Distributed optical fiber early warning regulation method for underground structure leakage hidden danger distribution

Abstract

The application discloses a distributed optical fiber early warning and regulation method for underground structure leakage hidden danger, and relates to the technical field of leakage monitoring.The application constructs an adaptive multi-parameter distributed optical fiber sensing network with FBG enhanced nodes, dynamically adjusts sensing parameters in combination with environmental parameters, and collects temperature, vibration and strain data; data is processed through a wavelet transform, adaptive filtering and EMD fusion algorithm, a multidimensional space-time feature matrix is constructed, an improved CNN-LSTM model is used to output preliminary identification results of leakage, fine data such as seepage velocity is collected based on the preliminary results, an improved Transform-BiLSTM model is used to input, and leakage probability, grade and type are obtained; a leakage grading algorithm based on multi-factor weighted fusion is used to calculate risk indexes, and corresponding grade early warning and prevention and control are triggered.The application realizes whole-process, high-precision and intelligent early warning and regulation of underground structure leakage hidden danger, and is suitable for accurate identification, risk quantification and grading prevention and control of underground structure leakage hidden danger.

Description

Technical Field

[0001] This invention relates to the field of leakage monitoring technology, specifically to a distributed optical fiber early warning and control method for potential leakage hazards in underground structures. Background Technology

[0002] Underground structures (including tunnels, underground utility tunnels, culverts, subway stations, etc.) are the core carriers of urban infrastructure and transportation lifelines. Their long-term safe and stable operation is directly related to urban operational safety and the safety of life and property. During their long service life, these structures are subject to the combined effects of multiple factors, including geological changes, groundwater pressure erosion, construction process deviations, and changes in external environmental temperature and humidity. This makes them highly susceptible to micro-cracks and leakage hazards at the lining structure, structural joints, and seepage paths. Industry statistics show that a large number of existing underground structures across the country suffer from varying degrees of leakage. Severe leakage can directly lead to steel corrosion, concrete deterioration, and load-bearing capacity reduction, potentially causing major safety accidents such as localized spalling, water inrush, structural instability, and even collapse. This poses a serious threat to personnel passage, equipment operation, and the surrounding surface environment in underground spaces. Currently, with the continuous expansion of urban underground space development, the operation and maintenance of structures face severe challenges of "numerous points, long lines, wide areas, and complex environments." Therefore, researching and developing a high-precision and intelligent early warning and control technology that can realize the whole process of leakage hazards in underground structures has become a key issue and urgent need to ensure the safe operation and maintenance of underground engineering.

[0003] Chinese patent (publication number CN119688177A) discloses a tunnel leakage identification method based on fiber optic sensing. The method acquires temperature data collected by a fiber optic sensing network deployed along the tunnel wall; identifies temperature anomalies based on the temperature data; determines potential leakage points based on the location and temperature change characteristics of the anomalies; and generates a leakage identification result containing the location information of the potential leakage points. However, this method uses single-parameter data, does not construct an adaptive multi-parameter sensing network, cannot dynamically adjust the collected parameters according to the underground environment, is susceptible to environmental interference, and has limited monitoring accuracy. Furthermore, the analytical performance of a single parameter is limited, making it impossible to analyze leakage probability, level, and type.

[0004] Chinese patent (publication number CN121561822A) discloses a signal processing method for quantum-enhanced distributed optical fiber monitoring. This method injects quantum-enhanced probe light into the optical fiber and collects echo signals. The differential echo signal is obtained by subtracting signal values ​​from adjacent time points. Subsequently, the differential echo signal and the echo signal are normalized separately, and signal enhancement is achieved by combining a signal enhancement coefficient, resulting in a differential enhanced signal and an original enhanced signal. Then, the two types of enhanced signals are divided into multiple sub-signals and labeled as normal sub-signals. Four variant vectors are then obtained based on the differential main amplitude, the original main amplitude, the difference between the differential high-frequency energy and the original high-frequency energy. Finally, a penetration identification neural network is used to process the four variant vectors and output the seepage status result. However, this method only focuses on signal enhancement and is difficult to accurately extract deep spatiotemporal features related to leakage. Furthermore, this method only achieves leakage identification or location, lacking a complete hierarchical algorithm and closed-loop control mechanism. This results in wasted computing power and the inability to convert the identification results into executable control commands, making it difficult to form a complete monitoring-identification-hierarchy-control link.

[0005] In view of the shortcomings of the existing patents, there is an urgent need for a method for early warning and control of leakage in underground structures that can achieve adaptive multi-parameter monitoring, accurate feature extraction, progressive intelligent identification, and closed-loop hierarchical prevention and control. Summary of the Invention

[0006] Based on the aforementioned technical problems, this application discloses a distributed optical fiber early warning and control method for potential leakage hazards in underground structures, specifically including:

[0007] An adaptive multi-parameter distributed optical fiber sensor network is constructed to dynamically adjust the sensing parameters in real time according to the underground environment and collect temperature, vibration, and strain data.

[0008] Based on the collected data, a multi-dimensional spatiotemporal feature matrix is ​​constructed by combining feature processing algorithms such as wavelet transform, adaptive filtering, and empirical mode decomposition.

[0009] An improved CNN-LSTM hybrid deep learning model is constructed, with a multi-dimensional spatiotemporal feature matrix as input, and preliminary identification results of potential leakage hazards as output.

[0010] Based on the preliminary identification results, detailed data is collected at the potential leakage sites, input into the improved Transformer-BiLSTM model, and the leakage probability, leakage level and leakage type are output to obtain the target identification results.

[0011] Based on the target identification results, leakage risk indicators are calculated using a leakage classification algorithm, and early warning and prevention are carried out based on these leakage risk indicators.

[0012] Preferably, the adaptive multi-parameter distributed optical fiber sensing network specifically comprises: a three-mode integrated armored waterproof optical fiber with built-in FBG enhancement nodes is deployed along the inner wall of the underground structure lining, structural joints and seepage path, using a three-dimensional topology structure of main cable, branches and crosses.

[0013] The completed tri-mode integrated fiber optic connection with the adaptive sensing gain adjustment module of the edge computing gateway and cloud platform will collect underground environmental parameters through the environmental sensing sensors mounted on the edge computing gateway.

