Pipeline intrusion event identification method based on distributed optical fiber sound wave sensing system

Abstract

The invention discloses a pipeline intrusion event identification method based on a distributed optical fiber sound wave sensing system, and the method comprises the steps: collecting intrusion sound wave signals of different positions and types of pipelines in actual engineering, carrying out the data preprocessing, and constructing a pipeline intrusion event data set; a multi-scale prime number convolution network model based on dynamic convolution and global attention is constructed, features of different time scales and frequency components are extracted through a multi-scale prime number convolution module, and key intrusion features are highlighted and noise interference is suppressed in combination with dynamic convolution capable of adaptively correcting a convolution kernel and an enhanced global attention module. Outputting an event recognition result; according to the method, the problems that the processing capacity of long sequence signals is insufficient and engineering background noise covers target signal features in pipeline intrusion identification are solved, and the monitoring success rate and accuracy of pipeline intrusion events are improved.

Description

Technical Field

[0001] The present invention relates to the field of distributed fiber optic acoustic sensing, and particularly to a method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system. Background Art

[0002] In recent years, the distributed fiber optic acoustic sensing technology based on a phase-sensitive optical time domain reflectometer has gradually become the focus of research at home and abroad. This technology has strong anti-electromagnetic interference ability, strong environmental adaptability, high detection sensitivity, low energy loss during operation, can achieve long-distance monitoring, and performs well in multi-point distributed detection, showing great application potential. This technology has currently been applied to perimeter security systems, pipeline safety, natural disaster warning, underwater cables, and structural health and other fields, and has become one of the main research and development directions for future intelligent perimeter security and intelligent scenario monitoring.

[0003] In the field of pipeline safety, monitoring potential external intrusion behaviors through a distributed fiber optic acoustic sensing system has been widely used. However, in actual engineering applications, many challenges still exist. The complex and changeable natural environment, such as meteorological conditions like wind, rain, and snow, will interfere with the system. In addition, non-intrusive events, such as the passing of pedestrians and vehicles and the operation of machines, will also generate acoustic signals, and these signals may be similar to those generated by intrusion behaviors, resulting in misjudgment by the system. At the same time, pseudo-vibration events caused by interference factors such as noise also increase the false alarm rate of the system. These factors keep the false alarm rate of the system high, greatly restricting the popularization and application of distributed fiber optic sensing technology in the field of pipeline safety.

[0004] With the continuous progress of data resources and computer computing levels, the recognition method based on the convolutional neural network of deep learning has been widely used in the monitoring of pipeline intrusion events of distributed fiber optic acoustic sensing. Its end-to-end feature learning ability can effectively simplify the processing process and improve the recognition accuracy. However, in the current monitoring methods based on distributed fiber optic acoustic sensing systems, there are still certain problems when using the convolutional neural network recognition method: (1) Traditional convolutional neural networks (CNNs) are widely used in pattern recognition tasks. However, when dealing with long sequence signals, it is often difficult to effectively capture long-distance dependencies, and the subtle features in long sequence signals are easily ignored or submerged in a large amount of data, further affecting the accuracy and stability of recognition; (2) Secondly, in actual applications, distributed fiber optic acoustic sensing systems are often in complex noise environments. Strong background noise not only masks the features of target signals but may also introduce pseudo-signals, seriously affecting the accuracy and reliability of pattern recognition, and further improvement of the feature extraction ability of the model is required. Summary of the Invention

[0005] To effectively solve the problems in the background art, the object of the present invention is to propose a method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system. The method uses the distributed fiber optic acoustic sensing system to collect acoustic signals of pipeline intrusion events, and constructs a multi-scale prime number convolutional network model based on dynamic convolution and global attention, enabling the model to effectively capture multi-level features in the signals, efficiently process long sequence signals, and maintain high recognition accuracy under the actual engineering conditions of high noise, and finally realizing the efficient monitoring of external intrusion events of the pipeline.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system specifically includes the following steps: S1. Use the distributed fiber optic acoustic sensing system to collect various pipeline intrusion event signals, and perform normalization processing on the collected original signals; S2. Segment and downsample the preprocessed different intrusion event signals, label the corresponding intrusion event labels, and construct a data set of pipeline intrusion events; S3. Construct a multi-scale prime number convolutional network model based on dynamic convolution and global attention; S4. Divide the data set constructed in S2 into a training set, a test set, and a validation set, adopt a mixed-precision training strategy, combine an adaptive learning rate scheduler, and use the training set to train on the model constructed in S3 to obtain an event recognition model.

