A tunnel environment monitoring and disaster early warning method based on Internet of Things

Abstract

The application discloses a tunnel environment monitoring and disaster early warning method based on an Internet of Things, and comprises the following steps: arranging multiple types of Internet of Things sensor nodes in a tunnel to form an original sensor dataset; performing spatial division and category division on the original sensor dataset to form a partitioned sub-dataset; performing fuzzy processing on the partitioned sub-dataset to generate a regional-level fuzzy feature set; in a hierarchical space-time fuzzy neural network model, generating a regional-level fuzzy risk output set; performing defuzzy processing on the regional-level fuzzy risk output set and performing hierarchical judgment to obtain a risk level result; generating an early warning control instruction according to the risk level result; based on disposal feedback information, constructing an online feedback dataset, and updating the hierarchical space-time fuzzy neural network model. The application adopts a hierarchical space-time fuzzy neural network to realize tunnel environment monitoring and disaster early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel safety monitoring, and particularly relates to a tunnel environment monitoring and disaster early warning method based on Internet of Things. BACKGROUND

[0002] The existing tunnel environment monitoring system mainly uses sensors to collect temperature, humidity, gas concentration, smoke and wind speed data, and makes risk judgments based on fixed thresholds. This kind of method can realize basic environmental early warning, but it is insufficient in the fusion of multi-source data, and is prone to false positives or false negatives. In complex tunnel scenarios, environmental factors show multi-dimensional superposition effect, and single-point threshold judgment is difficult to accurately reflect the overall risk level.

[0003] Existing research attempts to introduce fuzzy logic methods to conduct risk assessment through fuzzy membership and rule-based reasoning, but most of them are static processing, lack of modeling of the evolution law of risk over time, and fail to reflect the risk diffusion relationship between regions. In dynamic scenarios such as fire or toxic gas diffusion, this kind of method often cannot meet the requirements of timeliness and accuracy.

[0004] With the development of artificial intelligence, some deep learning models have been applied to risk prediction, but most of the models only process numerical features, do not combine the explainability of fuzzy logic, and do not divide the tunnel into spaces to support subsequent spatial propagation analysis. In addition, existing models usually use fixed parameters, lack adaptive optimization mechanism based on feedback data, and are difficult to adapt to complex and variable operating environments.

[0005] Therefore, how to provide a tunnel environment monitoring and disaster early warning method based on Internet of Things is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a tunnel environment monitoring and disaster early warning method based on Internet of Things. The present application uses multi-type Internet of Things sensors, hierarchical spatio-temporal fuzzy neural networks and adaptive optimization mechanism to collect, spatially divide, fuzzily process, spatio-temporally model and risk level determine multi-source environmental data in the tunnel, and realizes dynamic updating of model parameters combined with feedback information, which has the advantages of high data fusion degree, accurate risk assessment, timely early warning response and strong system adaptability.

[0007] According to the tunnel environment monitoring and disaster early warning method based on Internet of Things, the following steps are included:

[0008] Multiple types of Internet of Things sensor nodes are arranged in the tunnel to collect environmental monitoring data in the tunnel to form an original sensor data set;

[0009] The original sensor data set is spatially divided and classified to form a partitioned sub-data set;

[0010] performing fuzzification processing on each item of data in the sub-data set of the partition to generate a regional-level fuzzy feature set;

[0011] inputting the regional-level fuzzy feature set into the hierarchical spatio-temporal fuzzy neural network model to generate a regional-level fuzzy risk output set;

[0012] performing defuzzification processing on the regional-level fuzzy risk output set to form a partition risk score set, and performing grading judgment to obtain a corresponding risk level result;

[0013] triggering a disaster early warning response mechanism according to the risk level result, generating an early warning control instruction, and issuing the early warning control instruction to a linkage device;

[0014] collecting disposal feedback information after the risk level result and the early warning control instruction are executed, constructing an online feedback data set, and updating the hierarchical spatio-temporal fuzzy neural network model.

[0015] Optionally, the original sensor data set includes temperature data collected by a temperature sensor, humidity data collected by a humidity sensor, carbon monoxide concentration data collected by a carbon monoxide sensor, methane concentration data collected by a methane sensor, smoke concentration data collected by a smoke sensor, wind speed data collected by a wind speed sensor, and vibration amplitude data collected by a vibration sensor.

