Optical fiber distributed sensing signal identification method and system with low false detection rate

Abstract

The invention relates to the technical field of electric data digital processing, in particular to an optical fiber distributed sensing signal identification method and system with a low false drop rate, and the method comprises the following steps: S1, collecting vibration signal data of a sensing optical fiber through a distributed vibration sensing system, and carrying out the preprocessing of the vibration signal data to obtain a vibration signal data set; s2, an improved YOLOv9 network model is built; s3, dividing the collected vibration signal data set into a training set, a verification set and a test set, and performing iterative training on the improved YOLOv9 network model by using the training set until the model converges to obtain a trained improved YOLOv9 network model; and S4, inputting vibration signal data acquired in real time into the trained improved YOLOv9 network model, and performing forward propagation reasoning recognition prediction to obtain a recognition result containing category information and position information of the target. According to the method, the cross-scale target feature expression capability of the network is remarkably improved, the false detection probability in a complex scene is effectively reduced, and the application requirement of cross-scale target detection in an optical fiber environment is practically met.

Description

Technical Field

[0001] This invention relates to the field of vibration signal digital data processing technology, specifically to a method and system for identifying fiber optic distributed sensor signals with a low false detection rate. Background Technology

[0002] Distributed fiber optic sensing technology, with its high sensitivity, wide-area continuous monitoring capabilities, and excellent adaptability to complex environments, has attracted widespread research in many fields. Its derivative, distributed vibration sensing systems, construct a continuous vibration signal acquisition network along the fiber optic cable path to achieve real-time monitoring of vibration signals and simultaneously generate time-series datasets containing rich vibration characteristic information. However, due to the complexity of practical application scenarios, the time-series data output by distributed vibration sensing systems inevitably contains a large amount of background noise and non-target interference signals. How to achieve accurate extraction and feature analysis of effective vibration signals in strong noise backgrounds, especially achieving high-precision pattern recognition of intrusion targets under complex environmental interference conditions, has become a key technical bottleneck restricting the engineering application and industrial promotion of this technology.

[0003] In recent years, deep learning technology has demonstrated tremendous potential in signal processing due to its powerful feature extraction capabilities. Convolutional neural networks, in particular, have achieved breakthroughs in time-series signal recognition tasks. The published paper, "Spatiotemporal image-based method for external breakage event recognition in long-distance distributed fiber optic sensing," proposes the ASED algorithm to address strong noise interference in complex environments, specifically for identifying fiber optic disturbances. By employing a sliding window on a spatial scale and automatically calculating the optimal segmentation threshold for each local region based on variance differences, it effectively separates background noise from potentially valid signals, solving the problem of uneven spatial distribution of noise levels in monitoring. This lays a solid foundation for subsequent recognition. However, the algorithm has several disadvantages: it implicitly assumes that the amplitude difference between signal and noise must be significantly higher than the amplitude of the valid event signal. This assumption may fail when facing weak-intensity breakage events, leading to the signal being misjudged as noise and missed detection. Conversely, high-amplitude environmental interference may be misjudged as valid events, causing false alarms. Equalization can over-enhance; as a global contrast adjustment method, it significantly amplifies even the few remaining strong noise points after denoising, forming artifacts or false bright spots, interfering with the judgment of subsequent recognition models, and introducing new noise sources. The published paper "Distributed Fiber Optic Sensing Signal Processing Method and System" proposes a multi-receptive-field attention mechanism (MRFA). By setting standard convolutional branches and depthwise separable convolutional branches in parallel, it can capture spatial features at different scales, effectively enhancing the model's ability to perceive cross-scale anomalies. Simultaneously, this mechanism utilizes channel grouping, spatial pooling, and dynamic weighted fusion strategies to achieve adaptive focusing on key regions, significantly improving feature representation capabilities and suppressing environmental noise interference in fiber optic sensing data. This allows the model to maintain high detection accuracy and robustness even in complex backgrounds. However, the MRFA module structure is relatively complex, containing multiple parallel convolutional branches, multiple pooling operations, and nonlinear activation functions, resulting in lengthy internal computational paths. This high complexity may increase the latency during model inference, affecting the overall system response speed.

[0004] YOLO, a classic single-stage object detection algorithm, is designed to achieve a dynamic balance between detection speed and accuracy by optimizing the network structure and detection mechanism. It solves the joint problem of object localization and classification through a single forward propagation. Taking the YOLOv8 to YOLOv11 series of iterations as examples, YOLOv8 innovatively integrated instance segmentation and pose estimation modules while inheriting basic detection functions, expanding the algorithm's application boundaries. However, in benchmark tests, its overall performance in terms of detection accuracy and inference speed lagged behind later versions. YOLOv10 achieved a breakthrough in inference efficiency through lightweight network architecture design and optimized feature extraction strategies, but its limited feature representation capabilities made it difficult to meet the high-precision application requirements when facing object detection tasks across scales and complex backgrounds. The latest YOLOv11 version, especially the YOLO11m model, has built a more balanced performance system between detection accuracy, computational efficiency and model parameter quantity through techniques such as depthwise separable convolution and feature fusion. However, the deployment of this model in resource-constrained environments is limited by its high computational complexity and memory consumption, which poses challenges in ensuring real-time performance.

