Cavern group distributed optical fiber temperature field dynamic monitoring system based on reinforcement learning driving

Abstract

The invention discloses a reinforcement learning driving-based cavern group distributed optical fiber temperature field dynamic monitoring system, and relates to the technical field of underground engineering safety monitoring, and the system comprises a distributed optical fiber sensor array which is crosswise and annularly arranged along the inner wall of a cavern group and is used for collecting original optical signals containing temperature and strain coupling. According to the cavern group distributed optical fiber temperature field dynamic monitoring system based on reinforcement learning driving, through a dynamic coupling and decoupling mechanism of reinforcement learning and a physical model, the precision bottleneck of a traditional optical fiber temperature monitoring technology under complex multi-physics field interference is broken through. A dual-channel sensing network and a self-evolution interference feature library are constructed, real-time adaptive separation of strain-temperature components is realized, a dual heterogeneous verification mechanism is introduced, and the credibility of a decoupling result is guaranteed from two dimensions of a physical rule and data redundancy; and meanwhile, through closed-loop optimization and redundant fault-tolerant design, the monitoring robustness and continuity in a complex engineering environment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underground engineering safety monitoring, and specifically to a distributed optical fiber temperature field dynamic monitoring system for a cavern group driven by reinforcement learning. Background Art

[0002] In the field of safety monitoring of cavern group projects, the distributed optical fiber temperature field dynamic monitoring technology has been widely applied due to its advantages of long distance and high resolution. However, the internal environment of the cavern group is complex, and factors such as construction disturbances, equipment operation, and geological activities will cause dynamic mechanical vibrations, electromagnetic noises, and multi-physical field coupling effects, resulting in the superposition of temperature and strain components in the optical fiber sensing signal, forming a composite interference that is difficult to distinguish. Existing technologies usually use static calibration or preset filtering algorithms to denoise the signal, but these methods rely on prior environmental parameter assumptions and cannot adapt to the dynamic evolution of noise characteristics in the cavern group over time and space. For example, fixed-threshold filtering is difficult to cope with transient high-frequency interference caused by sudden construction vibrations, and traditional decoupling models are prone to temperature drift errors in scenarios where strain slowly accumulates due to continuous deformation of the rock mass. In addition, when facing unknown interference patterns, existing systems need to rely on manual experience to repeatedly adjust parameters and lack the ability of self-adaptation, resulting in a significant decrease in monitoring accuracy under complex working conditions. This limitation makes it difficult for the temperature field inversion results to truly reflect the thermodynamic state of the cavern group, directly affecting the timeliness and reliability of engineering safety early warning. Summary of the Invention

[0003] (1) Technical Problems to be Solved Aiming at the deficiencies of the existing technology, the present invention provides a distributed optical fiber temperature field dynamic monitoring system for a cavern group driven by reinforcement learning, which solves the problem that traditional monitoring systems are difficult to achieve real-time adaptive decoupling and accurate inversion of temperature signals under multi-physical field dynamic interference.

[0004] (2) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: A distributed optical fiber temperature field dynamic monitoring system for a cavern group driven by reinforcement learning, comprising: A distributed optical fiber sensor array, which is cross-annularly arranged along the inner wall of the cavern group and is used to collect the original optical signal containing temperature and strain coupling; the distributed optical fiber sensor array adopts multimode bending-resistant optical fiber and is cross-annularly arranged along the monitoring sections of the crown arch, side wall, and bottom plate of the inner wall of the cavern group to form a double-ring staggered redundant acquisition structure.

[0005] The dynamic data processing module is connected to the distributed optical fiber sensor array through an optical fiber signal demodulator, extracts the spatio-temporal characteristics of the original optical signal, and outputs the noise spectrum time series data. After receiving the original optical signal of the distributed optical fiber sensor array through the optical fiber signal demodulator, the dynamic data processing module first preprocesses the signal, including eliminating the baseline drift caused by the optical path attenuation, compensating the intensity noise caused by the light source fluctuation, and correcting the modal dispersion effect of the multimode optical fiber.

[0006] The reinforcement learning decoupling module is communicatively connected to the dynamic data processing module, and includes a physical prior model unit and a reinforcement learning policy unit. The physical prior model unit constructs a dynamic graph network based on the fiber strain-temperature coupling equation, and the reinforcement learning policy unit deploys a deep policy gradient network to generate decoupling parameters according to the noise spectrum time series data. After receiving the noise spectrum time series data output by the dynamic data processing module through the communication interface, the physical prior model unit first discretizes the cavern group into spatial micro-element nodes, establishes a dynamic graph network based on the fiber strain-temperature coupling equation, where each node represents the strain transfer path of the rock mass micro-element, and the connection weights between the nodes are dynamically adjusted according to the real-time rock mass deformation monitoring data, and calculates the initial contribution weight of the strain component to the temperature signal.

[0007] The cross-modal fusion unit is respectively connected to the physical prior model unit and the reinforcement learning policy unit, and fuses the outputs of the two through an attention gating mechanism to generate the decoupling result of the temperature signal. After receiving the strain contribution weight output by the physical prior model unit and the adaptive decoupling matrix generated by the reinforcement learning policy unit, the cross-modal fusion unit realizes dynamic fusion through the attention gating mechanism.

[0008] The dual verification module is connected to the cross-modal fusion unit, and includes a thermodynamic simulation verification unit and a multi-fiber path redundancy verification unit, which are used to perform physical law and data consistency verification on the decoupled temperature signal. After receiving the decoupled temperature signal output by the cross-modal fusion unit, the thermodynamic simulation verification unit first maps the temperature signal to the three-dimensional thermodynamic simulation model of the cavern group according to the spatial micro-element nodes, calculates the theoretical heat flow distribution based on the surrounding rock heat conduction coefficient and the internal heat source distribution, and makes a node-by-node comparison with the measured heat flow data of the distributed optical fiber sensor.

[0009] The temperature field inversion module is connected to the dual verification module, generates a three-dimensional temperature field distribution map based on the temperature signal that passes the verification, and outputs the safety status information through the early warning interface. After receiving the temperature signal that passes the verification of the dual verification module, the temperature field inversion module first maps the temperature data of the discrete micro-element nodes to the three-dimensional grid model of the cavern group according to the spatial coordinates, and generates a continuous spatial temperature field based on the Kriging interpolation algorithm combined with the temperature gradient distribution of adjacent nodes.

[0010] Preferably, the reinforcement learning policy unit includes: A self-evolving interference feature library that stores the construction logs of the cavern group, geological activity records, and historical interference pattern data; An adversarial generation network, connected to the self-evolving interference feature library, for dynamically synthesizing training data for unknown interference scenarios; A deep policy gradient network that generates an adaptive decoupling matrix based on the training data and updates the network weights according to the cross-domain reward function.