[0016] Optionally, the spatial division and category division of the original sensor data set to form a sub-data set of the partition specifically includes:

[0017] dividing the original sensor data set according to the spatial positions of the sensor nodes, and dividing the tunnel into three spatial regions, i.e., a front region, a middle region and a rear region, according to the actual total length of the tunnel, wherein the data of the sensor nodes located between the starting point of the tunnel and the first division boundary is divided into the front region, the data located between the first division boundary and the second division boundary is divided into the middle region, and the data located between the second division boundary and the ending point of the tunnel is divided into the rear region, to finally obtain a spatial division result region set;

[0018] classifying the original sensor data set according to the functional types of the sensors, and classifying the data collected by the temperature sensor, the humidity sensor and the wind speed sensor into environmental data, classifying the data collected by the carbon monoxide sensor, the methane sensor and the smoke sensor into gas data, and classifying the data collected by the vibration sensor into structural data, to finally obtain a category division result category set;

[0019] combining the spatial division result region set and the category division result category set to construct a plurality of region-category pairs;

[0020] According to the region-class pair, the corresponding data in the original sensor data set is classified and arranged, and a partitioned sub-data set is generated as input data set for subsequent fuzzification processing.

[0021] Optionally, the fuzzification processing is performed on each data in the partitioned sub-data set to generate a region-level fuzzy feature set, specifically including:

[0022] The original monitoring values in the partitioned sub-data set are normalized to obtain a normalized data set;

[0023] On the basis of the normalized data set, a preset fuzzy membership function is selected for each type of sensor data, the fuzzy membership function includes a Gaussian function, a triangular function or a trapezoidal function, and according to the definition of the fuzzy membership function, the membership degree value corresponding to each piece of sensor data under different fuzzy sets is calculated, and finally an initial membership matrix is generated.

[0024] The initial membership matrix is arranged and aggregated according to the division results of the spatial region and the functional category, the membership degree values belonging to the same functional category in the same spatial region are collected together to form a corresponding region-level membership matrix, and finally a region-level membership matrix set covering all spatial regions and functional categories is obtained.

[0025] On the basis of the region-level membership matrix, a preset fuzzy rule is introduced, each membership degree value is substituted into the corresponding fuzzy rule condition, the activation strength of each fuzzy rule is calculated, and according to the combination relationship of different rules, the result is inferred to obtain the region-level fuzzy activation result corresponding to each spatial region and functional category.

[0026] The preset fuzzy rule is established based on the tunnel operation environment safety specification, historical accident statistical data and the change characteristics of various sensor monitoring indicators;

[0027] The region-level fuzzy activation result is arranged and combined in time sequence at each sampling time, the fuzzy activation results under the same functional category in the same spatial region are summarized at the current time step to form a corresponding region-level fuzzy feature vector.

[0028] The region-level fuzzy feature vectors corresponding to each spatial region and each functional category at the same time step are uniformly collected and set to obtain a complete region-level fuzzy feature set.

[0029] Optionally, the region-level fuzzy feature set is input into a hierarchical spatio-temporal fuzzy neural network model to generate a region-level fuzzy risk output set, specifically including:

[0030] The regional fuzzy feature set is input as input data into a hierarchical spatio-temporal fuzzy neural network model, the hierarchical spatio-temporal fuzzy neural network model comprising an input access module, a local fuzzy rule processing layer, a time sequence fuzzy memory unit, a graph rule propagation layer and an output aggregation module;

[0031] In the local fuzzy rule processing layer, a preset fuzzy rule mapping matrix is called based on the input data, the fuzzy feature vectors under the same spatial region and the same functional category are combined and reasoned, an initial risk activation value is calculated, and the activation value is summarized as an initial risk representation tensor;

[0032] In actual operation, the input regional fuzzy feature vector and the fuzzy rule mapping matrix are operated to obtain the activation strength of each fuzzy rule, and finally an initial risk representation tensor is output;

[0033] The initial risk representation tensor is input into the time sequence fuzzy memory unit and fused with the time sequence risk representation tensor of the previous time step, and a time sequence risk representation tensor is output;

[0034] In the graph rule propagation layer, a graph structure is constructed based on the spatial division result region set, the time sequence risk representation tensor is subjected to graph propagation calculation, and a spatial propagation risk representation tensor is generated;

[0035] The spatial propagation risk representation tensor is input into the output aggregation module, the risk representations of all regions are merged and processed, and a regional fuzzy risk output set is generated, which is used for defuzzification processing and risk level determination.

