A pipeline leakage identification method based on distributed optical fiber acoustic wave sensing system

Abstract

The present invention relates to the field of distributed optical fiber acoustic wave sensing, and specifically to a pipeline leakage identification method based on a distributed optical fiber acoustic wave sensing system, comprising the following steps: S1: using a distributed optical fiber acoustic wave sensing system to collect data, and downsampling the collected original signal data; S2: performing an image operation on the downsampled signal, denoising the converted image through ADNet, highlighting the leakage signal characteristics, marking the denoised image with a frame, and establishing a data set with the final image; S3: constructing an improved YOLOv8 network model, namely RAL-YOLOv8; S4: dividing the data set into a training set, a test set, and a verification set, and using the training set to train the improved YOLOv8 network to obtain a leakage identification model. The present invention greatly reduces the underreporting rate of gas pipeline leakage event identification and ensures the accuracy of identification.

Description

Technical Field

[0001] The invention relates to the field of distributed optical fiber acoustic wave sensing, in particular to a pipeline leakage identification method based on a distributed optical fiber acoustic wave sensing system. Background Art

[0002] The distributed fiber acoustic sensing (DAS) system is based on a phase-sensitive optical time-domain reflectometer (φ-OTDR) to detect the backward Rayleigh scattered signal generated when the coherent pulse light propagates in the optical fiber, and detect and reconstruct the leakage event accordingly. It is mainly composed of optical fiber, optical signal amplifier / demodulator, high coherence pulse laser source and data processing analyzer. At present, in the field of distributed optical fiber sensing (DOFS), there are two main types of demodulation technologies for sensing optical fiber signals: scattering type and interference type. The φ-OTDR technology that demodulates the signal according to the phase of scattered light is currently one of the most sensitive technical means in distributed optical fiber sensing.

[0003] The occurrence of an external pipeline leakage event will affect the phase of the backward Rayleigh scattered light in the optical fiber, and the optical signal demodulator can detect the phase change of the scattered light based on the change in the optical path before and after the leakage event. Since the sensing optical fiber can perform distributed detection, it allows the sensing channel that monitors the leakage to accurately locate the location of the leakage event.

[0004] Through research and investigation, it is found that the use of machine learning and deep learning recognition methods to identify gas pipeline leakage events has gradually replaced the traditional gas pipeline leakage detection method. However, in the current monitoring method based on distributed optical fiber acoustic wave sensing system, the use of machine learning and deep learning recognition methods still has certain problems:

[0005] (1) The gas pipeline laying environment is complex and the noise interference is serious. Small leakage signals are easily buried in the noise, resulting in a high system underreporting rate, which has not yet been solved;

[0006] (2) Previous recognition algorithms only cared about accuracy. However, since gas pipeline leakage has serious consequences and great harm, it is even more important to reduce the false negative rate in practical applications, that is, to improve the recall rate of the model.

[0007] (3) Due to the complexity of the external environment, there are many types of disturbance signals acting on the optical fiber, which makes it difficult to accurately identify the disturbance signals, resulting in a high false alarm rate for events that has not yet been solved;

[0008] (4) Long-distance pipelines are prone to multiple events occurring simultaneously, resulting in the problem of untimely monitoring response, which still has no good solution. Summary of the invention

[0009] In order to overcome the technical problems, the present invention provides a pipeline leakage identification method based on a distributed fiber optic acoustic wave sensing system. The method is improved based on the advanced YOLOv8 network as the basic framework, and uses the distributed fiber optic acoustic wave sensing system to collect data when the pipeline leaks, and inputs the spatiotemporal image of the gas pipeline leakage data into an attention-oriented denoising convolutional neural network (ADNet) for denoising and enhancement, and then uses the processed image to train the improved YOLOv8 model, ultimately realizing real-time monitoring of gas pipeline leakage events.