[0011] Preferably, the cross-domain reward function includes: A temperature inversion error reward term that calculates the measurement difference between the decoupled temperature signal and the calibrated temperature sensor; A strain residual consistency reward term that compares the decoupled strain components with the measurement data of the independent strain sensors; A spectral entropy stability reward term that analyzes the entropy value change of the signal spectrum after decoupling to suppress high-frequency noise overfitting.

[0012] Preferably, the method for constructing the dynamic graph network of the physical prior model unit is as follows: Discretize the rock mass of the cavern group into multiple microelement nodes, and establish the dynamic connection weights between the nodes according to the fiber microelement strain transfer law; Update the dynamic connection weights based on the real-time rock mass deformation monitoring data, and calculate the contribution degree of the strain component to the temperature signal.

[0013] The physical prior model unit discretizes the rock mass of the cavern group into multiple microelement nodes according to the spatial grid, and each node corresponds to a specific acquisition position of the distributed optical fiber sensor. Based on the fiber microelement strain transfer law, the dynamic connection weights between adjacent nodes are established, including: if the adjacent nodes are located in the same rock mass structural layer, the initial weight is determined according to the elastic modulus of the rock mass and the fiber layout direction; if they span different structural layers, the weight is dynamically adjusted through the interface strain transfer coefficient; where the same rock mass structural layer includes intact surrounding rock or fault fracture zones, and different structural layers are the interfaces between the support structure and the surrounding rock.

[0014] Preferably, the thermodynamic simulation verification unit performs the following steps: Input the decoupled temperature signal into the thermodynamic simulation model of the cavern group to calculate the theoretical value of the heat flux distribution; Compare the deviation between the theoretical value and the measured heat flux data. If the deviation exceeds the threshold, trigger the decoupling parameter rollback and re-optimize the network parameters of the reinforcement learning policy unit. After receiving the decoupled temperature signal, the thermodynamic simulation verification unit maps it to the pre-constructed three-dimensional thermodynamic simulation model of the cavern group according to the spatial micro-element nodes. The model calculates the theoretical heat flux distribution based on the surrounding rock thermal conductivity, internal heat source distribution, and boundary conditions. During theoretical calculation, the model preferentially loads the heat source parameters of the current construction stage of the cavern group. The heat source parameters include mechanical equipment power and ventilation volume, and dynamically correct the thermal conductivity in combination with the rock mass moisture content.

[0015] Preferably, the multi-fiber path redundancy verification unit performs the following steps: Obtain the decoupled temperature data of the redundant fibers arranged in a cross pattern at the same spatial position; Adopt the Bayesian evidence fusion algorithm, combined with the spatial gradient constraint of the temperature field, to eliminate the abnormal data whose temperature change rate exceeds the upper limit of the geological heat conductivity coefficient among adjacent nodes.

[0016] Preferably, in the Bayesian evidence fusion algorithm, the determination threshold of the spatial gradient constraint of the temperature field is 1.2 - 1.5 times the thermal conductivity coefficient of the surrounding rock of the cavern group.

[0017] Preferably, the unknown interference scenario data synthesized by the generative adversarial network includes: transient high-frequency noise caused by blasting vibration, humidity-temperature coupling noise caused by seepage, and periodic electromagnetic interference caused by the start and stop of equipment.

[0018] Preferably, the cross-ring layout method of the distributed optical fiber sensor array is: in the monitoring sections of the crown, side walls, and floor of the cavern group, fiber paths arranged in a double-ring staggered pattern are laid at intervals of 0.5 - 1.0 meters.

[0019] Preferably, the threshold of the thermodynamic simulation model is set as: the deviation of the heat flux distribution does not exceed 8% of the average value of the measured data, and the local peak deviation does not exceed 15%.

[0020] (III) Beneficial effects The present invention provides a dynamic monitoring system for the distributed optical fiber temperature field of a cavern group driven by reinforcement learning, which has the following beneficial effects: (1). The dynamic monitoring system for the distributed optical fiber temperature field of the cavern group driven by reinforcement learning breaks through the accuracy bottleneck of traditional optical fiber temperature monitoring technology under the interference of complex multi-physical fields through the dynamic coupling and decoupling mechanism of reinforcement learning and physical models. A dual-channel sensing network and a self-evolving interference feature library are constructed to realize the real-time adaptive separation of strain-temperature components, and a dual heterogeneous verification mechanism is introduced to ensure the credibility of the decoupling results from the two dimensions of physical laws and data redundancy. Compared with traditional technologies, the temperature measurement error of the system is reduced in strong interference scenarios such as blasting vibration and equipment start-stop, and the false alarm rate in unknown interference modes is reduced. At the same time, through closed-loop optimization and redundant fault tolerance design, the monitoring robustness and continuity in complex engineering environments are significantly improved.

[0021] (2). The dynamic monitoring system for the distributed optical fiber temperature field of the cavern group driven by reinforcement learning not only solves the problem of high-precision inversion of the dynamic temperature field of the cavern group, but also provides intelligent decision-making support for engineering safety warning through self-evolving learning ability and thermodynamic state evolution analysis. Its adaptive anti-interference characteristics can reduce the need for manual parameter adjustment, while the three-dimensional temperature field visualization and multi-level warning mechanism improve the engineering risk response time to the minute level, providing a new generation of technical paradigms for the long-term operation and maintenance of underground engineering and disaster prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic framework diagram of the whole invention; Figure 2 is the control logic timing diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] Please refer to Figure 1 and Figure 2 , the present invention provides a technical solution: a dynamic monitoring system for the distributed optical fiber temperature field of the cavern group driven by reinforcement learning, including: The distributed optical fiber sensor array is arranged in a cross-ring shape along the inner wall of the cavern group and is used to collect the original optical signals containing temperature and strain coupling. The distributed optical fiber sensor array uses multimode bend-resistant optical fiber and is arranged in a cross-ring shape along the monitoring sections of the crown arch, side walls and floor of the inner wall of the cavern group to form a double-ring staggered redundant acquisition structure. During specific implementation, the main ring optical fiber paths are symmetrically laid on both sides of the crown arch center line with a spacing of 0.8 meters. The vertical optical fiber paths are arranged along the side walls at intervals of 1.0 meter in the vertical direction and intersect with the main ring of the crown arch at the shoulder of the arch. The optical fiber path on the floor extends parallel to the axis of the cavern and is connected to the vertical path of the side wall at the corner of the wall to form a closed loop. Signal coupling nodes are set at the intersection points of the optical fiber paths, and multi-path optical signal synchronous acquisition is realized through wavelength division multiplexing technology.