[0036] Optionally, the initial risk representation tensor is input into the time sequence fuzzy memory unit and fused with the time sequence risk representation tensor of the previous time step to output the time sequence risk representation tensor, specifically comprising:

[0037] The initial risk representation tensor is taken as the input of the current time step, and the time sequence risk representation tensor of the previous time step is called as the historical input;

[0038] A fuzzy update gate and a fuzzy forgetting gate are set in the time sequence fuzzy memory unit, the fuzzy update gate is generated according to the fuzzy membership degree distribution in the initial risk representation tensor of the current time step, and the fuzzy forgetting gate is generated according to the fuzzy activation degree in the time sequence risk representation tensor of the previous time step;

[0039] The fuzzy membership degree is calculated by a preset membership function from the numerical data collected by a sensor, and the membership degree value is obtained after the sensor numerical value is input into the membership function;

[0040] The fuzzy activation degree is calculated by combining the membership vector with the fuzzy rule mapping matrix, each fuzzy rule includes a premise condition and a conclusion, the premise condition is a combination of multi-dimensional membership, and the conclusion is a corresponding risk level label;

[0041] When the sensor membership satisfies the rule premise condition, the activation strength is calculated according to the preset operation operator, for example, when the minimum operator is used, the fuzzy activation degree takes the minimum membership value in the premise condition, and when the product operator is used, the fuzzy activation degree is the product of the membership degrees of the premise condition;

[0042] Based on the initial risk representation tensor of the current time step and the time sequence risk representation tensor of the previous time step, weighted operation and nonlinear transformation are performed to obtain a candidate state tensor;

[0043] Based on the fuzzy update gate, the fuzzy forget gate and the candidate state tensor, the time sequence risk representation tensor is updated.

[0044] Optionally, in the graph rule propagation layer, a graph structure is constructed based on the spatial division result region set, and graph propagation calculation is performed on the time sequence risk representation tensor to generate a spatial propagation risk representation tensor, which specifically includes:

[0045] A graph structure is constructed based on the spatial division result region set, wherein the node set is composed of each region in the spatial division result region set, and the edge set is determined according to the physical topological relationship and airflow channel relationship between the regions;

[0046] An adjacency matrix is obtained according to the graph structure, and a unit matrix is introduced based on the adjacency matrix, then a degree matrix is calculated, and the adjacency matrix is normalized according to the degree matrix, thereby forming a normalized adjacency matrix;

[0047] The time sequence risk representation tensor is input into the graph rule propagation layer as an initial node feature matrix, wherein each row corresponds to a spatial division result region, and each column corresponds to a risk feature dimension, the normalized adjacency matrix and the first layer trainable weight matrix are used for operation, and the first layer spatial propagation representation is obtained through a nonlinear activation function.

[0048] In the spatial propagation process, the first layer propagation representation is sequentially executed by iteration operation, each layer takes the propagation result of the previous layer as input, combines the normalized adjacency matrix and the corresponding trainable weight matrix for feature transmission and aggregation, and obtains a new propagation representation through a nonlinear activation function, and when the number of iterations reaches the preset number of layers, the spatial propagation risk representation tensor is formed.

[0049] Optionally, the defuzzification processing is performed on the region-level fuzzy risk output set to form a partition risk score set, and a hierarchical judgment is performed to obtain a corresponding risk level result, which specifically includes:

[0050] performing a defuzzification operation on the regional fuzzy risk output set, using a preset weighted average method or a centroid method, to obtain an initial risk score value of each region at the current time step by weighting and combining the membership values corresponding to the regional fuzzy risk output set and the risk output scalar;

[0051] normalizing the initial risk score value to uniformly map the risk score results of different regions into a standard interval of zero to one, so that the risk score values of the regions have comparability in the same dimension range, to obtain a standardized risk score value;

[0052] arranging the standardized risk score values of the regions in order according to the spatial division result region set to form a partition risk score set at the same time step;

[0053] comparing the partition risk score set with a preset risk level mapping table to determine the corresponding risk level label according to the threshold interval in which each risk score value is located, and to aggregate the risk level labels of all regions to form a risk level result set.

