Wide-area environment intelligent monitoring and early warning method based on fiber sensing integration and federated learning

Abstract

The application discloses a wide-area environment intelligent monitoring and early warning method based on optical fiber sensing integration and federated learning, first, frequency domain division communication signals and sensing signals; then, self-adaptive integrated transmission signal is generated and preprocessed, slicing is carried out according to a preset time / space window, time-space sample blocks are generated, and each is input into a deep neural network deployed in advance, and each monitored event is output; further division is carried out into local early warning events or events to be reviewed, and corresponding feature vectors are constructed; each deep neural network generates global model parameters based on federated learning to update a local model; finally, adaptive routing switching is realized through a link health degree comprehensive index, reliable transmission of federated learning parameters and early warning data is ensured, local early warning information and feature vectors of events to be reviewed of edge nodes with reliable transmission are converged through a center server, global early warning decision is realized after multi-source fusion and arbitration, and high reliability of early warning information feedback in a complex environment is ensured.

Description

Technical Field

[0001] This invention relates to the field of optical fiber communication and sensing, specifically to a wide-area environment intelligent monitoring and early warning method based on the integration of optical fiber communication and sensing and federated learning. Background Technology

[0002] Currently, in the field of wide-area environmental monitoring, such as large infrastructure clusters, national security perimeters, and long-distance oil and gas pipelines, the demand for distributed, real-time, multi-parameter environmental status perception and intelligent early warning is becoming increasingly urgent.

[0003] Existing technologies mainly fall into two categories:

[0004] One type is the monitoring network that relies on a large number of discrete sensors, which has problems such as high deployment costs, difficult maintenance, and difficulty in achieving true continuous spatial coverage;

[0005] Another type is distributed sensing technology based on dedicated optical fibers. Although it achieves continuous measurement, the system has a single function and usually only monitors one physical quantity. It cannot meet the complex needs of collaborative sensing and correlation analysis of multi-dimensional environmental parameters such as temperature, strain, and vibration.

[0006] More importantly, existing "sensing" and "communication" systems are often separate:

[0007] Sensor data needs to be transmitted back to the central processing center through an independent communication network, which not only increases the complexity and construction cost of the system, but also results in a high delay from the occurrence of an event to the generation of an early warning due to the long data transmission path and multiple processing steps, making it difficult to meet the real-time response requirements for security events such as intrusion, leakage, and deformation.

[0008] Furthermore, within the vast monitoring area, communication links are susceptible to interruption or performance degradation due to environmental factors, and there is a lack of effective guarantee for the reliable and stable transmission of early warning information, resulting in insufficient robustness of the overall system.

[0009] Meanwhile, the existing system is weak in data processing and intelligent analysis, and the data between monitoring nodes is isolated, making it difficult to form global knowledge sharing and model evolution. In addition, the uploading of massive amounts of sensor data also brings significant communication overhead and data privacy leakage risks. Summary of the Invention

[0010] To address the drawbacks of the aforementioned discrete systems, this invention proposes a wide-area intelligent environmental monitoring and early warning method based on fiber optic sensing integration and federated learning. This method constructs a new generation of wide-area intelligent environmental monitoring system that is characterized by wide coverage, low cost, fast response, high reliability, and continuous learning capabilities.

[0011] The specific steps of the wide-area environment intelligent monitoring and early warning method based on fiber optic sensing integration and federated learning are as follows:

[0012] Step 1: Build a communication scenario for wide-area environmental intelligent monitoring based on long-distance fiber optic links, and divide the communication signals and sensor signals in the frequency domain on the spectrum resources;

[0013] Total bandwidth Divided into communication sub-bands and perception subband And satisfy: , .

[0014] Step 2: Adaptively generate the communication quality index and perceived signal-to-noise ratio requirements based on the real-time channel assessment feedback. Integrated transmission signal in time slots ;

[0015] Integrated transmission signal The frequency domain expression is:

[0016]

[0017] in, For communication signal spectrum, For vibration sensing subcarrier set, For temperature / strain sensing subcarrier set, and These are the complex amplitudes of the corresponding sensing subcarriers. It is the Dirac function. Indicates frequency Discrete spectral lines (subcarriers) at the location. For frequency variables;

[0018] Step 3: Integrated signal transmission The backscattered light signal generated by optical fiber transmission is converted into an electrical signal through photoelectric conversion. And the vibration response signal is separated. and temperature response signal .

[0019] The electrical signal is calculated as follows:

[0020]

[0021] in, This represents the complex amplitude of the emitted light field. Represents convolution. It is a function characterizing the distributed time-varying impulse response of optical fibers. Indicates position Due to changes in physical quantities (temperature ,strain Sound pressure The fiber refractive index response function caused by ) For noise;

[0022] Through matched filtering, frequency domain segmentation, and coherent demodulation algorithms, from electrical signals... Separate from the corresponding and Vibration response signal and temperature response signal .

[0023] Step 4: Analyze the vibration response signal and temperature response signal Preprocessing and calibration are performed to obtain enhanced spatiotemporal data. and decoupling physical quantities :

[0024] Temperature response signal Cross-sensitivity decoupling between temperature and strain is achieved by solving the following matrix equations:

[0025]

[0026] in, The change in Brillouin divergence frequency shift, This represents the change in optical power. The decoupled physical quantities are obtained from the calibrated sensitivity coefficient matrix. ;

[0027] Vibration response signal Wavelet thresholding denoising and bandpass filtering are performed to obtain the enhanced signal. .