[0025] Each section of optical fiber is encapsulated with armor and fixed in a card slot embedded in the cave wall. The card slot is filled with elastic damping colloid to isolate the vibration conduction of the rock mass. Both ends of each optical fiber path in the array are connected to an optical fiber signal demodulator. The original optical signals containing temperature and strain coupling are obtained through bidirectional Raman scattering optical time domain reflectometry technology and are transmitted to the dynamic data processing module in real time. For key risk areas, a third layer of redundant ring path is added with the spacing reduced to 0.5 meters to form a high-density signal coverage network.

[0026] During the acquisition process, at least two different paths of optical fibers cross-cover the same spatial position. When the signal of a certain path is interrupted due to mechanical damage, it automatically switches to the redundant path to complete data complementation, ensuring the continuity and spatial resolution of the temperature and strain signals.

[0027] The dynamic data processing module is connected to the distributed optical fiber sensor array through an optical fiber signal demodulator, extracts the spatio-temporal characteristics of the original optical signals, and outputs the time series data of the noise spectrum. After receiving the original optical signals of the distributed optical fiber sensor array through the optical fiber signal demodulator, the dynamic data processing module first preprocesses the signals, including eliminating the baseline drift caused by optical path attenuation, compensating for the intensity noise caused by light source fluctuations, and correcting the modal dispersion effect of multimode optical fiber. The preprocessed optical signals are segmented and intercepted in the time domain and mapped in the spatial domain, and the continuous optical signals are divided into discrete micro-element nodes according to the spatial coordinates of the cavern group. Each node corresponds to the temperature and strain coupling data at a specific position.

[0028] Based on the short-time Fourier transform and empirical mode decomposition algorithms, the time-frequency characteristics of the signals of each micro-element node are extracted, and the transient high-frequency components related to construction vibration, the periodic intermediate-frequency components caused by equipment start-stop, and the low-frequency strain components caused by rock mass deformation are separated to generate a time series feature vector containing noise intensity, spectral peak frequency and energy distribution.

[0029] Regarding the spatio-temporal correlation of dynamic interference, the module adopts a sliding window mechanism to analyze the noise spectrum correlation of adjacent micro-element nodes, identify the interference sources propagating across regions, and mark their propagation paths and attenuation coefficients. The finally output noise spectrum time-series data is encoded according to the spatial grid, including the timestamp, frequency-domain energy distribution of each micro-element node, and the cross-node interference correlation mark, providing a quantitative input of dynamic environmental noise for the reinforcement learning decoupling module.

[0030] The reinforcement learning decoupling module, which is communicatively connected to the dynamic data processing module, includes a physical prior model unit and a reinforcement learning policy unit. The physical prior model unit constructs a dynamic graph network based on the fiber optic strain-temperature coupling equation. The reinforcement learning policy unit deploys a deep policy gradient network to generate decoupling parameters according to the noise spectrum time-series data. After receiving the noise spectrum time-series data output by the dynamic data processing module through the communication interface, the physical prior model unit first discretizes the cavern group into spatial micro-element nodes, and establishes a dynamic graph network based on the fiber optic strain-temperature coupling equation. Each node represents the strain transfer path of the rock mass micro-element, and the connection weights between nodes are dynamically adjusted according to the real-time rock mass deformation monitoring data to calculate the initial contribution weight of the strain component to the temperature signal. At the same time, the reinforcement learning policy unit analyzes the interference characteristics in the noise spectrum time-series data through the deep policy gradient network, and combines the blasting vibration, seepage humidity, and electromagnetic interference patterns pre-stored in the self-evolving interference feature library to generate an adaptive decoupling matrix for the current environmental noise. This matrix includes the dynamic separation coefficients of the temperature and strain components and the noise suppression weights.

[0031] The outputs of the physical prior model and the reinforcement learning policy are fused through a cross-modal attention gating mechanism, including: the baseline weight provided by the physical model is used as the initial decoupling constraint, and the decoupling matrix generated by the reinforcement learning dynamically corrects the weight distribution deviation according to the real-time noise characteristics to form a closed-loop optimized decoupling parameter.

[0032] During the decoupling process, the weights of the dynamic graph network of the physical prior model are updated every five minutes based on the latest rock mass deformation data, and the reinforcement learning policy unit adjusts the decoupling matrix parameters in real time through a cross-domain reward function to ensure rapid convergence to the optimal decoupling state under sudden interference. Among them, the cross-domain reward function includes temperature error, strain consistency, and spectral entropy constraints. For unknown interference scenarios, the reinforcement learning policy unit calls the interference data synthesized by the adversarial generative network for online fine-tuning to enable the decoupling matrix to have cross-scenario generalization ability. The finally output decoupling parameters are synchronously transmitted to the dual verification module for credibility verification.

[0033] The cross-modal fusion unit is respectively connected to the physical prior model unit and the reinforcement learning policy unit, and fuses the outputs of the two through an attention gating mechanism to generate the decoupling result of the temperature signal. After receiving the strain contribution weight output by the physical prior model unit and the adaptive decoupling matrix generated by the reinforcement learning policy unit, the cross-modal fusion unit realizes dynamic fusion through the attention gating mechanism.

[0034] First, the strain contribution weight of the physical model serves as the baseline decoupling parameter, and the decoupling matrix of reinforcement learning provides the dynamic correction amount in the noise environment. The two are input into the multi-head attention layer to calculate the similarity score between the physical features and the reinforcement learning features, and generate the attention weight.

[0035] This weight reflects the balance ratio between the physical law and the real-time interference suppression requirement, including: under the working conditions of stable environmental noise, the weight of the physical model dominates; while in the case of sudden interference, the weight of the correction amount of reinforcement learning increases significantly. Subsequently, the two types of weights are normalized through a learnable gating function, and the fused dynamic decoupling parameters are output, including the temperature signal separation coefficient, the strain compensation factor, and the noise suppression threshold.

[0036] During the fusion process, the gating function dynamically adjusts the fusion ratio according to the real-time noise spectrum entropy value. If the spectrum entropy value exceeds the preset threshold, the proportion of the fixed weight of the physical model is reduced, and the anti-interference parameters generated by reinforcement learning are preferentially adopted. The finally generated decoupling parameters are updated in a rolling manner in units of spatial micro-element nodes according to the time window, and are synchronously transmitted to the dual verification module for verification. This process forms a closed loop of "physical baseline constraint - reinforcement learning dynamic correction - feedback iterative optimization" to ensure the accurate separation of the temperature and strain components.