[0054] Optionally, the triggering of the disaster early warning response mechanism according to the risk level result, the generation of the early warning control instruction, and the issuance of the early warning control instruction to the linkage device specifically include:

[0055] taking the risk level result set as input, and calling a corresponding early warning rule table according to different risk level labels;

[0056] The early warning rule table is established according to tunnel operation safety specifications, historical accident disposal experience, and emergency plans of the management party;

[0057] The early warning rule table takes the risk level label as an index, and stores a corresponding early warning action set for each risk level, wherein the early warning action includes ventilation equipment start / stop, fire fighting system start, tunnel entrance control, and traffic signal adjustment operation;

[0058] When the risk level is low, the corresponding action is only an alarm prompt without triggering device action, when the risk level is medium, the corresponding action is to start the ventilation equipment to dilute harmful gas, when the risk level is high, the corresponding action is to close part of the tunnel entrance and start the fire fighting system, and when the risk level is extremely high, the corresponding action is to completely close the tunnel and simultaneously start all emergency linkage devices;

[0059] Matching the risk level labels of the regions with the early warning rule table one by one to determine the early warning action instruction set that needs to be executed by each region at the current time step, and aggregating and integrating the corresponding early warning action instruction sets of all regions to form a global early warning instruction set;

[0060] The global early warning instruction set is matched according to the binding relationship between the region and the device, and the abstract early warning action in the instruction set is mapped into a numerical or symbolic control parameter executable by the linkage device, and then the control vector is formed by arranging the control parameter according to the device number or category in sequence;

[0061] The early warning control instruction is generated according to the control vector, and is sent to the corresponding linkage device through the communication module, wherein the linkage device includes a ventilation device, a fire fighting system, a tunnel entrance control system and a traffic signal control system.

[0062] Optionally, the collection of risk level results and the disposal feedback information after the execution of the early warning control instruction construct an online feedback data set, and the hierarchical spatio-temporal fuzzy neural network model is updated, and the hierarchical spatio-temporal fuzzy neural network model is updated.

[0063] The collection of risk level results and the execution of the early warning control instruction obtain disposal feedback information, and are associated with the region-level fuzzy feature set, the risk representation tensor after spatial propagation, the partition risk score set and the control vector to construct an online feedback data set.

[0064] The target label vector is generated based on the disposal feedback information, and the corresponding confidence weight is assigned to the feedback result of each region to form a sample confidence weight vector.

[0065] The target label vector is generated based on the disposal feedback information; the disposal feedback information records the real execution situation after the risk level result and the early warning control instruction are sent, including the return result of the environmental monitoring data, the execution state of the linkage device and the manual intervention record.

[0066] The environmental monitoring data is compared with the preset safety threshold interval, and when the value exceeds the threshold, the risk label of the corresponding region is assigned as high risk, otherwise it is assigned as low risk.

[0067] The event disposal situation recorded in the feedback is corresponded with the preset risk level standard to determine the actual risk level of the region at the current time step.

[0068] The continuous index (such as gas concentration, smoke intensity) in the feedback data is normalized to form the risk intensity label.

[0069] In the above manner, each spatial division result region can obtain an actual risk label at the current time step, and the risk label combination of all regions forms a target label vector.

[0070] The target function is defined, the partition risk score set and the target label vector are compared region by region, the difference value between the predicted risk score and the actual risk label is calculated, and the difference value is weighted by combining the sample confidence weight vector to finally obtain a comprehensive loss result.

[0071] The gradient direction of the parameter set of the hierarchical spatio-temporal fuzzy neural network model is calculated based on the online feedback data set and the comprehensive loss result, and the fuzzy rule mapping matrix, the input weight matrix and the state weight matrix of the time sequence fuzzy memory unit and the bias vector thereof, and the trainable weight matrix of each layer of the graph rule propagation layer are adaptively adjusted according to the set learning rate, so that the updated parameter set is obtained.

[0072] The updated parameter set is written into the hierarchical spatio-temporal fuzzy neural network model, and an updated hierarchical spatio-temporal fuzzy neural network model is generated.

