Distributed optical fiber sensing system based on model fusion recognition algorithm

Abstract

The invention discloses a distributed optical fiber sensing system based on a model fusion recognition algorithm, and aims to solve the technical problems of insufficient external damage prevention early warning, low fault positioning efficiency, poor interference recognition accuracy and the like in operation and maintenance of infrastructures such as urban underground cables, oil pipelines, urban pipe galleries and the like. Through innovative combination of a dendritic optical network access framework, a spatial-temporal feature reconstruction technology and a multi-model fusion recognition algorithm, accurate sensing, interference distinguishing and anomaly positioning of vibration signals in a monitoring area are realized. The early warning effective rate reaches 85% or above, the third-party damage identification AUC value is 0.9667, the positioning precision reaches the centimeter level, the method can be widely applied to infrastructure operation and maintenance management in the fields of electric power, oil gas, municipal administration and the like, and the operation and maintenance efficiency and the safety guarantee capacity are remarkably improved.

Description

Technical Field

[0001] This invention relates to the intersection of fiber optic sensing technology, intelligent identification technology and digital twin technology, and specifically to a distributed fiber optic sensing system based on a model fusion identification algorithm. Background Technology

[0002] With the acceleration of urbanization, the scale of underground infrastructure such as power lines, oil pipelines, and urban utility tunnels in urban core areas continues to expand. Currently, the undergrounding rate of power lines in urban core areas has reached over 95%, posing a severe challenge to the safe operation of thousands of kilometers of underground cables. There are several technical pain points in the current operation and maintenance management of underground infrastructure: First, the technology for preventing external damage is limited. Traditional methods such as construction unit reporting, on-site inspections, and key area control are insufficient for real-time early warning and accurate location of external damage to underground facilities, leading to frequent accidents such as large-scale power outages and pipeline leaks caused by construction excavation and heavy vehicle traffic. Second, fault location efficiency is low. Current fault finding mainly relies on segmented testing of faulty lines, which is inefficient, results in frequent power outages, and is difficult to investigate. There is a lack of effective technical means to quickly locate fault points. Third, the accuracy of interference identification is poor. The environment along underground infrastructure is complex, with strong interference signals such as traffic horns, factory noise, and vehicle start-stop. Traditional feature extraction schemes of sensor systems are unable to effectively distinguish between third-party damage and strong interference events, resulting in high false alarm and false negative rates. Fourth, the level of informatization in operation and maintenance management is insufficient. Existing data collection and information management systems are difficult to integrate with existing business systems, lack visualization technology support, and the collected data cannot provide effective support for operation and maintenance decisions.

[0003] To address these issues, existing technologies have developed monitoring solutions based on distributed optical fiber sensing, such as using OTDR (Optical Time Domain Reflectometer) technology for fiber loss and fault detection, or classifying vibration signals using a single machine learning algorithm. However, these technologies still have several shortcomings: traditional OTDR technologies often employ linear monitoring architectures, resulting in blind spots and an inability to achieve comprehensive coverage of branching pipe networks; feature extraction relies solely on a single dimension, either time or frequency domain, without fully integrating geographic and geological information, leading to insufficient interference discrimination capabilities in complex environments; single recognition models have learning blind spots, making it difficult to adapt to diverse vibration event characteristics, resulting in limited recognition accuracy and generalization ability; and low visualization levels prevent deep integration of monitoring data and geographic information, making it difficult for maintenance personnel to intuitively grasp the facility's operational status.

[0004] Therefore, developing a distributed fiber optic sensing system with branched coverage, high anti-interference capability, high recognition accuracy, and panoramic visualization has become a key technological requirement for solving the pain points of underground infrastructure operation and maintenance management. It is of great significance for improving the level of infrastructure safety operation and reducing operation and maintenance costs. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a distributed optical fiber sensing system based on model fusion recognition algorithm. Through hardware architecture innovation, feature extraction optimization and recognition algorithm upgrade, it can realize real-time early warning, accurate positioning and effective differentiation of interference of external force damage to underground infrastructure, and improve the intelligence and visualization level of operation and maintenance management.

[0006] This invention provides the following technical solution: a distributed optical fiber sensing system based on a model fusion recognition algorithm, comprising a laser module, a laser modulation module, a sensing and transmission module, a signal receiving module, a data processing module, an intelligent recognition module, and a visualization management module. The laser module provides an ultra-narrow linewidth optical signal; the laser modulation module modulates continuous laser light into a pulse signal with a set pulse width and amplifies it; the sensing and transmission module uses a branched topology G652 single-mode optical fiber, converting external vibrations into optical signal phase changes based on the photoelastic effect; the signal receiving module converts backscattered Rayleigh light signals into electrical signals; the data processing module optimizes signal characteristics through spatiotemporal feature reconstruction technology; the intelligent recognition module uses a multi-model fusion algorithm to classify and identify vibration events; and the visualization management module displays monitoring data and early warning information based on digital twin technology. All modules interact with each other via an Ethernet interface and the MQTT protocol.

