A warning method based on dynamic separation of multiple physical quantities of optical fiber

Abstract

The application discloses a warning method based on optical fiber multi-physical quantity dynamic separation, relates to the field of optical fiber communication and sensing technology, and realizes dynamic separation and high-precision sensing of multi-physical quantities on the same optical fiber medium through the organic combination of a sensing-integrated transmission signal design and a deep prior decoupling network, and effectively solves the cross-sensitivity problem; through the cooperative intelligent architecture of "edge fast sensing + center deep fusion", the real-time performance of event identification and warning is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of optical fiber communication and sensing technology, and in particular to an early warning method based on the dynamic separation of multiple physical quantities in optical fiber. Background Technology

[0002] Currently, the field of wide-area environmental monitoring faces significant technical bottlenecks in addressing the monitoring needs of large-scale infrastructure such as bridges, tunnels, oil and gas pipelines, power corridors, and urban communities. First, traditional monitoring systems employ a separate architecture of "independent sensing + independent communication," leading to redundant hardware deployment, low utilization of fiber optic resources, high maintenance costs, and difficulty in achieving true continuous coverage across all times and spaces. Second, existing distributed fiber optic sensing technologies are functionally limited, mostly monitoring only single physical quantities such as temperature or vibration, lacking the ability to collaboratively sense and cross-analyze multiple physical quantities such as temperature, strain, and vibration. This makes them inadequate for diagnosing faults involving multiple coupled factors in complex environments. Furthermore, traditional methods for processing multi-physical quantity sensing data largely rely on traditional signal decoupling algorithms, which struggle to effectively address the cross-sensitivity issues between temperature, strain, and vibration, and generally lack quantification mechanisms for sensing uncertainties, resulting in high false alarm and false negative rates. Although there have been attempts in recent years to combine communication and sensing functions through wavelength division multiplexing or frequency division multiplexing technologies, these methods only share optical fiber media at the physical layer and have not achieved integrated design and dynamic resource allocation at the signal layer. The system efficiency and flexibility are still insufficient. Summary of the Invention

[0003] The purpose of this application is to provide an early warning method based on the dynamic separation of multiple physical quantities in optical fiber, so as to realize the dynamic separation and high-precision sensing of multiple physical quantities on the same optical fiber medium and solve the problem of cross-sensitivity.

[0004] To achieve the above objectives, this application provides the following solution.

[0005] In a first aspect, this application provides an early warning method based on dynamic separation of multiple physical quantities in optical fiber. The early warning method is applied to a distributed optical fiber system, which includes: a front-end processing module, optical fibers, multiple edge nodes, and a cloud; the early warning method includes: The front-end processing module generates an integrated transmission signal that meets the spectrum planning constraints based on the dynamic separation of multiple physical quantities, and then transmits it using optical fiber. At the edge nodes, a deep prior decoupling network and a lightweight intelligent model are used to make edge decisions and obtain edge decision results for various events. If the edge decision result is an immediate warning, a local warning is triggered immediately. If the edge decision result is to be reviewed, a feature vector of the event to be reviewed is generated and sent to the cloud. In the cloud, a weighted voting model is used to make cloud decisions based on the feature vectors of the events to be reviewed sent by each edge node, and the cloud decision results are obtained. If the cloud decision result is a confirmation warning, the warning level is determined and a cloud warning is issued based on the warning level.

[0006] Optionally, an integrated transmit signal satisfying spectrum planning constraints is generated based on the dynamic separation of multiple physical quantities, specifically including: The optical fiber spectrum resources are allocated by solving the objective function of compromising the inductive performance, and the allocation results are obtained. Based on communication signals and sensing signals, modulation parameters are generated by solving the fusion modulation objective function. Based on the allocation results and the modulation parameters, an integrated transmission signal that satisfies the spectrum planning constraints is generated.

[0007] Optionally, the allocation result includes: a communication subband, a vibration sensing subband, a temperature sensing subband, and a strain sensing subband; the optical fiber's spectrum resources are allocated by solving a trade-off optimization objective function for synergistic performance to obtain the allocation result, specifically including: The communication subband and sensing subband are obtained by solving the objective function that compromises the synergistic performance. The objective function for the trade-off optimization of synesthesia performance is: ; The constraints for optimizing the objective function by compromising synesthetic performance are: ; ; ; ; in, To optimize the objective function with a trade-off for synesthetic performance, and These are the communication weight coefficient and the perception weight coefficient, respectively. For communication sub-band, To sense sub-bands, This is a function for calculating the subband capacity of communication. , The function for calculating the sensitivity of the sensing subband is... The total bandwidth of the optical fiber. The signal-to-noise ratio of the communication signal. , The power of the communication signal. The noise power spectral density of the communication signal. To sense the signal-to-noise ratio of the signal. , To sense the transmission power of the signal, To sense the noise power spectral density of the signal, and These are the signal-to-noise ratio (SNR) thresholds for communication signals and sensing signals, respectively. It is an empty set; The sensing sub-bands are divided according to the following formula to obtain vibration sensing sub-bands, temperature sensing sub-bands, and strain sensing sub-bands; ; ; in, , , These are dedicated sub-bands for vibration sensing, temperature sensing, and strain sensing, obtained by dividing the sensing sub-bands.

[0008] Optionally, based on communication signals and sensing signals, modulation parameters are generated by solving a fusion modulation objective function, specifically including: Construct an adaptive modulation parameter selection objective function: ; ; in, This is the optimal modulation order for the communication signal. The optimal transmission power for sensing signals. Let M be the combined signal-to-noise ratio when the modulation order of the communication signal is M and the transmission power of the sensing signal is P. This represents the signal-to-noise ratio when the modulation order of the communication signal is M. , The signal-to-noise ratio of the communication signal. This indicates that the transmission power of the sensed signal is At that time, the signal-to-noise ratio of the received signal after transmission through optical fiber, , To sense the noise power spectral density of the signal, To sense sub-bands, This indicates the received power of the sensed signal after transmission through optical fiber. , For fiber optic channel gain, and Select weighting coefficients in the objective function for adaptive modulation parameters; Construct constraints for the objective function of adaptive modulation parameter selection: ; in, For total power, The power of the communication signal. For communication bit error rate threshold, To detect the threshold of the probability of missed detection. Let M be the bit error rate of the communication signal when the modulation order is M. , It is a standard Gaussian right-tailed function. Let P represent the probability of a missed detection when the transmitted power of the sensed signal is P. ; Based on the constraints of the adaptive modulation parameter selection objective function, and using the communication signal and the sensing signal, the adaptive modulation parameter selection objective function is solved to obtain the modulation parameters, which include the optimal modulation order of the communication signal and the optimal transmit power of the sensing signal.