[0008] Preferably, in step S1, in the actual engineering environment, collect acoustic signals of various external intrusion events against the pipeline. The acoustic signals cover diverse intrusion events at different positions of the pipeline and different mechanical and manual excavations. By collecting intrusion events at different positions and types, the generalization ability of the model in the actual scenario is improved. At the same time, in step S1, multiple normalization methods are used to normalize the acoustic signals, mainly including two methods: L2 norm normalization and min-max standardization. The calculation formula of min-max standardization is as follows:

[0009] In the formula, is the maximum value of the input signal sequence, is the minimum value of the input signal sequence, is the original input signal sequence, is the output signal sequence after min-max standardization; the calculation formula of L2 norm normalization is as follows:

[0010] In the formula, is the original input signal sequence, is the L2 norm of the signal, is the output signal sequence after normalization. After the acoustic wave signal is normalized, the influence of dimension and variation range is eliminated. According to the influence of different normalization methods used in subsequent model construction and recognition on the recognition accuracy, a more suitable normalization method can be flexibly selected.

[0011] Preferably, in step S2, the preprocessed signal is segmented in the time domain by a sliding window with a time unit of 1 s to obtain one-dimensional acoustic wave signal segments. Then, the sampling rate of the one-dimensional spatio-temporal signal segments is reduced to 5000 Hz to obtain downsampled data, which not only ensures the integrity of the intrusion signal data but also conforms to the input size of the model, balancing real-time performance and signal integrity. At the same time, by adjusting the step size of the sliding window, data augmentation and enhancement can also be achieved.

[0012] Preferably, in step S3, the multi-scale prime number convolutional network model mainly includes a multi-scale prime number convolutional module, a dynamic convolutional module (ODConv1D), an enhanced global attention module (GAM), and a feature fusion classification module; among them, the multi-scale prime number convolutional module sets multiple prime number convolutional kernel branches in parallel in the same network layer to achieve multi-scale feature extraction; the dynamic convolutional module generates or corrects the convolutional kernel weights based on the adaptive attention mechanism to achieve multi-scale dynamic representation of the input pipeline acoustic wave signal; the global attention module weights the feature map output by the dynamic convolution in the channel and spatial dimensions to highlight the significant feature regions of the target event; the feature fusion classification module integrates multi-scale features using the channel dimension splicing strategy, compresses the features through the adaptive global pooling layer, and finally completes event classification through the fully connected layer; the following is a detailed introduction to the four modules: S3-1. Multi-scale prime number convolutional branch module: The multi-scale prime number convolutional branch module sets multiple prime number convolutional kernel branches in parallel in the same network layer to achieve multi-scale feature extraction. Its specific implementation method includes the following steps: S3-1-1. Select the convolutional kernel size from the pre-set prime number set P, where the prime number set P = [3, 5, 7, 11, 13, 17, 19, 23...], which can be added, deleted, or adjusted according to actual needs; S3-1-2. Input the input features into the prime number convolutional kernels of each branch for parallel convolution operations. Each branch sequentially executes the dynamic convolutional module and the enhanced global attention module to capture the pipeline intrusion signals with different time scales and different frequency components, perform batch normalization (Batch Normalization) processing on the weighted features, and achieve non-linear mapping through the ReLU activation function to obtain the branch output; S3-1-3. Concatenate and fuse all branch outputs along the channel dimension to form multi-scale fusion features. The number of channels is the sum of the output channels of each branch. The fused features are used as the input to the next-level multi-scale convolution module, enabling further stacking or interaction of multi-scale features in subsequent levels to achieve rich representation of the pipeline intrusion acoustic signal.