[0005] YOLOv9, based on the information bottleneck principle and reversible function architecture, maintains network depth while efficiently preserving key data, significantly improving model convergence efficiency and detection performance, and demonstrating excellent accuracy in complex scene target detection tasks. However, in cross-scale target detection applications of fiber optic distributed vibration sensing systems (DVS), the YOLOv9 model reveals inherent limitations. Although the model has certain advantages in detecting small and overlapping targets, its fixed feature extraction paradigm is insufficient to meet the needs of effectively capturing multi-scale detailed information when faced with noise interference unique to fiber optic sensing data and monitoring scenarios with significant differences in target feature distribution across scales. This results in a high risk of missed detections and false detections in cross-scale target detection tasks. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a low-false-detection-rate fiber optic distributed sensor signal recognition method and system. Based on an improved YOLOv9 model design, it innovatively embeds LDConv and MSAA modules, and achieves deep optimization of the network structure by optimizing the information transmission path between convolutional layers and enhancing feature fusion efficiency. This improvement significantly enhances the network's feature representation capability for cross-scale targets, effectively reduces the false-detection probability in complex scenarios, and greatly improves the stability and accuracy of the model in actual monitoring, truly meeting the application requirements of cross-scale target detection in fiber optic environments.

[0007] This invention is achieved through the following technical solution: A method for identifying fiber optic distributed sensor signals with a low false detection rate is provided, comprising the following steps: S1. Vibration signal data of the sensing optical fiber is collected through a distributed vibration sensing system, and the vibration signal data is preprocessed to obtain a vibration signal dataset. S2. Build an improved YOLOv9 network model, the improvements of which include: Replace the first and second Conv modules in the original YOLOv9 network with the first and second LDConv modules; Add the first to sixth MSAA modules after the fourth to ninth RepNCSPELAN4 modules; The probes connected to the fifth, sixth, seventh, eighth, and ninth RepNCSPELAN4 modules in the original YOLOv9 network are changed to be connected to the second, third, fourth, fifth, and sixth MSAA modules respectively. S3. Divide the collected vibration signal dataset into a training set, a validation set, and a test set. Use the training set to iteratively train the improved YOLOv9 network model until the model converges, and obtain the trained improved YOLOv9 network model. S4. Input the real-time collected vibration signal data into the trained improved YOLOv9 network model, perform forward propagation inference for recognition and prediction, and obtain the recognition result containing the target's category information and location information.

[0008] Furthermore, step S1 specifically includes: S101. A fiber optic distributed vibration sensing system and a fiber optic data acquisition system are used. S102. Create cross-scale events at different locations and collect vibration data; S103. Preprocess the data, including differencing and standardization, to construct the dataset; Furthermore, in step S103, the collected vibration data is first processed into a two-dimensional array in terms of time and space, and then the two-dimensional array is processed by difference and standardization in sequence. The formula used for difference processing is: ; in: This represents the vibration intensity in the i-th row and j-th column; The formula used for standardization is: ; ; ; in: These are the numerical values ​​from the original dataset, representing the elements in the array. This represents the average value. It represents the standard deviation.

[0009] Furthermore, in step S2, the LDConv module is a convolutional module designed based on the deformable convolution concept. It generates a spatial offset of the input feature vector through sub-modules. This offset determines the position of subsequent sampling. The module generates a basic sampling grid, constructs regular sampling coordinates, and calculates the final sampling position by combining the offset. To ensure sampling stability, the module uses rounding to generate four adjacent integer coordinate points and calculates the corresponding weights through bilinear interpolation. Then, it extracts feature values ​​based on these coordinate points, performs weighted summation based on the weights to obtain resampled features, and finally rearranges the features to adapt to standard convolution operations, and outputs the final result through the main convolutional structure conv.

[0010] Furthermore, in step S2, the MSAA module is an innovative feature processing structure that enhances the expressiveness of deep learning models through multi-layered feature extraction and attention mechanisms. Input features undergo channel dimensionality reduction via convolutional layers, compressing high-dimensional features to low dimensions to reduce computational complexity. The dimensionality-reduced features are then fed in parallel into three convolutional layers of different sizes to capture cross-scale features. These cross-scale features are then element-wise added to obtain fused features, which are then spatially weighted using spatial attention. Simultaneously, the original dimensionality-reduced features are enhanced with channel attention. Finally, the spatially attention-enhanced cross-scale features and the channel-attention-enhanced original features are added and fused, and the features are restored to the target dimension via convolutional layers, completing the entire forward propagation process.

[0011] Furthermore, in step S3, the training process uses the PyTorch framework, with a learning rate of 0.01, a batch size of 16, 50 iterations, and is accelerated using CUDA.

[0012] Furthermore, in step S4, when the spatial dimension (length of spatial sampling points) of the input vibration signal exceeds a preset length, a sliding window method is used to segment it. The steps are as follows: Starting from the left end of the vibration signal, extract signals of size [value missing]. x·y Submatrices of size ′ are created until they cover the entire width of the matrix; where the width of each submatrix is ​​. y ',high x The height remains consistent with the original matrix, and the width of the overlapping region between adjacent sub-matrices is [missing information]. x sliding step size S for y ′- x Record the horizontal starting position of each submatrix within the original matrix. x offset = k·S ,k For submatrix index; The improved YOLOv9 network model is used to infer each submatrix to obtain the detection results in the local coordinate system. The detection results include the coordinates of the center point of the bounding box. ; The detection results in local coordinates are transformed to the original matrix global coordinate system, with the global center x-coordinate... y-axis ; Group the detection boxes of all sub-matrices by category and represent the bounding boxes in center point form. Convert to corner form ,in: ; Non-maximum suppression (NMS) is performed on the detection boxes for each category, based on a set IoU threshold. θ Eliminate duplicate detection boxes, retain the detection result with the highest confidence, and finally output the deduplication detection result in the global coordinate system.