[0014] Based on real-time acquired underground environmental parameters, the sampling rate and bandwidth of three types of sensor signals—temperature, vibration, and strain—are adjusted through an adaptive acquisition parameter adjustment algorithm.

[0015] The tri-mode integrated fiber optic demodulation terminal collects temperature, vibration, and strain data along the underground structure. The collected raw data is transmitted to the edge computing gateway in real time, where it is initially cached locally and then synchronously uploaded to the cloud platform.

[0016] Preferably, the adaptive acquisition parameter adjustment algorithm specifically involves: acquiring the temperature, humidity, and surrounding rock stress of the underground environment, and calculating the adaptive sensing gain adjustment coefficient based on preset baseline values ​​for temperature, humidity, and surrounding rock stress. The formula is:

[0017]

[0018] in, These are preset weighting coefficients based on the degree of influence of the underground environment on the sensor signals. , , To measure the ambient temperature, humidity, and surrounding rock stress, , , The reference values ​​are temperature, humidity, and surrounding rock stress.

[0019] exist A value greater than 1 increases the sampling rate, accuracy, and bandwidth, enhancing signal capture capabilities; When the value is less than 1, appropriately reduce the parameters to optimize system energy consumption;

[0020] Environmental data collection and adaptive collection parameter adjustments are performed periodically.

[0021] Preferably, the construction of the multi-dimensional spatiotemporal feature matrix specifically involves: synchronously inputting the collected temperature, vibration, and strain data into the feature processing algorithm module, performing preliminary normalization on the input raw data, and using a wavelet transform algorithm to perform preliminary denoising on the normalized raw data;

[0022] The wavelet transform-processed signal is input into an adaptive filtering algorithm. This algorithm, based on the minimum mean square error (LMS) criterion, adaptively adjusts the filtering coefficients in real time to selectively remove remaining low-frequency interference signals, resulting in a clean signal after double denoising. The formula is as follows:

[0023]

[0024]

[0025] in, For adaptive filtering output signal, Let the filter order be . For the first The first moment Each filter coefficient

[0026] For the first The wavelet-denoised signal at time 10:00 These are the adaptively adjusted filter coefficients. This is the step size factor that controls the update speed of the coefficients;

[0027] The purified signal after double denoising is decomposed using the EMD algorithm, which decomposes the time-series signal into several intrinsic mode functions (IMFs) and one residual component. The IMF components related to leakage characteristics are then selected using the following formula:

[0028]

[0029] in, For the first One IMF component, For the first Signals after secondary screening for The mean of the upper and lower envelopes; the intrinsic mode function (IMF) screening must simultaneously meet the following two core conditions: the difference between the number of extreme points and the number of zero crossings of the signal does not exceed 1, and the mean of the upper and lower envelopes of the signal at any time is 0;

[0030] Extract the time-domain peak value, mean, variance, mutation rate, frequency-domain dominant frequency, and spectral peak value of each component to form a single-parameter feature set;

[0031] Based on single-parameter feature sets of three types of parameters—temperature, vibration, and strain—the feature change trends at different time points are extracted in the time domain; and the feature distribution differences at different fiber optic deployment locations are extracted in the spatial domain. These features are then organized according to a time dimension × spatial dimension × parameter dimension structure to construct a spatiotemporal feature matrix.

[0032] Preferably, the construction of the improved CNN-LSTM hybrid deep learning model specifically involves: constructing a model training sample library, which includes two types of spatiotemporal feature matrix samples: normal without leakage and leakage of different degrees; the sample library is divided into a training set, a validation set, and a test set for model training, parameter optimization, and performance verification.

[0033] A four-layer architecture is adopted, consisting of a CNN feature extraction layer, an LSTM temporal analysis layer, an attention layer, and an output layer. The training set is input, and the model predicts the output through forward propagation. This prediction is then compared with the sample labels to calculate the loss function suitable for both binary classification and location regression. The formula is as follows:

[0034]

[0035]

[0036]

[0037] in, Total loss function, For classifying losses, For real labels, To predict the probability of leakage, To balance classification loss and location loss, For position loss, Mean absolute error;

[0038] The Adam optimization algorithm is used to update the parameters of each layer of the model through backpropagation. Hyperparameters are adjusted in real time using the validation set. An early stopping strategy is employed to avoid overfitting until the model converges. The formula is as follows:

[0039]

[0040] in, , For the first time, Model parameters at time 10:00 For learning rate, , For the estimation of first and second moments, To avoid tiny constants with a denominator of 0.

[0041] Preferably, the improved CNN-LSTM hybrid deep learning model specifically adopts a four-layer architecture: CNN feature extraction layer - LSTM temporal analysis layer - Attention layer - output layer. The spatiotemporal feature matrix is ​​input, spatial distribution features are extracted through the CNN layer, key leakage-related features are focused through the Attention layer, and temporal change features are mined through the LSTM layer. Finally, the preliminary identification results of leakage hazards are output through the output layer.

[0042] Preferably, the step of collecting detailed data at the potential leakage site specifically involves: receiving the preliminary identification results output by the improved CNN-LSTM model, filtering out samples with potential leakage, accurately extracting the preliminary location of the potential leakage, delineating a detailed collection area with the preliminary location of the potential leakage as the core, optimizing the sensing parameters of the detailed collection area through the adaptive sensing gain adjustment module of the edge computing gateway, and collecting high-frequency, high-precision seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement data to form a high-precision dataset.

[0043] Preferably, the improved Transformer-BiLSTM model specifically employs a three-layer architecture: a Transformer global feature extraction layer, a BiLSTM bidirectional temporal analysis layer, and a fully connected output layer. It takes a high-precision dataset as input, captures global correlation features of the fine-grained data through the Transformer layer, mines bidirectional temporal variation features through the BiLSTM layer, and outputs the target recognition result, including leakage probability, leakage level, and leakage type, via the fully connected output layer. The formula is as follows:

[0044]

[0045] in, For output layer output, This represents the probability of leakage. It is the sigmoid activation function. Leakage is categorized into three levels: blue alert, yellow alert, and red alert. Leakage can be categorized into three types: point seepage, line seepage, and surface seepage. , , These are the output layer weight matrices, , , These are the output layer bias terms.