[0028] Step 5: Process the preprocessed spatiotemporal data and decoupled Spatiotemporal slicing is performed according to a preset time window / spatial window, and temporal alignment and multi-channel feature organization are completed to continuously generate multiple fixed-length spatiotemporal sample blocks. ;

[0029] Step Six: Deploy a distributed data acquisition and processing unit in each of the different fiber optic monitoring areas or access locations. Each unit serves as an edge node, and a deep neural network model with the same structure but independently updated parameters is deployed on each edge node. ;

[0030] Step 7: Spatiotemporal sample block Each sample is input into a deep neural network model, and the output is the event probability vector corresponding to that sample block. From vector Select the event corresponding to the highest probability value, which is the spatiotemporal sample block. The events to be monitored corresponding to each deep neural network model;

[0031] Event probability vector ;in, For background event probabilities, Corresponding to The pre-defined probability of identifying the event to be monitored;

[0032] Step 8: Assign the probability of the event to be monitored. The event to be monitored is judged against a threshold and marked as a local early warning event or an event to be reviewed.

[0033] If it exists ( , If the threshold for high confidence in local decision-making is reached, a local early warning flag will be generated immediately.

[0034] like ( If the threshold for low confidence in local decision-making is set, then a flag for review is generated.

[0035] Step 9: Construct spatiotemporal sample blocks separately. Local early warning information and the feature vector of the event to be reviewed ;

[0036] Local early warning information is calculated as follows:

[0037]

[0038] For the first The types of events to be monitored corresponding to each deep neural network model: ;

[0039] For the first The location of the event to be monitored corresponding to each deep neural network model; For the first The confidence level of the event to be monitored corresponding to each deep neural network model: ; This is the warning time;

[0040] Feature vector of the event to be reviewed Similarly, this includes: the first The type, location, warning time, and recognition confidence level of the event to be reviewed corresponding to each deep neural network model;

[0041] Step 10: Each edge node trains its own deep neural network model based on locally collected spatiotemporal data, obtains updated model weights, and uploads them to the central server. Global model parameters are then generated based on federated learning. :

[0042]

[0043] in, The total number of edge nodes participating in federated learning. For the first The amount of local data at each edge node This represents the total amount of data across all nodes. For the first The updated model weights for each edge node;

[0044] Step 11: The central server will aggregate the global model. Distribute the data to each edge node and update the local model;

[0045] Step 12: Define a comprehensive link health index to achieve adaptive route switching and ensure reliable transmission of federated learning parameters and early warning data.

[0046] The comprehensive health index of the link is:

[0047]

[0048] in Indicates the link identifier. For signal-to-noise ratio, For bit error rate, For time delay, These are the weighting coefficients; This represents the maximum delay.

[0049] Adaptive route switching specifically refers to:

[0050] First, maintain a set of candidate routes for each data stream. ;

[0051] Indicates the primary (default) route. and This indicates two backup routes;

[0052] Then, when the health of the primary router is detected Less than the health threshold At that time, Within a given timeframe, based on the real-time health status of each backup router. The system selects the backup route with the highest health score and switches the data stream accordingly.

[0053] Step 13: By aggregating local early warning information and feature vectors of events awaiting review from multiple reliably transmitted edge nodes through the central server, and performing multi-source fusion and arbitration, a global early warning decision is achieved.

[0054] First, receive and associate reliable transmissions. The local early warning information of each edge node and the feature vector of the event to be reviewed are used as input. :

[0055]

[0056] in, For the first The feature vector of the event to be reviewed corresponding to each deep neural network model. For the first The video data stream corresponding to each deep neural network model For the first External meteorological and geological data of each edge node;

[0057] Then, the central server based on the input dataset The information uploaded by each edge node is spatiotemporally correlated and clustered to form a set of events to be decided.

[0058] For any event S to be decided, its fusion confidence level is calculated using a weighted voting model. :

[0059]

[0060] in, For from Or to The perceived confidence level obtained from the secondary AI analysis. In order to pass through The confidence level of the video analysis is obtained. For based on And the historical and situational confidence scores for the matching degree of the historical event database at that location; , , For adaptive weights;

[0061] Finally, the high and low confidence thresholds for global early warning decisions were set as follows: If the confidence level is fused Then confirm a global early warning for the event S to be decided; if If so, then the event S is marked as a suspected event requiring manual review; if If so, the event S to be decided is determined to be a false alarm and filtered out;

[0062] Global Alert Generation and Issuance: For confirmed global alerts, generate structured alert commands:

[0063] ;

[0064] Among them, the warning level according to Dynamically classify values ​​and event propagation speed; structured early warning instructions. Publish to designated terminals, display devices, and linkage systems.

[0065] The advantages of this invention are:

[0066] (1) This invention solves the problems of separation of sensing and communication systems, single function and poor real-time performance in existing wide-area monitoring. Through the integrated design of fiber optic sensing, it realizes the synchronous sensing and data transmission of multiple physical quantities on a single optical fiber, which greatly reduces the system complexity and deployment cost, and expands the monitoring dimensions, providing effective technical support for building a new generation of wide-area environmental intelligent monitoring system with wide coverage, low cost, fast response and high reliability.

