specific rules and output the weighted feature map;

[0025] The standard convolution operation in the convolution layer is replaced by a depth-wise separable convolution operation, and the result of the convolution operation is input into the multi-receptive field attention module before being output to the next layer, thereby obtaining the DwConv_MRFA layer;

[0026] The training module divides the spatiotemporal data image X into a normal spatiotemporal data image and an abnormal spatiotemporal data image containing events that affect the cable by manual comparison, and then selects and marks the events in the abnormal spatiotemporal data image according to the types of different events; trains the target detection model through the normal spatiotemporal data image and the abnormal spatiotemporal data image to obtain a trained target detection model;

[0027] Prediction module, target detection model application: The spatiotemporal data image X0 of the sensor signal collected in real time by the distributed optical fiber sensor is input into the trained target detection model to obtain the spatiotemporal data image with the event location marked, and then the event location in the spatiotemporal data image is converted into the actual cable location.

[0028] Compared with the prior art, the beneficial technical effects of the present invention are:

[0029] 1. The target detection model in the present invention is based on the improved version of the YOLOv8 model. It pre-trains the target detection model to obtain the model weights suitable for the custom monitoring data set scenario, and then uses the trained target detection model for event monitoring to identify the type and location information of the threat event. Compared with traditional monitoring methods, the method and system proposed in the present invention have lower costs and higher efficiency. It can reduce the demand for expensive equipment and human resources, while realizing real-time monitoring and early warning of transmission cables, reducing maintenance costs and risk management costs.

[0030] 2. The present invention proposes a novel simplified Multi-Receptive Field Attention (MRFA) module. The introduction of the multi-receptive field attention mechanism can enhance the network's feature extraction capability, reduce the impact of signal noise, and make the model pay more attention to the signals in the key areas, thereby improving the performance and generalization ability of the target detection model. Specifically, the multi-receptive field attention module can achieve the following functions: (1) Feature weighting: The multi-receptive field attention module can learn the importance weights of different positions or channels to weight the features of the input data. This can make the network more focused on information that is useful for the current task and suppress the response to irrelevant information, thereby improving the performance of the model. (2) Multi-scale perception: The multi-receptive field attention module can help the neural network perceive the input data at different receptive fields. By introducing the multi-scale attention mechanism, the model can better process features at different scales, thereby improving the robustness and generalization ability of the model. (3) Enhanced feature representation: The multi-receptive field attention mechanism helps to extract useful features from the input data and effectively represent them. By focusing on key areas or features, important patterns and structures in the data can be better captured, thereby improving the performance of the model. (4) Suppressing noise and interference: The multi-receptive field attention module can help the network suppress noise and interference, thereby improving the robustness of the model. By focusing on useful features, the model can reduce its sensitivity to noise and interference and improve its adaptability to complex environments.

[0031] 3. Replace the Bottleneck module (a special residual structure used to reduce the number of parameters, first reducing the dimension and then increasing the dimension) in the C2F module of the YOLOv8 model with a combination of the FasterNet Block and the MRFA module. The FasterNet Block is used to lighten the structure. The core of the FasterNet Block is the Partial Conv (PConv). PConv has lower floating point operations per second (FLOPs) than conventional convolution Conv and higher floating point operations per second than deep separable convolution (DWConv), which can better utilize the computing power of the device. MRFA is used to extract feature weights efficiently and lightly to make up for the information loss caused by the FasterNet Block.

[0032] 4. The depthwise separable convolution and multi-receptive field attention mechanism are introduced into the YOLOv8 model backbone network. The standard convolution operation in the convolution layer is replaced by the depthwise separable convolution operation, and the result of the convolution operation is input into the multi-receptive field attention module before being output to the next layer, resulting in the DwConv_MRFA layer, which will greatly reduce the computational complexity of network learning, reduce the number of parameters, and improve network efficiency. Specifically, the DwConv_MRFA layer can achieve the following functions: (1) Improve parameter efficiency: Since the depthwise separable convolution decomposes the convolution operation into two steps, it has fewer parameters. Compared with the standard convolution, the depthwise separable convolution can significantly reduce the number of model parameters, thereby reducing the risk of overfitting. However, the simple depthwise separable convolution cannot achieve efficient adjustment of the network parameter weights. The multi-receptive field attention mechanism realizes this adjustment process at the cost of very little parameter redundancy. (2) Improve computational efficiency: Since the depthwise separable convolution has fewer parameters, it requires less computation. This makes the depthwise separable convolution have better performance in scenarios with limited computing resources such as mobile devices. (3) Better representation capability: The DwConv_MRFA layer can combine features from different channels through a multi-receptive field attention mechanism, thereby improving the representation capability of the network. This combination can better capture the spatial and channel correlations in the input data, thereby improving the model’s representation capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of a multi-receptive field attention module in the present invention;

