A multi-modal fusion fire early detection and intelligent situation deduction system for deep underground space

Abstract

The application relates to the technical field of deep underground space safety monitoring and emergency management, and particularly discloses a multi-modal fusion fire early detection and intelligent situation deduction system for deep underground space, which comprises a multi-modal sensing module, a data fusion module, an intelligent deduction module, a decision linkage module, a guarantee module and a data management module; physical, optical, gaseous, acoustic and environmental auxiliary units are arranged in the system to comprehensively collect temperature, smoke, characteristic gas, abnormal sound and environmental parameters. The data fusion module adopts Kalman filtering for noise reduction, the feature level fusion dynamically adjusts the weight according to real-time reliability, the decision level fusion outputs the fire state based on a Bayesian network, different combustible gas pyrolysis gases are matched in combination with a characteristic spectrum library, dust, moisture and other interference are effectively filtered out, early fire conditions such as smoldering and overheating are accurately identified, and the problems of high false alarm and missing report rates of single sensors are solved.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and emergency management technology for deep underground spaces, and in particular to a multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces. Background Technology

[0002] As urban spaces extend underground, fire safety in deep underground spaces such as deep mines and integrated utility tunnels faces severe challenges. Existing fire prevention and control technologies are clearly inadequate.

[0003] First, the detection methods are limited, relying mainly on independent sensors such as temperature or smoke sensors. In deep environments with high dust, high humidity, and strong electromagnetic interference, they are prone to false alarms or missed alarms, and it is difficult to capture early weak signals such as smoldering.

[0004] Secondly, the lack of effective integration of multi-source data and the independent operation of each device create data silos. Without dynamic calibration based on environmental parameters, the system has poor anti-interference capabilities and is prone to alarm fatigue.

[0005] In addition, the situation simulation models mostly copy ground building scenarios, failing to consider characteristics such as limited ventilation and significant heat accumulation in deep spaces, and neglecting structural collapse and multi-hazard chain reactions, resulting in large deviations between the simulation results and reality.

[0006] Furthermore, detection, simulation, and decision-making are disconnected and lack closed-loop coordination. Simulations are often initiated only after the fire is confirmed, missing the opportunity for initial response, and the simulation results cannot be fed back to optimize front-end monitoring.

[0007] Therefore, a multimodal fusion fire early detection and intelligent situation simulation system adapted to deep underground environments is needed. Summary of the Invention

[0008] The purpose of this invention is to provide a multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces, in order to solve the problems mentioned in the background art.

[0009] A multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces, the system includes a multimodal perception module, a data fusion module, an intelligent simulation module, a decision linkage module, a support module, and a data governance module; The multimodal sensing module is installed in deep underground space to collect multi-dimensional physical and chemical signals in the early stages of a fire. The data fusion module is communicatively connected to the multimodal perception module and is used to preprocess, extract features, and perform fusion analysis on multidimensional physical and chemical signals, outputting fire identification results and core feature vectors. The intelligent simulation module is communicatively connected to the data fusion module. It is used to construct a simulation model driven by a combination of data and physical rules based on fire identification results and core feature vectors, combined with preset deep underground space parameters. The model dynamically simulates the development of the fire and outputs the simulation data. The decision-making linkage module is communicatively connected to the intelligent simulation module and the multimodal perception module. It is used to generate emergency response instructions based on fire identification results and situation simulation data, and to feed back risk area information to the multimodal perception module, driving the multimodal perception module to adjust the sampling strategy of the corresponding area. The protection module is connected to the multimodal perception module, data fusion module, intelligent inference module and decision linkage module respectively, and is used to provide power supply protection, communication protection and equipment self-diagnosis protection for the system in the harsh environment of deep underground. The data governance module is connected to the multimodal perception module, data fusion module, intelligent inference module, and decision linkage module, respectively, and is used to perform unified management and iterative optimization of data across the entire system.

[0010] Preferably, the multimodal sensing module includes a physical sensing unit, an optical sensing unit, a gas sensing unit, an acoustic sensing unit, and an environmental auxiliary unit. The physical sensing unit includes a temperature sensor and a vibration sensor, used to collect temperature signals and structural vibration signals; The optical sensing unit includes a smoke sensor and a flame radiation sensor, used to collect smoke signals and flame spectral signals. The gas sensing unit includes characteristic gas sensors for CO, CO2, CH4, and H2S, used to collect the concentration signals of characteristic gases in a fire. The acoustic sensing unit includes an acoustic sensor array for collecting local abnormal noise signals; The environmental auxiliary unit includes a humidity sensor and an electromagnetic interference monitoring module, which are used to collect humidity signals and electromagnetic interference signals and output environmental calibration data. The multimodal sensing module transmits the collected temperature signal, smoke signal, flame spectrum signal, characteristic gas concentration signal, abnormal noise signal, vibration signal, humidity signal, and electromagnetic interference signal as multidimensional physical and chemical signals to the data fusion module.