[0067] (2) This invention utilizes a collaborative intelligent architecture of "rapid edge perception + deep central fusion" that integrates deep neural networks and federated learning. This architecture enables high-precision event recognition and rapid local decision-making at the edge, while at the central level, a federated aggregation mechanism facilitates continuous evolution and knowledge sharing across node models. This significantly improves the real-time performance, accuracy, and adaptability of monitoring, while simultaneously ensuring data privacy and communication efficiency. Furthermore, the system's built-in link self-perception and anti-interruption transmission mechanism ensures high reliability of early warning information feedback in complex environments, achieving end-to-end optimization from perception to decision-making. Attached Figure Description

[0068] Figure 1 This is a flowchart of the wide-area environment intelligent monitoring and early warning method based on fiber optic sensing integration and federated learning, as described in this invention.

[0069] Figure 2 This is a schematic diagram of the internal logic of the "deep neural network model - local fast decision-making - central fusion decision-making" of the present invention;

[0070] Figure 3 This is a flowchart of a wide-area environment intelligent monitoring and early warning method in a specific embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram illustrating the monitoring and early warning results of five groups of wide-area environmental risk events using the present invention; Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0073] The wide-area environment intelligent monitoring and early warning method based on fiber optic sensing integration and federated learning proposed in this invention has important application value and necessity for realizing low-cost, high-efficiency and intelligent wide-area environment monitoring; its application scope is wide and can serve multiple national-level major safety monitoring scenarios such as smart city utility tunnels, highway slopes, power transmission corridors, and border security.

[0074] This invention deeply integrates wide-area environmental information perception, reliable data transmission, and edge intelligent decision-making: by embedding perception functionality deep into the communication fiber optic network itself, utilizing the same fiber optic medium and spectrum resources, it simultaneously achieves high-quality data communication and high-precision distributed perception of multiple physical quantities, thereby greatly simplifying the system architecture and reducing deployment and maintenance costs. Simultaneously, through an edge-cloud collaborative intelligent framework that combines deep neural networks and federated learning, data preprocessing and rapid event identification are completed at the edge, while multi-source information fusion and global decision-making are performed at the central side. The federated learning mechanism enables continuous evolution of models at each edge node and cross-domain knowledge sharing, effectively protecting data privacy and reducing communication overhead. The system's built-in link self-sensing and anti-interruption transmission mechanisms further ensure the high reliability of early warning information feedback in complex environments, ultimately constructing a new generation of wide-area environmental intelligent monitoring system with wide coverage, low cost, fast response, high reliability, and continuous learning capabilities.

[0075] The wide-area environment intelligent monitoring and early warning method based on fiber optic sensing integration and federated learning, such as Figure 1 and Figure 2 As shown, the specific steps are as follows:

[0076] Step 1: Build a communication scenario for wide-area environmental intelligent monitoring based on long-distance fiber optic links, and divide the communication signals and sensor signals in the frequency domain on the spectrum resources;

[0077] Total bandwidth Divided into communication sub-bands and perception subband And satisfy: , .

[0078] Step 2: Based on the spectrum resource allocation scheme, and according to the communication quality index and perceived signal-to-noise ratio requirements from real-time channel assessment feedback, adaptively select the modulation format combination of the integrated signal to generate... Integrated transmission signal in time slots ;

[0079] Modulation format combination ( , ),in The modulation order of the communication signal. The peak power of the sensing signal;

[0080] Integrated transmission signal The frequency domain expression satisfies:

[0081]

[0082] in, For communication signal spectrum, For vibration sensing subcarrier set, For temperature / strain sensing subcarrier set, and These are the complex amplitudes of the corresponding sensing subcarriers. It is the Dirac function. Indicates frequency Discrete spectral lines (subcarriers) at the location. For frequency variables;

[0083] Step 3: Integrated signal transmission The backscattered light signal generated by optical fiber transmission is converted into an electrical signal through photoelectric conversion. And the vibration response signal is separated. and temperature response signal .

[0084] The electrical signal is calculated as follows:

[0085]

[0086] in, This represents the complex amplitude of the emitted light field. Represents convolution. It is a function characterizing the distributed time-varying impulse response of optical fibers. Indicates position Due to changes in physical quantities (temperature ,strain Sound pressure The fiber refractive index response function caused by ) For noise;

[0087] Through matched filtering, frequency domain segmentation, and coherent demodulation algorithms, from electrical signals... Separate from the corresponding and Vibration response signal and temperature response signal .

[0088] Step 4: Analyze the vibration response signal and temperature response signal Preprocessing and calibration are performed to obtain enhanced spatiotemporal data. and decoupling physical quantities :

[0089] Edge-side data preprocessing and calibration: response to temperature variations Cross-sensitivity decoupling between temperature and strain is achieved by solving the following matrix equations:

[0090]

[0091] in, The change in Brillouin divergence frequency shift is linearly related to temperature and strain. This represents the change in optical power. The decoupled physical quantities are obtained from the calibrated sensitivity coefficient matrix. ;

[0092] Vibration response signal Wavelet thresholding denoising and bandpass filtering are performed to obtain the enhanced signal. .