[0037] The dual verification module is connected to the cross-modal fusion unit, including a thermodynamic simulation verification unit and a multi-fiber path redundancy verification unit, which are used to verify the physical law and data consistency of the decoupled temperature signal. After receiving the decoupled temperature signal output by the cross-modal fusion unit, the thermodynamic simulation verification unit first maps the temperature signal to the three-dimensional thermodynamic simulation model of the cavern group according to the spatial micro-element nodes, calculates the theoretical heat flow distribution based on the surrounding rock thermal conductivity and the internal heat source distribution, and compares it with the measured heat flow data of the distributed optical fiber sensor node by node.

[0038] If the average deviation between the theoretical value and the measured value in a certain area exceeds 8% or the local peak deviation exceeds 15%, the decoupling result of the area is judged to be abnormal, and the decoupling parameter rollback instruction is triggered, including: rolling back to the effective decoupling parameters of the previous time window, and linking the reinforcement learning strategy unit to re-optimize the network weight. At the same time, the multi-fiber path redundancy verification unit retrieves the decoupling temperature data of the cross-laid redundant optical fibers at the same spatial node, and uses the Bayesian evidence fusion algorithm to calculate the consistency probability of each path data. The temperature field spatial gradient threshold is set in combination with the geological thermal conductivity coefficient of the cave group, that is, the temperature difference change rate of adjacent nodes does not exceed 1.3 times the thermal conductivity of the surrounding rock. If the data of a node exceeds the threshold and is marked as abnormal by at least two redundant paths at the same time, the node data is judged to be invalid and the data completion mechanism is started, including: interpolation and reconstruction based on the temperature gradient distribution of adjacent normal nodes and the trend of historical data. After the double verification results are passed, the effective temperature signal is transmitted to the temperature field inversion module; if any verification fails, the current decoupling parameters are frozen and the system self-check process is triggered until both double verifications are passed before the final result is allowed to be output.

[0039] The temperature field inversion module is connected to the double verification module, and generates a three-dimensional temperature field distribution map based on the temperature signal that has passed the verification, and outputs the safety status information through the early warning interface; after receiving the temperature signal that has passed the verification of the double verification module, the temperature field inversion module first maps the temperature data of the discrete micro-element node to the three-dimensional grid model of the cavern group according to the spatial coordinates, and generates a continuous spatial temperature field based on the Kriging interpolation algorithm combined with the temperature gradient distribution of adjacent nodes. In view of the structural characteristics of the cavern group, the module introduces an adaptive weight adjustment mechanism to smoothly optimize the temperature field in key areas such as fault zones and support structure junctions to ensure the physical consistency of the thermodynamic conduction law. The structural characteristics of the cavern group include differences in rock thermal conductivity and heat source distribution on the construction surface.

[0040] Subsequently, the temperature change rate of each area is analyzed through the dynamic temperature gradient tracking algorithm to identify abnormal temperature rise or temperature drop areas, including local temperature change rates exceeding 1.5 times the thermal conductivity of the surrounding rock, and their spatial positions and diffusion trends are marked. The three-dimensional temperature field distribution map is superimposed on the BIM model of the cavern group in the form of a heat map, showing the spatial range and severity of the temperature anomaly area in real time. The early warning interface generates multi-level early warning signals based on the temperature gradient threshold, duration and diffusion speed of the abnormal area, including: if the temperature change rate of a certain area exceeds the standard for three consecutive time windows and the diffusion radius is greater than 2 meters, a red warning is triggered, and a safety status report containing risk coordinates and recommended disposal measures is pushed through the human-computer interaction interface, and the engineering control system is linked to respond to emergencies, where a time window is 15 minutes. During the inversion process, the module synchronously records the historical temperature field data, establishes a time series-based thermodynamic state evolution map of the cavern group, and provides data support for long-term safety assessment.

[0041] The reinforcement learning policy unit includes: A self-evolving interference feature library that stores the construction logs of the cavern group, geological activity records, and historical interference pattern data; An adversarial generative network, connected to the self-evolving interference feature library, for dynamically synthesizing training data for unknown interference scenarios; A deep policy gradient network that generates an adaptive decoupling matrix based on the training data and updates the network weights according to a cross-domain reward function.

[0042] It should be further noted that in the specific implementation process, the reinforcement learning policy unit dynamically stores the blasting records, equipment operation schedule, and geological activity monitoring data in the construction logs of the cavern group through the self-evolving interference feature library, and establishes a historical association map of multi-source interference patterns according to the time stamp and spatial position index.

[0043] The adversarial generative network generates synthetic training data based on the known interference patterns in the feature library, including the transient high-frequency noise waveform caused by simulated blasting, the coupled distortion characteristics of the humidity fluctuation and temperature signal in the seepage area, and the spectral template of the periodic electromagnetic interference during equipment startup and shutdown, so that the generated unknown interference scenario data is statistically consistent with the real working conditions in terms of time-frequency characteristics. Among them, the known interference patterns include the vibration frequency range and the electromagnetic noise intensity distribution.

[0044] After receiving the noise spectrum time series data output by the dynamic data processing module, the deep policy gradient network first extracts the dominant interference features in the current environment, performs similarity matching with the patterns in the self-evolving interference feature library, and screens out the most relevant subset of historical interference as prior knowledge; among them, the dominant interference features include the proportion of high-frequency vibration energy and the fundamental frequency of electromagnetic noise. Subsequently, the network generates an adaptive decoupling matrix based on the matching results. Each parameter in the matrix corresponds to the suppression weight of the noise in a specific frequency band and the dynamic separation coefficient of the strain-temperature component. During the training process, the network evaluates the decoupling effect through a cross-domain reward function, including: calculating the absolute difference between the decoupled temperature and the calibrated sensor for the temperature inversion error reward, comparing the decoupled strain component with the measurement residuals of the independent strain sensor for the strain residual consistency reward, and analyzing the flatness of the spectrum of the decoupled signal for the spectral entropy stability reward to avoid overfitting. The network weights are updated every ten minutes. If the adversarial generative network detects a new interference pattern, synthetic data is immediately injected and online fine-tuning is started to ensure the continuous optimization of the generalization ability of the decoupling matrix; among them, the new interference pattern includes unrecorded seepage noise. The finally generated decoupling parameters are jointly output with the physical prior model through the cross-modal fusion unit to form a dynamic anti-interference closed loop.

[0045] The cross-domain reward function includes: The temperature inversion error reward term calculates the measurement difference between the decoupled temperature signal and the calibrated temperature sensor; The strain residual consistency reward term compares the decoupled strain components with the measurement data of the independent strain sensors; The spectral entropy stability reward term analyzes the change in the entropy value of the spectrum of the decoupled signal to suppress the overfitting of high-frequency noise.