[0009] Optionally, the integrated transmission signal is: ; in, For integrated signal transmission, For communication signal spectrum, , Based on Modulated communication signals This is the optimal modulation order for the communication signal. The center frequency of the communication subcarrier. For communication sub-band, It is the Dirac function. For frequency variables, To detect the center frequency of the vibration-sensing subcarrier, A dedicated sub-band for vibration sensing. The center frequency of the temperature sensing subcarrier, A dedicated sub-band for temperature sensing. To sense the center frequency of the subcarrier, For strain sensing, a dedicated sub-band , , These are the complex amplitudes of the vibration-sensing subcarrier, the temperature-sensing subcarrier, and the strain-sensing subcarrier, respectively. , The optimal transmission power for sensing signals.

[0010] Optionally, a deep prior decoupling network and a lightweight intelligent model are used for edge decision-making to obtain edge decision results for various events, specifically including: Receive the backscattered signal of the integrated transmitted signal after transmission via optical fiber; The backscattered signal is demodulated to obtain the response signals of each physical quantity, including vibration, temperature, and strain. The response signals to temperature and strain are decoupled using a deep prior decoupling network to obtain estimates of temperature change, strain change, and uncertainty. The vibration response signal is enhanced to obtain a vibration estimate; A lightweight intelligent model is used to perform fusion analysis on temperature change estimation, strain change estimation, uncertainty estimation and vibration estimation to obtain the recognition probability of various events; Decisions are made based on the recognition probability of various events to obtain edge decision results for each type of event; the edge decision results are immediate warning, pending review, or normal.

[0011] Optionally, the deep prior decoupling network includes an input layer, a physical information embedding layer, a feature extraction module, an adaptive calibration module, and an output layer connected in sequence. The physical information embedding layer is represented as follows: ; in, For physical embedding matrix, This represents the bias vector. For physical embedding vectors, For the input vector, , Let be the Brillouin frequency shift at time t. Let be the change in optical power at time t. The strain response signal at time t. This represents the number of spatial sampling points on the optical fiber. For the spatial location of the optical fiber; The feature extraction module includes multiple feature extraction layers connected in sequence; the feature extraction layers are represented as follows: ; ; in, The feature vector output by the k-th feature extraction layer. This is the intermediate vector of the k-th feature extraction layer. This is the feature vector output by the (k-1)th feature extraction layer. , This represents the number of feature extraction layers. The input to the first feature extraction layer is the physical embedding vector. This is a multi-head self-attention mechanism. For layer-normalized networks, It is a feedforward neural network; The adaptive calibration module is represented as follows: ; in, The feature vector after automatic calibration. For the first The feature vectors output by each feature extraction layer For element-wise multiplication, and For affine transformation parameters, , For conditional vectors, It is a multilayer perceptron network. , Let be the signal-to-noise ratio of the sensed signal at time t. To sense the bandwidth of the subband, For ambient temperature, This refers to the aging factor of optical fibers. The output layer is represented as: ; ; ; in, Estimate the temperature change at time t. The strain change at time t is estimated. For the uncertainty estimate at time t, , The mean and variance of the temperature change at time t in the automatically calibrated eigenvector are given. and These represent the mean and variance of the strain change at time t in the automatically calibrated eigenvector.

[0012] Optionally, the deep prior decoupling network is obtained by training a multi-task loss function; ; ; ; ; in, For multi-task loss function, For mission losses, For physical constraint loss, For uncertainty regularization loss, , To obtain temperature change estimates and strain change estimates by decoupling the response signals of temperature and strain at the i-th spatial sampling point, and For the i-th spatial sampling point, the actual temperature change and the actual strain change are... and The variances of the temperature change estimates and strain change estimates obtained by decoupling the temperature and strain response signals at the i-th spatial sampling point are given. This represents the number of spatial sampling points on the optical fiber. Let Brillouin shift be the temperature response signal at the i-th spatial sampling point. The change in optical power of the strain response signal at the i-th spatial sampling point Let i be the strain response signal at the i-th spatial sampling point. For the physical mapping matrix, and These are the weighting coefficients in the multi-task loss function.

[0013] Optionally, the vibration response signal is enhanced to obtain the vibration estimation formula: ; in, For the vibration estimate at time t, Let be the vibration response signal at time t. For wavelet transform, This is the inverse wavelet transform. Let be the impulse response of the bandpass filter at time t. For threshold function, To set a threshold.

[0014] Optionally, a lightweight intelligent model is used to perform fusion analysis on temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation to obtain the recognition probability of various events, specifically including: Based on temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation, a lightweight intelligent model is used to determine the initial recognition probability of various events. ; ; in, spatiotemporal sample blocks The initial recognition probability, For lightweight intelligent models, For the parameters of the lightweight intelligent model, This represents a spatiotemporal sample block extracted from an input feature set, which includes spatiotemporal data for temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation. , Indicates the first Starting from a spatial sampling point, continuously take... One spatial sampling point, Indicates from the first Starting from a certain time point, continuously take... At a certain point in time, , , They are respectively Class 0, Class 1 and Class 2. The initial recognition probability of class events, The number of event types; The initial recognition probabilities of various events are corrected using the following formula to obtain the recognition probabilities of various events; ; in, For the first The probability of identifying class events For the first The initial recognition probability of class events, The average uncertainty of the spatiotemporal sample block. This represents the uncertainty penalty coefficient.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects.

[0016] This application provides an early warning method based on the dynamic separation of multiple physical quantities in optical fiber. By organically combining the integrated design of sensing and transmitting signals with a deep prior decoupling network, this application achieves dynamic separation and high-precision sensing of multiple physical quantities on the same optical fiber medium, effectively solving the problem of cross-sensitivity. The application adopts a collaborative intelligent architecture of "rapid edge sensing + deep center fusion", which significantly improves the real-time performance of event recognition and early warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an early warning method based on dynamic separation of multiple physical quantities in optical fiber, provided as an embodiment of this application.

[0019] Figure 2 A schematic diagram of an early warning method based on dynamic separation of multiple physical quantities in optical fiber, provided in an embodiment of this application.

[0020] Figure 3 The flowchart illustrates the algorithm of an early warning method based on dynamic separation of multiple physical quantities in optical fiber, as provided in an embodiment of this application.

[0021] Figure 4 This is a schematic diagram of data processing logic based on a deep prior decoupling network, provided as an embodiment of this application.

[0022] Figure 5 This application provides a schematic diagram illustrating the results of monitoring and early warning of five typical risks in one embodiment. Detailed Implementation

[0023] 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, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] In one exemplary embodiment, an early warning method based on dynamic separation of multiple physical quantities in optical fiber is provided. This method is applied to a distributed optical fiber system, which includes: a front-end processing module, optical fibers, multiple edge nodes, and a cloud, such as... Figure 1 As shown, the early warning method includes the following steps 101-103.