[0093] Step 5: Process the preprocessed spatiotemporal data and decoupled Spatiotemporal slicing is performed according to a preset time window / spatial window, and temporal alignment and multi-channel feature organization are completed to continuously generate multiple fixed-length spatiotemporal sample blocks. ;

[0094] Step Six: Deploy a distributed data acquisition and processing unit in each of the different fiber optic monitoring areas or access locations. Each unit serves as an edge node, and a deep neural network model with the same structure but independently updated parameters is deployed on each edge node. ;

[0095] Step 7: Spatiotemporal sample block Each sample is input into a deep neural network model, and the output is the event probability vector corresponding to that sample block. From vector Select the event corresponding to the highest probability value, which is the spatiotemporal sample block. The events to be monitored corresponding to each deep neural network model;

[0096] Deep neural network model It employs an architecture that integrates convolutional neural networks with temporal networks (such as LSTM or Transformer) to automatically extract spatiotemporal features, and its output is an event probability vector. ,

[0097] For background event probability, Corresponding to The probability of identifying a preset event to be monitored (such as intrusion, excavation, leakage, etc.);

[0098] Step 8: Assign the probability of the event to be monitored. The event to be monitored is judged against a threshold and marked as a local early warning event or an event to be reviewed.

[0099] If it exists ( , If the threshold for high confidence in local decision-making is reached, a local early warning flag will be generated immediately.

[0100] like ( If the threshold for low confidence in local decision-making is set, then a flag for review is generated.

[0101] Step 9: Construct spatiotemporal sample blocks separately. Local early warning information and the feature vector of the event to be reviewed ;

[0102] For satisfying Event category, take

[0103]

[0104] The local early warning information is constructed as follows:

[0105]

[0106] For the first The types of events to be monitored corresponding to each deep neural network model; For the first The location of the event to be monitored corresponding to each deep neural network model; For the first The recognition confidence of the event to be monitored corresponding to each deep neural network model; This is the warning time;

[0107] Feature vector of the event to be reviewed Similarly, this includes: the first The type, location, warning time, and recognition confidence level of the event to be reviewed corresponding to each deep neural network model;

[0108] Step 10: Each edge node trains its own deep neural network model using stochastic gradient descent based on locally collected spatiotemporal data to obtain updated model weights. Uploaded to the central server, global model parameters are generated based on federated learning. ;

[0109] Under the federated learning framework, each edge node only uploads model parameter updates. Alternatively, gradient information is sent to the central server, rather than the raw data. The central server uses a federated averaging algorithm to aggregate the updates from each node and generate global model parameters. :

[0110]

[0111] in, The total number of edge nodes participating in federated learning. For the first The amount of local data at each edge node This represents the total amount of data across all nodes. For the first The updated model weights for each edge node;

[0112] Step 11: The central server will aggregate the global model. Distribute the data to each edge node and update the local model;

[0113] A closed-loop optimization mechanism of "local training - central aggregation - model distribution" is formed, enabling the model to continuously adapt to environmental changes in different regions and time periods, while ensuring data privacy and communication efficiency.

[0114] Step 12: Define a comprehensive link health index to achieve adaptive route switching and ensure reliable transmission of federated learning parameters and early warning data;

[0115] Fiber optic communication links may experience transmission quality degradation due to environmental disturbances (such as construction or geological disasters). Without proactive link health monitoring and adaptive routing switching, early warning information uploaded by edge nodes may be lost or severely delayed, causing the entire system to fail. Therefore, implementing adaptive routing switching is an essential engineering step to ensure the reliable arrival of federated learning parameters and early warning data at the center.

[0116] The comprehensive health index of the link is:

[0117]

[0118] in Indicates the link identifier. For signal-to-noise ratio, For bit error rate, For time delay, These are the weighting coefficients; This represents the maximum delay.

[0119] Adaptive route switching specifically refers to:

[0120] First, maintain a set of candidate routes for each data stream. ;

[0121] Indicates the primary (default) route. and This indicates two backup routes;

[0122] Then, when the health of the primary router is detected Less than the health threshold At that time, Within a given timeframe, based on the real-time health status of each backup router. Select the backup route with the highest health score and switch the data stream to it. .

[0123] Step 13: By aggregating local early warning information and feature vectors of events awaiting review from multiple reliably transmitted edge nodes through the central server, and performing multi-source fusion and arbitration, a global early warning decision is achieved.

[0124] Local early warnings from a single edge node may result in false alarms (e.g., a vibration event is mistakenly identified as an intrusion). However, by aggregating information from multiple nodes at the central point (including the perception results of neighboring nodes, video data, and meteorological and geological data) and performing multi-source fusion and arbitration, the accuracy of early warnings can be greatly improved, and unified global commands can be output. This is the ultimate value of achieving "wide-area intelligent monitoring".

[0125] First, receive and associate reliable transmissions. The local early warning information of each edge node and the feature vector of the event to be reviewed are used as input. :

[0126]

[0127] in, For the first The feature vector of the event to be reviewed corresponding to each deep neural network model. For the first The video data stream corresponding to each deep neural network model For the first External meteorological and geological data of each edge node;

[0128] Then, the central server based on the input dataset The information uploaded by each edge node is spatiotemporally correlated and clustered to form a set of events to be decided.