[0046] It should be further noted that in the specific implementation process, the cross-domain reward function realizes the optimization guidance of the decoupling parameters through a multi-dimensional constraint mechanism in the reinforcement learning policy unit. The temperature inversion error reward term obtains the reference temperature value in real time through the calibrated temperature sensor, and calculates the absolute difference with the decoupled temperature signal node by node. If the difference is less than 1.0 °C, a positive reward is given, and for every 0.5 °C increase in the difference, the reward value is exponentially attenuated, forcing the intelligent agent to prioritize ensuring the core accuracy of temperature inversion; among them, the temperature sensors are arranged at key positions in the cavern group and have no cross-sensitivity with the optical fiber path.

[0047] The strain residual consistency reward term retrieves the measurement data of the independent strain sensors, performs spatio-temporal alignment and matching with the decoupled strain components, and calculates the root mean square error of the residuals between the two. For every 10% reduction in the error, the reward weight is increased, ensuring that the decoupling result of the strain components conforms to the physical measurement law; among them, the independent strain sensors are fiber Bragg grating strain gauges.

[0048] The spectral entropy stability reward term performs a fast Fourier transform on the decoupled temperature signal, analyzes its spectral energy distribution and calculates the normalized entropy value, including: if the spectral entropy value is higher than the historical normal operating condition baseline, indicating that the high-frequency noise has not been effectively suppressed, a punitive negative reward is imposed; if the entropy value is stable within the baseline fluctuation range, a reward is given to balance the signal fidelity and the noise reduction intensity.

[0049] The three rewards are fused through dynamic weighting coefficients. In the initial stage, the temperature error reward is dominant. As the training iteration progresses, the weight of the strain consistency reward is gradually increased and the weight of the spectral entropy stability is maintained, forcing the intelligent agent to gradually optimize the decoupling strategy from coarse-grained to fine-grained; among them, the weight of the temperature error reward accounts for 60%, the weight of the strain consistency reward accounts for 30%, and the weight of the spectral entropy stability accounts for 10%.

[0050] During the training process, if a certain reward fails to reach the threshold for three consecutive iterations, the reward weight adaptive adjustment mechanism is triggered to temporarily increase the weight ratio of this reward item, up to 80%, to enhance the convergence speed in a specific dimension. Finally, the output of the cross-domain reward drives the deep policy gradient network to update the decoupling matrix parameters, forming a decoupling ability that takes into account accuracy, physical rationality, and anti-overfitting.

[0051] The method for constructing the dynamic graph network of the physical prior model unit is as follows: Discretize the rock mass of the cavern group into multiple micro - element nodes, and establish the dynamic connection weights between nodes according to the strain transfer law of optical fiber micro - elements; Update the dynamic connection weights based on real - time rock mass deformation monitoring data, and calculate the contribution degree of strain components to the temperature signal.

[0052] It should be further noted that in the specific implementation process, the physical prior model unit discretizes the rock mass of the cavern group into multiple micro - element nodes according to the spatial grid, and each node corresponds to a specific acquisition position of the distributed optical fiber sensor. Based on the strain transfer law of optical fiber micro - elements, the dynamic connection weights between adjacent nodes are established, including: if adjacent nodes are located in the same rock mass structural layer, the initial weight is determined according to the elastic modulus of the rock mass and the optical fiber layout direction; if they span different structural layers, the weight is dynamically adjusted through the interface strain transfer coefficient. After the real - time rock mass deformation monitoring data is input, the connection weights between nodes are updated by the model every five minutes. For example, when the rock mass undergoes compressive deformation resulting in a shortening of the distance between adjacent nodes, the corresponding connection weight is increased to reflect the enhanced strain transfer effect. For each micro - element node, the model calculates the contribution degree of the strain component to the temperature signal under the current weight, including: in the stage of stable rock mass deformation, the strain contribution degree has a linear relationship with the weight; when sudden deformation is detected, the interference weight of the strain on the temperature is dynamically amplified through a non - linear correction factor. During the calculation process, the model preferentially processes the node data in high - risk areas. If the strain contribution degree of a certain node is updated more than the threshold three times continuously, local grid refinement is triggered, and the node is split into smaller sub - nodes to improve the calculation resolution.

[0053] The finally output strain contribution weight is encoded according to the spatial coordinates and synchronously input into the cross - modal fusion unit together with the decoupling matrix generated by the reinforcement learning strategy unit to form a dynamic decoupling baseline under the constraint of physical laws.

[0054] The thermodynamic simulation verification unit performs the following steps: Input the decoupled temperature signal into the thermodynamic simulation model of the cavern group to calculate the theoretical value of the heat flow distribution; Compare the deviation between the theoretical value and the measured heat flow data. If the deviation exceeds the threshold, trigger the rollback of the decoupling parameters and re - optimize the network parameters of the reinforcement learning strategy unit.

[0055] It should be further noted that in the specific implementation process, after receiving the decoupled temperature signal, the thermodynamic simulation verification unit maps it to the pre - constructed three - dimensional thermodynamic simulation model of the cavern group according to the spatial micro - element nodes. The model calculates the theoretical heat flow distribution based on the thermal conductivity of the surrounding rock, the internal heat source distribution and the boundary conditions. During theoretical calculation, the model preferentially loads the heat source parameters of the current construction stage of the cavern group. The heat source parameters include the power of mechanical equipment and the ventilation volume, and the thermal conductivity is dynamically corrected in combination with the water content of the rock mass. 3 After the theoretical heat flux calculation is completed, the unit retrieves the measured heat flux data of the distributed optical fiber sensor and analyzes the deviation between the theoretical value and the measured value node by node, including: if the average deviation of more than 80% of the nodes in a certain area is greater than 8%, or there is a local peak deviation exceeding 15%, it is determined that the decoupling result in this area is untrustworthy. When triggering the decoupling parameter rollback mechanism, the unit automatically retrieves the decoupling parameters that have been verified and effective 5 minutes ago to cover the current abnormal parameters, and generates an optimization instruction to the reinforcement learning strategy unit, forcing it to retrain the deep policy gradient network based on the current environmental noise data and adjust the noise suppression weight and strain separation coefficient in the decoupling matrix.