[0026] Step 101: The front-end processing module generates an integrated transmission signal that satisfies the spectrum planning constraints based on the dynamic separation of multiple physical quantities, and transmits it using optical fiber. Step 102: At the edge node, a deep prior decoupling network and a lightweight intelligent model are used to make edge decisions to obtain edge decision results for various events; if the edge decision result is an immediate warning, a local warning is triggered immediately; if the edge decision result is to be reviewed, a feature vector of the event to be reviewed is generated and the feature vector of the event to be reviewed is sent to the cloud. Step 103: In the cloud, based on the feature vectors of the events to be reviewed sent by each edge node, a weighted voting model is used to make cloud decisions and obtain cloud decision results; if the cloud decision result is a confirmed warning, the warning level is determined and a cloud warning is issued based on the warning level.

[0027] Implementing steps 101-103 above can achieve the following technical effects.

[0028] By organically combining the integrated sensing and transmitting signal design with the deep prior decoupling network, dynamic separation and high-precision sensing of multiple physical quantities are achieved on the same optical fiber medium, effectively solving the problem of cross-sensitivity; the collaborative intelligent architecture of "rapid edge sensing + deep center fusion" significantly improves the real-time performance of event recognition and early warning.

[0029] In another exemplary embodiment, the feature vector of the event to be reviewed includes at least the type, location, time, and recognition probability of the event to be reviewed.

[0030] In another exemplary embodiment, such as Figure 2 As shown, step 101 above includes the process of designing and frequency domain planning of integrated inductive transmission signals and adaptive fusion modulation and transmission.

[0031] In another exemplary embodiment, a specific implementation process for integrated inductive transmission signal design and frequency domain planning is provided.

[0032] First, determine the allocation scheme of communication signals and sensing signals in the spectrum resources, and set the total bandwidth of the optical fiber as [missing information]. Divide it into communication subbands and perception subband And satisfy: , .

[0033] The sensing subband is further divided into: , .

[0034] in, A dedicated sub-band for vibration sensing, bandwidth , Dedicated sub-band for temperature sensing, bandwidth , For strain sensing, dedicated sub-band bandwidth , It is an empty set.

[0035] In this embodiment, subband allocation follows an adaptive optimization principle. First, a trade-off optimization objective function that maximizes synesthetic performance is set: ; Then, set constraints for the objective function that maximizes synesthetic performance while compromising on performance: ; ; ; ; in, To optimize the objective function to compromise on synesthetic performance, under conditions of limited spectrum resources, by weighting... and This allows for the adjustment of strategies that prioritize either communication or sensing, aiming to achieve an optimal trade-off between communication capacity and sensing performance. and These are the communication weight coefficient and the perception weight coefficient, respectively. For communication sub-band, To sense sub-bands, This is a function for calculating the subband capacity of communication. = , The function for calculating the sensitivity of the sensing subband is... The total bandwidth of the optical fiber. The signal-to-noise ratio of the communication signal. , The power of the communication signal. The noise power spectral density of the communication signal. To sense the signal-to-noise ratio of the signal. , To sense the transmission power of the signal, To sense the noise power spectral density of the signal, and These are the signal-to-noise ratio (SNR) thresholds for communication signals and sensing signals, respectively. .

[0036] The above process determined the resource usage structure of various services in the frequency domain, providing a foundation for subsequent signal construction.

[0037] In another exemplary embodiment, a specific implementation process for adaptive fusion modulation and transmission is provided.

[0038] Based on the subband division results obtained from the above process, unified modulation coding and fusion construction are performed on communication and multi-type sensing signals to generate an integrated transmission signal that meets spectrum planning constraints. This results in a frequency domain structure for synesthetic fusion. Its frequency domain expression satisfies: ; in, For integrated signal transmission, For communication signal spectrum, , Based on Modulated communication signals This is the optimal modulation order for the communication signal. The center frequency of the communication subcarrier. For communication sub-band, It is the Dirac function. For frequency variables, To detect the center frequency of the vibration-sensing subcarrier, Special sub-band for vibration sensing The center frequency of the temperature sensing subcarrier, Dedicated sub-band for temperature sensing To sense the center frequency of the subcarrier, For strain sensing, a dedicated sub-band , , The complex amplitudes of the vibration-sensing subcarrier, temperature-sensing subcarrier, and strain-sensing subcarrier, respectively. ,Right now, , This represents the optimal transmit power for the sensing signal. The entire expression is a superposition structure of the communication signal (continuous spectrum) and the sensing signal (discrete spectrum).

[0039] Communication section Under the control of M, the complex amplitude of the sensing part is controlled by the power P. The sum of the squares of the moduli of the complex amplitudes of the sensing subcarriers is equal to the total power of the sensing signal, i.e. ,in, To sense the transmission power of the signal, To sense the complex amplitude of the subcarrier.

[0040] The adaptive modulation parameter selection uses an adaptive modulation parameter selection objective function: ; ; in, This is the optimal modulation order for the communication signal. The optimal transmission power for sensing signals. These represent the combined signal-to-noise ratios when the modulation order of the communication signal is M and the transmission power of the sensing signal is P, respectively. This represents the signal-to-noise ratio when the modulation order of the communication signal is M. , The signal-to-noise ratio of the communication signal. This indicates that the transmission power of the sensed signal is At that time, the signal-to-noise ratio of the received signal after transmission through optical fiber, , To sense the noise power spectral density of the signal, To sense sub-bands, This indicates the received power of the sensed signal after transmission through optical fiber. , For fiber optic channel gain, and To select weighting coefficients in the objective function for adaptive modulation parameters, .

[0041] Bit error rate affecting communication This, in turn, affects the actual attainability. Requirements (Higher-order modulation requires a higher signal-to-noise ratio to achieve the same bit error rate; Direct impact This can affect sensory sensitivity.

[0042] Constraints: in, For total power, The power of the communication signal. For communication bit error rate threshold, To detect the threshold of the probability of missed detection. Let M be the bit error rate of the communication signal when the modulation order is M. , It is a standard Gaussian right-tailed function. Let P represent the probability of a missed detection when the transmitted power of the sensed signal is P. For example, .

[0043] In another exemplary embodiment, such as Figure 2 As shown, step 102 includes the process of multi-physical quantity sensing and scattering signal demodulation, further processing of the above-mentioned physical quantity (vibration, temperature, strain) data, and edge intelligent decision-making based on uncertainty perception.