[0013] S3-2. The dynamic convolution module includes a dynamic convolution unit and a residual convolution unit. The former generates or corrects convolution kernel parameters through adaptive attention, and the latter adds a residual path before and after convolution to maintain the network depth while avoiding gradient vanishing. The specific implementation steps are as follows: S3-2-1. The input feature map first generates a channel statistical vector through global average pooling, which is a statistical description of the input acoustic features. Then, it is compressed and expanded through two layers of 1×1 convolution. The first layer compresses the channels to C / 4 (C is the number of channels), expands to 4K dimensions (K is the convolution kernel size) after passing through the ReLU activation function, and then uses the Sigmoid activation function to constrain the output to the range (0,1) to obtain the attention weight vector. S3-2-2. The generated attention weight vector is split and reshaped according to the pre-divided spatial weight, input channel weight, output channel weight, and convolution kernel weight, and the basic convolution kernel parameters are weighted and corrected, and the dynamically corrected convolution kernel is applied in the convolution calculation. S3-2-3. Perform one-dimensional convolution operations on a sample-by-sample basis. Through this process, the convolution kernel parameters of each sample can be adaptively adjusted with the change of the input signal, thereby enhancing the ability to capture variable acoustic features. After the convolution calculation, the intermediate output is restored to the corresponding output vector shape to form the main branch output of the dynamic convolution module. S3-2-4. Introduce a residual connection between the input and output of the convolution operation: if the input and output do not match in channels or length, perform projection conversion through 1×1 convolution, and finally add the residual and the main branch output element by element to achieve the retention of shallow features and the dynamic adjustment of deep features.

[0014] S3-3. The enhanced global attention module enhances the features of the pipeline acoustic signal through the following improvements. The specific implementation methods include: S3-3-1. Dual-channel attention mechanism: The input features adopt two paths of average pooling and max pooling to globally pool the input features to obtain two different channel description vectors; perform convolution or fully connected mapping and activation operations on the above two vectors respectively to obtain two sets of channel weighting coefficients; add the two sets of weighting coefficients point by point and map them to the interval [0,1] to form the final channel attention map. S3-3-2, Lightweight Spatial Attention Calculation: Use depthwise separable convolution instead of ordinary convolution to perform depth convolution and pointwise convolution operations on the original multi-channel features in sequence to reduce the number of parameters. Finally, after batch normalization and ReLU activation function processing on the convolution output, obtain the spatial attention map through the Sigmoid function; S3-3-3, Dynamic Weight Fusion: Set a learnable fusion coefficient between the channel attention map and the spatial attention map, and calculate the dynamic fusion weight through Softmax; Multiply the fused attention matrix and the input features point by point to highlight the significant features of pipeline intrusion events and suppress noise or interference information.

[0015] S3-4, Feature Fusion and Classification Module: Compress the multi-scale fusion features into a fixed-dimension global vector through adaptive global average pooling, set a fully connected layer with the output dimension according to the number of pipeline intrusion event categories, map the global feature vector, and output the probabilistic prediction scores through Softmax and complete classification in combination with the cross-entropy loss function.

[0016] Preferably, in step S4, the process of training the multi-scale prime number convolutional network model includes the following steps: S4-1, Divide the dataset described in step S2 into a training set, a test set, and a validation set according to the ratio of 7:2:1, and use the training set to train the multi-scale prime number convolutional network model; S4-2, Under the guidance of the model cross-entropy loss function, select the Adam optimizer, combine the mixed-precision training strategy and the adaptive learning rate scheduling to optimize the network parameters. At the same time, by monitoring indicators such as accuracy and loss curves, obtain the multi-scale prime number convolutional network recognition model based on dynamic convolution and global attention.

[0017] Compared with the prior art, the multi-scale prime number convolutional network recognition model based on dynamic convolution and global attention adopted by the present invention has the following beneficial effects: (1) By selecting a set of prime numbers (such as 3, 5, 7, 11, 13, 17, etc.) as the convolution kernel size and parallelly setting multi-scale prime number convolution branches in the same network layer, the present invention can more fully capture the subtle features and long-distance dependence information in the pipeline intrusion signal from different scales and frequency levels; Compared with the design of traditional fixed or conventional integer convolution sizes, the "relatively prime" property between prime number convolution kernels can reduce the overlap and redundancy of the receptive fields of different convolution kernels to a certain extent, helping the model to more clearly separate the key features of acoustic signals in different time periods or different frequency domains; (2) The dynamic convolution module used in the present invention corrects the convolution kernel parameters by means of adaptive attention, realizes the real-time adjustment of the convolution kernel to the input features, can adaptively generate or correct the convolution kernel for complex and variable actual acoustic signals, effectively enhances the generalization ability and robustness of the model to intrusion events in different scenarios, and reduces the influence of noise and external interference on the recognition results; (3) The enhanced global attention module used in the present invention is modified on the basis of the global attention (GAM). By combining dual-channel attention and lightweight spatial attention and introducing dynamic weight fusion, it further highlights the core feature region of the pipeline intrusion event, suppresses irrelevant or redundant information, and can still accurately focus on the target region in a high-noise or interference environment, thus significantly improving the detection performance of intrusion events.