[0034] Figure 2 is a schematic diagram of the C2F_FB_MRFA module in the present invention;

[0035] Figure 3 is a schematic diagram of the FB_MRFA module in the present invention;

[0036] Figure 4 This is the structural diagram of the existing YOLOv8 model;

[0037] Figure 5 It is a structural diagram of the improved target detection model in the present invention;

[0038] Figure 6 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0039] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0040] like Figure 6 As shown, this embodiment discloses a distributed optical fiber sensing signal processing method, including steps S1 to S4, which are as follows:

[0041] S1. Collect distributed optical fiber sensor signal sets and perform preprocessing.

[0042] S11. Build a distributed fiber optic sensor signal acquisition platform to obtain sensor signals of events such as machine excavation (such as excavators digging at different locations), manual excavation, and vehicle passing, and obtain the original sensor signal set.

[0043] S12, preprocessing the collected original sensor signal set, including but not limited to compressing the data and performing spatiotemporal data slicing to obtain a large number of spatiotemporal data images X.

[0044] S2. Build a target detection model based on the YOLOv8 model. The improvements include customizing the multi-receptive field attention module, replacing the Bottleneck module in the C2F module of the YOLOv8 model with a combination of the FasterNet Block and the multi-receptive field attention module, and introducing the depthwise separable convolution and multi-receptive field attention mechanism into the improved YOLOv8 model backbone.

[0045] S21. Define a multi-receptive field attention module.

[0046] The function of the custom multi-receptive field attention module is to re-weight the input feature map according to a certain learnable rule. The structure of the multi-receptive field attention module is as follows: Figure 1 As shown in the figure, the English explanation in the legend is shown in Table 1. The multi-receptive field attention module includes two 1x1 standard convolution branches and 3x3, 5x5, and 7x7 depth-separable convolution branches. In order to aggregate multi-scale spatial structure information and for fast response, the depth-separable convolution branches are placed in parallel with the standard convolution branches.

[0047] Table 1

[0048]

[0049]

[0050] For any given input feature map represents the real number domain, W, H, C represent the width, height, and number of channels of the feature map x, MRFA divides x into G sub-feature groups along the channel dimension to learn different semantics, and all sub-feature groups can be represented by x0, x1, ..., x g ,…,x G-1 , To express.

[0051] The multi-receptive field attention module encodes channels along two spatial directions in the 1x1 standard convolution branch, adopts two 1D global average pooling operations, and stacks 3x3, 5x5, and 7x7 depth-wise separable convolutions in the depth-wise separable convolution branch to capture multi-scale feature representations.

[0052] First, in the standard convolution branch, after average pooling in two spatial directions, the Sigmoid activation function is used as the weight to reassign the original sub-feature group, and the obtained feature map is then group normalized to obtain the feature map x 11 ; Then perform average pooling and pass through the Softmax activation function to obtain the feature map x 12 .

[0053] Secondly, in the depth-wise separable convolution branch, the feature map after a 3x3 convolution is denoted as x 21 , then perform average pooling and pass through the Softmax activation function, recorded as feature map x 22 ; The feature map after a 3x3, 5x5 depth-separable convolution is denoted as x 31 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 32 ; The feature map after a 3x3, 5x5, 7x7 depth separable convolution is denoted as x 41 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 42 .

[0054] Sequentially change x 11 Multiply by x 22 、x 11 Multiply by x 32 、x 11 Multiply by x 42 , x 12 Multiply by x 21 、x 12 Multiply by x 31 、x 12 Multiply by x 41 ; Then add the above results in sequence, and then pass through the Sigmoid activation function to finally form a weight parameter to reassign the sub-feature map and output a weighted feature map

[0055] S22. Replace the Bottleneck module in the C2F module of the YOLOv8 model with a combination of the FastNet module and the MRFA module.

[0056] The custom C2F_FB_MRFA module is a combination of the C2F module, the FastNet module, and the Multi-Receptive Field Attention module. Figure 2 As shown, Figure 2 The English explanations are shown in Table 2. The C2F_FB_MRFA module includes two Conv convolutional layers, a feature segmentation layer, an FB_MRFA module, and a feature splicing layer. The convolutional layer represents a set of operations, which makes the input feature map pass through a standard convolution Conv2d, a batch normalization operation BatchNorm2d, and an activation function SiLU in sequence; the feature segmentation layer divides the input feature map into two feature maps according to the number of channels, and the number of channels of each feature map is half of the total number; the FB_MRFA module is a fast feature fusion module, which quickly adjusts the weight of the input feature map in a way with lower computational cost; the feature map splices together several inputs connected to it to fuse several types of information with each other.