[0011] Preferably, the data fusion module includes a preprocessing unit, a feature extraction unit, a two-layer fusion unit, and an early warning unit; The preprocessing unit is used to perform noise reduction processing on multi-dimensional physical and chemical signals and remove invalid data using the Kalman filter algorithm, realize the synchronization and alignment of multi-source data through timestamps, and receive environmental interference calibration parameters transmitted by the protection module for environmental compensation. The feature extraction unit is used to extract time-domain features, frequency-domain features, and time-frequency-domain features for different data types, and to perform dimensionality reduction on the multi-dimensional features through a convolutional neural network to output the core feature vector. The dual-layer fusion unit includes a feature-level fusion subunit and a decision-level fusion subunit; The feature-level fusion subunit adopts a dynamic weighting algorithm, which adjusts the weight coefficients according to the real-time reliability of each sensor, and performs weighted fusion of multi-source core features to generate a unified feature vector. The feature-level fusion subunits are weighted and fused according to the following formula; ; in, The unified feature vector after fusion For the first The core feature vector extracted by each sensor For the first The fusion weighting coefficients of each sensor, The total number of sensors participating in the fusion; The weighting coefficient Based on the real-time reliability of the sensor Determine using the following formula: ; ; in, For the first The initial confidence level of each sensor, Given the current electromagnetic interference intensity, Given the current ambient humidity, This represents the current dust concentration. , and This is the preset environmental interference attenuation coefficient; The decision-level fusion subunit, based on a Bayesian inference network, calculates the probability of fire occurrence using the following formula: ; in, , and These represent three states: no fire, suspected fire, and confirmed fire, respectively. Unified feature vector, For the first Prior probabilities of various fire conditions. In the event of a fire The following feature vectors appear The conditional probability; The advanced early warning unit is used to extrapolate the trend of multi-sensor time-series data using a long short-term memory network or a Transformer model. Before the signal reaches the alarm threshold, it outputs a risk trend warning based on the rate of change and acceleration, forming a four-level warning gradient: trend warning, suspected fire, confirmed fire, and fire expansion.

[0012] Preferably, the intelligent simulation module includes a basic database unit, a hybrid-driven simulation unit, a dynamic update unit, a digital twin visualization unit, a structural coupling analysis unit, and a multi-hazard coupling assessment unit; The basic database unit is pre-set with structural parameters, ventilation parameters, combustible material types and calorific values, and safety facility location information for deep underground spaces; The hybrid-driven inference unit includes a data-driven subunit and a physical rule subunit; The data-driven subunit is based on a long short-term memory network and uses historical fire data and current fused data output by the data fusion module as input to predict the fire development trend. The physical rules subunit introduces a computational fluid dynamics model, which combines ventilation characteristics, heat conduction equations, and mass conservation equations to correct the output results of the data-driven subunit. The physical rules constrain the output boundary of the data model, and the data-driven model supplements the scenario adaptation of the physical rules. The dynamic update unit is used to receive the latest sensing data and update the inference results at preset time intervals, and adjust the model parameters in a timely manner when environmental parameters change. The digital twin visualization unit is used to construct a three-dimensional digital twin model based on building information model data or geographic information system data of deep underground space, and to map the situation simulation data to the three-dimensional digital twin model in real time for three-dimensional visualization display. The three-dimensional visualization display includes at least the dynamics of fire spread, dynamics of smoke diffusion, temperature field distribution and personnel location information. The structural coupling analysis unit is used to integrate distributed fiber optic sensors or vibration sensors to collect thermal stress and deformation data generated by fire on underground space structures, establish a thermo-mechanical coupling model from temperature field distribution to thermal stress to structural deformation to bearing capacity attenuation, and output collapse risk assessment results. The formula for calculating the thermo-mechanical coupling model using the structural coupling analysis unit is as follows: ; in, This represents the structural bearing capacity attenuation value. The initial bearing capacity of the structure, for The highest temperature value in the fire temperature field distribution at any given time. The thermal damage coefficient of the material. The moment the fire started. The current moment; The formula for calculating the collapse risk using the structural coupling analysis unit is as follows: ; in, The collapse risk level value. This represents the structural bearing capacity attenuation value. The initial bearing capacity of the structure, This is the structural safety redundancy factor; The multi-hazard coupling assessment unit is used to cross-couple the gas leakage diffusion model, hydraulic model, and power grid fault model with fire simulation to establish a coupling relationship model between hazards and output secondary disaster risk assessment results. The situation simulation data includes at least the fire spread rate, smoke diffusion path, temperature field distribution, toxic gas concentration distribution, structural collapse risk level, and secondary disaster risk assessment results.

[0013] Preferably, the decision-making linkage module includes a hierarchical early warning unit, an intelligent decision generation unit, a closed-loop feedback unit, a multi-agent collaborative fire extinguishing unit, a dynamic evacuation unit, and a response effect evaluation unit. The graded early warning unit is used to trigger corresponding level early warning signals and link alarm devices based on fire identification results and situation simulation data. The early warning signals include level one, level two, level three and level four early warnings, which correspond to four states: trend warning, suspected fire, confirmed fire and fire expansion, respectively. The intelligent decision generation unit is used to generate emergency response plans based on situation simulation data. The emergency response plan includes at least a fire extinguishing plan, an evacuation plan, and a rescue plan. The fire extinguishing plan includes information on the best fire extinguishing location, the evacuation plan includes information on the best escape route, and the rescue plan includes information on key rescue areas. The closed-loop feedback unit is used to feed back the high-risk area information in the situation simulation data to the multimodal perception module, and drive the sensor corresponding to the high-risk area to increase the sampling frequency from the first frequency to the second frequency, where the second frequency is higher than the first frequency. The closed-loop feedback unit determines the second frequency based on the risk level value R of the high-risk area according to the following formula: ; in, For the first frequency, For the second frequency, For high-risk areas, the risk level value R ranges from 0 to 1. This is the preset frequency adjustment coefficient.

[0014] Preferably, the multi-agent collaborative fire extinguishing unit is used to incorporate fixed fire extinguishing devices and mobile fire extinguishing equipment into a unified scheduling framework, generate a collaborative fire extinguishing plan based on situational simulation data, and allocate tasks, timing coordination and coverage areas for each fire extinguishing unit. The dynamic evacuation unit is used to integrate UWB positioning technology or Bluetooth AOA positioning technology to obtain the location coordinates of trapped personnel and rescuers. Combined with the smoke diffusion path and temperature field distribution in the situation simulation data, it uses a deep reinforcement learning algorithm to calculate the escape path and outputs path guidance through smart terminals and electronic signs. The emergency response effectiveness evaluation unit is used to continuously monitor changes in fire field parameters during the execution of the emergency response plan, establish emergency response effectiveness evaluation indicators, and trigger the correction of the emergency response plan when the evaluation results do not meet the preset conditions.