[0056] During the rollback process, the unit synchronously records the spatial position, deviation amount and environmental noise characteristics of the abnormal nodes, which are used as the input of the synthetic data of the subsequent generative adversarial network to enhance the suppression ability of the intelligent agent against the same type of interference. If the excessive deviation still appears in the same area after two consecutive rollbacks, the manual intervention mode is started, the automatic decoupling process is paused, and the operation and maintenance personnel are prompted to check whether the sensor layout or the rock thermal physical parameters are abnormal. Through closed-loop verification and dynamic optimization, it is ensured that the decoupled temperature signal strictly satisfies the law of thermodynamics conservation.

[0057] The multi-fiber path redundancy verification unit performs the following steps: Obtain the decoupled temperature data of the redundant optical fibers arranged in a cross pattern at the same spatial position; Adopt the Bayesian evidence fusion algorithm, combined with the spatial gradient constraint of the temperature field, to eliminate the abnormal data whose temperature change rate exceeds the upper limit of the geological heat conduction coefficient among adjacent nodes.

[0058] It should be further noted that in the specific implementation process, after the multi-fiber path redundancy verification unit retrieves the decoupled temperature data of the redundant optical fibers arranged in a cross pattern at the same spatial position, it first aligns the data of each path according to the spatial grid and calculates the mean square deviation of the temperature values of different paths at the same node. If the mean square deviation exceeds the preset threshold of 1.5 °C, the Bayesian evidence fusion process is triggered, including: calculating the probability that the current data is abnormal based on the statistical distribution of the data of each path under the historical normal working conditions, and making a joint determination in combination with the spatial gradient constraint of the temperature field. The spatial gradient constraint of the temperature field is: the temperature difference change rate between adjacent nodes shall not exceed 1.3 times of the surrounding rock heat conduction coefficient.

[0059] For example, if the temperature data of a certain node shows a sudden increase of 2.8°C and a normal fluctuation of 0.5°C in two paths respectively, and the temperature gradient of its upstream node exceeds the upper limit of the heat conduction capacity, it is determined that the node data is abnormal. For abnormal data, the unit starts the interpolation reconstruction mechanism, including: generating substitute data using the spatio-temporal Kriging interpolation algorithm according to the temperature distribution trend of normal nodes within the adjacent three-hop range, and marking the status of this node as "pending manual review". If the same node is marked as abnormal in three consecutive time windows, the self-check program of the distributed fiber optic sensor array is automatically triggered, the potential fiber optic damage points are located through the optical time domain reflectometry technology, and the data is collected by switching to the standby redundant path. Here, each time window is 15 minutes.

[0060] The verified temperature data is integrated according to the spatial coordinates and transmitted to the temperature field inversion module, while the abnormal data and processing logs are synchronously stored in the self-evolving interference feature library for subsequent training and optimization of the reinforcement learning policy unit.

[0061] In the Bayesian evidence fusion algorithm, the determination threshold for the temperature field spatial gradient constraint is 1.2 - 1.5 times the thermal conductivity of the surrounding rock of the cavern group. It should be further noted that in the specific implementation process, the temperature field spatial gradient constraint threshold in the Bayesian evidence fusion algorithm is dynamically set according to the thermal physical property survey data of the surrounding rock of the cavern group, and the specific range is 1.2 to 1.5 times the thermal conductivity of the surrounding rock.

[0062] During implementation, the algorithm first retrieves the reference value of the thermal conductivity of the surrounding rock from the geological survey report, corrects the actual coefficient by combining the water content and fracture development degree of the rock mass during the construction stage, and multiplies it by the dynamic adjustment factor to generate the upper limit of the gradient threshold.

[0063] In the redundant verification, if the actual temperature gradient calculated by dividing the temperature difference between the node and its upstream adjacent node by their spatial distance exceeds the threshold of 3.0 W / m·K, the abnormal determination is triggered. At the same time, combined with Bayesian probability calculation, including: calculating the posterior probability of the current data being abnormal based on the distribution law of multi-path temperature data at the same location under historical normal working conditions. When both the gradient exceeding the standard and the high abnormal probability are satisfied, it is determined that the node data is invalid. For invalid data, the algorithm reconstructs a reasonable temperature value using the spatio-temporal weighted interpolation method according to the temperature gradient direction of adjacent nodes, and marks this node as "pending manual review". If the same node is excluded due to the gradient exceeding the standard in two consecutive verifications, the optical path self-check program of the distributed fiber optic sensor array is automatically triggered, the break point or bending fault is located through the optical time domain reflectometry curve, and the data is collected by switching to the standby redundant path. Finally, the effective temperature data and abnormal processing logs are synchronously updated to the self-evolving interference feature library for training data enhancement of the adversarial generation network of the reinforcement learning policy unit, improving the system's adaptability to gradient abnormal interference.

[0064] The unknown interference scenario data synthesized by the adversarial generative network includes: transient high-frequency noise caused by blasting vibration, humidity-temperature coupling noise caused by seepage, and periodic electromagnetic interference caused by equipment startup and shutdown. It should be further noted that in the specific implementation process, the unknown interference scenario data synthesized by the adversarial generative network generates interference characteristics with statistical authenticity based on the blasting parameters recorded in the construction log of the cavern group, the humidity change curve of the seepage monitoring points, and the startup and shutdown cycles in the equipment operation ledger. Among them, the blasting parameters include explosive equivalent and detonation timing sequence.

[0065] For blasting vibration, the network extracts the seismic wave propagation attenuation model of historical blasting events and synthesizes a noise waveform containing transient shock phases, high-frequency attenuation oscillations, and aftershock wave superposition. For seepage interference, combining the lag correlation between humidity sensor data and temperature signals, a coupling distortion mode of humidity gradient change and temperature signal drift is generated; for equipment startup and shutdown interference, the current surge characteristics during motor startup are simulated to generate a periodic electromagnetic noise template of the fundamental frequency and its 3-5th harmonics, and random pulse burrs are superimposed.

[0066] The synthesized data is enhanced by time-frequency domain confusion processing, including randomly scaling the noise duration and adjusting the spectral energy distribution ratio to ensure that the generated interference is consistent with the statistical distribution of the real scenario. When a new type of interference is detected, such as unrecorded seepage-electromagnetic composite noise, the network automatically extracts its frequency domain characteristics, such as the coexistence characteristics of humidity-related low-frequency drift and electromagnetic high-frequency pulses, and generates a mixed interference template based on feature interpolation, injecting it into the training dataset of the reinforcement learning strategy unit to drive the online update of the mixed noise suppression parameters in the decoupling matrix of the deep policy gradient network.