[0057] Table 2

[0058]

[0059] The customized FB_MRFA module is a combination of the FastNet module and the MRFA module. Figure 3 As shown, Figure 3 The English explanations are shown in Table 3. The FB_MRFA module includes partial convolution PConv 3×3, Conv 1×1, multi-receptive field attention module, batch normalization operation and ReLU activation function. The combination of the fast net module and the MRFA module is shown in the figure. It adds a multi-receptive field attention module to the partial convolution branch, and allows the feature map information to flow through the multi-receptive field attention module before concatenating the two parts of the feature map.

[0060] Table 3

[0061] English label Chinese explanation English label Chinese explanation Input / output Input / output feature map PConv 3×3 Partial convolution operator, convolution kernel size 3×3 MRFA Multi-receptive field attention module Conv Standard convolution BN 2D Batch Normalization Operation ReLU ReLU activation function

[0062] S23. Introduce the depthwise separable convolution and multi-receptive field attention mechanism into the backbone network of the YOLOv8 model.

[0063] The YOLOv8 model is the latest YOLO series target detection algorithm launched by Ultralvtics. It can be used for tasks such as image classification, object detection, and instance segmentation. Its main structure is as follows Figure 4 The algorithm mainly consists of a convolutional layer, a C2F module, a fast spatial pyramid pooling module (SPPF), an upsample layer (Upsample), a concatenation layer (Concat), and a detection layer (Detect). In order to make the model more lightweight, the present invention proposes a lightweight improvement scheme, the main structure of which is as follows Figure 5 As shown in the figure. First, replace all the original convolutional layers with DwConv_MRFA layers. The DwConv_MRFA layer replaces the standard convolution in the original Conv layer with a depth-separable convolution. In addition, the result of the depth-separable convolution flows through a custom multi-receptive field attention module before being output to the next layer. Secondly, replace all C2F modules in the network with a lighter custom C2F_FB_MRFA module, whose structure is as follows: Figure 5 shown.

[0064] S3. Label the data and perform network training.

[0065] By manual comparison, the large number of spatiotemporal data images X obtained in step S1 are divided into normal spatiotemporal data images and abnormal spatiotemporal data images containing special events such as machine mining, manual mining, and vehicle passing, and then the events in the spatiotemporal data images are boxed and marked according to different event types. After marking enough data samples, the improved target detection model is used to train the model. After the model training is completed, appropriate pruning operations are performed, and then the trained weights are used for offline data monitoring, and a part of the data with more accurate monitoring results is screened out, and then combined with this part of the data, the data set is expanded, and the model training is performed again for multiple times to obtain better training results.

[0066] S4. Model application.

[0067] The spatiotemporal data image X0 of the sensor signal collected in real time by the distributed optical fiber sensor is input into the trained target detection model to obtain the spatiotemporal data image with the event location marked, and then the location information in the predicted label is converted into the actual cable location.

[0068] The system in the present invention corresponds to the method, and the specific technical solutions applicable to the method are also applicable to the system.