[0044] In another exemplary embodiment, such as Figure 3 As shown, the specific implementation process of the above-mentioned multi-physical quantity sensing and scattering signal demodulation is as follows: The backscattered signal after the integrated transmission signal generated in step 101 is transmitted through optical fiber. Represented as: ; in, Indicates by The complex amplitude of the emitted light field, Represents convolution. It is a function of the distributed time-varying impulse response of optical fibers. It is additive noise.

[0045] It can be represented as: ; in, The spatial location of the r-th spatial sampling point is determined by the optical fiber. Changes in physical quantities The scattering coefficient affected For fiber optic spatial location The position of the r-th spatial sampling point The round-trip time between them, and the change of physical quantity at any spatial location of the optical fiber, are denoted as... , , , , These are temperature, strain, and sound pressure changes, respectively; through matched filtering and frequency domain segmentation, from... Separate from the corresponding Physical quantity related signals: ; ; ; in, For Fourier transform, This is the inverse Fourier transform. , , For the matched filter of the corresponding subband, , , The center frequencies of the vibration-sensing subcarrier, temperature-sensing subcarrier, and strain-sensing subcarrier. , , It represents the response signals to vibration, temperature, and strain.

[0046] Demodulation is performed using a coherent demodulation algorithm. , Brillouin frequency shift can be obtained and changes in optical power .

[0047] In another exemplary embodiment, such as Figure 3 and Figure 4 As shown, in the process of further processing the above physical quantity (vibration, temperature, strain) data, a deep prior decoupling network (DPDN) is used for processing the temperature and strain signals, and an enhancement processing method is used for the vibration signal.

[0048] In another exemplary embodiment, such as Figure 4 As shown, the data processing based on the DPDN network architecture can be described internally as a deep network structure of "input construction → physical information embedding → attention feature extraction → adaptive calibration → uncertainty estimation → loss optimization → online fine-tuning → output result". The corresponding deep prior decoupled network includes: an input layer, a physical information embedding layer, a feature extraction module, an adaptive calibration module, and an output layer connected in sequence.

[0049] In this embodiment of the application, input data is constructed at the input layer.

[0050] Define a deep prior decoupling network ( Network parameters: weights + biases, where... This represents the number of spatial sampling points on the optical fiber. The network input is a tensor composed of three observed physical quantities: ; in, For the input vector, Let be the Brillouin frequency shift at time t. Let be the change in optical power at time t. The strain response signal at time t. For the spatial location of the optical fiber, This represents the number of spatial sampling points on the optical fiber. There are physical coupling relationships between these observations: temperature affects the Brillouin frequency shift, strain affects both the frequency shift and power, and fiber aging and environmental noise introduce errors. Therefore, physical decoupling and calibration are required through a network.

[0051] In this embodiment, the network is introduced through the PhysEmb sensing mechanism.

[0052] First, construct the physical embedding matrix: ; in, For physical embedding matrix, The coefficient matrix, , , , , , The sensitivity coefficient, obtained through experimental calibration, represents the degree of linear influence of a physical quantity on an observed quantity. This represents the Kronecker product, used to extend the coefficient matrix to each spatial sampling point. for Identity matrix.

[0053] The aforementioned physical quantities are real-world environmental physical quantities, including temperature changes and strain changes. In this embodiment, these physical quantities are obtained through laboratory measurements. The observed quantities are response signals measured by the sensing system, including Brillouin frequency shift and optical power changes, etc. , , The physical embedding matrix is ​​used to characterize the linear influence relationship between physical quantities and observed quantities, thereby introducing physical mechanisms into the network modeling process.

[0054] Introducing sensing mechanisms into the network through a physical information embedding layer: ; in, For physical embedding vectors, Represents the bias vector (learnable parameters), with dimensions AND They are the same and are used to compensate for static errors in the system.

[0055] The feature extraction module consists of multiple feature extraction layers connected in sequence, and each feature extraction layer includes Attention and FFN.

[0056] The feature extraction layer is represented as follows: ; ; in, The feature vector output by the k-th feature extraction layer. This is the intermediate vector of the k-th feature extraction layer. This is the feature vector output by the (k-1)th feature extraction layer. , This represents the number of feature extraction layers. The input to the first feature extraction layer is the physical embedding vector. This is a multi-head self-attention mechanism. For layer-normalized networks, It is a feedforward neural network.

[0057] The function of this feature extraction layer is as follows: First, a multi-head self-attention mechanism is implemented: the correlation between different locations of the optical fiber and the dynamic changes at different times are modeled to capture long-distance spatial dependence and improve the accuracy of distributed sensing.

[0058] Then a feedforward neural network is executed: used for nonlinear feature mapping and decoupling complex coupling relationships.

[0059] In this embodiment, the execution process of the adaptive calibration module (AdaptCalib) is as follows: The network uses conditional vectors Perform dynamic calibration: ; in, Let be the signal-to-noise ratio of the sensed signal at time t. To sense subband bandwidth, For ambient temperature, This refers to the aging factor of optical fibers.

[0060] Input to calibration network: ; Affine transformation of features is used to dynamically compensate for system drift: ; in, The feature vector after automatic calibration. For the first The feature vectors output by each feature extraction layer For element-wise multiplication, and For affine transformation parameters, It is a multilayer perceptron (MLP) network.

[0061] In another exemplary embodiment, the deep prior decoupling network described above is obtained by training a multi-task loss function.

[0062] The multi-task loss function consists of three parts: task loss, physical consistency loss, and uncertainty regularization.

[0063] Task loss, used to account for prediction error: ; Physical constraint loss is used to ensure that the prediction results conform to the sensing mechanism: ; Uncertainty regularization loss is used to avoid unbounded growth of variance. ; The final multi-task loss function is: ; in, For multi-task loss function, For mission losses, For physical constraint loss, For uncertainty regularization loss, , To obtain temperature change estimates and strain change estimates by decoupling the response signals of temperature and strain at the i-th spatial sampling point, and For the i-th spatial sampling point, the actual temperature change and the actual strain change are... and The variances of the temperature change estimates and strain change estimates obtained by decoupling the temperature and strain response signals at the i-th spatial sampling point are given. This represents the number of spatial sampling points on the optical fiber. Let Brillouin shift be the temperature response signal at the i-th spatial sampling point. The change in optical power of the strain response signal at the i-th spatial sampling point Let i be the strain response signal at the i-th spatial sampling point. and These are the weighting coefficients in the multi-task loss function. This is the physical mapping matrix, the specific values ​​of which can be obtained from fiber optic calibration experiments, and is expressed as: .in, This describes a positive linear relationship between real physical quantities and observed quantities, as described in the above embodiments. for The inverse or approximate inverse mapping is used to map coupled signals in the observation space to the physical decoupled feature space. By using the formula It can be obtained through calculation, or through parameter learning in a data-driven manner.