[0010] The present invention is achieved through the following technical solutions:

[0011] A pipeline leakage identification method based on a distributed optical fiber acoustic wave sensing system specifically comprises the following steps:

[0012] S1: Use the distributed fiber optic acoustic wave sensing system to collect data and downsample the collected original signal data;

[0013] S2: Perform an image operation on the downsampled signal, denoise the converted image through ADNet, highlight the leakage signal characteristics, annotate the denoised image with a frame, and use the final image to establish a data set;

[0014] S3: Build an improved YOLOv8 network model;

[0015] S4: Divide the dataset into a training set, a test set, and a validation set. Use the training set to train the improved YOLOv8 network to obtain a leakage identification model.

[0016] Furthermore, in step S1, the distributed fiber optic acoustic wave sensing system uses a narrow line laser of a continuous laser as a laser light source, the continuous wave passes through an acousto-optic modulator to generate a light pulse sequence, the light pulse sequence is amplified by an erbium-doped amplifier, the backward Rayleigh scattering signal of the sensing fiber passes through a circulator and then enters a photodetector, the signal is collected by a data acquisition card, and the data is processed in parallel by a personal computer.

[0017] Furthermore, in step S1, the collected original signal data is downsampled, which specifically includes the following steps:

[0018] Step S1.1: compress the collected original signal by 256 times using the differential averaging method;

[0019] Step S1.2: Normalize the compressed data matrix.

[0020] Furthermore, the specific operations of step S2 are as follows:

[0021] Step S2.1: The downsampled signal is divided into 1 second units in the time domain, and the input size of the improved model is also required to obtain an image of 1024×96 pixels.

[0022] Step S2.2: input the image into ADNet;

[0023] Step S2.3: The input spatiotemporal image is subjected to noise removal by using dilated convolution and conventional convolution in the sparse module, achieving a balance between performance and effectiveness; the global and local feature information is then integrated through the long path of the feature enhancement module to enhance the expressiveness of the denoising model; the noise information hidden in the complex background is then extracted through the attention module; the feature enhancement module is integrated with the attention module to improve the effectiveness of the denoising model and reduce the computational complexity; finally, the reconstruction module generates a clear image through the obtained noise map and the input noise image;

[0024] Step S2.4: Use the image annotation software (LabelImg) to annotate the image with frames. Different events correspond to corresponding labels, and the labels are stored as XML files. The 1024×96 pixel image and its corresponding label together constitute the final data set.

[0025] Furthermore, in step S3, the improved YOLOv8 network model includes a backbone feature network, a feature extraction network and a prediction network, and the improvements include: using additional auxiliary training heads Aux in the 9th, 12th and 15th layers of the network; using the RepNCSPELAN structure to replace the original C2f convolution; and adding the LSKA attention mechanism after the third pooling layer in the original SPPF module to form an SPPF_LSKA module.

[0026] Furthermore, the backbone feature network includes:

[0027] After performing two convolution operations on the images in the dataset, the feature layer F1 is obtained;

[0028] Input the feature layer F1 into the RepNCSPELAN module to obtain the feature layer F2;

[0029] The feature layer F2 is subjected to a convolution operation again to obtain a second feature layer, and the second feature layer is input into a RepNCSPELAN module to obtain a feature layer F3;

[0030] The feature layer F3 is subjected to a convolution operation again to obtain a third feature layer, and the third feature layer is input into a RepNCSPELAN module to obtain a feature layer F4;

[0031] The feature layer F4 is subjected to a convolution operation again to obtain a fourth feature layer, and the fourth feature layer is input into a RepNCSPELAN module to obtain a feature layer F5;

[0032] The feature layer F5 is input into the SPPF module, and after three pooling operations, it is input into the LSKA attention mechanism module to obtain a feature vector L of a fixed size;

[0033] The feature extraction network includes:

[0034] Perform a first upsampling on the fixed-size feature vector L, perform feature fusion on the first upsampling output feature map and the feature layer F4 to obtain a first fused feature layer M1, and input the M1 into the RepNCSPELAN module to obtain a feature layer F6;