[0129] For any event S to be decided, its fusion confidence level is calculated using a weighted voting model. :

[0130]

[0131] in, For from Or to The perceived confidence level obtained from the secondary AI analysis. To use the central video analysis model to The confidence level of the video analysis is obtained. For based on And the historical and situational confidence scores for the matching degree of the historical event database at that location; , , For adaptive weights, and ;

[0132] Finally, the high and low confidence thresholds for global early warning decisions were set as follows: If the confidence level is fused Then confirm a global early warning for the event S to be decided; if If so, then the event S is marked as a suspected event requiring manual review; if If so, the event S to be decided is determined to be a false alarm and filtered out;

[0133] Global Alert Generation and Issuance: For confirmed global alerts, generate structured alert commands:

[0134] ;

[0135] Among them, the warning level according to Dynamically classify values ​​and event propagation speed; structured early warning instructions. Publish to designated terminals, display devices, and linkage systems.

[0136] Example:

[0137] This implementation case uses a risk event as a background and designs 5 sets of experiments; the experimental data comes from sensing and monitoring data of different scales in communities over a period of 6 months; the experimental groups and data details are shown in Table 1:

[0138] Table 1

[0139]

[0140] like Figure 3 As shown, the specific steps are as follows:

[0141] (1) Integrated inductive signal design and frequency domain planning:

[0142] Determine the allocation scheme of communication signals and sensing signals in spectrum resources, and set the total bandwidth to be [missing information]. ;

[0143] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): Communication Subband and perception subband ;

[0144] Experiment 2 (Mountain Slope Stability Monitoring): Communication Subband and perception subband ;

[0145] Experiment 3 (Health Monitoring of Urban Bridge Bearings): Communication Subband and perception subband ;

[0146] Experiment 4 (Important Perimeter Security Monitoring): Communication Subband and perception subband ;

[0147] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): Communication Subband and perception subband ;

[0148] (2) Adaptive fusion modulation and transmission:

[0149] Based on the determined spectrum resource allocation scheme and the communication quality index fed back from real-time channel assessment, and perceived signal-to-noise ratio requirements Adaptive selection of modulation format combination for integrated signal ( , );

[0150] Real-time feedback value:

[0151] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): , ;

[0152] Experiment 2 (Mountain Slope Stability Monitoring): , ;

[0153] Experiment 3 (Health Monitoring of Urban Bridge Bearings): , ;

[0154] Experiment 4 (Important Perimeter Security Monitoring): , ;

[0155] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): , ;

[0156] Adaptive modulation selection:

[0157] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): , ;

[0158] Experiment 2 (Mountain Slope Stability Monitoring): , ;

[0159] Experiment 3 (Health Monitoring of Urban Bridge Bearings): , ;

[0160] Experiment 4 (Important Perimeter Security Monitoring): , ;

[0161] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): , ;

[0162] in, The communication signal spectrum occupies the communication subband corresponding to each group of experiments. ,

[0163] For vibration sensing subcarrier set, For the set of temperature / strain sensing subcarriers:

[0164] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): , 15 GHz (Raman scattering effect senses temperature changes).

[0165] Experiment 2 (Mountain Slope Stability Monitoring): All Sensing Sub-bands were assigned to... 30 GHz, strain monitoring based on Brillouin scattering.

[0166] Experiment 3 (Health Monitoring of Urban Bridge Bearings): All Sensing Subbands are Assigned 25 GHz, strain monitoring based on Brillouin scattering.

[0167] Experiment 4 (Important Perimeter Security Monitoring): All Sensing Subbands are Assigned 15GHz, monitoring vibration.

[0168] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): 20GHz, 15GHz.

[0169] and These represent the complex amplitudes of the corresponding sensing subcarriers, derived from the values ​​of each experimental group. Power drive:

[0170] (3) Multi-physical quantity sensing and scattering signal demodulation:

[0171] Integrated signal reception The backscattered light signal generated by optical fiber transmission is converted into an electrical signal through photoelectric conversion. Separate the vibration response signal and temperature response signal

[0172] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): Vibration Response Signal The signal indicates that mechanical excavation (vibration event) occurred at a point 25.3 km along the pipeline; temperature change response signal. ,exist A leak occurred (local temperature drop).

[0173] Experiment 2 (Mountain Slope Stability Monitoring): At a point 10.5 km along the slope, slow creep occurred, causing cumulative strain in the optical fiber. Demodulation... .

[0174] Experiment 3 (Health Monitoring of Urban Bridge Bearings): A bearing underwent shear deformation due to load eccentricity.

[0175] Experiment 4 (Important Perimeter Security Monitoring): A person climbed over the fence at a distance of 1.2 km from the perimeter. The vibration signal was demodulated. Its waveform exhibits the typical characteristics of "impact-climbing-landing".

[0176] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridor): Overheating occurred at 15.7km of the cable due to joint aging. Meanwhile, mechanical construction (vibration) is underway at 16.1km along the corridor.

[0177] (4) Edge-side data preprocessing and calibration:

[0178] Temperature response signal Cross-sensitivity decoupling between temperature and strain is achieved by solving the following matrix equations:

[0179]

[0180] Vibration response signal Wavelet thresholding denoising and bandpass filtering are performed to obtain the enhanced signal. .

[0181] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): Vibration Signal Processing: Signal-to-Noise Ratio Improvement Amplified signal obtained Temperature signal calibration: Ambient temperature changes of less than 5℃ can be ignored; the Raman scattering anti-Stokes light intensity change can be directly read and calculated. .

[0182] Experiment 2 (Mountain Slope Stability Monitoring): Based on Brillouin Frequency Shift Calculate the strain. (Measured) According to the calibration formula (coefficient) The fiber strain was calculated using a constant of 0.05 MHz / με. =15 / 0.05=300με. Temperature change is negligible.