[0007] As a preferred embodiment of the present invention, the center wavelength of the laser module is 1550nm, the linewidth is ≤3kHz, the laser modulation module includes an acousto-optic modulator, an amplifier and a circulator, the width of the modulated pulse signal is 486ns, and the amplified optical signal is connected to the sensing and transmission module through the circulator.

[0008] As a preferred embodiment of the present invention, the sensing and transmission module uses a 1×8 MEMS optical switch to construct a branched topology network, including one 24-hour all-weather monitoring main channel and three rotating monitoring branch lines. The optical switch response time is ≤10ms, supports automatic channel switching, and the maximum monitoring length of a single device is up to 30 kilometers.

[0009] As a preferred embodiment of the present invention, the spatiotemporal feature reconstruction technology of the data processing module includes: based on the statistical regularity of 5-10 occasional spikes within 120 seconds, a truncated differential feature reconstruction is adopted, with the tenth largest value of the waveform as the upper limit to weaken the impact interference; a basic noise threshold is set, and signals below the threshold are uniformly set to zero; the geographic and geological information of the monitoring area is integrated, and the geographic correlation information of the defense zone is obtained through graph embedding learning to construct a spatiotemporal fusion feature set.

[0010] As a preferred embodiment of the present invention, the multi-model fusion algorithm of the intelligent recognition module includes a base learning layer and a decision layer; the base learning layer integrates random forest, GBDT, XGBoost, LightGBM and at least one deep learning model to learn different dimensional features in the spatiotemporal fusion feature set respectively; the decision layer adopts a fully connected cascaded neural network, and fuses the output of the base learning layer through the tanh activation function and the linear summation function to output the vibration event classification result, which includes third-party damage events, strong interference events and no interference events.

[0011] As a preferred embodiment of the present invention, the signal receiving module includes a photodetector and an amplifier. The photodetector converts the backscattered Rayleigh light signal into an electrical signal. The amplification factor of the amplifier is adaptively adjusted according to the signal strength. The converted electrical signal is input into the data processing module after A / D conversion. The sampling frequency is dynamically adjusted according to the monitoring distance.

[0012] As a preferred embodiment of the present invention, the visualization management module includes a GIS geographic information module, a digital twin modeling module, an early warning push module, and a report statistics module; the GIS geographic information module supports switching between satellite imagery and electronic maps, the digital twin modeling module realizes three-dimensional visualization and interactive operation of cable channels and cable wells, the early warning push module pushes abnormal information through sound and light alarms and SMS notifications, and the report statistics module supports historical event tracing and data statistical analysis.

[0013] As a preferred embodiment of the present invention, the laser modulation module further includes a laser pulse width automatic adjustment unit, which consists of a detector, a data acquisition card, a host computer algorithm, and a laser pulse width controller forming a feedback adjustment network. The laser pulse width is dynamically adjusted according to the signal strength of the sensing and transmission modules to suppress signal saturation caused by strong light interference.

[0014] As a preferred embodiment of the present invention, the model training sample library of the intelligent identification module includes 15,000 dangerous intrusion samples, 15,000 strong interference samples and 15,000 non-interference samples. AUC is used as the model evaluation index, the model identification AUC value is ≥0.96, and the false alarm rate of third-party sabotage events is <5%.

[0015] As a preferred embodiment of the present invention, the system has a positioning accuracy of ≤±10cm, an event blind zone of ≤10cm, an attenuation blind zone of ≤40cm, a dynamic range of ≥10dB, an operating temperature range of -20℃ to +70℃, and supports data interface with power distribution automation systems, PMS systems, and SCADA systems.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) More comprehensive monitoring coverage: The branch topology architecture and 1×8 MEMS optical switch are adopted to realize the monitoring network layout of 1 main channel + 3 branch lines, which solves the blind spot problem of traditional linear monitoring. A single device can cover 12 kilometers of cable network and supports automatic switching of multiple channels to meet the comprehensive monitoring needs of complex pipeline networks.

[0017] (2) Higher positioning accuracy: Based on Φ-OTDR technology and laser pulse width optimization design, the event blind zone is ≤10cm, the attenuation blind zone is ≤40cm, and the positioning accuracy reaches ±10cm, realizing the accurate positioning of third-party damage events and providing accurate basis for maintenance personnel to quickly handle the situation.