[0015] Preferably, the protection module includes a dual-power supply unit, a dual-mode communication unit, an emergency self-organizing network unit, a network outage recovery unit, and a system self-diagnosis unit; The dual power supply unit adopts a power supply mode combining mains power and backup lithium battery pack. The lithium battery pack has a capacity of not less than 10kWh and can support the core equipment to work continuously for no less than 8 hours. The dual-mode communication unit adopts a dual-mode design of industrial Ethernet and 433MHz low-frequency wireless communication, and the shielding effectiveness of the communication module is not less than 40dB. The emergency self-organizing network unit is used to enable each detection node to form a temporary multi-hop Mesh communication network and transmit key alarm information when the main communication link is interrupted. The network outage recovery unit is used to cache lightweight inference models and decision rules locally on the edge side, run independently when the cloud is disconnected, and synchronize the local data during the network outage after communication is restored. The system self-diagnostic unit is used to monitor the operating status of each layer of equipment in real time, and to trigger a fault alarm and start redundant equipment when equipment fails.

[0016] Preferably, the data governance module includes a weight adaptive learning unit and a feature spectrum library storage unit; The weighted adaptive learning unit is used to monitor the current environmental status in real time, including dust concentration, humidity level and electromagnetic interference intensity, and is used to combine historical false alarm data and missed alarm data to adjust the weight coefficients of each sensor in feature-level fusion through an online learning algorithm. The characteristic spectral library storage unit is used to store the pyrolysis gas characteristic spectral library for typical combustibles in deep space, including coal, cable insulation materials, oil and gas pipeline leaks, wood, and rubber.

[0017] Preferably, the data governance module further includes a federated learning transfer unit; The federated learning transfer unit is used to upload the gradient parameters of the local model trained on each deployment node to the cloud coordination server for collaborative training of the global model, provided that the local data on each deployment node does not leave the domain. The federated learning transfer unit is also used to fine-tune the global model when deploying in new scenarios, and to continuously receive real-world data from each node to update the global model.

[0018] Preferably, the data governance module further includes a blockchain evidence storage unit; The blockchain evidence storage unit is used to store the original multimodal perception data, fusion processing logs, inference process parameters, decision instruction records, and execution feedback results on the blockchain for evidence storage. The blockchain evidence storage unit adopts a consortium blockchain architecture, with data hashing and on-chain storage, and raw data distributed storage.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention comprehensively collects temperature, smoke, characteristic gases, abnormal sounds, and environmental parameters by setting up physical, optical, gas, acoustic, and environmental auxiliary units. The data fusion module uses Kalman filtering for noise reduction, feature-level fusion dynamically adjusts weights based on real-time reliability, and decision-level fusion outputs fire status based on a Bayesian network. It combines a feature spectrum library to match different combustible pyrolysis gases, effectively filtering out interference from dust, moisture, etc., and accurately identifying early fire conditions such as smoldering and overheating, thus solving the problem of high false alarm and false negative rates of single sensors.

[0020] 2. This invention combines a data-driven long short-term memory network with a computational fluid dynamics model based on physical rules through an intelligent simulation module. It uses ventilation thermodynamics laws to constrain the data prediction boundaries, achieving mechanism-based boundary definition and improved data accuracy. Simultaneously, a thermo-mechanical coupling model is established to assess structural collapse risk, and gas diffusion, hydraulic, and power grid fault models are cross-coupled to assess secondary disasters. The simulation results accurately reflect fire spread, smoke diffusion, and the chain reaction of multiple hazards, overcoming the shortcomings of traditional models that are disconnected from the ground scene, neglect deep-seated characteristics, and fail to address the coupling of multiple hazards.

[0021] 3. This invention utilizes a long short-term memory network to extrapolate time-series data through an advanced early warning unit. When the signal does not reach a threshold, it outputs a trend warning based on the rate of change, forming a four-level warning gradient. The decision-making linkage module automatically generates a response plan including the optimal fire extinguishing location and the best escape route based on the extrapolated data. The closed-loop feedback unit feeds back high-risk area information to the perception module, driving the sensors to increase the sampling frequency to achieve resource focusing. The multi-agent collaboration and dynamic evacuation unit further enhances the level of automation, solving the problems of fragmented processes and delayed response in traditional systems.

[0022] 4. This invention features a dual-power supply module using both mains power and a lithium battery pack, supporting continuous operation of core equipment for at least 8 hours. It supports dual-mode communication via industrial Ethernet and 433MHz low-frequency wireless, with a shielding effectiveness of at least 40dB. When the main link is interrupted, it automatically establishes a Mesh emergency self-organizing network. When the cloud connection is lost, the edge-side network interruption recovery unit operates independently, enabling the system to operate reliably and continuously in deep underground environments with high temperature, high humidity, high dust, and strong electromagnetic interference, demonstrating strong survivability. Furthermore, the weighted adaptive learning unit optimizes and merges weights online based on environmental conditions and historical data; the federated learning transfer unit encrypts and uploads gradients without data leaving the domain, collaboratively training the global model and supporting fine-tuning for new scenarios, enabling rapid deployment and model iteration across underground spaces; and the blockchain evidence storage unit uses hash algorithms and consensus mechanisms to store all data on the blockchain, ensuring that perception, inference, and decision-making data are tamper-proof and traceable, combining security with promotional value. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a structural diagram of the multimodal sensing module of the present invention; Figure 3 This is a diagram of the data fusion module architecture of the present invention; Figure 4 This is a diagram of the intelligent inference module architecture of the present invention; Figure 5 This is a structural diagram of the decision-making linkage module of the present invention; Figure 6 This is a diagram of the guarantee module architecture of the present invention; Figure 7 This is a diagram of the data governance module architecture of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described 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.