[0067] The cross-ring layout method of the distributed fiber optic sensor array is as follows: in the monitoring sections of the crown, side walls, and floor of the cavern group, fiber optic paths arranged in a double-ring staggered pattern are laid at intervals of 0.5 - 1.0 meters. It should be further noted that in the specific implementation process, the main ring fiber optic paths are symmetrically laid on both sides of the center line of the crown monitoring section of the cavern group, with a spacing of 0.8 meters, covering 90% of the full length of the crown; the secondary ring fiber optic paths are laid parallel and offset 0.4 meters from the main ring to form a double-ring staggered structure, ensuring that there is at least one redundant path within every 0.4-meter interval in the crown area.

[0068] Vertical optical fiber paths are arranged at 1.0-meter intervals along the vertical direction in the side wall area. The upper end is connected to the secondary ring path of the crown arch at the arch shoulder through an arc transition section, and the lower end extends to the junction of the floor. The floor optical fiber path is arranged parallel to the axis direction of the cavern, with a spacing of 0.6 meters, and intersects with the vertical path of the side wall at the corner to form a closed loop. For high-risk areas such as fault fracture zones and adjacent areas of the construction surface, a third layer of redundant ring path is additionally installed, with the spacing reduced to 0.5 meters, and high-density signal coupling nodes are set at the intersection points to separate the optical signals of different paths through wavelength division multiplexing technology.

[0069] All optical fibers are encapsulated with armor and fixed in U-shaped card slots embedded in the cave wall. The card slots are filled with silicon-based elastic damping colloid to isolate the vibration conduction of the rock mass. At the turning points of the optical fiber paths, bending protection sleeves with a radius of not less than 50 mm are used to avoid micro-bending loss. Both ends of each optical fiber in the array are connected to an optical fiber signal demodulator, and temperature and strain coupling signals are synchronously collected through bidirectional Raman scattering optical time domain reflectometry and transmitted to the dynamic data processing module in real time. When a certain optical fiber is interrupted due to mechanical damage, the system automatically switches to the adjacent redundant path, and data interpolation and reconstruction are carried out by combining the signals of the undamaged sections on both sides of the interruption point to ensure the continuity of monitoring.

[0070] The cross-ring layout enables at least two orthogonal paths to cover the same spatial position. For example, the nodes at the intersection of the crown arch and side wall paths can simultaneously obtain axial and radial strain-temperature coupling data, providing spatial correlation features for multi-physical field decoupling.

[0071] The threshold of the thermodynamic simulation model is set as follows: the deviation of the heat flux distribution does not exceed 8% of the average value of the measured data, and the local peak deviation does not exceed 15%. It should be further noted that in the specific implementation process, the threshold of the thermodynamic simulation model is set based on the thermal physical properties of the surrounding rock mass of the cavern group and dynamically adjusted during the construction stage. Specifically, in the implementation, the system loads the reference values of the thermal conductivity, density, and specific heat capacity of the surrounding rock in the survey report at the initialization stage, and dynamically corrects them by combining the moisture content of the rock mass monitored in real time through the humidity sensor and the degree of fracture development inverted from the optical fiber micro-strain data.

[0072] When calculating the theoretical heat flow, if the current stage is blasting construction, the upper threshold is relaxed to 10% of the average value of the measured data to tolerate short-term thermal shock disturbances; in the stable operation and maintenance stage, it is tightened to an average deviation of 6% and a peak of 12%. The deviation judgment adopts a sliding window mechanism, including: calculating the absolute difference between the theoretical value and the measured value of each node every five minutes. If the average deviation of a certain area exceeds the threshold in three consecutive 15-minute windows, it is judged as a systematic error and triggers the rollback of the decoupling parameters; if the local peak exceeds the standard in a single window, such as the node deviation near the construction surface suddenly increases to 20%, it is immediately marked as an emergency exception and the manual review process is started first. When the rollback operation is executed, the system automatically loads the decoupling parameters of the previous valid period to cover the abnormal data, and links the reinforcement learning strategy unit to inject the current noise spectrum characteristics and deviation data, driving the deep policy gradient network to complete online retraining within 10 minutes, and optimizing the frequency band suppression weights and strain-temperature separation coefficients in the decoupling matrix. The threshold setting log and deviation event record are synchronously stored in the self-evolving interference feature library, which is used for the subsequent adversarial generation network to synthesize more complex interference training scenarios and improve the system's adaptability to threshold boundary conditions.

[0073] A method for dynamic monitoring of distributed optical fiber temperature field in cavern groups based on reinforcement learning includes the following steps: Step S1: Distributed optical fiber signal acquisition and preprocessing: Armored optical fiber sensor arrays are arranged in a cross-ring along the top arch, side walls and bottom plate of the cave group, and the original optical signal of temperature and strain coupling is collected in real time through the bidirectional Raman scattered light time domain reflection technology; the optical path attenuation, light source fluctuation and modal dispersion noise are eliminated, the signal is divided according to the spatial micro-element node and the spatiotemporal characteristic data is output.

[0074] Step S2: Dynamic noise spectrum analysis and feature extraction: Perform short-time Fourier transform and empirical mode decomposition on the preprocessed optical signal to separate the transient vibration high-frequency component, periodic electromagnetic medium-frequency component and transient strain low-frequency component, generate noise spectrum time series data and mark the cross-region interference propagation path.

[0075] Step S3: Reinforcement learning driven dynamic decoupling of multi-physical fields: construct a strain-temperature coupling dynamic graph network based on the physical prior model, and calculate the baseline contribution weight of the strain component; the reinforcement learning strategy unit combines the noise spectrum data with the self-evolving interference feature library to generate an adaptive decoupling matrix, integrates the physical constraints and dynamic correction parameters through the cross-modal attention gating mechanism, and outputs the decoupled temperature signal.

[0076] Step S4: Dual heterogeneous verification and exception handling: Input the decoupled temperature signal into the thermodynamic simulation model, compare the deviation between the theoretical heat flow and the measured data, and if it exceeds the standard, roll back the decoupling parameters and trigger retraining; synchronously call the multi-fiber redundant path data, eliminate abnormal nodes through Bayesian evidence fusion and spatial gradient constraints, and reconstruct the missing data.

[0077] Step S5: 3D temperature field inversion and safety warning: Based on the temperature signals that pass the verification, the Kriging interpolation algorithm is used to generate a continuous 3D temperature field, and the distribution smoothness is optimized by combining the thermal physical parameters of the rock mass; the abnormal temperature rise area is identified by dynamic gradient tracking, and the warning signals are triggered by grading according to the threshold and diffusion speed, and the ventilation equipment is linked for emergency response.

[0078] Step S6: Self-evolving interference feature library update and adversarial training: Inject abnormal event data (such as decoupling parameters that fail to pass verification, new interference patterns) into the self-evolving interference feature library, synthesize statistically matching interference scenario data through the adversarial generation network, and drive the reinforcement learning strategy unit to online micro-adjust the decoupling matrix parameters.