[0046] Preferably, the improved Transformer-BiLSTM model is trained as follows: A high-resolution training sample library is constructed, containing three types of seepage samples (point seepage, line seepage, and surface seepage) and corresponding high-resolution feature sets. The core data of the samples includes high-resolution data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement. The samples are divided into training, validation, and test sets proportionally. The training set is input into the model, and the predicted output is calculated through forward propagation. A multi-task loss function is calculated based on the sample labels, using the following formula:

[0047]

[0048] in, For multi-task loss function, The mean squared error loss is the probability of leakage. The cross-entropy loss is the leakage level. For the cross-entropy loss of the leakage type, , , There are three types of loss weights;

[0049] The Adam optimization algorithm is used to update model parameters through backpropagation; hyperparameters are optimized through the validation set, and model performance is verified through the test set to ensure that the model output accuracy, false alarm rate, and inference latency meet the preset targets.

[0050] Preferably, the leakage classification algorithm specifically comprises: receiving three core results output by the fine identification model and performing standardized processing; calculating leakage risk indicators by combining the influence of probability, level, and type through multi-factor weighted fusion; comparing the calculated leakage risk indicators with preset classification thresholds to trigger the corresponding level's early warning mechanism.

[0051] Compared with the prior art, the technical solution of this application has the following technical effects:

[0052] This invention utilizes an adaptive multi-parameter distributed optical fiber sensor network to dynamically adjust the sampling rate, accuracy, and bandwidth of three types of signals—temperature, vibration, and strain—by ​​calculating the gain adjustment coefficient Ka in real time based on environmental parameters. This ensures that the system can automatically adapt to working conditions under complex underground environmental changes and can capture weak, high-frequency leakage characteristic signals in the early stages of potential hazards, greatly improving the comprehensiveness and anti-interference capability of data acquisition.

[0053] This invention employs a triple fusion algorithm combining wavelet transform, adaptive filtering, and EMD. First, it performs dual denoising to ensure signal purity. Then, it uses EMD decomposition to extract leakage-related intrinsic mode functions, constructing a multi-dimensional spatiotemporal feature matrix of time, space, and parameters. This matrix can accurately characterize the development trend and spatial distribution differences of potential leakage hazards.

[0054] This invention uses an improved CNN-LSTM to quickly locate leakage areas, enabling wide-area, real-time preliminary screening. Based on the preliminary results, it collects detailed data such as seepage velocity, Rayleigh scattering intensity, and surrounding rock displacement, inputs them into an improved Transformer-BiLSTM model, and simultaneously outputs leakage probability, level, and type. This achieves a leap from qualitative location to quantitative classification, providing refined and quantifiable core data for risk management.

[0055] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more easily understood, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0056] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0058] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0059] Figure 1 A flowchart of a distributed optical fiber early warning and control method for potential leakage hazards in underground structures;

[0060] Figure 2 A diagram illustrating the overall architecture of a distributed fiber optic early warning and control method for potential leakage hazards in underground structures.

[0061] Figure 3 This is an architecture diagram of the improved CNN-LSTM model;

[0062] Figure 4 Architecture diagram of the improved Transformer-BiLSTM model;

[0063] Figure 5 A data graph illustrating how an adaptive multi-parameter distributed fiber optic sensor network adapts to changes in environmental parameters.

[0064] Figure 6 For real-time, detailed statistical probability density plots;

[0065] Figure 7 A comparison chart of the overall performance data of each method over 72 hours. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0067] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0068] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0069] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0070] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0071] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0072] Example 1 mainly describes a distributed fiber optic early warning and control method for potential leakage hazards in underground structures, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0073] An adaptive multi-parameter distributed optical fiber sensor network is constructed to dynamically adjust the sensing parameters in real time according to the underground environment and collect temperature, vibration, and strain data.

[0074] Based on the collected data, a multi-dimensional spatiotemporal feature matrix is ​​constructed by combining feature processing algorithms such as wavelet transform, adaptive filtering, and empirical mode decomposition.

[0075] An improved CNN-LSTM hybrid deep learning model is constructed, with a multi-dimensional spatiotemporal feature matrix as input, and preliminary identification results of potential leakage hazards as output.

[0076] Based on the preliminary identification results, detailed data is collected at the potential leakage sites, input into the improved Transformer-BiLSTM model, and the leakage probability, leakage level and leakage type are output to obtain the target identification results.

[0077] Based on the target identification results, leakage risk indicators are calculated using a leakage classification algorithm, and early warning and prevention are carried out based on these leakage risk indicators.

[0078] Furthermore, the adaptive multi-parameter distributed optical fiber sensing network specifically involves: along the inner wall of the underground structure lining, structural joints, and seepage path, a three-dimensional topology of main cable, branches, and crossovers is used to deploy tri-mode integrated armored waterproof optical fibers with built-in FBG enhancement nodes.

[0079] The completed tri-mode integrated fiber optic connection with the adaptive sensing gain adjustment module of the edge computing gateway and cloud platform will collect underground environmental parameters through the environmental sensing sensors mounted on the edge computing gateway.

[0080] Based on real-time acquired underground environmental parameters, the sampling rate and bandwidth of three types of sensor signals—temperature, vibration, and strain—are adjusted through an adaptive acquisition parameter adjustment algorithm.

[0081] The tri-mode integrated fiber optic demodulation terminal collects temperature, vibration, and strain data along the underground structure. The collected raw data is transmitted to the edge computing gateway in real time, where it is initially cached locally and then synchronously uploaded to the cloud platform.

[0082] Furthermore, the deployment of the adaptive multi-parameter distributed optical fiber sensor network is differentiated according to the critical leakage risk section (0.3m spacing) → edge leakage transition section (0.6m spacing) → ordinary monitoring section (1m spacing), forming a sensor network with full coverage and key encryption.

[0083] Furthermore, the adaptive acquisition parameter adjustment algorithm specifically involves: acquiring the temperature, humidity, and surrounding rock stress of the underground environment, and combining this with preset baseline values ​​for temperature, humidity, and surrounding rock stress to calculate the adaptive sensing gain adjustment coefficient. The formula is:

[0084]

[0085] in, These are preset weighting coefficients based on the degree of influence of the underground environment on the sensor signals. , , To measure the ambient temperature, humidity, and surrounding rock stress, , , The reference values ​​are temperature, humidity, and surrounding rock stress.