[0026] Example 1, please refer to Figure 1 As shown, a multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces includes a multimodal perception module, a data fusion module, an intelligent simulation module, a decision linkage module, a support module, and a data governance module.

[0027] The multimodal sensing module is installed in deep underground space, specifically on the tunnel sidewalls, roof, and near key equipment. It establishes a communication connection with the data fusion module via industrial Ethernet and wireless communication links. This connection is used to continuously collect multi-dimensional physical and chemical signals during the early stages of a fire and transmit the collected signals to the data fusion module. The data fusion module, communicating with the multimodal sensing module, receives the multi-dimensional physical and chemical signals, performs preprocessing to eliminate noise and outliers, extracts feature information from various signals, performs multi-source fusion analysis, and finally outputs the fire identification result and core feature vector.

[0028] The data fusion module sends the fire identification results and core feature vectors to the intelligent simulation module via an internal communication bus. The intelligent simulation module is communicatively connected to the data fusion module and has pre-set structural parameters, ventilation parameters, combustible material types and calorific values, and safety facility location information for deep underground spaces. After receiving the fire identification results and core feature vectors, it constructs a simulation model driven by a hybrid of data and physical rules.

[0029] The model uses the preliminary prediction results output by the data-driven sub-unit as a benchmark. The physical rule sub-unit introduces a computational fluid dynamics model and combines ventilation characteristics, heat conduction equations, and mass conservation equations to solve for the theoretical boundary values ​​of temperature field and smoke diffusion. When the prediction value of the data-driven sub-unit exceeds the theoretical boundary value, the calculation result of the physical rule sub-unit is used as the final extrapolation output, thereby dynamically extrapolating the fire development trend and outputting situation extrapolation data including fire spread speed, smoke diffusion path, temperature field distribution, and toxic gas concentration distribution.

[0030] The intelligent simulation module sends situation simulation data to the decision-making linkage module. The decision-making linkage module communicates with the intelligent simulation module and the multimodal perception module. After receiving the situation simulation data, it generates emergency response instructions based on the fire identification results and the situation simulation data. The emergency response instructions include instructions to activate fire extinguishing devices, instructions to trigger evacuation broadcasts, and instructions to dispatch rescue forces.

[0031] At the same time, the decision-making linkage module extracts information on high-risk areas from the situation simulation data and feeds this information back to the multimodal perception module, which then drives the sensors in the multimodal perception module corresponding to the high-risk areas to increase their sampling frequency from the first frequency to the second frequency, which is higher than the first frequency.

[0032] The support module is connected to the multimodal perception module, data fusion module, intelligent inference module, and decision linkage module respectively, and is used to provide the system with continuous and stable power supply, communication support, and equipment self-diagnosis support in the harsh environment of high temperature, high humidity, high dust and strong electromagnetic interference in deep underground.

[0033] The data governance module is connected to the multimodal perception module, data fusion module, intelligent inference module, and decision linkage module respectively. It is used to uniformly manage the raw perception data, fused intermediate data, inference parameters, and decision command records generated by the entire system link, and to realize model iterative optimization across deployment nodes using federated learning transfer units.

[0034] Example 2, please refer to Figures 2 to 3 As shown, a multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces is proposed. The multimodal perception module includes a physical sensing unit, an optical sensing unit, a gas sensing unit, an acoustic sensing unit, and an environmental auxiliary unit.

[0035] The physical sensing unit includes a temperature sensor and a vibration sensor, which are used to collect temperature signals and structural vibration signals, respectively.

[0036] The optical sensing unit includes a smoke sensor and a flame radiation sensor, used to acquire smoke signals and flame spectral signals, respectively. The gas sensing unit includes a carbon monoxide sensor, a carbon dioxide sensor, a methane sensor, and a hydrogen sulfide sensor, used to acquire the concentration signals of various fire-characteristic gases, respectively.

[0037] The acoustic sensing unit includes an acoustic sensor array for acquiring local noise signals. The environmental auxiliary unit includes a humidity sensor and an electromagnetic interference monitoring module for acquiring humidity and electromagnetic interference signals and outputting environmental calibration data.

[0038] The multimodal sensing module collects temperature signals, smoke signals, flame spectrum signals, various characteristic gas concentration signals, abnormal noise signals, vibration signals, humidity signals, and electromagnetic interference signals, and then transmits them as multidimensional physical and chemical signals to the data fusion module.

[0039] The data fusion module includes a preprocessing unit, a feature extraction unit, a two-layer fusion unit, and an early warning unit.

[0040] The preprocessing unit uses the Kalman filter algorithm to denoise and remove invalid data from multidimensional physical and chemical signals. The Kalman filter algorithm eliminates random noise in the signal by establishing state equations and observation equations and performing iterative operations in prediction and update steps.

[0041] The preprocessing unit achieves multi-source data synchronization and alignment through timestamps, and receives environmental interference calibration parameters transmitted by the protection module for environmental compensation.

[0042] The feature extraction unit extracts time-domain features, frequency-domain features, and time-frequency-domain features in a differentiated manner for different data types. It then performs dimensionality reduction on the multi-dimensional features through a one-dimensional convolutional neural network, mapping the high-dimensional features into a core feature vector of fixed dimensions and outputting the core feature vector.