[0018] The pipeline intrusion event recognition method based on the distributed fiber optic acoustic sensing system proposed by the present invention can not only fully extract and distinguish multi-scale information, but also take into account the noise and variability in the actual environment. Through a complete set of multi-scale prime number convolution network models based on dynamic convolution and global attention, it improves the recognition accuracy and stability of pipeline intrusion events, and has good popularization value and application prospects in long-distance pipeline safety monitoring and perimeter security applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a step diagram of the recognition method of pipeline intrusion events of a multi-scale prime number convolution network based on dynamic convolution and global attention provided by an embodiment of the present invention;

[0020] Figure 2 The pipeline optical cable laying diagram provided by an embodiment of the present invention;

[0021] Figure 3 It is a schematic diagram of the network structure of a multi-scale prime number convolution network based on dynamic convolution and global attention provided by an embodiment of the present invention;

[0022] Figure 4 It is a comparison diagram of the confusion matrices of the training results of four models on the test set of the same data set. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To make the objectives, 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 with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0024] 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts fall within the scope of protection of the present invention.

[0025] As Figure 1 shown, the present invention provides a method for identifying pipeline intrusion events of a multi-scale prime number convolution network based on dynamic convolution and global attention, Figure 1 which is a schematic diagram of the implementation steps of this method, specifically including:

[0026] S1. Use a distributed fiber optic acoustic sensing system to collect various pipeline intrusion event signals, and perform normalization processing on the collected original signals;

[0027] S2. Segment and downsample the preprocessed different intrusion event signals, and label the corresponding intrusion event labels to construct a dataset of pipeline intrusion events;

[0028] S3. Construct a multi-scale prime number convolution network model based on dynamic convolution and global attention;

[0029] S4. Divide the dataset constructed in S2 into a training set and a test set, adopt a mixed-precision training strategy, combine with an adaptive learning rate scheduler, and use the training set to train on the model constructed in S3 to obtain an event recognition model.

[0030] In this embodiment, in step S1, the laying method of the pipeline optical cable for data collection in the actual engineering environment is as Figure 2 shown. The acoustic measurement optical cable is laid directly above the pipeline. Stretch the optical cable to make it closely adhere to the upper part of the pipeline and then fix the optical cable. Use stainless steel tie straps to bundle and fix it at a distance of 20 cm on both sides of the pipeline connection. When fixing, clamp a 6-cm-long rubber gasket between the optical cable and the tie strap. At the same time, use stainless steel tie straps and rubber gaskets to bundle and fix it in the middle of each pipeline; collect various acoustic signals of external intrusion events against the pipeline along the acoustic measurement optical cable. In this embodiment, the total length of the acoustic measurement optical cable is about 4.8 km, and a data collection point is set every 5 m, with a total of 960 collection points. The collection frequency is 5 KHz, and a total of 7 common intrusion events near the buried pipeline optical cable are collected, namely background noise, electric pickaxe digging on the padstone, excavator digging, shovel digging, electric pickaxe digging without the padstone, hammer knocking, and pickaxe digging. The acoustic signals cover different positions of the pipeline. At the same time, multiple normalization methods are used to normalize the acoustic signals, mainly including two methods: L2 norm normalization and min-max standardization. The calculation formula of min-max standardization is as follows:

[0031] Where, is the maximum value of the input signal sequence, is the minimum value of the input signal sequence, is the original input signal sequence, is the output signal sequence after min-max normalization; the L2 norm normalization calculation formula is as follows:

[0032] Where, is the original input signal sequence, is the L2 norm of the signal, The output signal sequence after normalization is that the acoustic wave signal is normalized to eliminate the influence of dimension and variation range. After experiments, this embodiment selects min-max normalization with higher accuracy as the data normalization method to preprocess the data.