[0064] In another exemplary embodiment, the obtained multi-task loss function is used in the following two phases. During the training phase, the multi-task loss function is calculated for each mini-batch. The gradient is calculated through backpropagation, and the network parameters of DPDN are updated using an optimizer (such as Adam (Adaptive Moment Estimation)) to minimize the loss; in the online fine-tuning phase, when the system detects a steady-state condition ( When using self-monitored loss, (can be combined) (Or used alone) Continue to perform small gradient updates to correct the model online to adapt to environmental changes.

[0065] In another exemplary embodiment, an online self-supervised fine-tuning process is also provided for the deep prior decoupling network described above.

[0066] In the process of deep prior decoupling network operation, it is necessary to detect whether it has entered a steady state. ; in, Indicates whether it is in a steady state. Indicates the length of the time window. Indicates input features (multiple physical quantities). It is any time variable within the time window. This represents the mean of the input features within the time window. The squared L2 norm represents the intensity of the fluctuation. This represents the steady-state determination threshold.

[0067] If satisfied If represents a steady state, then self-supervised learning is performed to adapt to changes in the environment.

[0068] In another exemplary embodiment, the self-supervised loss function described above is: ; in, For self-supervised loss function, This is the steady-state period.

[0069] In another exemplary embodiment, the output layer described above is used for uncertainty estimation.

[0070] The output layer is divided into two parts: ; For channel segmentation, the feature tensor is divided into two parts along the last dimension, corresponding to the mean and mean, respectively. and variance The mean is: ; ; Uncertainty: ; in, Estimate the temperature change at time t. The strain change at time t is estimated. For the uncertainty estimate at time t, The uncertainty estimate at time t represents the combined uncertainty of temperature and strain. , The mean and variance of the temperature change at time t in the automatically calibrated eigenvector are given. and These represent the mean and variance of the strain change at time t in the automatically calibrated eigenvector. This is the first average channel (temperature) output by the network. This is the second mean channel (strain) output by the network. For the spatial location of the optical fiber, specifically the location of the event, the purpose of this application is to [determine the location of the event]. The impact of an event on the physical quantities generated at spatial sampling points is sampled and calculated. The mean and variance in this process are the mean and variance of the results obtained at each spatial sampling point. For time.

[0071] This step not only outputs the predicted value, but also assesses the reliability of the prediction.

[0072] The final output tensor is: ; in, This is the final output tensor.

[0073] In another exemplary embodiment, such as Figure 4 As shown, the vibration signal processing procedure is as follows.

[0074] Parallel vibration signal enhancement: ; in, For the vibration estimate at time t, Let be the vibration response signal at time t. For wavelet transform, This is the inverse wavelet transform. Let be the impulse response of the bandpass filter at time t. For threshold function, To set a threshold, , The noise standard deviation is usually estimated using the high-frequency subband of the wavelet coefficients (e.g., using the median absolute deviation). The signal length (or the total number of wavelet coefficients) is the number of discrete sampling points involved in denoising.

[0075] In another exemplary embodiment, the purpose of the aforementioned uncertainty-aware edge intelligent decision-making is to generate local decisions.

[0076] To achieve real-time monitoring and rapid response, this application embodiment deploys a lightweight intelligent model at the edge node to perform fusion analysis on the monitoring results of multiple physical quantities and generate local intelligent decisions.

[0077] like Figure 3 As shown, the specific process of this edge intelligent decision-making is as follows: First, construct the edge decision input features. : ; in, The features at time t, where the optical fiber's spatial location z is input to the edge decision features.

[0078] Subsequently, the continuous monitoring data was segmented into fixed-length spatiotemporal sample blocks in both spatial and temporal dimensions: This formula represents the input features from edge decision-making. Extract a spatiotemporal sample block. Among them, The spatiotemporal data includes four channels: temperature change, strain change, vibration response, and uncertainty. The data is in the form of (total number of spatial points) × (total number of time points) × 4. Indicates the length of the spatial window. Indicates the length of the time window; Indicates the first Starting from a spatial sampling point, continuously take... One spatial sampling point; : Indicates from the first Starting from a certain time point, continuously take... At which point in time. Result It is a size of The tensor represents a block of multi-physical quantity data within the spatiotemporal window, used as input to the edge recognition model. .

[0079] Based on this, spatiotemporal samples are input into the edge-side lightweight recognition model. The output yields the recognition probability of various events: ; ; in, spatiotemporal sample blocks The initial recognition probability, For lightweight intelligent models, For the parameters of the lightweight intelligent model, , , They are respectively Class 0, Class 1 and Class 2. The initial recognition probability of class events, This represents the number of event types.

[0080] Considering that sensor data may contain noise and measurement uncertainties in complex environments, the uncertainty estimate obtained in the aforementioned steps is further utilized. Weighted correction is applied to the recognition probability: ; in, For the first The probability of identifying class events For the first The initial recognition probability of class events, The average uncertainty of the spatiotemporal sample block. As an uncertainty penalty coefficient, this weighting mechanism can automatically reduce the confidence probability of the corresponding event when the uncertainty of the monitoring data is high, thereby reducing the risk of false alarms and improving the reliability of decision-making. ; in, These represent the spatial and temporal window lengths, respectively. This indicates the position of the i-th spatial sampling point. This represents the j-th time point. This indicates single-point uncertainty.

[0081] Ultimately, based on the adjusted event probabilities Execute edge decision rules: ; in, and The threshold for early warning of edge decision-making. When the probability of recognizing a certain type of event exceeds a threshold When the probability is high, a local alert is immediately triggered; when the probability is low... and During this period, the event is marked as pending review, meaning it is identified as a review event. The feature vector of this event is uploaded to the cloud for further analysis. This feature vector must include at least: the type, location, time, and recognition probability of the event pending review. If the recognition probability of all events is below a threshold... If so, the current monitoring status is determined to be normal.

[0082] Through the above mechanism, rapid and reliable intelligent decision-making can be achieved at the edge, and critical events can be fed back to the subsequent monitoring and management system (i.e., the cloud) in a timely manner.

[0083] In another exemplary embodiment, network state awareness and reliable transmission technology is applied in the process of sending the feature vector of the event to be reviewed to the cloud in step 102 above.

[0084] Define a comprehensive health index for the link: ; in, Indicates link health Indicates link The transmission delay at time t, The maximum transmission delay across all links. For link The signal-to-noise ratio at time t The maximum signal-to-noise ratio across all links. For link The bit error rate at time t The maximum bit error rate across all links. , , These are weighting coefficients that can be dynamically adjusted based on business needs (such as prioritizing low latency or high reliability) through expert experience or adaptive optimization algorithms (such as reinforcement learning).