[0043] The two-layer fusion unit comprises a feature-level fusion subunit and a decision-level fusion subunit. The feature-level fusion subunit employs a dynamic weighting algorithm, adjusting the weight coefficients based on the real-time reliability of each sensor to weightedly fuse multi-source core features and generate a unified feature vector.

[0044] Feature-level fusion subunits are weighted and fused according to the following formula: ; in, The unified feature vector after fusion For the first The core feature vector extracted by each sensor For the first The fusion weighting coefficients of each sensor, The total number of sensors participating in the fusion; Weighting coefficient Based on the real-time reliability of the sensor Determine using the following formula: ; ; in, For the first The initial confidence level of each sensor, Given the current electromagnetic interference intensity, Given the current ambient humidity, This represents the current dust concentration. , and This is the preset environmental interference attenuation coefficient; The decision-level fusion subunit, based on a Bayesian inference network, calculates the probability of fire occurrence using the following formula: ; in, , and These represent three states: no fire, suspected fire, and confirmed fire, respectively. Unified feature vector, For the first The prior probability of a certain fire condition is preset based on historical fire statistics and normal operating condition data. In the event of a fire The following feature vectors appear The conditional probability is obtained by statistically analyzing a large number of labeled samples collected offline and stored in a lookup table.

[0045] The decision-level fusion subunit takes the state corresponding to the maximum probability as the final fire identification result output.

[0046] The advanced early warning unit uses a long short-term memory network to extrapolate the trend of multi-sensor time-series data. The long short-term memory network takes the data sequence of each sensor in multiple consecutive time steps in the past as input and outputs the predicted values ​​of temperature and gas concentration in the future time step. The advanced early warning unit calculates the rate of change and acceleration of the predicted values ​​and outputs a risk trend warning before the signal reaches the alarm threshold, forming a four-level warning gradient of trend warning, suspected fire, confirmed fire, and fire expansion.

[0047] Example 3, please refer to Figures 4 to 7 As shown, a multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces includes a basic database unit, a hybrid-driven simulation unit, a dynamic update unit, a digital twin visualization unit, a structural coupling analysis unit, and a multi-hazard coupling assessment unit.

[0048] The basic database unit is pre-set with structural parameters, ventilation parameters, combustible material types and calorific values, and safety facility location information for deep underground spaces.

[0049] The hybrid-driven simulation unit comprises a data-driven subunit and a physical rule subunit. The data-driven subunit, based on a long short-term memory network, uses historical fire data and current fused data output from the data fusion module as input to predict fire development trends.

[0050] The physical rules subunit introduces a computational fluid dynamics model and combines ventilation characteristics, heat conduction equations, and mass conservation equations to correct the output results of the data-driven subunit. The physical rules subunit uses the upper limit of the temperature field and the flue gas diffusion boundary obtained from the computational fluid dynamics model as constraints. When the predicted value of the data-driven subunit exceeds the constraint range, the predicted value is truncated or replaced by the physical rules calculation results, and the data-driven model is used to supplement the scenario adaptation of the physical rules.

[0051] The dynamic update unit receives the latest sensing data and updates the inference results at preset time intervals, and adjusts the model parameters in a timely manner when environmental parameters change.

[0052] The digital twin visualization unit constructs a three-dimensional digital twin model based on building information model data or geographic information system data of deep underground space. The situation simulation data is mapped to the three-dimensional digital twin model in real time for three-dimensional visualization display. The displayed content includes the dynamics of fire spread, smoke diffusion, temperature field distribution and personnel location information.

[0053] The structural coupling analysis unit integrates distributed fiber optic sensors or vibration sensors to collect data on the thermal stress and deformation of underground space structures caused by fire, establishes a thermo-mechanical coupling model from temperature field distribution to thermal stress to structural deformation to bearing capacity attenuation, and outputs collapse risk assessment results.

[0054] The formula for calculating the thermo-mechanical coupling model using the structural coupling analysis unit is as follows: ; in, This represents the structural bearing capacity attenuation value. The initial bearing capacity of the structure, for The highest temperature value in the fire temperature field distribution at any given time. The thermal damage coefficient is determined by referring to a table based on the type of concrete or steel used in the underground space structure. The moment the fire started. The current moment; The integral term is obtained by performing a trapezoidal discretization summation on the temperature time series, which is calculated as follows: ; in, From arrive The total number of discrete time steps at time 1. For the first The time corresponding to each time step For the first The step size of each time step is set to the update time interval set by the dynamic update unit.

[0055] The formula for calculating the collapse risk using a structural coupling analysis unit is as follows: ; in, The collapse risk level value. This represents the structural bearing capacity attenuation value. The initial bearing capacity of the structure, The structural safety redundancy factor is determined according to the underground space design code. If the value is greater than or equal to the preset threshold, it is judged as an extremely high risk of collapse.

[0056] The multi-hazard coupling assessment unit is used to cross-couple the gas leakage diffusion model, hydraulic model, and power grid fault model with fire simulation. The gas leakage diffusion model is calculated based on Fick's diffusion law, the hydraulic model is calculated based on Bernoulli's equation, and the power grid fault model is evaluated based on the short-circuit current calculation model. The multi-hazard coupling assessment unit exchanges data between the above models and the temperature field and smoke field of the fire simulation to establish a coupling relationship model between the hazards and output the secondary disaster risk assessment results.

[0057] The situation simulation data includes at least the fire spread rate, smoke diffusion path, temperature field distribution, toxic gas concentration distribution, structural collapse risk level, and secondary disaster risk assessment results. The decision-making linkage module includes a tiered early warning unit, an intelligent decision generation unit, a closed-loop feedback unit, a multi-agent collaborative firefighting unit, a dynamic evacuation unit, and a response effectiveness evaluation unit.