[0013] Furthermore, step S5 involves deploying the trained and improved YOLOv9 network model on the terminal system, connecting it to the real-time data stream collected by the distributed vibration sensing system, and performing real-time monitoring of disturbance events.

[0014] A system for identifying fiber-optic distributed sensor signals using a low false detection rate method includes: The vibration signal acquisition module is used to acquire vibration data and obtain a two-dimensional array; The data preprocessing module is used to differ, standardize, and generate a data matrix from a two-dimensional array; The target event detection module has a built-in pre-trained improved YOLOv9 network model for target recognition and localization on the input data matrix.

[0015] The beneficial effects of this invention are: This invention introduces the LDConv mechanism into the YOLOv9 network model. This module generates spatial offsets of input features through learnable convolutional layers, which guide subsequent adaptive sampling processes. Specifically, the module first generates a set of normalized basic sampling position patterns, then combines them with the spatial dimensions of the current features to construct a regular initial sampling coordinate grid. The output dynamic offsets are then superimposed onto the initial coordinates to obtain the final deformable sampling positions. To ensure the stability and gradient differentiability of the sampling process, a rounding strategy is used to determine the four nearest integer coordinate positions, and weight coefficients for each position are calculated based on bilinear interpolation. Feature values ​​at corresponding positions are extracted from the original features based on discrete coordinates, and weighted fusion is performed using interpolation weights to obtain the resampled offset features. Subsequently, the dimensions of these features are reorganized to meet the input format requirements of subsequent standard convolutional operations. Finally, the transformed feature response is output by the backbone convolutional layer Conv. Compared to traditional convolution, LDConv exhibits several significant advantages. Traditional convolutional algorithms employ a fixed and uniform sampling layout, limiting their receptive field to a rigid grid structure. LDConv introduces a learnable adaptive sampling mechanism that dynamically adjusts the sampling position based on the content of the input features, allowing the network to focus on more semantically meaningful key signal regions and significantly improving the discriminative power of feature representation. The introduction of bilinear interpolation enables LDConv to achieve fine-grained sampling, breaking through the limitation of traditional methods that only sample at integer coordinates, thus enhancing the continuity and accuracy of spatial localization. In target localization tasks, its learnable offset mechanism allows the receptive field to flexibly align with the key structures of the target, effectively improving the accuracy of boundary regression. Strong spatial awareness further optimizes the fineness of the localization boundary, and the adaptive sampling strategy gives the model stronger robustness, enabling it to maintain stable feature extraction performance even in complex scenarios such as target deformation, occlusion, or non-rigid geometric changes. The module effectively suppresses the instability caused by drastic fluctuations in offset parameters during the early stages of training by finely controlling the learning rate of the offset parameters, improving the convergence and reliability of the overall training process. LDConv, through its innovative deformable sampling architecture, significantly enhances the model's ability to perceive spatial structure while maintaining high computational efficiency. Its dynamic, continuous, and content-aware sampling method provides strong support for high-precision feature understanding tasks, demonstrating outstanding performance in improving the position sensitivity and localization robustness of object detection systems.

[0016] This invention introduces the MSAA mechanism into the YOLOv9 network model. This module is a highly efficient feature enhancement structure that integrates cross-scale feature extraction and dual-path attention mechanisms, aiming to improve the model's ability to perceive complex spatial patterns and the richness of feature representation. First, a convolutional layer performs channel dimensionality reduction on the input features, compressing the high-dimensional input into a lower-dimensional intermediate space, thereby reducing the computational burden and preserving key semantic information. The dimensionality-reduced features are then fed in parallel into three convolutional branches of different scales, capturing local details, intermediate-scale structure, and global contextual information, respectively. Subsequently, element-wise summation achieves preliminary cross-scale feature fusion. This parallel multi-branch strategy overcomes the limitation of the receptive field in traditional single convolutional kernels, significantly enhancing the model's adaptability to cross-scale content.

[0017] Building upon cross-scale fusion, MSAA introduces a refined dual-path attention enhancement mechanism. On one hand, the fused cross-scale features are processed by a spatial attention module. This module calculates the responses of average pooling and max pooling along the channel dimension, concatenates them, and generates a spatial attention vector through a 1×1 convolution and a sigmoid function, thus highlighting key spatial locations in the signal. On the other hand, the original dimensionality-reduced features are fed into a channel attention module. Channel-level statistical information is extracted through global average pooling and global max pooling, processed by a shared two-layer 1×1 convolutional network, summed, and activated to generate channel attention weights, thereby strengthening semantically important feature channels. This separate design, applying spatial attention to the cross-scale fused features and channel attention to the original features, achieves complementary enhancement, avoids information confusion, and improves the accuracy of feature selection. The two attention-weighted features are element-wise summed to complete the final fusion, and a 1×1 convolutional layer restores the number of channels to the target output dimension, forming a complete forward flow of "dimensionality reduction—cross-scale extraction—dual attention enhancement—fusion and dimensionality enhancement." Compared to the limitations of traditional convolutional modules that use only fixed-scale convolutional kernels, and the redundancy caused by serial processing in attention mechanisms such as CBAM, MSAA deeply embeds the attention mechanism into the cross-scale feature construction process, forming an efficient and collaborative processing flow. Its lightweight design of "dimensionality reduction first, then processing, then dimensionality increase" significantly reduces the computational cost of performing cross-scale convolutions in high-dimensional spaces, maintaining powerful expressive capabilities while possessing good engineering practicality. Through its innovative structural design, MSAA achieves a good balance between feature diversity, attention accuracy, and computational efficiency.