[0183] Experiment 3 (Health Monitoring of Urban Bridge Bearings): Strain readings were taken from the vertical section using fiber optic configurations deployed on the sides of the bearings. and strain in the 45° inclined section Calculate shear strain using formulas from mechanics of materials. ,have to Based on this, shear stress can be further calculated.

[0184] Experiment 4 (Important Perimeter Security Monitoring): After wavelet denoising and bandpass filtering, the signal-to-noise ratio is improved. of .

[0185] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): The system simultaneously demodulates the temperature signal. and vibration signals .

[0186] (5) Edge intelligent analysis and local rapid decision-making based on deep learning:

[0187] Preprocessed spatiotemporal data or after decoupling Divided into fixed-length spatiotemporal sample blocks ;Will Input a deep neural network model Output the event probability vector corresponding to this sample block. From vector Select the event corresponding to the highest probability value, which is the spatiotemporal sample block. For each deep neural network model, the monitored events are compared with the threshold, and the monitored events are marked as local early warning events or events to be reviewed.

[0188] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): Output Probability Vector ( ),judge Immediately generate a local intrusion alert. ,right Continuous monitoring.

[0189] Experiment 2 (Mountain Slope Stability Monitoring): Model analysis of the continuously increasing strain trend at 10.5km (cumulative strain over the past 24 hours) Output probability vector ( ),judge Generate local slope instability early warning.

[0190] Experiment 3 (Health Monitoring of Urban Bridge Bearings): The model identifies abnormal shear strain patterns and outputs a probability vector. ( ),judge Generate a local support health warning.

[0191] Experiment 4 (Important Perimeter Security Monitoring): The model will... The time-frequency features are compared with the behavior database. Output probability vector. ,( ),judge Instantly generate a local intrusion warning and trigger an audible and visual alarm.

[0192] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): Parallel processing by two edge nodes, outputting a probability vector. ( ),judge Generate a local overheat warning; ( ),judge Generate local construction vibration early warning.

[0193] (6) Optimization and knowledge update of edge-cloud collaborative model based on federated learning:

[0194] Experiment 1 (Long-Distance Pipeline Safety Monitoring): A total of 200 edge nodes (corresponding to 200 monitoring sections) were deployed along the pipeline. Each node locally stored historical vibration and temperature data for the past 30 days, with a sample size of approximately 5000 spatiotemporal sample blocks. Each node was trained locally using the SGD optimizer with a learning rate of 0.01, a batch size of 32, and 10 local iteration rounds. After local training, each node updated its model weights and uploaded them to the central server. The central server collected the updates from all 200 nodes and aggregated them using a federated averaging algorithm to generate global model parameters. The aggregated global model was then distributed to each node, replacing the original local model. After this round of federated learning, the event recognition accuracy of each node's model improved by an average of 3.2%. Furthermore, because only model parameters were uploaded instead of raw data, the privacy of sensitive data in each monitoring area was effectively protected.

[0195] Experiment 2 (Mountain Slope Stability Monitoring): Approximately 1000 sensing units along the route were divided into 50 edge nodes (one node per kilometer). Each node locally stored strain and vibration data for the past 60 days, with a sample size of approximately 8000 spatiotemporal sample blocks. Each node was trained locally using the SGD optimizer with a learning rate of 0.01, a batch size of 64, and 15 local iteration rounds. After local training, each node updated its model weights and uploaded them to the central server. The central server aggregated the updates from the 50 nodes to generate a global model. After aggregation, the average accuracy of each node in early identification of slow deformation events improved by 4.1%, and the model's adaptability to geological conditions was significantly enhanced.

[0196] Experiment 3 (Urban Bridge Bearing Health Monitoring): 2000 sensing points were divided into 40 edge nodes (one node every 250 meters). Each node locally stored strain and temperature data for the past 90 days, with a sample size of approximately 6000 spatiotemporal sample blocks. Each node was trained locally using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 12 local iteration rounds. After local training, each node updated its model weights and uploaded them to the central server. The central server aggregated the updates from the 40 nodes to generate a global model. After aggregation, the average accuracy of each node in calculating bearing shear deformation improved by 5.0%, indicating improved model generalization ability.

[0197] Experiment 4 (Important Perimeter Security Monitoring): A 5-kilometer perimeter was divided into 10 edge nodes (one node every 500 meters). Each node locally stored vibration and acoustic data from the past 15 days, with a sample size of approximately 3000 spatiotemporal sample blocks. Each node was trained locally using the SGD optimizer with a learning rate of 0.01, a batch size of 16, and 8 local iteration rounds. After local training, each node updated its model weights and uploaded them to the central server. The central server aggregated the updates from the 10 nodes to generate a global model. After aggregation, the average accuracy of each node in identifying intrusion behavior improved by 2.8%, and the false alarm rate decreased by 1.5%.

[0198] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): A 30-kilometer transmission line was divided into 15 edge nodes (one node every 2 kilometers). Each node locally stored temperature, stress, and vibration data for the past 45 days, with a sample size of approximately 7000 spatiotemporal data blocks. Each node was trained locally using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 10 local iteration rounds. After local training, each node updated its model weights and uploaded them to the central server. The central server aggregated the updates from the 15 nodes to generate a global model. After aggregation, the average accuracy of each node in identifying overheating events and external force damage events improved by 3.5%, and its robustness under strong electromagnetic interference environments was further improved.