[0079] Step S7: Closed-loop optimization and system self-check: For the nodes with continuous rollback or verification failure, start the optical fiber path self-check program, locate the damage point and switch to the redundant path; the updated decoupling parameters and threshold settings are synchronously fed back to the physical prior model and the reinforcement learning network to complete the dynamic anti-interference closed-loop optimization.

[0080] Step S8: Thermodynamic state evolution and long-term evaluation: Record the historical temperature field data and construct a spatio-temporal evolution map, analyze the long-term trend of the thermodynamics of the cavern group, generate a periodic safety assessment report, and provide decision-making support for the construction and operation and maintenance strategies.

[0081] Through the dynamic coupling and decoupling mechanism of reinforcement learning and physical models, the accuracy bottleneck of traditional optical fiber temperature monitoring technology under complex multi-physical field interference is broken through. A dual-channel sensing network and a self-evolving interference feature library are constructed to realize the real-time adaptive separation of strain-temperature components, and a dual heterogeneous verification mechanism is introduced to ensure the credibility of the decoupling results from both the physical law and data redundancy dimensions. Compared with traditional technologies, the temperature measurement error of the system is reduced in strong interference scenarios such as blasting vibration and equipment start-stop, and the false alarm rate in unknown interference patterns is reduced. At the same time, through closed-loop optimization and redundant fault tolerance design, the monitoring robustness and continuity in complex engineering environments are significantly improved.

[0082] It not only solves the problem of high-precision inversion of the dynamic temperature field of the cavern group, but also provides intelligent decision-making support for engineering safety warning through self-evolving learning ability and thermodynamic state evolution analysis. Its adaptive anti-interference characteristics can reduce the need for manual parameter adjustment, and the 3D temperature field visualization and multi-level warning mechanism improve the engineering risk response time to the minute level, providing a new generation of technical paradigm for the long-term operation and maintenance and disaster prevention and control of underground engineering.

[0083] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0084] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed optical fiber temperature field dynamic monitoring system for cavern groups based on reinforcement learning, characterized in that: include: Distributed optical fiber sensor arrays are arranged in a cross-ring pattern along the inner wall of the cavern group to collect raw optical signals containing temperature and strain coupling; A dynamic data processing module is connected to the distributed optical fiber sensor array through an optical fiber signal demodulator, extracts spatiotemporal features of the original optical signal, and outputs noise spectrum time series data; A reinforcement learning decoupling module is communicatively connected with the dynamic data processing module, and includes a physical prior model unit and a reinforcement learning strategy unit. The physical prior model unit constructs a dynamic graph network based on the optical fiber strain-temperature coupling equation, and the reinforcement learning strategy unit deploys a deep policy gradient network to generate decoupling parameters according to the noise spectrum time series data; A cross-modal fusion unit is connected to the physical prior model unit and the reinforcement learning strategy unit respectively, and fuses the outputs of the two through an attention gating mechanism to generate a temperature signal decoupling result; A dual verification module, connected to the cross-modal fusion unit, includes a thermodynamic simulation verification unit and a multi-fiber path redundancy verification unit, and is used to perform physical law and data consistency verification on the decoupled temperature signal; The temperature field inversion module is connected to the dual verification module, generates a three-dimensional temperature field distribution diagram based on the verified temperature signal, and outputs safety status information through the early warning interface.

2. According to claim 1, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The reinforcement learning strategy unit includes: Self-evolving disturbance feature library, which stores cavern group construction logs, geological activity records and historical disturbance pattern data; A generative adversarial network, connected to the self-evolving interference feature library, for dynamically synthesizing training data for unknown interference scenarios; A deep policy gradient network generates an adaptive decoupling matrix based on the training data and updates network weights according to a cross-domain reward function.

3. According to claim 2, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The cross-domain reward function includes: The temperature inversion error bonus term calculates the difference between the decoupled temperature signal and the measurement of the calibrated temperature sensor; Strain residual consistency bonus, comparing the decoupled strain components with the measurements of independent strain sensors; The spectral entropy stability bonus item analyzes the entropy change of the signal spectrum after decoupling to suppress overfitting of high-frequency noise.

4. According to claim 1, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The method for constructing the dynamic graph network of the physical prior model unit is as follows: The cavern group rock mass is discretized into multiple microelement nodes, and the dynamic connection weights between the nodes are established according to the optical fiber microelement strain transmission law. The dynamic connection weight is updated based on real-time rock mass deformation monitoring data, and the contribution of the strain component to the temperature signal is calculated.

5. According to claim 1, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The thermodynamic simulation verification unit performs the following steps: The decoupled temperature signal is input into the thermodynamic simulation model of the cavern group to calculate the theoretical value of the heat flux distribution; The deviation between the theoretical value and the measured heat flow data is compared. If the deviation exceeds a threshold, the decoupling parameter rollback is triggered and the network parameters of the reinforcement learning strategy unit are re-optimized.

6. According to claim 1, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The multi-fiber path redundancy verification unit performs the following steps: Obtain decoupled temperature data of cross-laid redundant optical fibers at the same spatial position; The Bayesian evidence fusion algorithm is used in combination with the spatial gradient constraint of the temperature field to eliminate abnormal data whose temperature change rate in adjacent nodes exceeds the upper limit of the geological heat conductivity coefficient.

7. The distributed optical fiber temperature field dynamic monitoring system for cavern groups based on reinforcement learning drive according to claim 6 is characterized by: In the Bayesian evidence fusion algorithm, the judgment threshold of the temperature field spatial gradient constraint is 1.2-1.5 times the thermal conductivity coefficient of the surrounding rock of the cavern group.

8. According to claim 2, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The unknown interference scene data synthesized by the adversarial generation network include: transient high-frequency noise caused by blasting vibration, humidity-temperature coupling noise caused by seepage, and periodic electromagnetic interference caused by equipment start-up and shutdown.

9. According to claim 1, a distributed optical fiber temperature field dynamic monitoring system for cavern groups driven by reinforcement learning is characterized in that: The cross-ring arrangement of the distributed optical fiber sensor array is as follows: in the monitoring sections of the cavern group's top arch, side walls and bottom plates, double-ring staggered optical fiber paths are arranged with a spacing of 0.5-1.0 meters.

10. The distributed optical fiber temperature field dynamic monitoring system for cavern groups based on reinforcement learning drive according to claim 5 is characterized by: The threshold of the thermodynamic simulation model is set as follows: the heat flux distribution deviation does not exceed 8% of the average value of the measured data, and the local peak deviation does not exceed 15%.

Images