[0018] (3) Stronger recognition accuracy: By integrating geographic information and signal features through spatiotemporal feature reconstruction technology and combining multi-model fusion recognition algorithm, the AUC value reaches 0.9667, which is more than 28% higher than the traditional single model. The false alarm rate in strong interference scenarios in factories is <5%, effectively distinguishing between third-party damage and strong interference events such as traffic flow, horn honking, and factory noise.

[0019] (4) More timely early warning response: The system's response time from vibration sensing to early warning push is ≤3 seconds, and the early warning effectiveness rate is over 85%, realizing the transformation from passive defense to active early warning and significantly reducing accident losses caused by external force damage.

[0020] (5) Smarter operation and maintenance management: Relying on digital twin technology, a panoramic visualization platform is built, and the data channel with the existing PMS and SCADA systems is opened up to realize the integrated management of monitoring data, equipment ledgers and geographic information, support fault tracing, operation and maintenance decision-making and process closure, and improve operation and maintenance efficiency by more than 30%.

[0021] (6) Wider range of applications: It is not only applicable to the monitoring of external damage to urban underground cables, but can also be extended to the operation and maintenance management of infrastructure such as oil pipelines and urban utility tunnels. It has good compatibility and scalability, and is applicable to a temperature range of -20℃ to +70℃, meeting the needs of use in different environments. Attached Figure Description

[0022] Figure 1 is a schematic diagram of the technical principle of the present invention; Figure 2 is a schematic diagram of the working principle of the OTDR of the present invention; Figure 3 shows the OTDR detection curve of the present invention; Figure 4 is a schematic diagram of the optical path for monitoring the branch-type optical cable of the present invention; Figure 5 is a schematic diagram of the detection principle of the vibration sensor of the present invention; Figure 6 is a detailed waveform comparison diagram of third-party sabotage events and occasional spikes in this invention; Figure 7 is a comparison chart of the peak situations of various sporadic peak events in this invention; Figure 8 is a comparison diagram of the distribution of the differential features of the present invention; Figure 9 is a schematic diagram of the multi-model fusion technology process of the present invention; Figure 10 is a comparison chart of the ROC curve and AUC area of ​​the present invention; Figure 11 is a comparison chart of the AUC index of the multi-model fusion technology of the present invention with other solutions; Figure 12 is a schematic diagram of the model recognition effect when the present invention is damaged by a third party; Figure 13 shows the recognition effect of the strong interference event model caused by the factory in this invention.

[0023] The above figures are only used to illustrate the structural principles of the present invention and do not constitute a limitation on the present invention. The proportions of each component in the figures can be adjusted according to actual production conditions. Detailed Implementation

[0024] To make the technical problems solved by the present invention, the technical solutions adopted, and the technical effects achieved clearer, the technical solutions of the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] A distributed fiber optic sensing system based on a model fusion recognition algorithm includes a laser module, a laser modulation module, a sensing and transmission module, a signal receiving module, a data processing module, an intelligent recognition module, and a visualization management module. Each module interacts with the MQTT message protocol via an Ethernet interface to form a closed-loop process of "perception-processing-recognition-control".

[0026] Laser module: As the system's light source, it provides an ultra-narrow linewidth optical signal with a center wavelength of 1550nm and a linewidth of approximately 3kHz, ensuring the detection sensitivity and stability of the sensing system. This wavelength features low transmission loss and strong anti-interference capabilities, making it suitable for long-distance monitoring scenarios.

[0027] The laser modulation module consists of an acousto-optic modulator, an amplifier, a circulator, and an automatic laser pulse width adjustment unit. The acousto-optic modulator modulates the continuous laser output from the laser module into a pulse signal with a pulse width of 486 ns. The amplifier amplifies the power of the pulse signal, and the circulator ensures unidirectional transmission of the optical signal, guaranteeing that the modulated pulsed light enters the sensing and transmission module unidirectionally, while preventing reflected light from interfering with the light source. The automatic laser pulse width adjustment unit, composed of a detector, a data acquisition card, a host computer algorithm, and a laser pulse width controller, forms a feedback adjustment network. It dynamically adjusts the laser pulse width based on the signal strength fed back from the sensing and transmission module, suppressing signal saturation caused by strong light interference and ensuring signal quality under different monitoring distances and environments.

[0028] Sensing and Transmission Module: Employing a branched topology architecture, this module integrates sensing and transmission functions based on standard G652 single-mode fiber. It constructs a monitoring network using a 1×8 MEMS optical switch, comprising one 24 / 7 main monitoring channel and three rotating monitoring branches. The optical switch response time is ≤10ms, supporting automatic channel switching, and the maximum monitoring length for a single device reaches 30 kilometers. When external vibrations act on the fiber, the photoelastic effect causes changes in the fiber's length and refractive index, leading to a phase change in the optical signal. Backscattered Rayleigh light carries this phase change information back along the fiber, enabling the sensing and transmission of vibration signals.