[0035] The feature map output after upsampling the feature layer F6 is fused with the feature layer F3 to obtain a second fused feature layer M2, and the M2 is input into the RepNCSPELAN module to obtain a feature layer F7;

[0036] The feature layer F7 is fused with the feature layer F6 to obtain a third fused feature layer M3, and the M3 is input into the RepNCSPELAN module to obtain a feature layer F8;

[0037] The feature layer obtained by performing a convolution operation on the feature layer F8 is fused with the feature layer F5 to obtain a fourth fused feature layer M4, and the M4 is input into the RepNCSPELAN module to obtain a feature layer F9;

[0038] The prediction network includes a guide head and an auxiliary head, and also includes:

[0039] The fixed-size feature vector L, the effective features of the feature layer F6, and the feature layer F7 are respectively input into different auxiliary heads Aux, and the effective features of the feature layer F7, the feature layer F8, and the feature layer F9 are respectively input into different guiding heads Detect; the auxiliary head and the guiding head learn the effective features during training, and only the guiding head Detect needs to output the prediction results during prediction.

[0040] Furthermore, in step S3, the RepNCSPELAN module combines CSPNet and Elan, two neural network structures using gradient path planning, and replaces the convolution group in the original ELAN with a CSPNet block with RepConv, taking into account light weight, inference speed and accuracy.

[0041] Furthermore, in step S3, the specific structure of the SPPF_LSKA module is: convolution layer c1--maximum pooling layer 1--maximum pooling layer 2--maximum pooling layer 3--LSKA attention mechanism--feature fusion layer--convolution layer c2, which enhances the multi-scale feature extraction capability of the model.

[0042] Furthermore, in step S3, the DetectAux module in the prediction network integrates the auxiliary head and the guide head together. The auxiliary head allows the middle layer of the network to learn more information and has richer gradient information to help training. During model inference, the auxiliary head is removed to reduce the inference time.

[0043] Furthermore, the step S4 divides the data set into a training set, a test set, and a validation set, and uses the training set to train the improved YOLOv8 network to obtain a leakage identification model, which specifically includes the following steps:

[0044] S4.1: Divide the data set described in step 2 into training set, test set and validation set in a ratio of 7:2:1, and use the training set to train the improved YOLOv8 network;

[0045] S4.2: Under the guidance of the loss function Loss of the existing YOLOv8 network, the optimizer SGD is selected, and the improved YOLOv8 network is trained using the training set to obtain an improved YOLOv8 network model.

[0046] Compared with the prior art, the present invention adopts the ADNet algorithm and the improved YOLOv8 network model, which has the following beneficial effects:

[0047] (1) The present invention introduces the attention-guided denoising CNN-ADNet into the distributed fiber optic sensor signal recognition, and uses the algorithm model to perform denoising on the spatiotemporal image of the leakage signal. The ADNet algorithm does not lose some shallow features, making the spatiotemporal image of the leakage event clearer and the features more obvious.

[0048] (2) Introduce the YOLOv8 network and add an additional auxiliary training head Aux to the network. Aux improves the recall rate by selecting more adjacent grids as positive samples in positive sample matching, thereby reducing the false negative rate. Aux allows the middle layer of the network to learn more information and has richer gradient information to help training. Aux participates in training during the training phase and is removed in the final inference phase, which does not affect the inference speed.

[0049] (3) The present invention adds the LSKA attention mechanism to the SPPF module, which enhances the multi-scale feature extraction capability of the model and thus improves the accuracy of event recognition. At the same time, the RepNCSPELAN module is used to replace the original C2f structure. This module combines CSPNet and Elan, two neural network structures that use gradient path planning, to achieve lightweight, inference speed, and accuracy.