[0085] : Measured in real time by the physical layer, obtained by the ratio of received signal strength to noise floor, and is usually reported by the monitoring register of the optical module or RF front end.

[0086] Statistics are collected by the communication link layer, such as the number of failed CRC (Cyclic Redundancy Check) checks and the number of FEC (Forward Error Correction) correction bits, and are reported periodically.

[0087] : Measured through the network layer or transport layer, for example by sending probe packets to calculate half of the RTT (Round-Trip Time), or by using timestamp differences to obtain the one-way delay.

[0088] These data are continuously collected and updated into the link state database by the network state awareness module for use in route selection. The specific process of route selection is as follows: For each data flow, maintain a candidate route set, which includes the primary route. and backup route set When detected At that time, Within a given timeframe, based on the real-time status of each backup route Value, switch the data stream to .

[0089] ; in, This represents the actual route selected at time t. Indicates the primary route. Represents the set of backup routes. Indicates the health of route R. Indicates the health threshold. This indicates selecting the route that maximizes the objective function. This is the formula for calculating the comprehensive health index of the link.

[0090] In another exemplary embodiment, such as Figure 2 As shown, step 103 above is the process of multi-source fusion and global decision-making. Specifically, as follows... Figure 3 As shown, it includes multi-source data association, fusion confidence of uncertainty perception, global decision rules, and dynamic classification of early warning levels.

[0091] In another exemplary embodiment, a specific implementation process for multi-source data association is provided.

[0092] For any event to be reviewed, hereinafter referred to as the target event, its fusion confidence level is... The calculation, based on a weighted voting model, is as follows: Collect feature vectors of all events to be reviewed that are of the same type as the target event, and fuse them using the following formula.

[0093] ; in, The fusion confidence of the target event. Let O be the weight of the edge node o with respect to the target event. Let be the probability of identifying the target event obtained from edge node o. The target event is determined based on its location and time. Here, location refers to the position between the location where the target event occurs and the edge node o, and time refers to the time when the edge node o detects the target event. Generally, the farther the location and the later the time, the smaller the weight. The value of is between 0 and 1, and .

[0094] In another exemplary embodiment, a specific method for setting global decision rules is provided.

[0095] The high and low confidence thresholds for global early warning decision-making are respectively ,like If so, then confirm the global alert; if If so, it is marked as a suspected event requiring manual review; if If so, it is judged as a false alarm and filtered; ; This is the result of the overall decision.

[0096] In another exemplary embodiment, a specific implementation method for dynamically classifying warning levels is provided.

[0097] For confirmed global alerts, their alert levels are as follows: according to Dynamic partitioning: ; According to the specific embodiments provided in this application, this application has the following technical effects.

[0098] (1) This application effectively solves the problems of single sensing function, cross-sensitivity of multiple physical quantities, and competition for sensing and communication resources in existing fiber optic monitoring systems. This application achieves high-precision collaborative sensing and dynamic separation of multi-dimensional physical quantities such as vibration, temperature, and strain within the same fiber optic channel through integrated sensing and transmitting signal design and deep prior decoupling network, which significantly improves the monitoring dimensions and decoupling accuracy and achieves global optimization of communication and sensing resources.

[0099] (2) This application significantly shortens the delay from the occurrence of an event to the early warning through the intelligent decision-making architecture and uncertainty perception mechanism of edge-cloud collaboration, and improves the response speed and decision reliability; combined with link self-sensing and dynamic routing technology, it ensures the efficient and reliable transmission of early warning information in complex environments, and provides full-coverage and highly reliable intelligent monitoring protection for major infrastructure such as oil and gas pipelines and power corridors.

[0100] In another exemplary embodiment, in order to illustrate the implementation and effects of the above-described solution of this application, the present invention will be further described in detail below through implementation examples.

[0101] This implementation case uses a dataset of 800 samples, covering five typical events (tapping, climbing, pedestrian passing, manual digging, and normal conditions), with 160 fiber optic monitoring data sets for each of the five events. A random sampling method was used to select 120 samples from each of the five datasets as the training set, and the remaining 40 sets as the test set. Ultimately, a total of 600 samples were used for training, and a total of 200 samples were used for testing.

[0102] The overall process of the method provided in this application is as follows: Figure 2 As shown, the specific steps are as follows: 1. Integrated Sensing and Transmitting Signal Design and Frequency Domain Planning Set master tape: Divide it into communication subbands and perception subband And satisfy: , ; The sensing subband is further divided into: , ; Among them, the vibration sensing sub-band: , Temperature sensing dedicated sub-band: , ; Strain sensing dedicated sub-band: , .

[0103] Subband allocation follows the principle of adaptive optimization: Maximizing synesthetic performance requires a trade-off in optimizing the objective function: ; Constraints: ; ; ⊆ ; ; Among them, communication subband capacity Sensing subband sensitivity , , These are weighting coefficients, measured in practice. , , All of these values ​​are higher than their preset thresholds. The optimized target value is calculated to be: .

[0104] 2. Adaptive Fusion Modulation and Transmission Generate integrated transmission signal Its frequency domain expression satisfies: ; in, The bandwidth is ,exist Ten vibration sensing subcarriers are set within the range, in Five temperature sensing subcarriers are set within the range, Five strain sensing subcarriers are set within the range. , , .

[0105] Adaptive modulation parameter selection: ; ; Pick According to calculations, .

[0106] Constraints: ; In this example, the following is adopted: Higher-order modulation to improve communication speed Total power A balanced allocation between communication and sensing functions.

[0107] 3. Multi-physical quantity sensing and scattering signal demodulation Backscattered signal after receiving integrated transmission signal transmitted through optical fiber Represented as: ; In this example, by position Taking the "knocking" incident as an example, this incident caused... Change, and thus through The reaction is reflected in the backscattered signal. It is additive noise, and its power spectral density is .

[0108] By using a set of matched filters, the mixed signal is segmented in the frequency domain to separate the signals related to physical quantities, providing clean input data for subsequent in-depth processing. Demodulation is performed using a coherent demodulation algorithm. , Brillouin frequency shift can be obtained and changes in optical power .

[0109] Here, vibration is used as an example. After processing, in Vibration response signal obtained at the location Its peak signal-to-noise ratio was calculated to be 18dB.

[0110] 4. Data processing based on Deep Prior Decoupling Network (DPDN) 4.1 DPDN Network Architecture Define a deep prior decoupling network ,in =5000 represents the number of spatial sampling points.

[0111] Generate a tensor consisting of three observed physical quantities: .

[0112] For a single "tapping" event sample, the location window is taken. 100 points.

[0113] : ; : ; : .