[0058] The tiered early warning unit triggers corresponding early warning signals and activates alarm devices based on fire identification results and situation simulation data. These early warning signals include Level 1, Level 2, Level 3, and Level 4, corresponding to four states: trend warning, suspected fire, confirmed fire, and fire expansion, respectively. The intelligent decision generation unit generates emergency response plans based on situation simulation data, including fire extinguishing plans, evacuation plans, and rescue plans.

[0059] The fire extinguishing plan includes information on the best fire extinguishing location, the evacuation plan includes information on the best escape route, and the rescue plan includes information on key rescue areas.

[0060] The intelligent decision generation unit also incorporates secondary disaster risk assessment results into the scheme verification. The closed-loop feedback unit feeds back high-risk area information from the situation simulation data to the multimodal perception module, driving the sensors corresponding to the high-risk areas to increase the sampling frequency from the first frequency to the second frequency.

[0061] The closed-loop feedback unit determines the second frequency based on the risk level value R of the high-risk area using the following formula: ; in, For the first frequency, For the second frequency, The risk level value for high-risk areas The value is normalized from the collapse risk level or fire spread probability output by the situation simulation module to a range of 0 to 1. This is the preset frequency adjustment coefficient.

[0062] The multi-agent collaborative fire suppression unit is used to integrate fixed fire suppression devices and mobile fire suppression equipment into a unified scheduling framework. Based on the contract network protocol or task allocation algorithm, it allocates the tasks, timing coordination and coverage area of ​​each fire suppression unit according to the location distance, remaining workload and current status of each fire suppression unit.

[0063] The dynamic evacuation unit integrates UWB positioning technology or Bluetooth AOA positioning technology to obtain the location coordinates of trapped personnel and rescuers. Combined with the smoke diffusion path and temperature field distribution in the situation simulation data, it constructs a Markov decision process model with personnel location, smoke concentration and temperature as the state space and the distance to the safe exit as the reward function. It uses a deep Q-network in the deep reinforcement learning algorithm to calculate the escape path and outputs path guidance through smart terminals and electronic signs.

[0064] The emergency response effectiveness assessment unit is used to continuously monitor changes in fire field parameters during the execution of the emergency response plan, establish assessment indicators for the effectiveness of the response, and trigger the revision of the emergency response plan when the assessment results do not meet the preset conditions.

[0065] The protection module includes a dual-power supply unit, a dual-mode communication unit, an emergency self-organizing network unit, a network outage recovery unit, and a system self-diagnosis unit.

[0066] The dual-power supply unit adopts a combined power supply mode of AC power and backup lithium battery packs, with a lithium battery pack capacity of no less than 10kWh, supporting continuous operation of core equipment for no less than 8 hours. The dual-mode communication unit adopts a dual-mode design of industrial Ethernet and 433MHz low-frequency wireless communication, with a communication module shielding effectiveness of no less than 40dB. The emergency self-organizing network unit is used to establish a temporary multi-hop mesh communication network between detection nodes and transmit critical alarm information when the main communication link is interrupted.

[0067] The network outage recovery unit is used to locally cache lightweight inference models and decision rules on the edge side, run independently when the cloud connection is lost, and synchronize the local data during the network outage after communication is restored. The system self-diagnosis unit is used to monitor the operating status of devices at each layer in real time, and trigger fault alarms and activate redundant devices when a device fails.

[0068] Example 4, please refer to Figure 7 As shown, a multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces includes a weighted adaptive learning unit, a feature spectrum library storage unit, a federated learning transfer unit, and a blockchain evidence storage unit.

[0069] The weighted adaptive learning unit is used to monitor the current environmental status in real time, including dust concentration, humidity level and electromagnetic interference intensity. It combines historical false alarm data and missed alarm data and adjusts the weight coefficients of each sensor in feature-level fusion through an online learning algorithm.

[0070] The characteristic spectrum library storage unit is used to store the characteristic spectrum library of pyrolysis gases for typical combustibles in deep space. Typical combustibles include coal, cable insulation materials, oil and gas pipeline leaks, wood and rubber. Each combustible corresponds to characteristic gas combinations and concentration ratio data at different temperature stages.

[0071] The federated learning transfer unit is used to train the model and calculate the gradient update amount on each deployment node using local data, provided that the local data on each deployment node does not leave the domain. The gradient update amount is then encrypted and uploaded to the cloud coordination server. The cloud coordination server aggregates the gradients using a weighted average method to generate global model parameters and distributes them to each node.

[0072] The Federated Learning Transfer Unit is also used for fine-tuning the global model when deploying in new scenarios, and continuously receives real-world data from each node to update the global model. The Blockchain Evidence Storage Unit is used to store the original multimodal perception data, fusion processing logs, inference process parameters, decision instruction records, and execution feedback results on the blockchain. The Blockchain Evidence Storage Unit adopts a consortium blockchain architecture, uses the SHA-256 algorithm to calculate the hash value of data records, and packages the hash value onto the blockchain through a practical Byzantine fault-tolerant consensus mechanism. The original data is distributed and stored in the independent databases of each participating node.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces, characterized in that: The system includes a multimodal perception module, a data fusion module, an intelligent inference module, a decision-making linkage module, a support module, and a data governance module; The multimodal sensing module is installed in deep underground space to collect multi-dimensional physical and chemical signals in the early stages of a fire. The data fusion module is communicatively connected to the multimodal perception module and is used to preprocess, extract features, and perform fusion analysis on multidimensional physical and chemical signals, outputting fire identification results and core feature vectors. The intelligent simulation module is communicatively connected to the data fusion module. It is used to construct a simulation model driven by a combination of data and physical rules based on fire identification results and core feature vectors, combined with preset deep underground space parameters. The model dynamically simulates the development of the fire and outputs the simulation data. The decision-making linkage module is communicatively connected to the intelligent simulation module and the multimodal perception module. It is used to generate emergency response instructions based on fire identification results and situation simulation data, and to feed back risk area information to the multimodal perception module, driving the multimodal perception module to adjust the sampling strategy of the corresponding area. The protection module is connected to the multimodal perception module, data fusion module, intelligent inference module and decision linkage module respectively, and is used to provide power supply protection, communication protection and equipment self-diagnosis protection for the system in the harsh environment of deep underground. The data governance module is connected to the multimodal perception module, data fusion module, intelligent inference module, and decision linkage module, respectively, and is used to perform unified management and iterative optimization of data across the entire system.