[0086] exist A value greater than 1 increases the sampling rate, accuracy, and bandwidth, enhancing signal capture capabilities; When the value is less than 1, appropriately reduce the parameters to optimize system energy consumption;

[0087] Environmental data collection and adaptive collection parameter adjustments are performed periodically.

[0088] Furthermore, an optional adaptive acquisition parameter adjustment algorithm is set as follows: The acquired signal parameters are adjusted as follows: , ,in, As the baseline sampling rate, To adjust the sampling rate after adjustment, The bandwidth of the reference sampling signal, To adjust the bandwidth of the post-sampled signal.

[0089] Furthermore, a multi-dimensional spatiotemporal feature matrix is ​​constructed. Specifically, the collected temperature, vibration, and strain data are synchronously input into the feature processing algorithm module to perform preliminary normalization on the input raw data, and wavelet transform algorithm is used to perform preliminary denoising on the normalized raw data.

[0090] The wavelet transform-processed signal is input into an adaptive filtering algorithm. This algorithm, based on the minimum mean square error (LMS) criterion, adaptively adjusts the filtering coefficients in real time to selectively remove remaining low-frequency interference signals, resulting in a clean signal after double denoising. The formula is as follows:

[0091]

[0092]

[0093] in, For adaptive filtering output signal, Let the filter order be . For the first The first moment Each filter coefficient

[0094] For the first The wavelet-denoised signal at time 10:00 These are the adaptively adjusted filter coefficients. This is the step size factor that controls the update speed of the coefficients;

[0095] The purified signal after double denoising is decomposed using the EMD algorithm, which decomposes the time-series signal into several intrinsic mode functions (IMFs) and one residual component. The IMF components related to leakage characteristics are then selected using the following formula:

[0096]

[0097] in, For the first One IMF component, For the first Signals after secondary screening for The mean of the upper and lower envelopes; the intrinsic mode function (IMF) screening must simultaneously meet the following two core conditions: the difference between the number of extreme points and the number of zero crossings of the signal does not exceed 1, and the mean of the upper and lower envelopes of the signal at any time is 0;

[0098] Extract the time-domain peak value, mean, variance, mutation rate, frequency-domain dominant frequency, and spectral peak value of each component to form a single-parameter feature set;

[0099] Based on single-parameter feature sets of three types of parameters—temperature, vibration, and strain—the feature change trends at different time points are extracted in the time domain; and the feature distribution differences at different fiber optic deployment locations are extracted in the spatial domain. These features are then organized according to a time dimension × spatial dimension × parameter dimension structure to construct a spatiotemporal feature matrix.

[0100] Furthermore, the input raw data undergoes preliminary standardization, specifically including: data alignment (synchronizing the three types of parameter data according to timestamps), missing value completion (filling in a small number of missing data using linear interpolation), and data normalization (mapping the data to the [0,1] interval) to avoid data deviation affecting the subsequent feature extraction effect and lay the foundation for algorithm fusion processing; the rows of the spatiotemporal feature matrix correspond to the time series (time nodes of continuous acquisition), the columns correspond to the spatial location (fiber optic deployment points), and the elements are the feature values ​​of various parameters, outputting a standardized multi-dimensional spatiotemporal feature matrix, which is directly input into the subsequent leakage analysis model.

[0101] Furthermore, an improved CNN-LSTM hybrid deep learning model is constructed. Specifically, a model training sample library is built, which includes two types of spatiotemporal feature matrix samples: normal without leakage and leakage of different degrees. The sample library is divided into a training set, a validation set, and a test set for model training, parameter optimization, and performance verification.

[0102] A four-layer architecture is adopted, consisting of a CNN feature extraction layer, an LSTM temporal analysis layer, an attention layer, and an output layer. The training set is input, and the model predicts the output through forward propagation. This prediction is then compared with the sample labels to calculate the loss function suitable for both binary classification and location regression. The formula is as follows:

[0103]

[0104]

[0105]

[0106] in, Total loss function, For classifying losses, For real labels, To predict the probability of leakage, To balance classification loss and location loss, For position loss, Mean absolute error;

[0107] The Adam optimization algorithm is used to update the parameters of each layer of the model through backpropagation. Hyperparameters are adjusted in real time using the validation set. An early stopping strategy is employed to avoid overfitting until the model converges. The formula is as follows:

[0108]

[0109] in, , For the first time, Model parameters at time 10:00 For learning rate, , For the estimation of first and second moments, To avoid tiny constants with a denominator of 0.

[0110] Furthermore, the improved CNN-LSTM hybrid deep learning model specifically involves: inputting a spatiotemporal feature matrix, and extracting spatial distribution features through CNN layers, as shown in the formula:

[0111]

[0112] in, This represents the T×S×K dimensional spatial feature map output by the CNN layer, where T and S represent the temporal and spatial dimensions, respectively. The number of convolution kernels, For the first A two-dimensional convolutional kernel, For the first The bias term of each convolution kernel, It is the ReLU activation function;

[0113] The Attention layer focuses on key features related to leakage, and the formula is:

[0114]

[0115] in, The weighted spatial features are the output of the Attention layer. Here is the weight matrix of the Attention layer. For feature dimensions , This is the normalization function;

[0116] The formula for mining temporal variation features using LSTM layers is:

[0117]

[0118]

[0119] in, , For the first time, LSTM cell state at time -1 , , For forget gate, input gate, and output gate. For the first The features output by the Attention layer at each time step For the first Temporal characteristics of the LSTM layer output at any given time;

[0120] The preliminary identification results of potential leakage are output through the output layer, using the following formula:

[0121]

[0122] in, Preliminary identification results of potential leakage risks. This is the output layer weight matrix. The LSTM output features at the last time point For output layer bias terms, For vector concatenation, The output layer attention weight matrix. The Attention feature at the last time point is used to map the initial location of leakage.