[0033] In this embodiment, in step S2, the preprocessed signal is segmented in the time domain using a sliding window with a time unit of 1 s to obtain one-dimensional acoustic wave signal segments, which are then labeled. This not only ensures the integrity of the intrusion signal data, but also meets the input size of the model, balancing real-time performance and signal integrity. At the same time, by adjusting the step size of the sliding window, data expansion and enhancement can also be achieved.

[0034] By collecting intrusion events of different locations and types, we formed the dataset shown in the following table:

[0035] This example collected 8,081 sets of disturbance event samples for seven types of disturbance events: background noise, shovel excavation, electric pick excavation with stone padding, excavator excavation, electric pick excavation without stone padding, hammer strikes, and pickaxe excavation. Each set of samples contained 5,000 × 1 data, which fully reflects the information of the disturbance event. In the experiment, 56 sets of samples were selected as the training dataset, accounting for approximately 70% of the total number of samples, for training and validating the pattern recognition algorithm. At the same time, 1,616 sets of samples were selected as the test dataset, accounting for approximately 20% of the total number of samples, and 808 sets of samples were selected as the validation dataset, accounting for approximately 10% of the total number of samples, to evaluate the generalization ability of the model.

[0036] In this embodiment, in step S3, the network structure diagram of the constructed multi-scale prime convolutional network model is as follows: Figure 3As shown in the figure, it mainly includes a multi-scale prime number convolution module, a dynamic convolution module (ODConv1D), an enhanced global attention module (GAM), and a feature fusion classification module. Among them, the multi-scale prime number convolution module sets multiple prime number convolution kernel branches in parallel in the same network layer to achieve multi-scale feature extraction; the dynamic convolution module generates or corrects the convolution kernel weights based on the adaptive attention mechanism to realize the multi-scale dynamic representation of the input pipeline acoustic signal; the global attention module weights the feature map output by the dynamic convolution in the channel and spatial dimensions to highlight the significant feature regions of the target event; the feature fusion classification module integrates multi-scale features using the channel dimension splicing strategy, compresses the features through the adaptive global pooling layer, and finally completes event classification through the fully connected layer. The following is a detailed introduction to the three modules:

[0037] S3-1. Multi-scale prime number convolution branch module: The multi-scale prime number convolution branch module sets multiple prime number convolution kernel branches in parallel in the same network layer to achieve multi-scale feature extraction. Its specific implementation method includes the following steps:

[0038] S3-1-1. Select the convolution kernel size from the pre-set prime number set P, where the prime number set P = [3, 5, 7, 11, 13, 17, 19, 23…], etc., which can be added, deleted or adjusted according to actual needs;

[0039] S3-1-2. Input the input features into the prime number convolution kernels of each branch for parallel convolution operations. Among them, each branch sequentially executes the dynamic convolution module and the enhanced global attention module to capture the pipeline intrusion signals of different time scales and different frequency components, perform batch normalization (Batch Normalization) processing on the weighted features, and realize non-linear mapping through the ReLU activation function to obtain the branch output;

[0040] S3-1-3. Concatenate and fuse all branch outputs along the channel dimension to form multi-scale fusion features. The number of channels is the sum of the output channels of each branch. The fused features are used as the input of the next-level multi-scale convolution module, so that the multi-scale features are further stacked or interacted in the subsequent levels to realize the rich representation of the pipeline intrusion acoustic signal.

[0041] S3-2. The dynamic convolution module (ODConv1D) includes a dynamic convolution unit and a residual convolution unit. The former generates or corrects the convolution kernel parameters through adaptive attention, and the latter adds a residual path before and after the convolution unit to maintain the network depth while avoiding gradient disappearance. Its specific implementation steps include:

[0042] S3-2-1. The input feature map first generates a channel statistical vector through global average pooling, which is the statistical description of the input acoustic wave features. Then, it is compressed and expanded through two layers of 1×1 convolutions. The first layer compresses the channels to C / 4 (C is the number of channels), and after passing through the ReLU activation function, it is expanded to 4K dimensions (K is the convolution kernel size). Then, the Sigmoid activation function is used to constrain the output to the range of (0,1) to obtain the attention weight vector;