[0050] Through the improvements, the present invention significantly reduces the underreporting rate of gas pipeline leakage events while ensuring the accuracy of identification, providing a new idea for subsequent distributed optical fiber acoustic wave sensing systems to identify intrusion events. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A step diagram of a gas pipeline leakage signal recognition method based on an attention-guided deformable convolutional network and an improved YOLOv8 network provided in an embodiment of the present invention;

[0052] Figure 2 A structural diagram of a pipeline leakage monitoring system based on distributed optical fiber acoustic wave sensing provided by an embodiment of the present invention;

[0053] Figure 3 A pipeline optical cable laying diagram provided for an embodiment of the present invention;

[0054] Figure 4 ADNet structure diagram provided by an embodiment of the present invention;

[0055] Figure 5 A comparison diagram of a spatiotemporal image provided in an embodiment of the present invention before (top) and after (bottom) ADNet denoising;

[0056] Figure 6 The overall model framework diagram of the improved YOLOv8 network provided by the embodiment of the present invention;

[0057] Figure 7 The overall structure diagram of RepNCSPELAN provided by the embodiment of the present invention;

[0058] Figure 8 The SPPF_LSKA structure diagram provided by the embodiment of the present invention;

[0059] Fig. 9 A learning flow chart of an auxiliary head and a guide head provided in an embodiment of the present invention;

[0060] Fig.10 The validation set accuracy curves of the four models during training on the same dataset;

[0061] Fig.11 The validation set recall curves of the four models during training on the same dataset;

[0062] Fig.12 The improved YOLOv8 network model and YOLOv5n model of this embodiment use the ten-fold cross-validation method to evaluate the result diagram. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0064] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] like Figure 1 As shown, the present invention provides a gas pipeline leakage signal recognition method based on an attention-guided deformable convolutional network and an improved YOLOv8 network. Figure 1 This is a schematic diagram of the implementation steps of this method, which specifically includes:

[0066] S1: Use the distributed fiber optic acoustic wave sensing system to collect data and downsample the collected original signal data;

[0067] S2: Perform an image operation on the downsampled signal, denoise the converted image through ADNet, highlight the leakage signal characteristics, annotate the denoised image with a frame, and use the final image to establish a data set;

[0068] S3: Build an improved YOLOv8 network model;

[0069] S4: Divide the data set into a training set, a test set, and a validation set, and use the training set to train the improved YOLOv8 network to obtain a leakage identification model;

[0070] The principle of the distributed optical fiber acoustic wave sensing system in step S1 is as follows: Figure 2As shown, the main equipment used in the system includes a narrow line laser of continuous laser, an acousto-optic modulator, an erbium-doped amplifier, a circulator, a balanced photodetector, a data acquisition card and a host computer. The system of this embodiment uses a narrow line laser with a line width of 5kHz as a laser light source. The continuous wave passes through the acousto-optic modulator to generate a light pulse sequence with a pulse width of 100ns. The light pulse sequence is amplified by an erbium-doped amplifier. The backward Rayleigh scattered signal of the sensing fiber passes through the circulator and enters the balanced photodetector. The signal is collected by a 100 MHZ data acquisition card, and the data is processed in parallel by the host computer.

[0071] Single-mode optical fiber is used as a detection cable to be laid along the pipeline to collect event samples such as overhead gas pipeline leakage, underground gas pipeline leakage, manual excavation, background noise, etc. to form event data collection. Figure 3 The pipeline is laid in a certain way, that is, the pipeline is divided into an overhead pipeline and an underground pipeline. For the underground pipeline part, the pipeline is buried 10 cm underground. The optical cable is laid in a straight line along the pipeline and finally connected to the DAS host. The transmission optical cable is composed of an optical fiber, a reinforcement layer, and a protective layer. After the optical cable is laid, the optical fiber distributed acoustic wave sensing system collects data, including but not limited to events such as overhead pipeline leakage, underground pipeline leakage, manual excavation, background noise, etc., and saves the data in a list in a numerical format. Each list saves the single collection data of a single event and is stored in a mobile hard disk for easy storage. The sampling frequency of the optical fiber distributed acoustic wave sensing system in this embodiment is 10kHz, and the collection time of each event is 5 to 10 minutes.