[0018] This invention constructs a concise data processing architecture. By implementing a time-domain feature extraction strategy based on differential operations and a standardized data preprocessing strategy, it effectively ensures the real-time processing performance of the data stream while avoiding computational delays introduced by high-order complex algorithms. Experimental results demonstrate that this scheme significantly improves the average accuracy of cross-scale event recognition in complex noisy environments, showcasing good engineering application value and technical superiority. Attached Figure Description

[0019] Figure 1 This is a flowchart of the fiber optic distributed sensor signal recognition method with low false detection rate according to the present invention.

[0020] Figure 2 This is the dataset collected by the fiber optic distributed vibration sensing system in this invention.

[0021] Figure 3 This is a structural diagram of the improved YOLOv9 network in this invention.

[0022] Figure 4 This is a structural diagram of the existing YOLOv9 network.

[0023] Figure 5 This is a schematic diagram of the LDConv module in this invention.

[0024] Figure 6 This is a schematic diagram of the MSAA module in this invention.

[0025] Figure 7 This is a confusion matrix diagram of the improved model in this invention.

[0026] Figure 8 The results are the test results of the three detection methods YOLOv10, YOLOv9 and YOLOv9-improved used in this invention.

[0027] Figure 9 This is a diagram showing the results of the ablation experiment in this invention.

[0028] Figure 10 This is a graph showing the test results of the two detection models, YOLOv9 and YOLOv9-improved, on the terminal system test set in this invention. Detailed Implementation

[0029] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0030] like Figure 1 As shown, a method for identifying fiber optic distributed sensor signals with a low false detection rate includes the following steps: S1. Vibration signal data of the sensing optical fiber is collected through a distributed vibration sensing system, and the vibration signal data is preprocessed to obtain a vibration signal dataset. In this embodiment, vibration information of the optical fiber under urban noise for 17.76km is collected, and the vibration information is preprocessed to obtain a vibration signal dataset.

[0031] S101. A fiber optic distributed vibration sensing system and a fiber optic data acquisition system are used. The fiber optic distributed vibration sensing system includes a narrow-linewidth laser (NLL), an acousto-optic modulator (AOM), an erbium-doped fiber amplifier (EDFA), an optical circulator, a photodetector, a data acquisition card, and host computer software. Continuous light emitted from the NLL is modulated into pulsed light by the AOM, amplified by the EDFA, and then enters the sensing fiber via the optical circulator (CIR). Backscattered Rayleigh light generated during the transmission of the pulsed light through the fiber passes through the circulator again to the photodetector (PD). Data is collected by the data acquisition card (DAQ) and transmitted to the data processing system for processing, obtaining vibration data at various points along the fiber. In this embodiment, the fiber optic monitoring signals include two-dimensional vibration signals from four types of cross-scale target events: impact (positive sample), sedimentation (positive sample), drilling (positive sample), and traffic flow (negative sample).

[0032] S102. Initiate cross-scale events at different locations and collect vibration data; such as... Figure 2 As shown, four types of cross-scale target events—impact, sedimentation, drilling, and traffic flow—were generated at different locations on the optical fiber, with 1,000 samples for each event.

[0033] S103. Preprocess the data, including differencing and standardization, and construct the dataset; First, the collected vibration data is divided into a two-dimensional array in terms of time and space. Then, the two-dimensional array is subjected to difference and standardization processes in sequence to construct a dataset. The formula used for difference processing is: ; in: This represents the vibration intensity in the i-th row and j-th column; The formula used for standardization is: ; ; ; in: These are the numerical values ​​from the original dataset, representing the elements in the array. This represents the average value. It represents the standard deviation.

[0034] S2. Build an improved YOLOv9 network model, the improvements of which include: Replace the first and second Conv modules in the original YOLOv9 network with the first and second LDConv modules; Add the first to sixth MSAA modules after the fourth to ninth RepNCSPELAN4 modules.

[0035] The probes connected to the fifth, sixth, seventh, eighth, and ninth RepNCSPELAN4 modules in the original YOLOv9 network are changed to be connected to the second, third, fourth, fifth, and sixth MSAA modules respectively.

[0036] In addition to the improvements made to the existing YOLOv9 network as described above, other modules and their functions in the existing YOLOv9 network are also adopted in the improved YOLOv9 network. Therefore, the functions of the existing modules in the existing YOLOv9 network adopted in the improved YOLOv9 network will not be described in detail in this application.

[0037] The improved YOLOv9 network model in this application improves the Head and Backbone parts.

[0038] like Figure 3 As shown, the improved Head section includes the first LDConv module, the second LDConv module, the first ELAN1 module, the first AConv module, the first RepNCSPELAN4 module, the second AConv module, the second RepNCSPELAN4 module, the third AConv module, and the third RepNCSPELAN4 module connected in sequence.

[0039] The improved backbone consists of the following modules connected in sequence: the first SPPELAN module, the first Upsample module, the first Concat module, the fourth RepNCSPELAN4 module, the first MSAA module, the second Upsample module, the second Concat module, the fifth RepNCSPELAN4 module, the second MSAA module, the fourth AConv module, the third Concat module, the sixth RepNCSPELAN4 module, the third MSAA module, the fifth AConv module, the fourth Concat module, the seventh RepNCSPELAN4 module, the fourth MSAA module, the second SPPELAN module, the third Upsample module, the fifth Concat module, the eighth RepNCSPELAN4 module, and the fifth... The system consists of an MSAA module, a fourth Upsample module, a sixth Concat module, a ninth RepNCSPELAN4 module, and a sixth MSAA module. The first RepNCSPELAN4 module is also connected to the second and sixth Concat modules; the second RepNCSPELAN4 module is also connected to the first and fifth Concat modules; the first MSAA module is also connected to the third Concat module; the first Upsample module is also connected to the fourth Concat module; the second MSAA module is also connected to the first probe; the third MSAA module is also connected to the second probe; the fourth MSAA module is also connected to the third probe; the second Upsample module is also connected to the fourth probe; the fifth MSAA module is also connected to the fifth probe; and the sixth MSAA module is also connected to the sixth probe.