2. The multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The multimodal sensing module includes a physical sensing unit, an optical sensing unit, a gas sensing unit, an acoustic sensing unit, and an environmental auxiliary unit. The physical sensing unit includes a temperature sensor and a vibration sensor, used to collect temperature signals and structural vibration signals; The optical sensing unit includes a smoke sensor and a flame radiation sensor, used to collect smoke signals and flame spectral signals. The gas sensing unit includes characteristic gas sensors for CO, CO2, CH4, and H2S, used to collect the concentration signals of characteristic gases in a fire. The acoustic sensing unit includes an acoustic sensor array for collecting local abnormal noise signals; The environmental auxiliary unit includes a humidity sensor and an electromagnetic interference monitoring module, which are used to collect humidity signals and electromagnetic interference signals and output environmental calibration data. The multimodal sensing module transmits the collected temperature signal, smoke signal, flame spectrum signal, characteristic gas concentration signal, abnormal noise signal, vibration signal, humidity signal, and electromagnetic interference signal as multidimensional physical and chemical signals to the data fusion module.

3. The multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces according to claim 1, characterized in that: The data fusion module includes a preprocessing unit, a feature extraction unit, a two-layer fusion unit, and an early warning unit; The preprocessing unit is used to perform noise reduction processing on multi-dimensional physical and chemical signals and remove invalid data using the Kalman filter algorithm, realize the synchronization and alignment of multi-source data through timestamps, and receive environmental interference calibration parameters transmitted by the protection module for environmental compensation. The feature extraction unit is used to extract time-domain features, frequency-domain features, and time-frequency-domain features for different data types, and to perform dimensionality reduction on the multi-dimensional features through a convolutional neural network to output the core feature vector. The dual-layer fusion unit includes a feature-level fusion subunit and a decision-level fusion subunit; The feature-level fusion subunit adopts a dynamic weighting algorithm, which adjusts the weight coefficients according to the real-time reliability of each sensor, and performs weighted fusion of multi-source core features to generate a unified feature vector. The feature-level fusion subunits are weighted and fused according to the following formula; ； in, The unified feature vector after fusion For the first The core feature vector extracted by each sensor For the first The fusion weighting coefficients of each sensor, The total number of sensors participating in the fusion; The weighting coefficient Based on the real-time reliability of the sensor Determine using the following formula: ； ； in, For the first The initial confidence level of each sensor, Given the current electromagnetic interference intensity, Given the current ambient humidity, This represents the current dust concentration. , and This is the preset environmental interference attenuation coefficient; The decision-level fusion subunit, based on a Bayesian inference network, calculates the probability of fire occurrence using the following formula: ； in, , and These represent three states: no fire, suspected fire, and confirmed fire, respectively. Unified feature vector, For the first The prior probability of a fire state is given by the fact that in fire state... The following feature vectors appear The conditional probability; The advanced early warning unit is used to extrapolate the trend of multi-sensor time-series data using a long short-term memory network or a Transformer model. Before the signal reaches the alarm threshold, it outputs a risk trend warning based on the rate of change and acceleration, forming a four-level warning gradient: trend warning, suspected fire, confirmed fire, and fire expansion.

4. The multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The intelligent simulation module includes a basic database unit, a hybrid-driven simulation unit, a dynamic update unit, a digital twin visualization unit, a structural coupling analysis unit, and a multi-hazard coupling assessment unit. The basic database unit is pre-set with structural parameters, ventilation parameters, combustible material types and calorific values, and safety facility location information for deep underground spaces; The hybrid-driven inference unit includes a data-driven subunit and a physical rule subunit; The data-driven subunit is based on a long short-term memory network and uses historical fire data and current fused data output by the data fusion module as input to predict the fire development trend. The physical rules subunit introduces a computational fluid dynamics model, which combines ventilation characteristics, heat conduction equations, and mass conservation equations to correct the output results of the data-driven subunit. The physical rules constrain the output boundary of the data model, and the data-driven model supplements the scenario adaptation of the physical rules. The dynamic update unit is used to receive the latest sensing data and update the inference results at preset time intervals, and adjust the model parameters in a timely manner when environmental parameters change. The digital twin visualization unit is used to construct a three-dimensional digital twin model based on building information model data or geographic information system data of deep underground space, and to map the situation simulation data to the three-dimensional digital twin model in real time for three-dimensional visualization display. The three-dimensional visualization display includes at least the dynamics of fire spread, dynamics of smoke diffusion, temperature field distribution and personnel location information. The structural coupling analysis unit is used to integrate distributed fiber optic sensors or vibration sensors to collect thermal stress and deformation data generated by fire on underground space structures, establish a thermo-mechanical coupling model from temperature field distribution to thermal stress to structural deformation to bearing capacity attenuation, and output collapse risk assessment results. The formula for calculating the thermo-mechanical coupling model using the structural coupling analysis unit is as follows: ； in, This represents the structural bearing capacity attenuation value. The initial bearing capacity of the structure, for The highest temperature value in the fire temperature field distribution at any given moment. The thermal damage coefficient of the material. The moment the fire started. The current moment; The formula for calculating the collapse risk using the structural coupling analysis unit is as follows: ； in, The collapse risk level value. Here, represents the structural bearing capacity attenuation value, and represents the initial structural bearing capacity. This is the structural safety redundancy factor; The multi-hazard coupling assessment unit is used to cross-couple the gas leakage diffusion model, hydraulic model, and power grid fault model with fire simulation to establish a coupling relationship model between hazards and output secondary disaster risk assessment results. The situation simulation data includes at least the fire spread rate, smoke diffusion path, temperature field distribution, toxic gas concentration distribution, structural collapse risk level, and secondary disaster risk assessment results.