[0072] In this embodiment, the step S1 performs downsampling preprocessing on the collected original signal data, specifically including: compressing the collected original signal by 256 times using the differential averaging method; and normalizing the compressed data matrix.

[0073] In this embodiment, the downsampled signal is imaged in step S2, the converted image is denoised by ADNet, the leakage signal feature is highlighted, and a data set is established using the final image, specifically including:

[0074] Step S2.1: Segment the collected signals of leakage and human interference events in the time domain into 2 s units, and at the same time conform to the input size of the improved model, and finally obtain an image of 1024×96 pixels;

[0075] Step S2.2: Input the image into ADNet, where the ADNet network structure is as follows: Figure 4 As shown;

[0076] Step S2.3: The input spatiotemporal image is subjected to noise removal by using dilated convolution and conventional convolution through the sparse module, achieving a balance between performance and effectiveness; then the global and local feature information is integrated through the long path of the feature enhancement module to enhance the expressiveness of the denoising model; then the noise information hidden in the complex background is extracted through the attention module; the integration of the feature enhancement module and the attention module improves the effectiveness of the denoising model and reduces the computational complexity; finally, the reconstruction module generates a clear image through the obtained noise map and the input noise image. The comparison of the spatiotemporal image before and after ADNet denoising is shown in the figure below. Figure 5 As shown;

[0077] Step S2.4: Use LabelImg to annotate the image with frames. Different events correspond to corresponding labels. The labels are stored as XML files. The 1024×96 pixel images and their corresponding labels together constitute the final data set.

[0078] In this embodiment, the final data set obtained includes 2750 samples of overhead pipeline leakage events, 2407 samples of underground pipeline leakage events, 2353 samples of manual excavation events, and 1500 negative samples, for a total of 9010 samples.

[0079] In step S3 of this embodiment, the improved YOLOv8 network model includes a backbone feature network, a feature extraction network and a prediction network. The improved YOLOv8 network of this application includes using an additional auxiliary training head Aux in the 9th, 12th and 15th layers of the network, using the RepNCSPELAN structure to replace all the original C2f convolutions, and adding the LSKA attention mechanism after the third pooling layer in the original SPPF module to form an SPPF_LSKA module. The overall network structure is as follows: Figure 6 Specifically:

[0080] The backbone feature network includes:

[0081] After performing two convolution operations on the images in the dataset, the feature layer F1 is obtained;

[0082] Input the feature layer F1 into the RepNCSPELAN module to obtain the feature layer F2;

[0083] The feature layer F2 is subjected to a convolution operation again to obtain a second feature layer, and the second feature layer is input into a RepNCSPELAN module to obtain a feature layer F3;

[0084] The feature layer F3 is subjected to a convolution operation again to obtain a third feature layer, and the third feature layer is input into a RepNCSPELAN module to obtain a feature layer F4;

[0085] The feature layer F4 is subjected to a convolution operation again to obtain a fourth feature layer, and the fourth feature layer is input into a RepNCSPELAN module to obtain a feature layer F5;

[0086] The feature layer F5 is input into the SPPF module, and after three pooling operations, it is input into the LSKA attention mechanism module to obtain a feature vector L of a fixed size.

[0087] The feature extraction network includes:

[0088] Perform a first upsampling on the fixed-size feature vector L, perform feature fusion on the first upsampling output feature map and the feature layer F4 to obtain a first fused feature layer M1, and input the M1 into the RepNCSPELAN module to obtain a feature layer F6;

[0089] The feature map output after upsampling the feature layer F6 is fused with the feature layer F3 to obtain a second fused feature layer M2, and the M2 is input into the RepNCSPELAN module to obtain a feature layer F7;

[0090] The feature layer F7 is fused with the feature layer F6 to obtain a third fused feature layer M3, and the M3 is input into the RepNCSPELAN module to obtain a feature layer F8;

[0091] The feature layer obtained by performing a convolution operation on the feature layer F8 is fused with the feature layer F5 to obtain a fourth fused feature layer M4, and the M4 is input into the RepNCSPELAN module to obtain a feature layer F9.