[0069] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0070] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A distributed optical fiber sensing signal processing method, characterized in that: The following steps are involved: Step 1: collect sensor signals of different events through distributed optical fiber, pre-process the sensor signals, and obtain multiple spatiotemporal data images X; Step 2: Improve the YOLOv8 model to obtain a target detection model; the improvements to the YOLOv8 model include replacing the C2F module in the YOLOv8 model with a custom C2F_FB_MRFA module, and replacing the convolutional layer of the YOLOv8 model backbone network with a custom DwConv_MRFA layer; The Bottleneck module in the C2F module of the YOLOv8 model is replaced with a combination of the FastNet module and the Multi-Receptive Field Attention module to obtain the C2F_FB_MRFA module; the Multi-Receptive Field Attention module is used to weight the input feature map according to specific rules and output the weighted feature map; The standard convolution operation in the convolution layer is replaced by a depth-wise separable convolution operation, and the result of the convolution operation is input into the multi-receptive field attention module before being output to the next layer, thereby obtaining the DwConv_MRFA layer; Step 3: divide the spatiotemporal data image X into a normal spatiotemporal data image and an abnormal spatiotemporal data image containing events that affect the cable by manual comparison, and then select and mark the events in the abnormal spatiotemporal data image according to the types of different events; The target detection model is trained by using normal spatiotemporal data images and abnormal spatiotemporal data images to obtain a trained target detection model; Step 4: Application of target detection model: Input the spatiotemporal data image X0 of the sensor signal collected in real time by the distributed optical fiber sensor into the trained target detection model to obtain the spatiotemporal data image with the event location marked, and then convert the event location in the spatiotemporal data image into the real cable location; The C2F_FB_MRFA module includes two convolutional layers Conv1 and Conv2, a feature segmentation layer, a plurality of FB_MRFA modules connected in series, and a feature splicing layer; wherein the convolutional layer Conv1 represents a set of operations, so that the feature map input to the C2F_FB_MRFA module undergoes a standard convolution operation, a batch normalization operation, and a SiLU activation function in sequence; the feature segmentation layer divides the feature map output by the convolutional layer Conv1 into two feature maps according to the number of channels, and the number of channels of each feature map is half of the feature map output by the convolutional layer Conv1, and one of the two feature maps output by the feature segmentation layer is input to the first FB_MRFA module, and the other is input to the feature splicing layer; the FB_MRFA module is a fast feature fusion module for adjusting the weight of the feature map; the feature splicing layer splices a feature map separated by the feature segmentation layer and the feature maps output by each FB_MRFA module together to obtain the feature map output by the C2F_FB_MRFA module; The FB_MRFA module is obtained by combining a fast net module and a multi-receptive field attention module; specifically, the FB_MRFA module includes a partial convolution operator, a convolution layer, a batch normalization operation, a ReLU activation function and a multi-receptive field attention module, wherein the partial convolution operator, the convolution layer, the batch normalization operation and the ReLU activation function constitute the fast net module; the feature map input to the FB_MRFA module is sequentially passed through the partial convolution operator, the convolution layer, the batch normalization operation, the ReLU activation function and the multi-receptive field attention module, and then the obtained feature map tensor is added to the feature map tensor input to the FB_MRFA module to obtain the feature map output by the FB_MRFA module; The multi-receptive field attention module includes two 1x1 standard convolution branches and 3x3, 5x5, and 7x7 depth-separable convolution branches, and the depth-separable convolution branches are placed in parallel with the standard convolution branches; For any given input to the feature map of the multi-receptive field attention module represents the real number domain, W, H, C represent the width, height and number of channels of the feature map x, and divide the feature map x into G sub-feature map groups x0, x1, ..., x along the channel dimension. g ,…,x G-1 , to learn different semantics, The multi-receptive field attention module encodes channels along two spatial directions in two 1x1 standard convolution branches, adopts two 1D global average pooling operations, and stacks 3x3, 5x5, and 7x7 depth-wise separable convolutions in the depth-wise separable convolution branch to capture multi-scale feature representations; In the two standard convolution branches, after average pooling in two spatial directions, they are respectively subjected to the Sigmoid activation function, and the obtained results are weighted to re-assign the sub-feature map groups. The obtained feature maps are then group normalized to obtain the feature map x 11 ; Then perform average pooling and pass through the Softmax activation function to obtain the feature map x 12 ; In the depthwise separable convolution branch, the feature map after a 3x3 depthwise separable convolution is denoted as x 21 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 22 ; The feature map that has undergone 3x3 and 5x5 depth-separable convolutions is denoted as x 31 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 32 ; The feature map that has been sequentially processed by 3x3, 5x5, and 7x7 depth-separable convolutions is denoted as x 41 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 42 ; Sequentially change x 11 Multiply by x 22 、x 11 Multiply by x 32 、x 11 Multiply by x 42 , x 12 Multiply by x 21 、x 12 Multiply by x 31 、x 12 Multiply by x 41 ; Then add the above results in sequence, and then pass through the Sigmoid activation function to finally form a weight parameter to reassign the sub-feature map and output a weighted feature