5. The multimodal fusion fire early detection and intelligent situation simulation system for deep underground spaces according to claim 1, characterized in that: The decision-making linkage module includes a hierarchical early warning unit, an intelligent decision generation unit, a closed-loop feedback unit, a multi-agent collaborative fire extinguishing unit, a dynamic evacuation unit, and a response effect evaluation unit. The graded early warning unit is used to trigger corresponding level early warning signals and link alarm devices based on fire identification results and situation simulation data. The early warning signals include level one, level two, level three and level four early warnings, which correspond to four states: trend warning, suspected fire, confirmed fire and fire expansion, respectively. The intelligent decision generation unit is used to generate emergency response plans based on situation simulation data. The emergency response plan includes at least a fire extinguishing plan, an evacuation plan, and a rescue plan. The fire extinguishing plan includes information on the best fire extinguishing location, the evacuation plan includes information on the best escape route, and the rescue plan includes information on key rescue areas. The closed-loop feedback unit is used to feed back the high-risk area information in the situation simulation data to the multimodal perception module, and drive the sensor corresponding to the high-risk area to increase the sampling frequency from the first frequency to the second frequency, where the second frequency is higher than the first frequency. The closed-loop feedback unit determines the second frequency based on the risk level value R of the high-risk area according to the following formula: ； in, For the first frequency, For the second frequency, For high-risk areas, the risk level value R ranges from 0 to 1. This is the preset frequency adjustment coefficient.

6. A multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 5, characterized in that: The multi-agent collaborative fire extinguishing unit is used to incorporate fixed fire extinguishing devices and mobile fire extinguishing equipment into a unified scheduling framework, generate collaborative fire extinguishing plans based on situational simulation data, and allocate tasks, timing coordination and coverage areas for each fire extinguishing unit. The dynamic evacuation unit is used to integrate UWB positioning technology or Bluetooth AOA positioning technology to obtain the location coordinates of trapped personnel and rescuers. Combined with the smoke diffusion path and temperature field distribution in the situation simulation data, it uses a deep reinforcement learning algorithm to calculate the escape path and outputs path guidance through smart terminals and electronic signs. The emergency response effectiveness evaluation unit is used to continuously monitor changes in fire field parameters during the execution of the emergency response plan, establish emergency response effectiveness evaluation indicators, and trigger the correction of the emergency response plan when the evaluation results do not meet the preset conditions.

7. The multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The protection module includes a dual-power supply unit, a dual-mode communication unit, an emergency self-organizing network unit, a network outage recovery unit, and a system self-diagnosis unit; The dual power supply unit adopts a power supply mode combining mains power and backup lithium battery pack. The lithium battery pack has a capacity of not less than 10kWh and can support the core equipment to work continuously for no less than 8 hours. The dual-mode communication unit adopts a dual-mode design of industrial Ethernet and 433MHz low-frequency wireless communication, and the shielding effectiveness of the communication module is not less than 40dB. The emergency self-organizing network unit is used to enable each detection node to form a temporary multi-hop Mesh communication network and transmit key alarm information when the main communication link is interrupted. The network outage recovery unit is used to cache lightweight inference models and decision rules locally on the edge side, run independently when the cloud is disconnected, and synchronize the local data during the network outage after communication is restored. The system self-diagnostic unit is used to monitor the operating status of each layer of equipment in real time, and to trigger a fault alarm and start redundant equipment when equipment fails.

8. The multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The data governance module includes a weighted adaptive learning unit and a feature spectrum library storage unit; The weighted adaptive learning unit is used to monitor the current environmental status in real time, including dust concentration, humidity level and electromagnetic interference intensity, and is used to combine historical false alarm data and missed alarm data to adjust the weight coefficients of each sensor in feature-level fusion through an online learning algorithm. The characteristic spectral library storage unit is used to store the pyrolysis gas characteristic spectral library for typical combustibles in deep space, including coal, cable insulation materials, oil and gas pipeline leaks, wood, and rubber.

9. A multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The data governance module also includes a federated learning transfer unit; The federated learning transfer unit is used to upload the gradient parameters of the local model trained on each deployment node to the cloud coordination server for collaborative training of the global model, provided that the local data on each deployment node does not leave the domain. The federated learning transfer unit is also used to fine-tune the global model when deploying in new scenarios, and to continuously receive real-world data from each node to update the global model.

10. A multimodal fusion fire early detection and intelligent situational simulation system for deep underground spaces according to claim 1, characterized in that: The data governance module also includes a blockchain evidence storage unit; The blockchain evidence storage unit is used to store the original multimodal perception data, fusion processing logs, inference process parameters, decision instruction records, and execution feedback results on the blockchain for evidence storage. The blockchain evidence storage unit adopts a consortium blockchain architecture, with data hashing and on-chain storage, and raw data distributed storage.

Images