[0040] like Figure 5As shown, the core structure of the LDConv module consists of three key components: a learnable offset generation network, a dynamic sampling coordinate calculation unit, and a resampled feature reconstruction and convolution processing path. This structure, while maintaining compatibility with standard convolution, introduces flexible geometric transformation capabilities. LDConv first generates a set of normalized base coordinate patterns based on the specified number of sampling points N, then maps them to the actual coordinate space in conjunction with the spatial dimensions of the current features, forming an initial regular sampling grid. The output offsets are superimposed on this initial coordinate to obtain deformable dynamic sampling positions. To ensure the continuity and differentiability of the sampling operation, the module adopts a bilinear interpolation strategy, selecting four nearest-neighbor integer positions around each non-integer coordinate and assigning weights based on distance. Weighted fusion is used to resample the feature values, extracting the responses from these positions from the original features, and redimensionalizing the resampled features to adapt them to the input format of standard convolution operations. Finally, the backbone convolutional layer completes the feature transformation and outputs the results. LDConv breaks the dual limitations of traditional convolution, namely fixed sampling structure and quadratic growth of the number of parameters. On the one hand, it proposes a general initial coordinate generation algorithm that supports an arbitrary number of sampling points, thereby achieving linear expansion of the convolution kernel parameters and significantly improving the model's flexibility in adjusting the balance between parameter efficiency and performance. On the other hand, unlike standard deformable convolution which is only applicable to regular grids, LDConv allows the construction of irregular initial sampling shapes and, combined with learnable offsets, achieves truly "arbitrary shape" convolution operations. This design enhances the model's adaptability to irregular target structures. Compared to traditional convolution and standard deformable convolution, LDConv demonstrates better parameter utilization and computational efficiency while improving detection accuracy. Its adaptive sampling mechanism significantly enhances the model's ability to focus on key target regions and improves the accuracy of bounding box regression, especially in the detection of small targets, occluded targets, and deformed targets. Through structural innovation and process optimization, LDConv achieves more flexible and accurate spatial feature extraction while maintaining high computational efficiency.

[0041] like Figure 6As shown, the MSAA module is an innovative structure designed to enhance the feature representation capabilities of deep neural networks. Its core lies in achieving refined modeling of spatial and channel-dimensional information through the synergistic effect of cross-scale convolution and a dual-path attention mechanism. The module adopts a "dimensionality reduction, fusion, and enhancement" processing flow: First, the input feature vector undergoes channel compression through a 1×1 convolutional layer, reducing high-dimensional features to an intermediate dimension, effectively reducing the computational complexity of subsequent operations. The dimensionality-reduced features are simultaneously fed into three parallel branches, using 3×3, 5×5, and 7×7 convolutional kernels for feature extraction, capturing local details, intermediate-range structure, and global contextual information in parallel. Cross-scale fused features are then generated through element-wise addition. This design overcomes the bottleneck of the limited receptive field of traditional single convolutional kernels, enabling the model to perceive semantic content at different scales more comprehensively. Based on cross-scale feature fusion, MSAA introduces a separate dual-attention enhancement mechanism, significantly improving the accuracy of feature selection. The module employs two independent paths to apply spatial attention and channel attention: On one hand, the cross-scale fused features are input to the spatial attention module, where statistical operations of average pooling and max pooling are performed on the channel dimensions. After concatenation, the vectors are generated through 1×1 convolution and sigmoid activation, thus highlighting key spatial regions in the data matrix. On the other hand, the original dimensionality-reduced features are processed by the channel attention module, which extracts channel-level responses using global average pooling and global max pooling. These responses are then transformed by a shared two-layer 1×1 convolutional network, added together, and activated to generate channel attention weights, which are used to strengthen the most discriminative feature channels. This strategy of applying spatial attention to cross-scale fused features and channel attention to the original features achieves functional complementarity, avoids information redundancy or confusion, and enhances the representation quality of features. The parallel cross-scale convolutional structure of the MSAA module breaks the dependence of traditional convolution on a fixed receptive field, giving the model stronger scale adaptability. Unlike serial attention mechanisms such as CBAM, MSAA deeply integrates the attention module with the cross-scale extraction process, forming an efficient workflow of "cross-scale extraction → attention enhancement → feature fusion," improving processing efficiency and feature utilization.

[0042] S3. Divide the collected vibration signal dataset into a training set, a validation set, and a test set. Use the training set to iteratively train the improved YOLOv9 network model until the model converges, and obtain the trained improved YOLOv9 network model.

[0043] The vibration signal dataset was divided into training, validation, and test sets in a 7:2:1 ratio using random allocation.

[0044] Based on the existing YOLOv9 model, the model structure is optimized and parameters are tuned using the training dataset to obtain an optimized model.

[0045] Specifically, the learning rate of the model was set to 0.01, the training batch size was 16, the improved YOLOv9 network model was implemented with GPU acceleration supported by CUDA, and the number of iterations of the improved YOLOv9 vibration signal neural network model was set to 50 according to the gradient change of the loss function. The training of the entire improved network was based on the PyTorch deep learning framework.