[0123] Furthermore, such as Figure 3 The diagram shows the architecture of the improved CNN-LSTM model (@Height×Width in the diagram is a demonstration value; the actual Height×Width used in the actual working condition is the corresponding value). The CNN feature extraction layer of the improved CNN-LSTM model adopts a design of two convolutional layers and pooling layers in series. The first convolutional layer is responsible for the initial extraction of local spatial features, and the second convolutional layer deepens feature mining. Each convolutional layer is followed by a max pooling layer to compress the feature dimension. The core parameters are set as follows: the first convolutional layer has 32 kernels, the second layer has 64 kernels, the kernel size is 3×3, the stride is 1×1, and Same padding is used to ensure that the feature map size after convolution is consistent with the input, avoiding the loss of edge features. The activation function is ReLU, which introduces non-linear transformation to accelerate model convergence. The pooling layer uses a 2×2 max pooling kernel with a stride of 2×2, which can compress the feature space size by 50%, effectively reducing the risk of overfitting.

[0124] The attention layer structure is built on a multi-head self-attention mechanism. By assigning different weights to the spatial features output by the CNN, the importance of key features related to leakage is highlighted. The core parameters are set as follows: the number of attention heads is 8, which can capture the correlation between features in multiple dimensions; the feature dimension is set to 64, which is consistent with the feature dimension of the second layer output of the CNN to ensure data dimension matching; a scaling factor of 8 is introduced to scale the calculation results of attention weights to avoid saturation of the softmax function; the activation function is selected as Softmax to normalize the attention weights so that the sum of the weights of each feature is 1, accurately focusing on key features.

[0125] The LSTM temporal analysis layer adopts a single-layer LSTM structure, using a forget gate, input gate, and output gate to collaboratively control information transmission, retaining key temporal features and forgetting useless historical information. The core parameters are set as follows: the number of hidden units is 128, which can fully capture the long-term dependencies of leaked features in the time series; the input dimension is 64, matching the output dimension of the Attention layer; the activation function is Tanh, which normalizes the LSTM cell state in the range of [-1,1] to accurately characterize the changes in temporal features; a Dropout mechanism is introduced after the LSTM layer with a probability of 0.2 to randomly block 20% of neurons to reduce the risk of overfitting.

[0126] The multi-task output layer structure adopts a dual-branch parallel design: the classification branch is responsible for outputting the binary classification result of whether leakage exists, and the location regression branch is responsible for outputting the two-dimensional coordinates of the initial location of leakage; the core parameters are set as follows: the classification branch adopts a design of 64→32→1 fully connected layer neurons, the activation function is Sigmoid, and the output value is mapped to the interval [0,1], where 0 represents no leakage and 1 represents leakage; the location regression branch adopts a design of 128→64→2 fully connected layer neurons, the activation function is a linear function, and the output two-dimensional value corresponds to the x and y coordinates of the spatial point of the potential leakage; the loss function adopts a weighted fusion of classification loss (cross-entropy) and location loss (MAE), with the weight coefficient λ set to 0.5 to balance the loss ratio of the two tasks.

[0127] Furthermore, detailed data is collected at potential leakage sites. Specifically, the preliminary identification results output by the improved CNN-LSTM model are received, samples with potential leakage are selected, the preliminary location of the potential leakage is accurately extracted, and a detailed collection area is delineated with the preliminary location of the potential leakage as the core. The sensor parameters of the detailed collection area are optimized through the adaptive sensor gain adjustment module of the edge computing gateway, and high-frequency, high-precision data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement are collected.

[0128] Furthermore, the improved Transformer-BiLSTM model employs a three-layer architecture: a Transformer global feature extraction layer, a BiLSTM bidirectional temporal analysis layer, and a fully connected output layer. It takes a high-precision dataset as input, captures the global correlation features of the fine-grained data through the Transformer layer, and mines the bidirectional temporal variation features through the BiLSTM layer. The formula is as follows:

[0129]

[0130]

[0131]

[0132] The target identification results, including leakage probability, leakage level, and leakage type, are output through the fully connected output layer, as shown in the formula:

[0133]

[0134] in, For output layer output, This represents the probability of leakage. It is the sigmoid activation function. Leakage is categorized into three levels: blue alert, yellow alert, and red alert. Leakage can be categorized into three types: point seepage, line seepage, and surface seepage. , , These are the output layer weight matrices, , , These are the output layer bias terms.

[0135] Furthermore, the improved Transformer-BiLSTM model is trained as follows: a refined training sample library is constructed, containing three types of seepage samples—point seepage, line seepage, and surface seepage—and corresponding refined data feature sets. The core of the samples consists of refined data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement. These samples are divided into training, validation, and test sets in a 7:2:1 ratio. The training set is input into the model, and the predicted output is calculated through forward propagation. A multi-task loss function is calculated based on the sample labels, using the following formula:

[0136]

[0137] in, For multi-task loss function, The mean squared error loss is the probability of leakage. The cross-entropy loss is the leakage level. For the cross-entropy loss of the leakage type, , , There are three types of loss weights;

[0138] The Adam optimization algorithm is used to update model parameters through backpropagation; hyperparameters are optimized through the validation set, and model performance is verified through the test set to ensure that the model output accuracy is ≥98%, false positive rate is <0.5%, and inference latency is <1s.

[0139] Furthermore, such as Figure 4 The diagram shows the architecture of the improved Transformer-BiLSTM model. The input layer time series length of the improved Transformer-BiLSTM model is T=432 (72 hours × 6 times / hour), and the feature dimension is 3 (seepage velocity, Rayleigh scattering intensity, and surrounding rock displacement).

[0140] The embedding layer has an embedding dimension of 128 and uses a fully connected layer for embedding with the ReLU activation function. The Transformer encoder layer consists of two stacked Transformer encoder layers, each containing a multi-head self-attention sub-layer and a feedforward network sub-layer. It has 8 attention heads and a feature dimension of 128. The fully connected layer of the feedforward network has a dimension of 128→512→128 and uses the GELU activation function. Residual connections and LayerNorm are added after each sub-layer, and Dropout=0.1.

[0141] The network structure of the BiLSTM bidirectional temporal layer is a single-layer BiLSTM (forward LSTM, backward LSTM), with 128 hidden units in each direction. The cell state activation is Tanh, the gate activation is Sigmoid, and Dropout=0.2 is added to the output layer. By concatenating the outputs of the forward and backward LSTMs at the last moment, a temporal feature with a dimension of 256 is obtained.