[0029] The signal receiving module, consisting of a photodetector and an amplifier, is responsible for converting the backscattered Rayleigh light signal into an electrical signal and performing signal amplification and filtering. The photodetector uses a high-sensitivity avalanche photodiode to ensure effective conversion of weak light signals. The amplifier's amplification factor is adaptively adjusted according to the signal strength. A low-pass filter is used to remove high-frequency noise. The converted electrical signal is then input to the data processing module after A / D conversion. The sampling frequency is dynamically adjusted according to the monitoring distance to ensure signal integrity and real-time performance.

[0030] Data Processing Module: The core employs spatiotemporal feature reconstruction technology to optimize the extraction of vibration signal features. This technology comprises three parts: First, truncated differential feature reconstruction, based on the statistical regularity of 5-10 occasional spikes within 120 seconds, uses the tenth largest value of the waveform as the upper limit, uniformly weakening any excess to the upper limit value, effectively avoiding the impact of impactful interference events; Second, background noise filtering, setting a basic background noise threshold, uniformly setting signals below the threshold to zero, reducing interference from low and weak noise; Third, geographic information fusion, acquiring geographic association information of the monitoring area through graph embedding learning, fusing geographic geological data (such as the distribution of highways, suburbs, and factories) with time-domain and frequency-domain features to construct a spatiotemporal fusion feature set, providing high-quality data support for subsequent identification.

[0031] The intelligent recognition module employs a multi-model fusion algorithm, including a base learning layer and a decision layer. The base learning layer integrates Random Forest (RF), GBDT, XGBoost, LightGBM, and deep learning models to learn different dimensions of features from the spatiotemporal fusion feature set. Random Forest and GBDT focus on statistical feature learning, XGBoost and LightGBM enhance feature weight optimization capabilities, and the deep learning model captures complex nonlinear features. The decision layer uses a fully connected cascaded neural network (FCC-Net), which performs a nonlinear transformation on the outputs of the first four base models using the tanh activation function. Finally, it fuses the outputs of all base models using a linear summation function to obtain the final vibration event classification result, which includes third-party sabotage events, strong interference events, and no interference events. The model training sample library for this module includes 15,000 dangerous intrusion samples, 15,000 strong interference samples, and 15,000 no interference samples. AUC is used as the model evaluation metric; the model's AUC value is ≥0.96, and the false alarm rate for third-party sabotage events is <5%.

[0032] Visual Management Module: Based on digital twin technology and GIS (Geographic Information System), a panoramic visual management platform is constructed, including a GIS module, a digital twin modeling module, an early warning push module, and a report statistics module. The GIS module supports switching between satellite imagery and electronic maps, enabling accurate display of geographic information in the monitored area. The digital twin modeling module enables 3D visual modeling of facilities such as cable channels, cable wells, and pipelines, supporting interactive operations such as rotation, zooming, and simulated excavation, and displaying the status of key monitoring areas through heat maps. The early warning push module immediately triggers audible and visual alarms upon detecting third-party sabotage events and pushes abnormal information, including event type, latitude and longitude coordinates, and defense zone location, to maintenance personnel via SMS. The report statistics module supports historical event tracing, data statistical analysis, and report generation, providing data support for maintenance decision-making.

[0033] Example 1 System hardware deployment In this embodiment, the hardware deployment of the distributed optical fiber sensing system based on the model fusion recognition algorithm includes the following steps: Laser module selection and installation: A narrow linewidth laser with a center wavelength of 1550nm and a linewidth of 2.8kHz was selected and installed in the equipment room of the center's office building. The working environment temperature was kept between 15℃ and 25℃ to avoid direct sunlight and vibration interference.

[0034] Laser modulation module assembly: The acousto-optic modulator, erbium-doped fiber amplifier (EDFA), circulator, and automatic laser pulse width adjustment unit are assembled into one unit. The modulation frequency of the acousto-optic modulator is set to 1MHz, and the output pulse width is 486ns. The output power of the EDFA amplifier is adjusted to 15dBm. Port 1 of the circulator is connected to the output of the laser module, port 2 is connected to the input of the EDFA amplifier, and port 3 is connected to the optical switch input of the sensing and conduction module. The detector of the automatic laser pulse width adjustment unit is an InGaAs photodetector, and a 16-bit high-speed acquisition card is used. The sampling frequency is set to 100MHz. The dynamic adjustment of the laser pulse width is achieved through a host computer algorithm, with an adjustment range of 100ns to 1μs.