[0092] The prediction network includes a guide head and an auxiliary head, and also includes:

[0093] The fixed-size feature vector L, the effective features of the feature layer F6, and the feature layer F7 are respectively input into different auxiliary heads Aux, and the effective features of the feature layer F7, the feature layer F8, and the feature layer F9 are respectively input into different guiding heads Detect; the auxiliary head and the guiding head learn the effective features during training, and only the guiding head Detect needs to output the prediction results during prediction.

[0094] Furthermore, the RepNCSPELAN module in step three of the present invention combines CSPNet and Elan, two neural network structures that use gradient path planning, and replaces the convolution group in the original ELAN with CSPNetblock with RepConv, taking into account light weight, inference speed and accuracy.

[0095] Furthermore, the SPPF_LSKA module in step three of the present invention has a specific structure: convolution layer c1--maximum pooling layer 1--maximum pooling layer 2--maximum pooling layer 3--LSKA attention mechanism--feature fusion layer--convolution layer c2, which strengthens the multi-scale feature extraction capability of the model.

[0096] Furthermore, in the DetectAux module of the prediction network in step three of the present invention, the auxiliary head and the guide head are integrated together. The auxiliary head allows the middle layer of the network to learn more information and has richer gradient information to help training. During model inference, the auxiliary head is removed to reduce the inference time.

[0097] All RepNCSPELAN module structures used in the improved YOLOv8 network of the present invention are the same, as shown in the structural diagram Figure 7 As shown; all RepNCSPELAN module structures used in the improved YOLOv8 network in this application are a combination of CSPNet and Elan, two neural network structures using gradient path planning, and the convolution group in the original ELAN is replaced with a CSPNet block with RepConv, taking into account lightweight, inference speed and accuracy.

[0098] The SPPF_LSKA module used in the improved YOLOv8 network in the present invention has a specific structure: convolution layer c1--max pooling layer 1--max pooling layer 2--max pooling layer 3--LSKA attention mechanism--feature fusion layer--convolution layer c2, which enhances the multi-scale feature extraction capability of the model. The structure diagram is as follows: Figure 8 shown.

[0099] The DetectAux module used in the improved YOLOv8 network in the present invention integrates the auxiliary head and the guide head together. The auxiliary head allows the middle layer of the network to learn more information and has richer gradient information to help training. During model inference, the auxiliary head is removed to reduce the inference time. The learning process diagram of the auxiliary head and the guide head is as follows: Fig. 9 shown.

[0100] In step 4 of the present invention, the data set is divided into a training set, a test set, and a validation set, wherein the background noise is input as a negative sample and is not considered a class of events. The training set is used to train the improved YOLOv8 network to obtain a leakage recognition model, which specifically includes:

[0101] S4.1: Divide the data set described in step 2 into training set, test set and validation set in a ratio of 7:2:1, and use the training set to train the improved YOLOv8 network;

[0102] S4.2: Under the guidance of the loss function Loss of the existing YOLOv8 network, the optimizer SGD is selected, and the improved YOLOv8 network is trained using the training set to obtain an improved YOLOv8 network model.

[0103] The improved YOLOv8 network is constructed using the PyTorch platform, which provides a rich neural network function interface, making the network construction process more convenient. When in use, the acquisition card in the DAS system collects the signal, and the image conversion program, sliding window segmentation program, and denoising algorithm integrated in the algorithm reasoning program are reasoned in sequence. Finally, the image of appropriate size passes through the backbone feature network and feature extraction network in sequence, and the prediction network finally outputs the prediction result. In the experiment, the SGD optimizer is used to train the network model, with a learning rate (LR) of 0.001, a batch size of 32, and a total of 200 iterations. The model training process effectively utilizes the powerful computing power of three NVIDIA A100 Tensor Core GPUs, which greatly improves the efficiency of model training.