[0043] S3-2-2. The generated attention weight vector is split and reshaped according to the pre-divided spatial weight, input channel weight, output channel weight, and convolution kernel weight, and the basic convolution kernel parameters are weighted and corrected, and the dynamically corrected convolution kernel is applied in the convolution calculation;

[0044] S3-2-3. Perform one-dimensional convolution operations on a sample-by-sample basis. Through this process, the convolution kernel parameters of each sample can be adaptively adjusted with the change of the input signal, so as to enhance the ability to capture variable acoustic wave features. After the convolution calculation is completed, the intermediate output is restored to the corresponding output vector shape to form the main branch output of the dynamic convolution module;

[0045] S3-2-4. Introduce a residual connection between the input and output of the convolution operation: if the input and output do not match in channels or length, then a 1×1 convolution is used for projection conversion, and finally the residual is added to the main branch output element by element to achieve the retention of shallow features and the dynamic adjustment of deep features.

[0046] S3-3. Enhanced global attention module (GAM) realizes the feature enhancement of pipeline acoustic signals through the following improvements, and its specific implementation methods include:

[0047] S3-3-1. Dual-channel attention mechanism: The input features adopt two paths of average pooling and max pooling to perform global pooling on the input features to obtain two different channel description vectors; respectively perform convolution or fully connected mapping and activation operations on the above two vectors to obtain two groups of channel weighting coefficients; the two groups of weighting coefficients are added point by point and then mapped to the interval [0,1] to form the final channel attention map;

[0048] S3-3-2. Lightweight spatial attention calculation: Use depthwise separable convolution instead of ordinary convolution to perform depthwise convolution and pointwise convolution operations on the original multi-channel features in turn to reduce the number of parameters. Finally, after batch normalization and ReLU activation function processing on the convolution output, the spatial attention map is obtained through the Sigmoid function;

[0049] S3-3-3. Dynamic weight fusion: Set a learnable fusion coefficient between the channel attention map and the spatial attention map, and calculate the dynamic fusion weight through Softmax; Multiply the fused attention matrix point by point with the input features to highlight the significant features of the pipeline intrusion event and suppress noise or interference information.

[0050] S3-4. The feature fusion and classification module compresses the multi-scale fusion features into a fixed-dimension global vector through adaptive global average pooling, sets a fully connected layer with the output dimension according to the number of pipeline intrusion event categories, maps the global feature vector, and outputs the probabilistic prediction scores through Softmax and combines with the cross-entropy loss function to complete the classification.

[0051] In this embodiment, in step S4, the process of training the multi-scale prime number convolutional network model includes the following steps:

[0052] S4-1. Divide the data set consisting of 8081 groups of samples in step S2 into a training set and a test set according to a ratio of 7:2:1, and use the training set to train the multi-scale prime number convolutional network model;

[0053] S4-2. Under the guidance of the model cross-entropy loss function, select the Adam optimizer, combine the mixed-precision training strategy and the adaptive learning rate scheduling to optimize the network parameters. At the same time, by monitoring indicators such as accuracy and loss curves, obtain the multi-scale prime number convolutional network recognition model based on dynamic convolution and global attention.

[0054] Use the PyTorch deep learning framework to construct a multi-scale prime number convolutional network recognition model. When in use, collect signals through a distributed fiber optic acoustic sensing system. The data transposition program, sliding window segmentation program, and normalization algorithm integrated in the algorithm inference program are carried out in sequence, and finally the time series signals of appropriate size are sequentially passed through the dynamic convolution module and the global attention module, and finally the prediction results are output by the prediction network.

[0055] In the experiment, the Adam optimizer was used to train the network model, the learning rate (LR) was 0.001, the batch size was 32, and a total of 200 iterations were performed. The overall recognition results of the test set and the validation set of the model are shown in the following table:

[0056] Among 2425 groups of samples, the recognition accuracy of 7 types of events reached 99.13%, verifying that the model has a high recognition accuracy.

[0057] The ablation experiment of the multi-scale prime number convolutional network based on dynamic convolution and global attention in the present invention is as Figure 4As shown, under the same hyperparameter settings, after 200 iterations, the multi-scale prime number convolutional network recognition model that combines both the dynamic convolution and global attention modules achieved an accuracy of 99.13%, the multi-scale prime number convolutional network that combines the global attention module achieved an accuracy of 98.88%, the multi-scale prime number convolutional network that combines the dynamic convolution module achieved an accuracy of 97.53%, and the multi-scale prime number convolutional network achieved an accuracy of 95.96%, verifying the effectiveness of the dynamic convolution and global attention modules, and further verifying the accuracy and effectiveness of the model.