[0035] Sensing and conduction module deployment: G652 single-mode optical fiber is used as the sensing fiber, laid parallel to the underground cable in the same trench. The fiber optic connectors are fusion spliced, with a splice loss ≤0.1dB. A branched monitoring network is constructed using 1×8 MEMS optical switches. The optical switches are installed inside the cable manhole and encapsulated in a waterproof and moisture-proof shell. The input port of the optical switch is connected to the output of the circulator, and the 8 output ports are respectively connected to 1 main channel optical fiber and 7 branch optical fibers (4 are selected in this embodiment, including 1 main channel and 3 branch fibers). The main channel optical fiber covers the core monitoring area with a length of 5 kilometers, and the branch optical fibers cover different zones with a length of 2-3 kilometers each. The total monitoring length of a single device is 12 kilometers.

[0036] Signal receiving module installation: The photodetector of the signal receiving module is an avalanche photodiode (APD) with a response wavelength range of 1520nm~1580nm and a responsivity ≥0.8A / W. The signal amplifier is a low-noise operational amplifier with an adjustable amplification factor range of 10 to 100 times. The filter cutoff frequency is set to 50kHz. The signal receiving module and the laser modulation module are installed in the same equipment room and connected to the feedback port of the optical switch via fiber optic patch cords.

[0037] Deployment of the data processing and intelligent recognition module: An industrial control computer is used as the hardware carrier for the data processing and intelligent recognition module. It is equipped with an Intel Core i7 processor, 32GB of memory, and a 1TB solid-state drive. The Ubuntu 20.04 operating system and Python 3.8 programming environment are installed. Algorithm libraries such as Scikit-learn and TensorFlow are integrated to realize the operation of spatiotemporal feature reconstruction and multi-model fusion algorithms.

[0038] Visualization Management Module Deployment: The visualization management module uses a rack-mount server configured with dual Intel Xeon processors, 64GB of memory, and a 4TB hard drive. It is equipped with the Windows Server 2019 operating system, and deploys Mapbox GIS services and digital twin modeling software. It interacts with the data processing module via an Ethernet interface. The operation and maintenance terminal accesses the visualization platform through a browser, supporting access from both PC and mobile devices.

[0039] Example 2 System software implementation In this embodiment, the system software includes data acquisition software, signal processing software, intelligent recognition software, and visual management software, which are implemented as follows: Data acquisition software: Developed in C++, employing multi-threaded programming technology, it enables parameter configuration and data acquisition for the laser module, laser modulation module, and signal receiving module. The software supports remote configuration of parameters such as laser power, pulse width, and sampling frequency. Acquired data is stored in binary format and transmitted to the data processing module in real time, with an acquisition latency of ≤1ms. Data transmission uses the TCP / IP protocol to ensure data integrity.

[0040] Signal processing software: Developed using Python, its core function is spatiotemporal feature reconstruction. First, the acquired raw electrical signals are preprocessed, including detrending, filtering, and normalization. Then, a truncated differential feature reconstruction algorithm is used to extract the tenth largest value of the waveform as an upper limit, truncating signals exceeding this limit. Next, a basic noise floor threshold is set (50 in this embodiment), and signals below this threshold are set to zero. Finally, a graph embedding learning algorithm is used to fuse geographic information data of the monitoring area. This geographic information data is imported through a GIS system and includes parameters such as zone type (highway, suburbs, factory, school), latitude and longitude coordinates, and distance. A spatiotemporal fusion feature set is constructed, comprising 12 dimensions including energy value, maximum value, variance value, sequence sum, and geographic correlation coefficient.

[0041] Intelligent recognition software: Developed using Python and the TensorFlow framework, this software implements a multi-model fusion recognition algorithm. The base learning layer uses a random forest model with 100 decision trees and a maximum depth of 10; the GBDT model has a learning rate of 0.1 and 200 iterations; the XGBoost model has a tree depth of 8 and a learning rate of 0.05; and the LightGBM model has 31 leaf nodes and a learning rate of 0.08. The deep learning model employs a 3-layer convolutional neural network (CNN), with 12-dimensional features in the input layer, 64 and 32 hidden nodes respectively, and 3 classes in the output layer (third-party interference, strong interference, and no interference). The fully connected cascaded neural network in the decision layer consists of 5 input neurons, 4 hidden neurons (using the tanh activation function), and 1 output neuron (linear summation function). The final classification result is obtained by fusing the outputs of the base learning layer. The model training uses cross-validation, with a training set to test set ratio of 7:3, 500 training iterations, and a batch size of 64. After training, the model is embedded into the intelligent recognition module, supporting online updates and parameter tuning.