[0104] The improved YOLOv8 network model ablation experiment in the present invention is shown in the following table:

[0105] .

[0106] Four models, RAL-YOLOv8, YOLOv5n, ResNET and RT-DERT, were trained on the same dataset to obtain the validation set accuracy curve ( Fig.10 ) and validation set recall curve ( Fig.11 ), it can be found that the model method proposed in this embodiment has achieved the best in both precision and recall. Overall, the model of this embodiment has achieved the best classification effect, which ultimately proves the effectiveness of the method proposed in this embodiment.

[0107] In order to make better use of limited data sets and reduce occasional problems caused by unreasonable data division, the improved YOLOv8 network model and YOLOv5n model in the present invention are evaluated using the ten-fold cross-validation method. Fig.12 The accuracy of the model on each fold of data after ten cross-evaluations of the improved YOLOv8 network model and the YOLOv5n model in the present invention is shown. It can be clearly seen that the recognition accuracy of the method proposed in this paper can stably reach more than 97% on each fold of data, which is far better than the performance of the YOLOv5n model, further verifying the performance of the improved YOLOv8 network model in the present invention.

Claims

1. A pipeline leakage identification method based on a distributed optical fiber acoustic wave sensing system, characterized in that: The specific steps include: S1: Use the distributed fiber optic acoustic wave sensing system to collect data and downsample the collected original signal data; S2: Perform an image operation on the downsampled signal, denoise the converted image through ADNet, highlight the leakage signal characteristics, annotate the denoised image with a frame, and use the final image to establish a data set; S3: Build an improved YOLOv8 network model, namely RAL-YOLOv8; S4: Divide the data set into a training set, a test set, and a validation set, and use the training set to train the improved YOLOv8 network to obtain a leakage identification model; The specific operation of step S2 is as follows: Step S2.1: The downsampled signal is divided into 1 second units in the time domain, and the input size of the improved model is also required to obtain an image of 1024×96 pixels. Step S2.2: input the image into ADNet; Step S2.3: The input spatiotemporal image is subjected to noise removal by using dilated convolution and conventional convolution in the sparse module, achieving a balance between performance and effectiveness; the global and local feature information is then integrated through the long path of the feature enhancement module to enhance the expressiveness of the denoising model; the noise information hidden in the complex background is then extracted through the attention module; the feature enhancement module is integrated with the attention module to improve the effectiveness of the denoising model and reduce the computational complexity; finally, the reconstruction module generates a clear image through the obtained noise map and the input noise image; Step S2.4: Use LabelImg to annotate the image with frames. Different events correspond to corresponding labels. The labels are stored as XML files. The 1024×96 pixel images and their corresponding labels together constitute the final data set. In step S1, the distributed optical fiber acoustic wave sensing system uses a narrow line laser of a continuous laser as a laser light source, the continuous wave passes through an acousto-optic modulator to generate an optical pulse sequence, the optical pulse sequence is amplified by an erbium-doped amplifier, the backward Rayleigh scattered signal of the sensing optical fiber passes through a circulator and enters a photodetector, the signal is collected by a data acquisition card, and the data is processed in parallel by a personal computer; In step S3, the improved YOLOv8 network model includes a backbone feature network, a feature extraction network and a prediction network, and the improvements include: using additional auxiliary training heads Aux in the 9th, 12th and 15th layers of the network; using the RepNCSPELAN structure to replace the original C2f convolution; adding the LSKA attention mechanism after the third pooling layer in the original SPPF module to form an SPPF_LSKA module; The backbone feature network includes: After performing two convolution operations on the images in the dataset, the feature layer F1 is obtained; Input the feature layer F1 into the RepNCSPELAN module to obtain the feature layer F2; The feature layer F2 is subjected to a convolution operation again to obtain a second feature layer, and the second feature layer is input into a RepNCSPELAN module to obtain a feature layer F3; The feature layer F3 is subjected to a convolution operation again to obtain a third feature layer, and the third feature layer is input into a RepNCSPELAN module to obtain a feature layer F4; The feature layer F4 is subjected to a convolution operation again to obtain a fourth feature layer, and the fourth feature layer is input into a RepNCSPELAN module to obtain a feature layer F5; The feature layer F5 is input into the SPPF module, and after three pooling operations, it is input into the LSKA attention mechanism module to obtain a feature vector L of a fixed size; The feature extraction network includes: Perform a first upsampling on the fixed-size feature vector L, perform feature fusion on the first upsampling output feature map and the feature layer F4 to obtain a first fused feature layer M1, and input the M1 into the RepNCSPELAN module to obtain a feature layer F6; The feature map output after upsampling the feature layer F6 is fused with the feature layer F3 to obtain a second fused feature layer M2, and the M2 is input into the RepNCSPELAN module to obtain a feature layer F7; The feature layer F7 is fused with the feature layer F6 to obtain a third fused feature layer M3, and the M3 is input into the RepNCSPELAN module to obtain a feature layer F8; The feature layer obtained by performing a convolution operation on the feature layer F8 is fused with the feature layer F5 to obtain a fourth fused feature layer M4, and the M4 is input into the RepNCSPELAN module to obtain a feature layer F9; The prediction network includes a guide head and an auxiliary head, and also includes: The fixed-size feature vector L, the effective features of the feature layer F6, and the feature layer F7 are respectively input into different auxiliary heads Aux, and the effective features of the feature layer F7, the feature layer F8, and the feature layer F9 are respectively input into different guiding heads Detect; the auxiliary head and the guiding head learn the effective features during training, and only the guiding head Detect needs to output the prediction results during prediction.