[0046] The existing YOLOv9 network's loss function, Loss, consists of localization loss (loc), classification loss (cls), confidence loss (obj), and cross-level consistency loss (cc), as shown below: Loss=λloc·Lloc+λcls·Lcls+λobj·Lobj+λcc·Lcc; λ is the weighting coefficient. In this application, λloc is 0.75, λcls is 0.25, λobj is 0.25, and λcc is 0.1.

[0047] S4. Input the real-time collected vibration signal data into the trained improved YOLOv9 network model, perform forward propagation inference for recognition and prediction, and obtain the recognition result containing the target's category information and location information.

[0048] S5. Deploy the trained and improved YOLOv9 network model on the terminal system, connect it to the real-time data stream collected by the distributed vibration sensing system, and perform real-time monitoring of disturbance events.

[0049] The initial configuration of the data receiving environment is completed by establishing a TCP (Transmission Control Protocol) network communication link. Specifically, a TCP socket instance is created in the terminal system, bound to a specified local IP address and port, and a connection request is initiated to the remote distributed vibration sensing (DVS) device. After the TCP connection is successfully established, the data buffer and sensor data storage array are initialized.

[0050] A fixed-length data packet reception mechanism based on the TCP protocol is adopted to parse sensor data packets according to a preset data format: the TCP data stream is sliced ​​using a fixed-length window to parse structured fields such as the packet header, packet type field, payload data, and packet trailer. A sliding buffer technique is used to dynamically assemble data packets, accurately identifying the complete data packet structure. When a packet type identifier conforming to the system protocol specification is detected, the data packet is determined to be a valid sensor data carrier. The payload portion is obtained through the data extraction module, and a format conversion operation from binary raw data to a decimal numerical sequence is performed.

[0051] Organize sensor data and trigger data processing flow: Store the parsed decimal sensor data row by row into a NumPy two-dimensional array, and adopt a rolling update strategy, with new data overwriting the oldest row.

[0052] When the spatial dimension (length of spatial sampling points) of the input vibration signal exceeds a preset length, a sliding window method is used to segment it. The steps are as follows: The vibration signal matrix is ​​divided into 200×8000 pixels for segmentation and identification. A sliding window segmentation is performed on the original matrix. Each sub-matrix is ​​1000 pixels wide, with an overlap of 200 pixels, and a sliding step size S of 800 pixels. Starting from the left end of the matrix, sub-matrices of size 200×1000 are extracted sequentially until the entire 8000-pixel matrix width is covered. Each block generated in this process retains its original height (i.e., time dimension information) to ensure the temporal structure is not disrupted. The horizontal starting position of each sub-matrix within the original matrix is ​​recorded. x offset = k·S , k This is the submatrix index, used for subsequent coordinate reconstruction.

[0053] Object detection and inference are performed independently on each submatrix to obtain the detection results in the local coordinate system. The detection results include the coordinates of the center point of the bounding box. .

[0054] The detection results in local coordinates are transformed to the original matrix global coordinate system, with the global center x-coordinate... y-axis .

[0055] Group the detection boxes of all sub-matrices by category and represent the bounding boxes in center point form. Convert to corner form ,in: .

[0056] Non-maximum suppression (NMS) is performed on the detection boxes for each category, based on a set IoU threshold. θ The algorithm eliminates duplicate detection boxes, retaining only the detection result with the highest confidence, and finally outputs the deduplicated detection result in the global coordinate system. Since the same target may appear in two adjacent overlap matrices, thus being detected twice or multiple times, Non-Maximum Search (NMS) determines whether they are duplicates by comparing the Intersection over Union (IoU) between detection boxes of the same category. Given an IoU threshold θ of 0.5, the algorithm retains the detection box with the highest confidence and suppresses other boxes with an IoU exceeding the threshold. This process is performed independently for each category, ensuring that targets of different categories do not affect each other. After NMS, only one optimal detection result is retained for each true target, effectively avoiding the recognition redundancy problem caused by block segmentation while solving the recognition problem.

[0057] The output is a final detection set after deduplication and coordinate restoration, forming a complete detection result. This enables efficient and accurate detection of the vibration signal matrix.

[0058] While generating the vibration signal matrix, an asynchronous task is started using a thread pool to call a pre-deployed model to perform real-time analysis on the generated vibration signal matrix, identify the disturbance event areas and their categories, and ensure that the vibration signal matrix processing does not block the reception and parsing of subsequent data packets, thus guaranteeing the real-time performance and continuity of the system.

[0059] For each successfully parsed valid data packet, a 10-byte acknowledgment response packet is returned to the fiber optic distributed vibration sensing system (DVS) device to notify the server that the data packet has been correctly received and processed.

[0060] A system for identifying fiber-optic distributed sensor signals using a low false detection rate method includes: The vibration signal acquisition module is used to acquire vibration data and obtain a two-dimensional array in terms of time and space. The data preprocessing module is used to differ, standardize, and generate a data matrix from a two-dimensional array; The target event detection module has a built-in pre-trained improved YOLOv9 network model for target recognition and localization on the input data matrix.

[0061] To verify the effectiveness of this application in cross-scale target event classification and localization, the detection performance of the method in this application (hereinafter referred to as YOLOv9-improved), the YOLOv10 target detection method (from "YOLOv10: Real-Time End-to-End Object Detection"), and the YOLOv9 detection method (from "YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information") on the validation set of this application were compared.

[0062] To ensure fairness in the testing, this application trains the network model using the same training strategy on the training set of this embodiment, based on the existing detection methods YOLOv10, YOLOv9, and the improved YOLOv9 method described in this application. Then, it tests the network model using the same testing strategy on the validation set of this application. The test results are as follows: Figure 8 As shown.