[0042] Visual management software: Developed based on the Vue.js framework and Mapbox GIS services, it realizes a digital twin panoramic control interface. The software includes a map display module, an equipment management module, a real-time monitoring module, an early warning and alarm module, a historical query module, and a report statistics module. The map display module supports switching between satellite imagery and electronic maps, and allows zooming, panning, and rotation, displaying information such as fiber optic routes, cable well locations, and equipment distribution in the monitored area. The equipment management module displays the basic attributes, operating status, and maintenance records of devices such as optical switches and sensors. The real-time monitoring module displays vibration signal intensity and characteristic value changes in the form of waveform graphs and heat maps, supporting real-time updates (refresh frequency 1 second / time). When a third-party sabotage event is detected, the early warning and alarm module pops up an alarm window, displaying information such as event type, latitude and longitude, defense zone, and occurrence time, while simultaneously triggering an audible and visual alarm and SMS notification. The SMS notification is sent to the mobile phone of designated maintenance personnel via the GSM module. The historical query module supports querying historical events by time, event type, defense zone, and other conditions, displaying the waveform data, identification results, and handling records of the events. The report statistics module supports generating daily, weekly, and monthly reports, statistically analyzing indicators such as the number of events, early warning accuracy, and timely handling rate. Reports can be exported to Excel format.

[0043] Example 3 System Workflow In this embodiment, the system's workflow includes the following steps: System initialization: The laser module, laser modulation module, signal receiving module, data processing module, intelligent identification module, and visual management module are started. After each module completes its self-test, a communication connection is established through the MQTT protocol. The laser module outputs continuous laser light, and the laser modulation module modulates the laser light into a 486ns pulse signal. After being amplified by an amplifier, the signal is input to a 1×8 MEMS optical switch through a circulator. The optical switch switches to the main channel by default, and the system enters the standby monitoring state.

[0044] Vibration sensing and signal transmission: When external forces damage the monitoring area (such as digging or crushing) or cause interference (such as traffic or honking), vibration acts on the sensing optical fiber. Through the photoelastic effect, the fiber length and refractive index change, which in turn causes a phase change in the optical signal. The backscattered Rayleigh light carries the phase change information back along the optical fiber and is input to the signal receiving module via the circulator.

[0045] Signal conversion and preprocessing: The photodetector of the signal receiving module converts the backscattered Rayleigh light signal into an electrical signal. The signal amplifier amplifies the electrical signal (the amplification factor is automatically adjusted to 20 times according to the signal strength). The low-pass filter removes high-frequency noise (cutoff frequency 50kHz). After A / D conversion (sampling frequency 100MHz), the digital signal is transmitted to the data processing module.

[0046] Spatiotemporal feature reconstruction: The data processing module performs spatiotemporal feature reconstruction on the digital signal. First, it extracts the energy value, maximum value, variance value, sequence and other time-domain features of the original signal. Then, it uses a truncated difference feature reconstruction algorithm with the tenth largest value of the waveform as the upper limit (the upper limit value in this embodiment is 2000). Signals exceeding the upper limit are truncated. Then, signals below the noise floor threshold (50) are set to zero. Finally, the geographical information of the monitoring area is fused (e.g., the area is along a highway, and the geographical correlation coefficient is marked as 0.8) to construct a spatiotemporal fusion feature set, which is then transmitted to the intelligent recognition module.

[0047] Multi-model fusion recognition: The base learning layer of the intelligent recognition module processes the spatiotemporal fusion feature set in parallel. Random forest, GBDT, XGBoost, LightGBM and CNN models output classification probabilities respectively. The fully connected cascaded neural network of the decision layer fuses the classification probabilities. When the probability of a third-party sabotage event after fusion is ≥0.8, it is judged as a third-party sabotage event; when the probability of a strong interference event is ≥0.8, it is judged as a strong interference event; otherwise, it is judged as a non-interference event. The recognition result is transmitted to the visualization management module.

[0048] Early warning and visualization: If the event is determined to be a third-party sabotage incident, the visualization management module immediately triggers an audible and visual alarm (alarm sound intensity 80dB, alarm light flashing frequency 2 times / second), and simultaneously sends an SMS notification to maintenance personnel via the GSM module (content: "A third-party sabotage incident has occurred in XX defense zone (latitude and longitude: 38.0°N, 114.5°E), please handle it promptly"). The digital twin interface marks the event location with a red heat map and displays the event waveform data and identification results. If the event is determined to be a strong interference incident, the visualization management module only records the event information in the background and does not trigger an alarm. If the event is determined to be a non-interference incident, the system continues to maintain monitoring status.