2. The pipeline leakage identification method based on the distributed optical fiber acoustic wave sensing system according to claim 1 is characterized in that: In step S1, the collected original signal data is downsampled, which specifically includes the following steps: Step S1.1: compress the collected original signal by 256 times using the differential averaging method; Step S1.2: Normalize the compressed data matrix.

3. The pipeline leakage identification method based on the distributed optical fiber acoustic wave sensing system according to claim 1 is characterized in that: In step S3, the RepNCSPELAN module combines CSPNet and Elan, two neural network structures using gradient path planning, and replaces the convolution group in the original ELAN with a CSPNet block with RepConv, taking into account light weight, inference speed and accuracy.

4. The pipeline leakage identification method based on the distributed optical fiber acoustic wave sensing system according to claim 1 is characterized in that: In step S3, the specific structure of the SPPF_LSKA module is: convolution layer c1--max pooling layer 1--max pooling layer 2--max pooling layer 3--LSKA attention mechanism--feature fusion layer--convolution layer c2, which enhances the multi-scale feature extraction capability of the model.

5. The pipeline leakage identification method based on the distributed optical fiber acoustic wave sensing system according to claim 1 is characterized in that: In step S3, the DetectAux module in the prediction network integrates the auxiliary head and the guide head together. The auxiliary head allows the middle layer of the network to learn more information and has richer gradient information to help training. During model inference, the auxiliary head is removed to reduce the inference time.

6. The pipeline leakage identification method based on the distributed optical fiber acoustic wave sensing system according to claim 1 is characterized in that: The step S4 divides the data set into a training set, a test set, and a validation set, and uses the training set to train the improved YOLOv8 network to obtain a leakage identification model, which specifically includes the following steps: S4.1: Divide the data set described in step S2 into a training set, a test set, and a validation set in a ratio of 7:2:1, and use the training set to train the improved YOLOv8 network; S4.2: Under the guidance of the loss function Loss of the existing YOLOv8 network, the optimizer SGD is selected, and the improved YOLOv8 network is trained using the training set to obtain an improved YOLOv8 network model.

Images