[0114] The network was trained using 600 pre-prepared training samples to learn complex mappings that decouple independent physical quantities from mixed observations. The network outputs a decoupled tensor as follows: ; Network structure: ; ; ; ; ; in, As a physical information embedding layer, For adaptive calibration module, For conditional vectors, The output header contains the mean branch and the variance branch.

[0115] 4.2 Physical Information Embedding Layer Physical embedding matrix construction: ; Embedding operation: .

[0116] 4.3 Adaptive Calibration Module Condition vector ; in: Let be the signal-to-noise ratio of the sensed signal at time t. To sense subband bandwidth, For ambient temperature, This refers to the aging factor of optical fibers.

[0117] Calibration transformation: ; ; In this example, , .

[0118] 4.4 Uncertainty Estimation Modeling uncertainty using heteroscedasticity: ; ; ; 4.5 Loss Function Multi-task loss function: ; Among the losses incurred: ; In this example, and All are 0. , , ; Physical constraint loss is approximately zero. ; Uncertainty regularization: ; Calculations show that .

[0119] 4.6 Online self-monitoring fine-tuning Steady-state detection: ; Self-monitoring loss: ; in, This is the steady-state period.

[0120] 4.7 Vibration Signal Processing And on Vibration signal enhancement was performed to obtain : .

[0121] 5. Edge intelligent decision-making based on uncertainty perception Edge Lightweight Model Input feature construction: ; Spatiotemporal segmentation, taking time windows The 10 time points have a feature dimension of 4×10: ; Calculate the initial predicted probabilities for 5 types of events by analyzing 200 samples in the test set. : ; These correspond to [normal, knocking, climbing, pedestrian, digging] respectively; make Calculate the uncertainty-weighted decision for this spatiotemporal region: ; ; Decision-making rules: ; decision making: But smaller than Therefore, the decision was "pending review", and the event was identified as an event pending review.

[0122] 6. Network status awareness and reliable transmission Link health metrics: .

[0123] Actual measurement: , , , , , , , In this example, the link Set it as the primary route.

[0124] Routing selection: Threshold ; because Therefore, choose .

[0125] 7. Multi-source fusion and global decision-making Combining the above The final fusion confidence score, obtained from the edge decision results of other edge nodes, is as follows: .

[0126] 7.1 Global Decision-Making Rules The high and low confidence thresholds for global early warning decision-making are respectively ; ; because The overall decision-making process is "manual review".

[0127] 7.2 Dynamic Classification of Early Warning Levels .

[0128] After running the complete process on the test set of 200 samples, the statistical model prediction results are as follows: Correctly predicted sample size: 196; Early warning accuracy: .

[0129] In another exemplary embodiment, to verify the overall performance of the method proposed in this application, five sets of targeted experiments were designed to test the early warning accuracy. The experimental results are summarized as follows: Figure 5 As shown. Analysis Figure 5 As can be seen, the average early warning accuracy of the technical solution in this application reaches an excellent level of 99%. The results show that, while ensuring the long-term stable and reliable operation of the system, this application achieves high-precision, low-latency sensing and accurate early warning of multiple physical quantities such as vibration, temperature, and strain, significantly outperforming traditional single-function sensing systems. This not only fully verifies the effectiveness and advancement of the fiber-optic multi-physical quantity dynamic separation and early warning system established in this invention, but also provides solid technical support and practical engineering solutions for building a new generation of intelligent monitoring and early warning system for infrastructure that is wide-coverage, cost-effective, fast-responding, and highly reliable.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An early warning method based on dynamic separation of multiple physical quantities in optical fiber, characterized in that, The early warning method is applied to a distributed optical fiber system, which includes: a front-end processing module, optical fibers, multiple edge nodes, and a cloud; the early warning method includes: The front-end processing module generates an integrated transmission signal that meets the spectrum planning constraints based on the dynamic separation of multiple physical quantities, and then transmits it using optical fiber. At the edge nodes, a deep prior decoupling network and a lightweight intelligent model are used to make edge decisions and obtain edge decision results for various events. If the edge decision result is an immediate warning, a local warning is triggered immediately. If the edge decision result is to be reviewed, a feature vector of the event to be reviewed is generated and sent to the cloud. In the cloud, a weighted voting model is used to make cloud decisions based on the feature vectors of the events to be reviewed sent by each edge node, and the cloud decision results are obtained. If the cloud decision result is a confirmation warning, the warning level is determined and a cloud warning is issued based on the warning level.

2. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 1, characterized in that, The method of generating an integrated transmit signal that satisfies spectrum planning constraints based on the dynamic separation of multiple physical quantities specifically includes: The optical fiber spectrum resources are allocated by solving the objective function of compromising the inductive performance, and the allocation results are obtained. Based on communication signals and sensing signals, modulation parameters are generated by solving the fusion modulation objective function. Based on the allocation results and the modulation parameters, an integrated transmission signal that satisfies the spectrum planning constraints is generated.

3. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 2, characterized in that, The allocation results include: communication subband, vibration sensing subband, temperature sensing subband, and strain sensing subband; the optical fiber spectrum resources are allocated by solving a trade-off optimization objective function for sensing performance, and the allocation results specifically include: The communication subband and sensing subband are obtained by solving the objective function that compromises the synergistic performance. The objective function for the trade-off optimization of synesthesia performance is: ； The constraints for optimizing the objective function by compromising synesthetic performance are: ； ； ； ； in, To optimize the objective function with a trade-off for synesthetic performance, and These are the communication weight coefficient and the perception weight coefficient, respectively. For communication sub-band, To sense sub-bands, This is a function for calculating the subband capacity of communication. , The function for calculating the sensitivity of the sensing subband is... The total bandwidth of the optical fiber. The signal-to-noise ratio of the communication signal. , The power of the communication signal. The noise power spectral density of the communication signal. To sense the signal-to-noise ratio of the signal. , To sense the transmission power of the signal, To sense the noise power spectral density of the signal, and These are the signal-to-noise ratio (SNR) thresholds for communication signals and sensing signals, respectively. It is an empty set; The sensing sub-bands are divided according to the following formula to obtain vibration sensing sub-bands, temperature sensing sub-bands, and strain sensing sub-bands; ； ； in, , , These are dedicated sub-bands for vibration sensing, temperature sensing, and strain sensing, obtained by dividing the sensing sub-bands.

4. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 2, characterized in that, Based on communication signals and sensing signals, modulation parameters are generated by solving a fusion modulation objective function, specifically including: Construct an adaptive modulation parameter selection objective function: ； ； in, This is the optimal modulation order for the communication signal. The optimal transmission power for sensing signals. Let M be the combined signal-to-noise ratio when the modulation order of the communication signal is M and the transmission power of the sensing signal is P. This represents the signal-to-noise ratio when the modulation order of the communication signal is M. , The signal-to-noise ratio of the communication signal. This indicates that the transmission power of the sensed signal is At that time, the signal-to-noise ratio of the received signal after transmission through optical fiber, , To sense the noise power spectral density of the signal, To sense sub-bands, This indicates the received power of the sensed signal after transmission through optical fiber. , For fiber optic channel gain, and Select weighting coefficients in the objective function for adaptive modulation parameters; Construct constraints for the objective function of adaptive modulation parameter selection: ； in, For total power, The power of the communication signal. For communication bit error rate threshold, To detect the threshold of the probability of missed detection. Let M be the bit error rate of the communication signal when the modulation order is M. , It is a standard Gaussian right-tailed function. Let P represent the probability of a missed detection when the transmitted power of the sensed signal is P. ; Based on the constraints of the adaptive modulation parameter selection objective function, and using the communication signal and the sensing signal, the adaptive modulation parameter selection objective function is solved to obtain the modulation parameters, which include the optimal modulation order of the communication signal and the optimal transmit power of the sensing signal.

5. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 2, characterized in that, The integrated transmission signal is: ； in, For integrated signal transmission, For communication signal spectrum, , Based on Modulated communication signals This is the optimal modulation order for the communication signal. The center frequency of the communication subcarrier. For communication sub-band, It is the Dirac function. For frequency variables, To detect the center frequency of the vibration-sensing subcarrier, A dedicated sub-band for vibration sensing. The center frequency of the temperature sensing subcarrier, A dedicated sub-band for temperature sensing. To sense the center frequency of the subcarrier, For strain sensing, a dedicated sub-band , , These are the complex amplitudes of the vibration-sensing subcarrier, the temperature-sensing subcarrier, and the strain-sensing subcarrier, respectively. , The optimal transmission power for sensing signals.

6. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 1, characterized in that, Edge decision-making is performed using a deep prior decoupling network and a lightweight intelligent model to obtain edge decision results for various events, including: Receive the backscattered signal of the integrated transmitted signal after transmission via optical fiber; The backscattered signal is demodulated to obtain the response signals of each physical quantity, including vibration, temperature, and strain. The response signals to temperature and strain are decoupled using a deep prior decoupling network to obtain estimates of temperature change, strain change, and uncertainty. The vibration response signal is enhanced to obtain a vibration estimate; A lightweight intelligent model is used to perform fusion analysis on temperature change estimation, strain change estimation, uncertainty estimation and vibration estimation to obtain the recognition probability of various events; Decisions are made based on the recognition probability of various events to obtain edge decision results for each type of event; the edge decision results are immediate warning, pending review, or normal.

7. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 6, characterized in that, The deep prior decoupling network comprises an input layer, a physical information embedding layer, a feature extraction module, an adaptive calibration module, and an output layer connected in sequence. The physical information embedding layer is represented as follows: ； in, For physical embedding matrix, This represents the bias vector. For physical embedding vectors, For the input vector, , Let be the Brillouin frequency shift at time t. Let be the change in optical power at time t. The strain response signal at time t. This represents the number of spatial sampling points on the optical fiber. For the spatial location of the optical fiber; The feature extraction module includes multiple feature extraction layers connected in sequence; the feature extraction layers are represented as follows: ； ； in, The feature vector output by the k-th feature extraction layer. This is the intermediate vector of the k-th feature extraction layer. This is the feature vector output by the (k-1)th feature extraction layer. , This represents the number of feature extraction layers. The input to the first feature extraction layer is the physical embedding vector. This is a multi-head self-attention mechanism. For layer-normalized networks, It is a feedforward neural network; The adaptive calibration module is represented as follows: ； in, The feature vector after automatic calibration. For the first The feature vectors output by each feature extraction layer For element-wise multiplication, and For affine transformation parameters, , For conditional vectors, It is a multilayer perceptron network. , Let be the signal-to-noise ratio of the sensed signal at time t. To sense the bandwidth of the subband, For ambient temperature, This refers to the aging factor of optical fibers. The output layer is represented as: ； ； ； in, Estimate the temperature change at time t. The strain change at time t is estimated. For the uncertainty estimate at time t, , The mean and variance of the temperature change at time t in the automatically calibrated eigenvector are given. and These represent the mean and variance of the strain change at time t in the automatically calibrated eigenvector.

8. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 6, characterized in that, The deep prior decoupling network is obtained through training using a multi-task loss function; ； ； ； ； in, For multi-task loss function, For mission losses, For physical constraint loss, For uncertainty regularization loss, , To obtain temperature change estimates and strain change estimates by decoupling the response signals of temperature and strain at the i-th spatial sampling point, and For the i-th spatial sampling point, the actual temperature change and the actual strain change are... and The variances of the temperature change estimates and strain change estimates obtained by decoupling the temperature and strain response signals at the i-th spatial sampling point are given. This represents the number of spatial sampling points on the optical fiber. Let Brillouin shift be the temperature response signal at the i-th spatial sampling point. The change in optical power of the strain response signal at the i-th spatial sampling point Let i be the strain response signal at the i-th spatial sampling point. For the physical mapping matrix, and These are the weighting coefficients in the multi-task loss function.

9. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 6, characterized in that, The vibration response signal is enhanced to obtain the vibration estimation formula: ； in, For the vibration estimate at time t, Let be the vibration response signal at time t. For wavelet transform, This is the inverse wavelet transform. Let be the impulse response of the bandpass filter at time t. For threshold function, To set a threshold.

10. The early warning method based on dynamic separation of multiple physical quantities in optical fiber according to claim 6, characterized in that, A lightweight intelligent model is used to fuse and analyze temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation to obtain the recognition probability of various events, specifically including: Based on temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation, a lightweight intelligent model is used to determine the initial recognition probability of various events. ； ； in, spatiotemporal sample blocks The initial recognition probability, For lightweight intelligent models, For the parameters of the lightweight intelligent model, This represents a spatiotemporal sample block extracted from an input feature set, which includes spatiotemporal data for temperature change estimation, strain change estimation, uncertainty estimation, and vibration estimation. , Indicates the first Starting from a spatial sampling point, continuously take... One spatial sampling point, Indicates from the first Starting from a certain time point, continuously take... At a certain point in time, , , They are respectively Class 0, Class 1 and Class 2. The initial recognition probability of class events, The number of event types; The initial recognition probabilities of various events are corrected using the following formula to obtain the recognition probabilities of various events; ； in, For the first The probability of identifying class events For the first The initial recognition probability of class events, The average uncertainty of the spatiotemporal sample block. This represents the uncertainty penalty coefficient.

Images