[0049] Operation and maintenance handling and recording: After receiving the warning notification, the operation and maintenance personnel can view the event details through the visual management module, go to the site to handle the situation, and enter the handling record (including handling time, handling personnel, and handling result) into the system after the handling is completed. The system will automatically update the event status to "handled" and store the relevant data in the database to support subsequent query and statistics.

[0050] Channel rotation: During the main channel monitoring, the optical switch rotates through the three branch lines at a cycle of 5 minutes. During each rotation, the optical switch switches to the branch line channel for 10 seconds of monitoring. After the monitoring is completed, it switches back to the main channel, achieving full-area monitoring without blind spots.

[0051] Example 4 System performance testing In this embodiment, system performance testing was conducted in the underground cable area of ​​Minjiang Road in Shijiazhuang. The test scenarios included third-party damage simulation (excavator knocking on the road), strong interference simulation (heavy truck start-stop, factory noise), and no-interference scenario. The test results are as follows: Positioning accuracy test: Excavator knocking was simulated at distances of 3 km, 5 km, and 8 km from the optical switch. The system positioning results were 3.005 km, 5.008 km, and 7.997 km, respectively, with positioning errors of ≤ ±10 cm, meeting the centimeter-level positioning requirements.

[0052] Recognition accuracy test: 1000 sets of third-party tampering samples, 1000 sets of strongly interfering samples, and 1000 sets of non-interfering samples were collected. The system's recognition accuracy for third-party tampering samples was 96.8%, for strongly interfering samples it was 95.2%, and for non-interfering samples it was 98.5%, with an overall recognition accuracy of 96.8% and an AUC value of 0.9667, which is 4.1% higher than the traditional XGBoost algorithm (AUC value 0.9286).

[0053] Warning response time test: The total time from when the excavator starts knocking on the road to when the system triggers the audible and visual alarm and SMS notification is 2.8 seconds, which meets the response requirement of ≤3 seconds.

[0054] False alarm rate test: In a strong interference scenario along the highway (1000 starts and stops of heavy trucks and 1000 factory noises), the system falsely reported 48 third-party sabotage events, with a false alarm rate of 4.8%, <5%.

[0055] Early warning effectiveness test: During the field test, a total of 200 simulated third-party sabotage events occurred, and the system successfully issued early warnings 172 times, with an early warning effectiveness rate of 86%, ≥85%.

[0056] Test results show that the distributed optical fiber sensing system of the present invention has high positioning accuracy, recognition accuracy, response speed and low false alarm rate, and can effectively meet the technical requirements for monitoring the external damage prevention of underground cables.

[0057] Industrial applicability The distributed optical fiber sensing system based on model fusion recognition algorithm of the present invention solves the technical pain points in the operation and maintenance management of existing underground infrastructure, such as insufficient early warning of external damage, low efficiency of fault location, and poor accuracy of interference identification, through the innovative combination of branched optical network access architecture, spatiotemporal feature reconstruction technology and multi-model fusion recognition algorithm. It has the following industrial applicability: High level of technological maturity: The Φ-OTDR technology, MEMS optical switch, photoelectric detection technology, machine learning algorithm and other technologies adopted by the system are all existing mature technologies. After optimization and integration, they form a stable and reliable technical solution. The hardware modules and software algorithms have been tested in the laboratory and verified in the field, meeting the requirements for industrial use.

[0058] Easy to install and deploy: The system's sensing optical fiber can be laid parallel to underground cables, oil pipelines, and urban utility tunnels in the same trench without additional excavation. The optical switches, laser modules, and other equipment are small in size and can be installed in existing equipment rooms or cable manholes, resulting in low deployment costs and minimal impact on the renovation of existing facilities.

[0059] Stable and reliable operation: The system adopts industrial-grade hardware design, with an operating temperature range of -20℃ to +70℃ and a storage temperature range of -40℃ to +85℃. It is waterproof, moisture-proof, and resistant to electromagnetic interference. The optical switch has a service life of 10^7 cycles, and the laser module has a continuous working time of ≥10,000 hours, meeting the requirements for long-term stable operation.

[0060] High compatibility: The system supports data exchange with power distribution automation systems, PMS systems, and SCADA systems. It adopts standard Ethernet interfaces and MQTT protocols, and can be integrated into existing operation and maintenance management systems without the need for large-scale modifications to existing systems.

[0061] Significant economic benefits: After the system is put into use, it can reduce the workload of manual inspection by more than 60%, reduce the accident losses caused by external damage by more than 80%, improve operation and maintenance efficiency by more than 30%, and has a high return on investment. It is suitable for large-scale promotion and application in State Grid, China Southern Power Grid, petrochemical enterprises, municipal engineering management departments and other units.