Claims

1. A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system, characterized in that, It includes the following steps: S1. Use a distributed fiber optic acoustic sensing system to collect various pipeline intrusion event signals, and perform normalization processing on the collected original signals; S2. Segment and downsample the preprocessed different intrusion event signals, and label the corresponding intrusion event labels to construct a dataset of pipeline intrusion events; S3. Construct a multi-scale prime convolutional network model based on dynamic convolution and global attention; S4. Divide the dataset constructed in S2 into a training set, a test set, and a validation set. Adopt a mixed-precision training strategy, combine an adaptive learning rate schedule, and use the training set to train on the model constructed in S3 to obtain an event recognition model.

2. The pipeline intrusion event recognition method based on a distributed fiber optic acoustic sensing system according to claim 1, wherein: In step S1, in the actual engineering environment, collect various acoustic signals of external intrusion events against the pipeline. The acoustic signals cover diverse intrusion events at different positions of the pipeline and different mechanical and manual excavations. By collecting intrusion events at different positions and types, the generalization ability of the model in the actual scenario is improved.

3. A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 1, characterized in that: In step S1, various normalization methods are used to normalize the acoustic signals, mainly including two methods: L2 norm normalization and min-max standardization. The calculation formula of min-max standardization is as follows: In the formula, is the maximum value of the input signal sequence, is the minimum value of the input signal sequence, is the original input signal sequence, is the output signal sequence after min-max normalization; The calculation formula for L2 norm normalization is as follows: In the formula, is the original input signal sequence, is the L2 norm of the signal, is the output signal sequence after normalization. After the acoustic wave signal is normalized, the influence of dimension and variation range is eliminated. According to the influence of different normalization methods used in subsequent model construction and recognition on the recognition accuracy, a more suitable normalization method can be flexibly selected.

4. The pipeline intrusion event recognition method based on a distributed fiber optic acoustic sensing system according to claim 1, wherein: In step S2, the preprocessed signal is segmented in the time domain with a time unit of 1 s through a sliding window to obtain one-dimensional acoustic signal segments. Then, the one-dimensional acoustic signal segments are downsampled to obtain the downsampled data and label them. This not only ensures the integrity of the intrusion signal data but also conforms to the input size of the model, balancing real-time performance and signal integrity. At the same time, by adjusting the step size of the sliding window, data augmentation and enhancement can also be achieved.

5. The pipeline intrusion event recognition method based on a distributed fiber optic acoustic sensing system according to claim 1, characterized in that: In step S3, the multi-scale prime convolutional network model mainly includes a multi-scale prime convolutional module, a dynamic convolution module (ODConv1D), an enhanced global attention module (GAM), and a feature fusion classification module. Among them, the multi-scale prime convolutional module parallelly sets multiple prime convolutional kernel branches in the same network layer to achieve multi-scale feature extraction; the dynamic convolution module generates or corrects the convolutional kernel weights based on the adaptive attention mechanism to achieve multi-scale dynamic characterization of the input pipeline acoustic signals; the global attention module performs weighted processing on the feature map output by the dynamic convolution in the channel and spatial dimensions to highlight the significant feature regions of the target event; the feature fusion classification module integrates multi-scale features using the channel dimension splicing strategy, compresses the features through an adaptive global pooling layer, and finally completes event classification through a fully connected layer.

6. The method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 5, wherein: The multi-scale prime convolutional branch module can efficiently cover all possible receptive field sizes to achieve multi-scale feature extraction by parallelly setting multiple prime convolutional kernel branches in the same network layer. Its specific implementation method includes: Select the convolutional kernel size from the pre-set prime number set P, where the prime number set P = [3, 5, 7, 11, 13, 17, 19, 23…], and it can be added, deleted, or adjusted according to actual needs; The input features are simultaneously input into the prime number convolutional kernels of each branch for parallel convolutional operations. In each branch, a dynamic convolutional module (ODConv1D) and an enhanced global attention module (GAM) are sequentially executed to capture the pipeline intrusion signals with different time scales and different frequency components, perform batch normalization (Batch Normalization) on the weighted features, and achieve non-linear mapping through the ReLU activation function to obtain the branch output; The outputs of all branches are concatenated and fused along the channel dimension to form multi-scale fusion features. The number of channels is the sum of the output channels of each branch. The fused features are used as the input of the next-level multi-scale convolutional module, enabling the multi-scale features to be further stacked or interacted in subsequent levels to achieve rich representation of the pipeline intrusion acoustic signals.