map 2. A distributed optical fiber sensing signal processing system, characterized in that: include: The acquisition module collects the sensor signals of different events through distributed optical fibers, pre-processes the sensor signals, and obtains multiple spatiotemporal data images X; The model building module improves the YOLOv8 model to obtain the target detection model; the improvements to the YOLOv8 model include replacing the C2F module in the YOLOv8 model with a custom C2F_FB_MRFA module, and replacing the convolutional layer of the YOLOv8 model backbone network with a custom DwConv_MRFA layer; The Bottleneck module in the C2F module of the YOLOv8 model is replaced with a combination of the FastNet module and the Multi-Receptive Field Attention module to obtain the C2F_FB_MRFA module; the Multi-Receptive Field Attention module is used to weight the input feature map according to specific rules and output the weighted feature map; The standard convolution operation in the convolution layer is replaced by a depth-wise separable convolution operation, and the result of the convolution operation is input into the multi-receptive field attention module before being output to the next layer, thereby obtaining the DwConv_MRFA layer; The training module divides the spatiotemporal data image X into a normal spatiotemporal data image and an abnormal spatiotemporal data image containing events that affect the cable by manual comparison, and then selects and marks the events in the abnormal spatiotemporal data image according to the types of different events; The target detection model is trained by using normal spatiotemporal data images and abnormal spatiotemporal data images to obtain a trained target detection model; Prediction module, target detection model application: input the spatiotemporal data image X0 of the sensor signal collected in real time by the distributed optical fiber sensor into the trained target detection model to obtain the spatiotemporal data image with the event location marked, and then convert the event location in the spatiotemporal data image into the real cable location; The C2F_FB_MRFA module includes two convolutional layers Conv1 and Conv2, a feature segmentation layer, a plurality of FB_MRFA modules connected in series, and a feature splicing layer; wherein the convolutional layer Conv1 represents a set of operations, so that the feature map input to the C2F_FB_MRFA module undergoes a standard convolution operation, a batch normalization operation, and a SiLU activation function in sequence; the feature segmentation layer divides the feature map output by the convolutional layer Conv1 into two feature maps according to the number of channels, and the number of channels of each feature map is half of the feature map output by the convolutional layer Conv1, and one of the two feature maps output by the feature segmentation layer is input to the first FB_MRFA module, and the other is input to the feature splicing layer; the FB_MRFA module is a fast feature fusion module for adjusting the weight of the feature map; the feature splicing layer splices a feature map separated by the feature segmentation layer and the feature maps output by each FB_MRFA module together to obtain the feature map output by the C2F_FB_MRFA module; The FB_MRFA module is obtained by combining a fast net module and a multi-receptive field attention module; specifically, the FB_MRFA module includes a partial convolution operator, a convolution layer, a batch normalization operation, a ReLU activation function and a multi-receptive field attention module, wherein the partial convolution operator, the convolution layer, the batch normalization operation and the ReLU activation function constitute the fast net module; the feature map input to the FB_MRFA module is sequentially passed through the partial convolution operator, the convolution layer, the batch normalization operation, the ReLU activation function and the multi-receptive field attention module, and then the obtained feature map tensor is added to the feature map tensor input to the FB_MRFA module to obtain the feature map output by the FB_MRFA module; The multi-receptive field attention module includes two 1x1 standard convolution branches and 3x3, 5x5, and 7x7 depth-separable convolution branches, and the depth-separable convolution branches are placed in parallel with the standard convolution branches; For any given input to the feature map of the multi-receptive field attention module represents the real number domain, W, H, C represent the width, height and number of channels of the feature map x, and divide the feature map x into G sub-feature map groups x0, x1, ..., x along the channel dimension. g ,…,x G-1 , to learn different semantics, The multi-receptive field attention module encodes channels along two spatial directions in two 1x1 standard convolution branches, adopts two 1D global average pooling operations, and stacks 3x3, 5x5, and 7x7 depth-wise separable convolutions in the depth-wise separable convolution branch to capture multi-scale feature representations; In the two standard convolution branches, after average pooling in two spatial directions, they are respectively subjected to the Sigmoid activation function, and the obtained results are weighted to re-assign the sub-feature map groups. The obtained feature maps are then group normalized to obtain the feature map x 11 ; Then perform average pooling and pass through the Softmax activation function to obtain the feature map x 12 ; In the depthwise separable convolution branch, the feature map after a 3x3 depthwise separable convolution is denoted as x 21 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 22 ; The feature map that has undergone 3x3 and 5x5 depth-separable convolutions is denoted as x 31 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 32 ; The feature map that has been sequentially processed by 3x3, 5x5, and 7x7 depth-separable convolutions is denoted as x 41 , then perform average pooling and pass through the Softmax activation function to obtain the feature map x 42 ; Sequentially change x 44 Multiply by x 22 、x 44 Multiply by x 32 、x 11 Multiply by x 42 , x 12 Multiply by x 21 、x 12 Multiply by x 31 、x 12 Multiply by x 41 ; Then add the above results in sequence, and then pass through the Sigmoid activation function to finally form a weight parameter to reassign the sub-feature map and output a weighted feature map

Images