[0142] The feature fusion layer uses splicing and fully connected dimensionality reduction to reduce the 256 (BiLSTM) + 128 (Transformer) = 384-dimensional features to 128-dimensional features through the fully connected layer. The fully connected layer has 128 neurons, the activation function is ReLU, and LayerNorm normalization is added.

[0143] Furthermore, the leakage classification algorithm specifically involves: receiving three core results from the refined identification model and standardizing them; calculating leakage risk indicators by combining the influence of probability, level, and type through multi-factor weighted fusion; comparing the calculated leakage risk indicators with preset classification thresholds to trigger the corresponding level's early warning mechanism.

[0144] Furthermore, the standardization process specifically involves:

[0145] The leakage probability is mapped to the [0,1] interval, and the closer the value is to 1, the higher the risk of leakage.

[0146] The blue alert (1), yellow alert (2), and red alert (3) of the leakage level are converted into numerical labels according to preset weights;

[0147] The leakage types—point seepage (1), line seepage (2), and surface seepage (3)—are converted into numerical labels based on their influence coefficients.

[0148] Furthermore, the formula for calculating the leakage risk index is as follows:

[0149]

[0150] in, This is a leakage risk indicator, with a value range of [0,1]. A larger value indicates a higher overall risk. , These are preset weighting coefficients, corresponding to the importance of probability, level, and type, respectively.

[0151] Furthermore, the calculated leakage risk index R is compared with a preset classification threshold, triggering the corresponding level of early warning mechanism as follows:

[0152] Low risk (blue alert): R∈[0,0.33), triggering a blue alert and initiating continuous monitoring;

[0153] Medium risk (yellow alert): R∈[0.33,0.66), triggering a yellow alert and initiating restrictive control measures;

[0154] High risk (red alert): R∈[0.66,1], triggering a red alert and initiating emergency prevention and control.

[0155] This embodiment details a distributed fiber optic early warning and control method for potential leakage hazards in underground structures. It constructs an adaptive multi-parameter distributed fiber optic sensor network with FBG enhancement nodes, dynamically adjusts sensor parameters based on environmental parameters, and collects temperature, vibration, and strain data. The data is processed using wavelet transform, adaptive filtering, and EMD fusion algorithms to construct a multi-dimensional spatiotemporal feature matrix. An improved CNN-LSTM model outputs preliminary leakage identification results. Based on these preliminary results, detailed data such as seepage velocity are collected and input into an improved Transformer-BiLSTM model to obtain leakage probability, level, and type. A multi-factor weighted fusion leakage grading algorithm calculates risk indicators, triggering corresponding level early warnings and control measures.

[0156] Example 2, based on Example 1, details the process of using this method for early warning analysis of leakage hazards in subway tunnels. During the process, existing methods, Tunnel Fiber Optic Leakage Identification Method (TFLI) and Quantum Enhanced Fiber Optic Monitoring Method (QEFMM), are used for comparative testing. The specific process is as follows:

[0157] The tunnel is 300m long and the lining is 28cm thick. Fiber optic cables are laid along the inner wall of the lining at 1m / point, for a total of 300 monitoring points. Typical leakage hazards exist in the section from K0+100 to K130, including three types of leakage: point seepage, line seepage, and surface seepage.

[0158] During implementation, the TFLI tunnel fiber optic leakage identification method, compared with the previous scheme, deployed a single-temperature-parameter fiber optic sensor network with fixed parameters to collect temperature data within a 300m range of the tunnel; the temperature data was processed using a wavelet denoising algorithm to extract simple temperature features; the basic identification model was input, and the output showed whether there was leakage and its location.

[0159] The comparative scheme, Quantum Enhanced Fiber Optic Monitoring (QEFMM), deploys a single fiber optic scattering signal sensing network to collect fiber optic scattering signals without multi-parameter acquisition. It uses a quantum enhancement algorithm to amplify the scattering signal intensity without multi-algorithm fusion feature processing. It determines whether leakage has occurred based on signal intensity and provides preliminary location. It also issues early warnings based on signal intensity levels, triggering simple warning prompts.

[0160] The implementation process of this method is as follows: Tri-mode integrated armored waterproof optical fibers with built-in FBG enhancement nodes are laid along the inner wall and structural joints of the tunnel lining, and connected to an edge computing gateway with an adaptive sensing gain adjustment module; real-time environmental parameters inside the tunnel are collected: temperature 27℃ (reference value Tcal=25℃), humidity 83% (reference value 70%), and surrounding rock stress 11.9MPa (reference value 10.8MPa); the adaptive gain adjustment coefficient is calculated to obtain...

[0161] Adjusting the acquisition parameters: Since ka < 1, energy consumption was appropriately reduced. The temperature sampling rate was set to 1.1 Hz (originally 1.2 Hz), the vibration bandwidth was set to 0.9 Hz (originally 1 Hz), and the strain accuracy was maintained at ±0.01 με. Temperature, vibration, and strain data were collected for 72 hours, and the ka coefficient was recalculated every hour. Some real-time environmental parameters are shown in Table 1.

[0162] Table 1. Some Real-Time Environmental Parameters

[0163] Monitoring time Temperature (°C) humidity(%) Surrounding rock stress (MPa) ka coefficient Hour 0 27.1 83.4 11.9 12th hour 26.8 81 10.8 0.525 36th hour 27.2 84 11.2 0.528 60th hour 27.4 85 11.4 0.561 72nd hour 27.1 82 10.9 0.508

[0164] According to Table 1 and Figure 5 As shown in the data graph of the adaptive multi-parameter distributed optical fiber sensor network of this method, which adaptively adjusts based on changes in environmental parameters, during the continuous 72-hour monitoring process, the environmental parameters inside the tunnel exhibit certain periodic and correlated changes over time. The adaptive adjustment coefficient can make a synchronous, smooth, and reasonable response based on the real-time changes in temperature, humidity, and surrounding rock stress, demonstrating that the sensor network of this invention has the ability to adaptively adjust to the environment.

[0165] The collected data were processed using wavelet transform, adaptive filtering, and EMD fusion algorithms to construct a spatiotemporal feature matrix (50 time dimension, 150 spatial dimension, and 9 parameter dimension). The matrix was then input into an improved CNN-LSTM model, and the following preliminary leakage identification results were output: Leakage hazard locations: K0+110 (point seepage), K0+118 (line seepage), K0+127 (area seepage); Preliminary identification accuracy: 98.6%.