[0199] (7) Network state awareness and uninterrupted reliable transmission:

[0200] Define a comprehensive health index for the link:

[0201]

[0202] in Represents link identifier, signal-to-noise ratio Bit error rate Delay , , These are the weighting coefficients;

[0203] Maintain a set of candidate routes for each data stream. ;

[0204] When detected hour( ,exist Within a given timeframe, based on the real-time status of each backup route Value, switch the data stream to .

[0205] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): The early warning data is stably transmitted back through the main router.

[0206] Experiment 2 (Mountain Slope Stability Monitoring): Data transmission is normal.

[0207] Experiment 3 (Health Monitoring of Urban Bridge Bearings): Warnings are uploaded.

[0208] Experiment 4 (Important Perimeter Security Monitoring): Data is transmitted back quickly.

[0209] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): The main route was damaged due to the excavation of optical cables during construction. ,exist Inside, the system switches to the backup route. ( This ensures that early warning information is not lost.

[0210] (8) Centralized multi-source fusion and global early warning decision-making:

[0211] Experiment 1 (Long-distance pipeline safety monitoring): The center receives early warnings and correlates them with video data of the area. (Excavator shown) and weather data (No abnormalities).

[0212] Experiment 2 (Mountain Slope Stability Monitoring): Central Correlation Geological Data (This area is prone to landslides) and recent rainfall data.

[0213] Experiment 3 (Urban Bridge Bearing Health Monitoring): Data from the centrally linked bridge health monitoring system.

[0214] Experiment 4 (Important Perimeter Security Monitoring): The center immediately retrieved the video footage from this location. confirm.

[0215] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): The center independently assessed the two events.

[0216] Fusion confidence calculation and arbitration: For any event to be decided, its fusion confidence score... Calculated using a weighted voting model:

[0217]

[0218] in, , , For adaptive weights;

[0219] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): , , (The construction in this area has been reported but is illegal and exceeds the permitted scope.)

[0220] Experiment 2 (Mountain Slope Stability Monitoring): , (The video has no visual distortion) .

[0221] Experiment 3 (Health Monitoring of Urban Bridge Bearings): , (Slight deviation during drone inspection) (This support has a historical record.)

[0222] Experiment 4 (Important Perimeter Security Monitoring): , (The intruder is clearly visible) .

[0223] 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;

[0224] Global alert generation and release:

[0225] Experiment 1 (Safety Monitoring of Long-Distance Pipelines): Confirm the overall early warning and issue instructions to stop illegal operations.

[0226] Experiment 2 (Mountain Slope Stability Monitoring): Events marked as suspected cases requiring manual review are subject to special investigation.

[0227] Experiment 3 (Health Monitoring of Urban Bridge Bearings): These are marked as suspected incidents requiring monitoring, prompting a request for enhanced investigation.

[0228] Experiment 4 (Important Perimeter Security Monitoring): Confirm the overall security alert and notify security personnel to take action.

[0229] Experiment 5 (Comprehensive Monitoring of High-Voltage Cable Corridors): Confirm the warning and dispatch the power company for maintenance; Confirm the warning and notify patrol personnel to stop it.

[0230] To verify the sensing response rate, comprehensive accuracy, and all-time availability of this invention for monitoring and early warning of multiple physical quantities in a wide-area environment, five sets of fiber optic monitoring and early warning experiments were conducted. The experimental results are as follows: Figure 3 As shown. By Figure 3 As can be seen, the wide-area environmental intelligent monitoring and early warning system based on fiber optic sensing integration and federated learning established in this invention achieves a perception response rate, comprehensive accuracy, and all-time availability of over 99%, demonstrating high accuracy while maintaining stability and good monitoring and early warning effects. This indicates that the wide-area environmental multi-physical quantity intelligent monitoring and early warning system established in this invention is effective, providing a better method for establishing accurate wide-area environmental intelligent monitoring and early warning models, and possesses certain practical value.