[0063] Precision represents the accuracy of each network model, i.e., the proportion of samples that are predicted to be positive but are actually positive; Recall represents the recall of each network model, i.e., the proportion of samples that are predicted to be positive but are actually positive out of the total number of positive samples; F1 represents the harmonic mean of the accuracy and recall of each network model; mAP@0.5 represents the average accuracy of each network model when IOU=0.5; FAR represents the false alarm rate, MDR represents the false negative rate, and FPS represents the inference speed; in this invention, the larger the values ​​of Precision, Recall, FPS, F1, and mAP@0.5, the better, while the smaller the values ​​of FAR and MDR, the better.

[0064] The improved YOLOv9 network model in this application is compared with the YOLOv9 network model, as follows: Figure 8 As shown: Precision Index Comparison: The monitoring method based on the improved YOLOv9 in this invention achieves a Precision index of 0.941, which is 3.4 percentage points higher than the existing YOLOv9 detection method.

[0065] Recall metric comparison: The monitoring method based on the improved YOLOv9 in this invention achieves a Recall metric of 0.931, which is 0.2 percentage points lower than the existing YOLOv9 detection method.

[0066] F1 score comparison: The monitoring method based on the improved YOLOv9 in this invention achieves an F1 score of 0.936, which is 1.7 percentage points higher than the existing YOLOv9 target detection method.

[0067] Comparison of mAP@0.5 index: The monitoring method based on the improved YOLOv9 in this invention achieves 0.978 in mAP@0.5, which is 1.0 percentage point higher than the existing YOLOv9 detection method.

[0068] The recall rate decreased slightly, but YOLOv9-improved outperformed the original model in core metrics, indicating that the detection method based on the improved YOLOv9 in this invention performs better in recognition and classification when detecting cross-scale events under complex noise conditions.

[0069] like Figure 9As shown in the ablation experiments, YOLOv9-improved exhibits the best performance across key metrics, demonstrating its significant advantages in target detection tasks. The introduction of the DConv module significantly enhances the model's ability to perceive detected targets and improves its modeling ability for geometric changes in complex scenes. This is reflected in the YOLOv9-LDConv model, where mAP@0.5 reaches 0.973, a 0.5 percentage point improvement over the original YOLOv9. While its accuracy and F1 score have improved, they are not optimal. The MSAA module enhances the dynamic fusion of cross-scale contextual information, strengthens key feature regions through an attention mechanism, and suppresses background noise. YOLOv9-MSAA achieves a recall of 0.939, the highest among all models, indicating stronger target detection capabilities in complex backgrounds. Its F1 score is 0.927, and its mAP@0.5 score is 0.973, demonstrating that this module effectively improves the comprehensiveness and stability of detection. Its accuracy is lower than YOLOv9-improved, indicating that although the attention mechanism enhances feature selection capabilities, it still carries a certain risk of false activations. YOLOv9-improved achieves a comprehensive performance improvement through the fusion of LDConv and MSAA modules for collaborative optimization. This model achieves the highest levels in accuracy, F1 score, and mAP@0.5, while maintaining a high recall of 0.931, demonstrating that it has achieved the best balance between reducing false positives and false negatives. By significantly improving accuracy and overall performance while maintaining high recall, it has become the most practically valuable detection model among the four methods currently available, suitable for real-world applications with extremely high requirements for accuracy and robustness.

[0070] from Figure 7 The confusion matrix shows that the improved model has a correct prediction rate of 0.91, 0.97, and 0.95 for settlement, impact, and electric drilling, respectively, demonstrating high accuracy.

[0071] like Figure 10 As shown, at a confidence level of 0.5, the performance of the improved model deployed on the terminal system test set was tested. In a real-world scenario, 1500 data signals generated from noise data collected by the monitoring system within 0.5 hours were added to the test set along with 400 data signals. The detection accuracy, false negative rate, and false positive rate were tested and recorded. The performance of YOLOv9 on key indicators was compared with other models, and the overall results showed that structural optimization significantly improved the comprehensive performance.

[0072] The performance comparison on the test set shows that the improved YOLOv9 (YOLOv9-improved) demonstrates significantly better overall performance than the original YOLOv9 in terms of detection accuracy, false positive control, and false negative suppression. Regarding detection accuracy, the improved model's accuracy increased from 0.912 to 0.974, indicating a substantial increase in the proportion of true targets in its output detection results, making predictions more reliable. The false alarm rate (FAR) decreased from 0.013 to 0.005, demonstrating that the model can more effectively suppress false activations under complex backgrounds or noisy interference, significantly enhancing its discriminative ability and robustness.

[0073] In terms of recall, the improved model's false negative rate (MDR) decreased from 0.021 to 0.012, reflecting a significant enhancement in its ability to capture true targets and reduce the occurrence of false negatives. While maintaining high accuracy, it also improved the completeness of detection, which is particularly beneficial for identifying occluded targets. The model's balance between accuracy and recall has been optimized, resulting in a significant leap in overall detection quality.

[0074] The improved model's FPS decreased slightly from 420 to 369, but it still maintains an extremely high level of real-time processing. Overall, YOLOv9-improved, through reasonable structural optimization, achieved a significant improvement in detection performance with almost no sacrifice in real-time performance. The performance of the improved YOLOv9 on the test set fully validates its effectiveness and advancement. Substantial progress has been made in reducing false positives, minimizing false negatives, and improving overall detection reliability, demonstrating stronger generalization ability and application potential.