[0166] Fine data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement were collected at the aforementioned potential hazard locations. The average real-time fine data is shown in Table 2.

[0167] Table 2 Average Real-Time Fine Data

[0168] Leakage type Monitoring points Seepage velocity (m / s) Rayleigh scattering intensity (dB) Surrounding rock displacement (mm) sampling frequency Pitting K0+110 0.022 18.2 0.11 10 minutes / session Linear infiltration K0+118 0.039 21.9 0.23 10 minutes / session Surface seepage K0+127 0.065 26.5 0.45 10 minutes / session

[0169] The refined data was input into the improved Transformer-BiLSTM model, and the output results are shown in Table 3:

[0170] Table 3 Output results of the improved Transformer-BiLSTM

[0171] Leakage location Leakage probability Leakage rating Leakage type Recognition time (s) K0+110 0.82 Huang Jing Pitting 0.7 K0+118 0.90 Red Alert Linear infiltration 0.65 K0+127 0.97 Red Alert Surface seepage 0.72

[0172] According to Table 2, Table 3 and Figure 6 As shown in the real-time detailed data statistical probability density diagram, the improved Transformer-BiLSTM model of the present invention can accurately identify the location, type and level of leakage, with high identification probability and short time consumption. The warning level in Table 3 is consistent with the actual leakage level in Table 2, which verifies the effectiveness, accuracy and engineering applicability of the high-frequency detailed monitoring and intelligent identification method of the present invention.

[0173] The output of Transformer-BiLSTM is standardized, and the leakage probability is mapped to [0,1]. The level weights are WL (yellow alert = 2, red alert = 3) and the type coefficients are CTp (point seepage = 1, line seepage = 2, surface seepage = 3).

[0174] The risk indicators were calculated, and the results are shown in Table 4:

[0175] Table 4 Risk Calculation Results for Each Seepage Location

[0176] Leakage location R value Warning Level Prevention and control measures Leakage probability 24 hours after prevention and control K0+110 0.58 Huang Jing Local grouting 0.11 K0+118 0.81 Red Alert Emergency sealing and structural reinforcement 0.07 K0+127 0.94 Red Alert Comprehensive drainage and emergency sealing 0.04

[0177] As shown in Table 4, the R value can accurately reflect the leakage risk level. The warning level is consistent with the on-site leakage degree. After the implementation of targeted prevention and control measures, the leakage probability at each location has decreased significantly.

[0178] The test results of the three methods were collected simultaneously, and the comprehensive performance data of each method within 72 hours were obtained as shown in Table 5 below:

[0179] Table 5. Overall performance data of each method within 72 hours.

[0180] Comparison indicators TFLI QEFMM This invention AMFEW Leakage location identification accuracy 91.2% 97.5% 98.6% Leakage type identification accuracy 94.1% 89.8% 98.5% False negative rate 3.1% 5.2% 0.1% False alarm rate 6.2% 1.8% 0.3% Average recognition time 3.1s 2.7s 0.7s Risk quantification capability None (qualitative only) None (signal strength rating only) Quantify R-values ​​for precise grading Early warning response time 14min 9min 2min

[0181] According to Table 5 and Figure 7 The comparison chart of the comprehensive performance data of each method over 72 hours shows that...

[0182] This embodiment details the entire process of precise location, fine identification, quantitative classification, and closed-loop prevention and control of seepage hazards using this method in a 300m short tunnel scenario. Through adaptive multi-parameter sensor networks, multi-algorithm fusion feature processing, progressive deep learning identification, and quantitative classification algorithms, this method achieves significant advantages in identification accuracy, false negative rate, and prevention and elimination rate, fully verifying the advanced nature and engineering practicality of the technical solution.

[0183] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A distributed optical fiber early warning and control method for potential leakage hazards in underground structures, characterized in that, include: An adaptive multi-parameter distributed optical fiber sensor network is constructed to dynamically adjust the sensing parameters in real time according to the underground environment and collect temperature, vibration, and strain data. Based on the collected data, a multi-dimensional spatiotemporal feature matrix is ​​constructed by combining feature processing algorithms such as wavelet transform, adaptive filtering, and empirical mode decomposition. An improved CNN-LSTM hybrid deep learning model is constructed, with a multi-dimensional spatiotemporal feature matrix as input, and preliminary identification results of potential leakage hazards as output. Based on the preliminary identification results, detailed data is collected at the potential leakage sites, input into the improved Transformer-BiLSTM model, and the leakage probability, leakage level and leakage type are output to obtain the target identification results. Based on the target identification results, leakage risk indicators are calculated using a leakage classification algorithm, and early warning and prevention are carried out based on these leakage risk indicators.

2. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The adaptive multi-parameter distributed optical fiber sensing network is specifically: along the inner wall of the underground structure lining, structural joints and seepage path, a three-dimensional topology structure of main cable, branch and cross is adopted, and a three-mode integrated armored waterproof optical fiber with built-in FBG enhancement nodes is deployed. The completed tri-mode integrated fiber optic connection with the adaptive sensing gain adjustment module of the edge computing gateway and cloud platform will collect underground environmental parameters through the environmental sensing sensors mounted on the edge computing gateway. Based on real-time acquired underground environmental parameters, the sampling rate and bandwidth of three types of sensor signals—temperature, vibration, and strain—are adjusted through an adaptive acquisition parameter adjustment algorithm. The tri-mode integrated fiber optic demodulation terminal collects temperature, vibration, and strain data along the underground structure. The collected raw data is transmitted to the edge computing gateway in real time, where it is initially cached locally and then synchronously uploaded to the cloud platform.

3. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 2, characterized in that, The adaptive acquisition parameter adjustment algorithm specifically involves: acquiring the temperature, humidity, and surrounding rock stress of the underground environment; and calculating the adaptive sensing gain adjustment coefficient based on preset baseline values ​​for temperature, humidity, and surrounding rock stress. The formula is: in, These are preset weighting coefficients based on the degree of influence of the underground environment on the sensor signals. , , To measure the ambient temperature, humidity, and surrounding rock stress, , , The reference values ​​are temperature, humidity, and surrounding rock stress. exist A value greater than 1 increases the sampling rate, accuracy, and bandwidth, enhancing signal capture capabilities; When the value is less than 1, appropriately reduce the parameters to optimize system energy consumption; Environmental data collection and adaptive collection parameter adjustments are performed periodically.

4. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The construction of the multi-dimensional spatiotemporal feature matrix specifically involves: synchronously inputting the collected temperature, vibration, and strain data into the feature processing algorithm module, performing preliminary normalization on the input raw data, and using wavelet transform algorithm to perform preliminary denoising on the normalized raw data; The wavelet transform-processed signal is input into an adaptive filtering algorithm. This algorithm, based on the minimum mean square error (LMS) criterion, adaptively adjusts the filtering coefficients in real time to selectively remove remaining low-frequency interference signals, resulting in a clean signal after double denoising. The formula is as follows: in, For adaptive filtering output signal, Let the filter order be . For the first The first moment Each filter coefficient For the first The wavelet-denoised signal at time 10:00 These are the adaptively adjusted filter coefficients. This is the step size factor that controls the update speed of the coefficients; The purified signal after double denoising is decomposed using the EMD algorithm, which decomposes the time-series signal into several intrinsic mode functions (IMFs) and one residual component. The IMF components related to leakage characteristics are then selected using the following formula: in, For the first One IMF component, For the first Signals after secondary screening for The mean of the upper and lower envelopes; the intrinsic mode function (IMF) screening must simultaneously meet the following two core conditions: the difference between the number of extreme points and the number of zero crossings of the signal does not exceed 1, and the mean of the upper and lower envelopes of the signal at any time is 0; Extract the time-domain peak value, mean, variance, mutation rate, frequency-domain dominant frequency, and spectral peak value of each component to form a single-parameter feature set; Based on single-parameter feature sets of three types of parameters—temperature, vibration, and strain—the feature change trends at different time points are extracted in the time domain; and the feature distribution differences at different fiber optic deployment locations are extracted in the spatial domain. These features are then organized according to a time dimension × spatial dimension × parameter dimension structure to construct a spatiotemporal feature matrix.

5. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The construction of the improved CNN-LSTM hybrid deep learning model specifically involves: constructing a model training sample library, which includes two types of spatiotemporal feature matrix samples: normal without leakage and leakage of different degrees; the sample library is divided into a training set, a validation set, and a test set for model training, parameter optimization, and performance verification. A four-layer architecture is adopted, consisting of a CNN feature extraction layer, an LSTM temporal analysis layer, an attention layer, and an output layer. The training set is input, and the model predicts the output through forward propagation. This prediction is then compared with the sample labels to calculate the loss function suitable for both binary classification and location regression. The formula is as follows: in, Total loss function, For classifying losses, For real labels, To predict the probability of leakage, To balance classification loss and location loss, For position loss, Mean absolute error; The Adam optimization algorithm is used to update the parameters of each layer of the model through backpropagation. Hyperparameters are adjusted in real time using the validation set. An early stopping strategy is employed to avoid overfitting until the model converges. The formula is as follows: in, , For the first time, Model parameters at time 10:00 For learning rate, , For the estimation of first and second moments, To avoid tiny constants with a denominator of 0.

6. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 5, characterized in that, The improved CNN-LSTM hybrid deep learning model specifically adopts a four-layer architecture: CNN feature extraction layer, LSTM temporal analysis layer, Attention layer, and output layer. The spatiotemporal feature matrix is ​​input, the spatial distribution features are extracted by the CNN layer, the Attention layer focuses on key features related to leakage, the LSTM layer mines temporal change features, and the output layer outputs the preliminary identification results of leakage risks.

7. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The process of collecting detailed data at potential leakage sites involves: receiving the preliminary identification results from the improved CNN-LSTM model, filtering out samples with potential leakage sites, accurately extracting the preliminary location of the potential leakage sites, defining a detailed collection area based on the preliminary location of the potential leakage sites, optimizing the sensor parameters of the detailed collection area through the adaptive sensor gain adjustment module of the edge computing gateway, and collecting high-frequency, high-precision data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement to form a high-precision dataset.

8. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The improved Transformer-BiLSTM model specifically employs a three-layer architecture: a Transformer global feature extraction layer, a BiLSTM bidirectional temporal analysis layer, and a fully connected output layer. It takes a high-precision dataset as input, captures global correlation features of the fine-grained data through the Transformer layer, mines bidirectional temporal variation features through the BiLSTM layer, and outputs the target recognition results, including leakage probability, leakage level, and leakage type, as shown in the formula: in, For output layer output, This represents the probability of leakage. It is the sigmoid activation function. Leakage is categorized into three levels: blue alert, yellow alert, and red alert. Leakage can be categorized into three types: point seepage, line seepage, and surface seepage. , , These are the output layer weight matrices, , , These are the output layer bias terms.

9. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 8, characterized in that, The improved Transformer-BiLSTM model is trained as follows: A high-resolution training sample library is constructed, containing three types of seepage samples (point seepage, line seepage, and surface seepage) and their corresponding high-resolution feature sets. The core data of the samples includes high-resolution data on seepage velocity, fiber Rayleigh scattering intensity, and surrounding rock displacement. The samples are divided into training, validation, and test sets proportionally. The training set is input into the model, and the predicted output is calculated through forward propagation. A multi-task loss function is calculated using the sample labels, as shown in the formula: in, For multi-task loss function, The mean squared error loss is the probability of leakage. The cross-entropy loss is the leakage level. For the cross-entropy loss of the leakage type, , , There are three types of loss weights; The Adam optimization algorithm is used to update model parameters through backpropagation; hyperparameters are optimized through the validation set, and model performance is verified through the test set to ensure that the model output accuracy, false alarm rate, and inference latency meet the preset targets.

10. The distributed optical fiber early warning and control method for potential leakage hazards in underground structures according to claim 1, characterized in that, The leakage classification algorithm specifically involves: receiving three core results from the refined identification model and standardizing them; calculating leakage risk indicators by combining the influence of probability, level, and type through multi-factor weighted fusion; comparing the calculated leakage risk indicators with preset classification thresholds to trigger the corresponding level's early warning mechanism.

Images