[0062] The system's hardware components all utilize existing mature technologies. Its core innovation lies in the integration of modules and the synergy of functions. The equipment's production cost is controllable, and the user interface is simple and intuitive, easily mastered by medical staff after minimal training. It is also readily accepted by patients and their families, possessing broad clinical application prospects and market promotion value. Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A distributed optical fiber sensing system based on a model fusion recognition algorithm, characterized in that, It includes a laser module, a laser modulation module, a sensing and transmission module, a signal receiving module, a data processing module, an intelligent identification module, and a visualization management module. The laser module provides ultra-narrow linewidth optical signals. The laser modulation module modulates continuous laser light into pulse signals with a set pulse width and amplifies them. The sensing and transmission module uses G652 single-mode fiber with a branched topology and converts external vibrations into phase changes of optical signals based on the photoelastic effect. The signal receiving module converts backscattered Rayleigh light signals into electrical signals. The data processing module optimizes signal characteristics through spatiotemporal feature reconstruction technology. The intelligent identification module uses a multi-model fusion algorithm to classify and identify vibration events. The visualization management module displays monitoring data and early warning information based on digital twin technology. All modules interact with each other via Ethernet interface and MQTT protocol.

2. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The laser module has a center wavelength of 1550nm and a linewidth of ≤3kHz. The laser modulation module includes an acousto-optic modulator, an amplifier, and a circulator. The modulated pulse signal width is 486ns. The amplified optical signal is connected to the sensing and transmission module through the circulator.

3. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The sensing and transmission module uses a 1×8 MEMS optical switch to construct a branched topology network, including one 24-hour all-weather monitoring main channel and three rotating monitoring branches. The optical switch response time is ≤10ms, supports automatic channel switching, and the maximum monitoring length of a single device is 30 kilometers.

4. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The spatiotemporal feature reconstruction technology of the data processing module includes: based on the statistical regularity of 5-10 occasional spikes within 120 seconds, a truncated differential feature reconstruction is adopted, with the tenth largest value of the waveform as the upper limit to weaken the impact interference; a basic noise threshold is set, and signals below the threshold are uniformly set to zero; the geographic and geological information of the monitoring area is integrated, and the geographic correlation information of the defense zone is obtained through graph embedding learning to construct a spatiotemporal fusion feature set.

5. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The multi-model fusion algorithm of the intelligent recognition module includes a base learning layer and a decision layer. The base learning layer integrates random forest, GBDT, XGBoost, LightGBM and at least one deep learning model to learn different dimensional features in the spatiotemporal fusion feature set. The decision layer adopts a fully connected cascaded neural network, which fuses the output of the base learning layer through the tanh activation function and the linear summation function to output the vibration event classification result. The classification result includes third-party damage events, strong interference events and no interference events.

6. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The signal receiving module includes a photodetector and an amplifier. The photodetector converts the backscattered Rayleigh light signal into an electrical signal. The amplifier's amplification factor is adaptively adjusted according to the signal strength. The converted electrical signal is then input into the data processing module after A / D conversion. The sampling frequency is dynamically adjusted according to the monitoring distance.

7. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The visualization management module includes a GIS geographic information module, a digital twin modeling module, an early warning push module, and a report statistics module. The GIS geographic information module supports switching between satellite imagery and electronic maps. The digital twin modeling module enables 3D visualization and interactive operation of cable channels and cable wells. The early warning push module pushes abnormal information through sound and light alarms and SMS notifications. The report statistics module supports historical event tracing and data statistical analysis.

8. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The laser modulation module also includes an automatic laser pulse width adjustment unit, which consists of a detector, a data acquisition card, a host computer algorithm, and a laser pulse width controller forming a feedback adjustment network. It dynamically adjusts the laser pulse width according to the signal strength of the sensing and transmission modules to suppress signal saturation caused by strong light interference.

9. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The training sample library of the intelligent recognition module includes 15,000 dangerous intrusion samples, 15,000 strong interference samples, and 15,000 interference-free samples. AUC is used as the model evaluation index. The model recognizes an AUC value ≥ 0.96, and the false alarm rate of third-party sabotage events is < 5%.

10. The distributed optical fiber sensing system based on model fusion recognition algorithm according to claim 1, characterized in that, The system has a positioning accuracy of ≤±10cm, an event blind zone of ≤10cm, an attenuation blind zone of ≤40cm, a dynamic range of ≥10dB, and an operating temperature range of -20℃ to +70℃. It supports data interface with power distribution automation systems, PMS systems, and SCADA systems.

Images