[0075] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A low false alarm rate method of optical fiber distributed sensing signal identification, characterized in that: Comprising the following steps: S1, collecting vibration signal data of the sensing optical fiber by a distributed vibration sensing system, and preprocessing the vibration signal data to obtain a vibration signal data set; S2, building an improved YOLOv9 network model, the improvements including: Replacing the first and second Conv modules in the original YOLOv9 network with the first and second LDConv modules; Adding the first to sixth MSAA modules after the fourth to ninth RepNCSPELAN4 modules; Connecting the detection heads of the fifth, sixth, seventh, eighth and ninth RepNCSPELAN4 modules in the original YOLOv9 network to the second, third, fourth, fifth and sixth MSAA modules, respectively; S3, dividing the collected vibration signal data set into a training set, a validation set and a test set, and iteratively training the improved YOLOv9 network model with the training set until the model converges, to obtain a trained improved YOLOv9 network model; S4, inputting the real-time collected vibration signal data into the trained improved YOLOv9 network model, performing forward propagation inference recognition prediction, and obtaining the recognition result containing the category information and position information of the target.

2. The low false alarm rate optical fiber distributed sensing signal identification method of claim 1, wherein: Step S1 specifically includes: S101, building a collection system using a fiber distributed vibration sensing system and an optical fiber; S102, manufacturing cross-scale events at different positions and collecting vibration data; S103, preprocessing the data, including difference and standardization processing, and constructing a data set.

3. The low false alarm rate optical fiber distributed sensing signal identification method of claim 2, wherein: In step S103, the collected vibration data is first obtained as a two-dimensional array in time and space dimensions, and then the two-dimensional array is sequentially subjected to difference processing and standardization processing; The formula used for difference processing is: ； wherein: represents the shock intensity of the ith row, jth column; The formula used for standardization processing is: ； ； ； where: is the value in the original data set, representing an element in the array, denotes the mean value, denotes the standard deviation.

4. The low false alarm rate optical fiber distributed sensing signal identification method of claim 1, wherein: In step S2, the LDConv module is a convolution module designed based on the deformable convolution idea, which generates a spatial offset of the input feature vector through a sub-module to determine the position of subsequent sampling, generates a basic sampling grid through the module, constructs a regular sampling coordinate, and calculates the final sampling position combined with the offset; four adjacent integer coordinate points are generated by rounding off, and the corresponding weights are calculated by bilinear interpolation, the feature values are extracted according to the coordinate points, and the weighted sum is obtained by weighting the weights, and finally the features are rearranged to adapt to the standard convolution operation, and the final result is output through the main convolution structure conv.

5. The low false alarm rate optical fiber distributed sensing signal identification method of claim 4, wherein: In step S2, the MSAA module is a feature processing structure with multi-level feature extraction and attention mechanism, the input features are subjected to channel dimension reduction through a convolution layer to compress high-dimensional features to low-dimensional features to reduce computational complexity, the reduced features are input in parallel to three convolution layers of different sizes to capture cross-scale features; The cross-scale features are obtained by element-wise addition to obtain fused features, and the spatial attention is weighted using spatial attention, while the original reduced features are enhanced by channel attention, and finally the cross-scale features enhanced by spatial attention and the original features enhanced by channel attention are added and fused, and the features are restored to the target dimension through a convolution layer to complete the forward propagation process.

6. The low false alarm rate optical fiber distributed sensing signal identification method of claim 1, wherein: In step S3, the training process uses the PyTorch framework, the learning rate is set to 0.01, the batch size is 16, and the iteration is 50 times, and CUDA acceleration is used.

7. The low false alarm rate optical fiber distributed sensing signal identification method of claim 1, wherein: In step S4, when the spatial dimension (spatial sampling point length) of the input vibration signal exceeds the preset length, the sliding window method is used for segmentation processing, specifically: Starting from the left end of the vibration signal, extract signals of size [value missing]. x·y Submatrices of size ′ are created until they cover the entire width of the matrix; where the width of each submatrix is ​​. y ',high x The height remains consistent with the original matrix, and the width of the overlapping region between adjacent sub-matrices is [missing information]. x sliding step size S for y ′- x Record the horizontal starting position of each submatrix within the original matrix. x offset = k·S , k For submatrix index; The improved YOLOv9 network model is used for inference on each sub-matrix to obtain a detection result in a local coordinate system, and the detection result includes a center point coordinate of a bounding box ; converting the detection results in local coordinates to the original matrix global coordinate system, global center horizontal coordinate , vertical coordinate ; grouping the bounding boxes of all sub-matrices by class, and converting the bounding box representation from a center point form to a corner point form wherein: ; Non-maximum suppression (NMS) is performed on each category of detection frame based on a set IoU threshold θ Duplicate detection frames are eliminated, and the detection result with the highest confidence is retained, and finally the de-duplicated detection result in the global coordinate system is output.

8. The low false alarm rate optical fiber distributed sensing signal identification method of claim 1, wherein: Further comprising step S5, deploying the trained improved YOLOv9 network model to the terminal system, interfacing with the real-time data stream collected by the distributed vibration sensing system, and performing real-time monitoring of the disturbance event.

9. A system for using the low false alarm rate optical fiber distributed sensing signal recognition method according to any one of claims 1 to 8, characterized in that: Comprise: a vibration signal acquisition module for collecting vibration data and obtaining a two-dimensional array; a data preprocessing module for differentiating, standardizing and generating a data matrix from the two-dimensional array; a target event detection module, which has a trained improved YOLOv9 network model built-in, for target recognition and positioning of the input data matrix.

Images