7. A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 5, characterized in that: The dynamic convolutional module (ODConv1D) includes a dynamic convolutional unit and a residual convolutional unit. The former generates or corrects the convolutional kernel parameters through adaptive attention, and the latter adds a residual path before and after convolution to maintain the network depth while avoiding gradient disappearance; Its specific implementation method includes: The input feature map first generates a channel statistical vector through global average pooling, the statistical description of the input acoustic features, and is compressed and expanded through two layers of 1×1 convolutions. The first layer compresses the channels to C / 4 (C is the number of channels), and after passing through the ReLU activation function, it is expanded to 4K dimensions (K is the convolutional kernel size). Then, the Sigmoid activation function is used to constrain the output to the range (0,1) to obtain the attention weight vector; The generated attention weight vector is split and reshaped according to the pre-divided spatial weight, input channel weight, output channel weight, and convolutional kernel weight, and the basic convolutional kernel parameters are weighted and corrected, and the dynamically corrected convolutional kernel is used in the convolutional calculation; Perform one-dimensional grouped convolutional operations. Through this process, the convolutional kernel parameters of each sample can be adaptively adjusted with the change of the input signal, thereby enhancing the ability to capture variable acoustic features. After the convolutional calculation is completed, the intermediate output is restored to the corresponding output vector shape to form the main branch output of the dynamic convolutional module; A residual connection is introduced between the input and output of the convolutional operation: if the input and output do not match in channels or length, projection conversion is performed through 1×1 convolution, and finally the residual is added to the main branch output element by element to achieve the retention of shallow features and the dynamic adjustment of deep features.

8. A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 5, characterized in that: The enhanced global attention module (GAM) enhances the features of the pipeline acoustic signals through the following improvements. Its specific implementation method includes: Dual-channel attention mechanism: The input features adopt two paths of average pooling and max pooling to globally pool the input features to obtain two different channel description vectors; convolutional or fully connected mapping and activation operations are respectively performed on the above two vectors to obtain two groups of channel weighting coefficients; the two groups of weighting coefficients are pointwise added and then mapped to the interval [0,1] to form the final channel attention map; Lightweight spatial attention calculation: Depthwise separable convolution is used instead of ordinary convolution. Depth convolution and pointwise convolution operations are sequentially performed on the original multi-channel features to reduce the number of parameters. Finally, after batch normalization and ReLU activation function processing on the convolution output, a spatial attention map is obtained through the Sigmoid function; Dynamic weight fusion: A learnable fusion coefficient is set between the channel attention map and the spatial attention map, and the dynamic fusion weight is calculated through Softmax; the fused attention matrix is multiplied pointwise with the input features to highlight the significant features of the pipeline intrusion event and suppress noise or interference information.

9. The method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 5, wherein: The feature fusion and classification module realizes multi-scale feature fusion and classification decision-making in the following way: The multi-scale fusion features are compressed into a fixed-dimensional global vector through adaptive global average pooling. A fully connected layer with the output dimension set according to the number of pipeline intrusion event categories is used to map the global feature vector, and probabilistic prediction scores are output through Softmax and combined with the cross-entropy loss function to complete classification.

10. A method for identifying pipeline intrusion events based on a distributed fiber optic acoustic sensing system according to claim 1, characterized in that: In step S4, the process of training the multi-scale prime number convolutional network model includes the following steps: S4-1, divide the dataset described in step S2 into a training set, a test set, and a validation set according to the ratio of 7:2:1, and use the training set to train the multi-scale prime number convolutional network model; S4-2, under the guidance of the model cross-entropy loss function, select the Adam optimizer, combine the mixed-precision training strategy and the adaptive learning rate scheduling to optimize the network parameters. At the same time, by monitoring indicators such as accuracy and loss curves, obtain a multi-scale prime number convolutional network recognition model based on dynamic convolution and global attention.

Images