Claims

1. A method for wide-area environment intelligent monitoring and early warning based on fiber sensing integration and federated learning, characterized in that, Includes the following steps: Step 1: Build a communication scenario for wide-area environmental intelligent monitoring based on long-distance fiber optic links, and divide the communication signals and sensor signals in the frequency domain on the spectrum resources; Step 2: Adaptively generate the communication quality index and perceived signal-to-noise ratio requirements based on the real-time channel assessment feedback. Integrated transmission signal in time slots The backscattered light signal generated by the optical fiber transmission is converted into an electrical signal through photoelectric conversion. ; Step 3: Further separate the vibration response signal from the electrical signal. and temperature response signal Preprocessing and calibration are performed to obtain enhanced spatiotemporal data. and decoupling physical quantities ; Step 4: Process the preprocessed spatiotemporal data and decoupled Spatiotemporal slicing is performed according to a preset time window / spatial window, and temporal alignment and multi-channel feature organization are completed to continuously generate multiple fixed-length spatiotemporal sample blocks. ; Step 5: Deploy a distributed data acquisition and processing unit in each of the different fiber optic monitoring areas or access locations. Each unit serves as an edge node, and a deep neural network model with the same structure but independently updated parameters is deployed on each edge node. ; Step six, the spatiotemporal sample block The input depth neural network model is input one by one, the monitored events of the spatiotemporal sample in each depth neural network model are obtained according to the output event probability, and after being judged with a threshold, the monitored events are marked as local early warning events or events to be reviewed. Specifically: spatiotemporal sample blocks Enter the first In a deep neural network model with edge nodes, spatiotemporal features are automatically extracted, and event probability vectors are output. From vector Select the event corresponding to the highest probability value, which is the spatiotemporal sample block. In the The events to be monitored corresponding to each edge node; Event probability vector ;in, For background event probabilities, Corresponding to The pre-defined probability of identifying the event to be monitored; Then, the probability of the event to be monitored Compare with the threshold: If there is then generate a local warning flag immediately; if then generate a pending review flag. , is a high confidence threshold for local decisions; is a low confidence threshold for local decisions; Step seven, constructing spatio-temporal sample blocks respectively local early warning information and feature vectors of events to be reviewed ; Step 8: Each edge node trains its own deep neural network model based on locally collected spatiotemporal data, obtains updated model weights, and uploads them to the central server. Global model parameters are then generated based on federated learning. And distribute it to each edge node to update the local model; Step 9: Define a comprehensive link health index to achieve adaptive route switching and ensure reliable transmission of federated learning parameters and early warning data; Step 10: The central server aggregates local early warning information and feature vectors of events awaiting review from multiple reliable edge nodes, performs multi-source fusion and arbitration, and then realizes global early warning decision-making.

2. The method of claim 1, wherein, In step one, the total bandwidth Divided into communication subbands and perception subband And satisfy: , .

3. The method of claim 1, wherein, The step two integrated transmit signal The frequency domain expression is: in, For communication signal spectrum, For vibration sensing subcarrier set, For temperature / strain sensing subcarrier set, and These are the complex amplitudes of the corresponding sensing subcarriers. It is the Dirac function. Indicates frequency Discrete spectral lines at that location, For frequency variables.

4. The method as described in claim 3, characterized in that, In step two, the electrical signal is calculated as follows: in, This represents the complex amplitude of the emitted light field. Represents convolution. It is a function characterizing the distributed time-varying impulse response of optical fibers. Indicates position Due to changes in physical quantities The resulting fiber refractive index response function, It is noise.

5. The method as described in claim 4, characterized in that, In step three, matched filtering, frequency domain segmentation, and coherent demodulation algorithms are used to analyze the electrical signal. Separate from the corresponding and Vibration response signal and temperature response signal .

6. The method of claim 5, wherein, In step three, the temperature change response signal Cross-sensitivity decoupling between temperature and strain is achieved by solving the following matrix equations: in, The change in Brillouin divergence frequency shift, This represents the change in optical power. The decoupled physical quantities are obtained from the calibrated sensitivity coefficient matrix. ; To the vibration response signal Wavelet threshold denoising and band-pass filtering are performed to obtain an enhanced signal .

7. The method of claim 1, wherein, In step seven, the local early warning information is calculated as follows: The first edge node corresponds to the to-be-monitored event type: ;​ For the first The location of the event to be monitored corresponding to each edge node; For the first The confidence level of the event to be monitored corresponding to each edge node: ; This is the warning time; Feature vector of the event to be reviewed Similarly, this includes: the first The type, location, warning time, and identification confidence level of the events to be reviewed for each edge node.

8. The method of claim 1, wherein, In step eight, in, The total number of edge nodes participating in federated learning. For the first The amount of local data at each edge node This represents the total amount of data across all nodes. For the first The weights of the deep neural network model after updating the edge nodes.

9. The method as described in claim 1, characterized in that, In step nine, the comprehensive health index of the link is: wherein represents a link identifier, is a signal-to-noise ratio, is a bit error rate, is a time delay, is a weight coefficient; is a maximum time delay; Adaptive route switching specifically refers to: First, a candidate route set is maintained for each data flow ; denotes a primary route, and denotes two backup routes; Then, when the health of the primary router is detected Less than the health threshold At that time, Within a given timeframe, based on the real-time health status of each backup router. The system selects the backup route with the highest health score and switches the data stream accordingly.

10. The method of claim 7, wherein, In step ten First, receive and associate reliable transmissions. The local early warning information of each edge node and the feature vector of the event to be reviewed are used as input. : in, For the first Feature vectors of events to be verified for each edge node. For the first The video data stream corresponding to each edge node For the first External meteorological and geological data of each edge node; Then, the central server based on the input dataset The information uploaded by each edge node is spatiotemporally correlated and clustered to form a set of events to be decided. For any event S to be decided, its fusion confidence level is calculated using a weighted voting model. : in, For from Or to The perceived confidence level obtained from the secondary AI analysis. In order to pass through The confidence level of the video analysis is obtained. For based on And the historical and situational confidence scores for the matching degree of the historical event database at that location; , , For adaptive weights; Finally, the high and low confidence thresholds of the global early warning decision are set as ; if the fusion confidence , the event S to be decided is confirmed as a global early warning; if , the event S to be decided is marked as a suspected event requiring manual review; if , the event S to be decided is determined as a false alarm and filtered. Global Alert Generation and Issuance: For confirmed global alerts, generate structured alert commands: ； wherein the early warning level According to value and event diffusion speed dynamic division; structured early warning instructions are issued to designated terminals, display devices and linkage systems.

Images