[0073] The beneficial effects of the present application are:

[0074] The present application can orderly organize the multi-source information in the complex tunnel environment, and distinguish the monitoring results of different regions, so as to provide unified data input for subsequent fuzzy processing and risk modeling.

[0075] In the data feature processing stage, the monitoring values in the partitioned sub-data set are converted into fuzzy membership degrees, and the regional-level fuzzy risk representation tensor is generated through the fuzzy rule mapping matrix. This design can effectively express the uncertainty in the sensor data, so that the risk analysis has the explainability and flexible expression ability of fuzzy logic. At the same time, the structured form of the fuzzy rule mapping matrix supports parallel computing of multi-source input, and allows optimization of rule weights during model training, so as to balance the transparency of rule reasoning and the dynamic adaptability of the model.

[0076] In terms of spatio-temporal modeling, the hierarchical spatio-temporal fuzzy neural network proposed by the present application introduces time sequence fuzzy memory units and graph rule propagation layers. The time sequence fuzzy memory unit dynamically models the regional-level fuzzy risk representation in the time dimension through fuzzy update gates and fuzzy forget gates, and can capture the evolution characteristics of the risk at consecutive time points. The graph rule propagation layer constructs a graph structure using the regional set of the tunnel space division result, its physical topology and airflow channel relationship, realizes the propagation and aggregation of the risk between adjacent regions, and can simulate the transmission law of fire and toxic gas diffusion in the tunnel. By combining time evolution and space diffusion, the modeling process of the present application breaks through the static and local limitations of existing methods, and realizes global and dynamic risk assessment results.

[0077] In terms of system optimization, the application establishes an adaptive optimization mechanism based on treatment feedback, collects risk level results, control vectors, early warning instructions and treatment information, constructs an online feedback dataset, and jointly updates the parameters of the fuzzy rule mapping matrix, the time sequence fuzzy memory unit and the graph rule propagation layer with a weighted loss function as the target. This mechanism enables the model to continuously correct the deviation between predicted and actual treatment effects in long-term operation, maintaining adaptability to environmental changes. BRIEF DESCRIPTION OF DRAWINGS

[0078] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not limit the application. In the drawings:

[0079] Figure 1 A flowchart of a tunnel environment monitoring and disaster warning method based on the Internet of Things is provided for the application;

[0080] Figure 2 A structure diagram of a hierarchical spatio-temporal fuzzy neural network model in a tunnel environment monitoring and disaster warning method based on the Internet of Things is provided for the application;

[0081] Figure 3 A time sequence fuzzy memory unit subgraph in a tunnel environment monitoring and disaster warning method based on the Internet of Things is provided for the application. DETAILED DESCRIPTION

[0082] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically illustrate the basic structure of the application, and therefore only show the components related to the application.

[0083] REFERENCE Figures 1-3 A tunnel environment monitoring and disaster warning method based on the Internet of Things, comprising the following steps:

[0084] Multiple types of Internet of Things sensor nodes are arranged in the tunnel to collect environmental monitoring data in the tunnel, forming an original sensor dataset;

[0085] The original sensor dataset is spatially and categorically divided to form a partitioned sub-dataset;

[0086] Each item of data in the partitioned sub-dataset is subjected to fuzzy processing to generate a regional-level fuzzy feature set;

[0087] The regional-level fuzzy feature set is input into a hierarchical spatio-temporal fuzzy neural network model to generate a regional-level fuzzy risk output set;

[0088] Performing deblurring processing on the regional-level fuzzy risk output set to form a partition risk score set, and performing hierarchical judgment to obtain a corresponding risk level result;

[0089] Triggering a disaster early warning response mechanism according to the risk level result, generating an early warning control instruction, and issuing the early warning control instruction to a linkage device;

[0090] Collecting disposal feedback information after the risk level result and the early warning control instruction are executed, constructing an online feedback data set, and updating the hierarchical spatiotemporal fuzzy neural network model.

[0091] In the embodiment, the multiple types of Internet of Things sensor nodes are arranged in the tunnel to collect environmental monitoring data in the tunnel to form an original sensor data set, which specifically includes:

[0092] A plurality of Internet of Things sensor nodes are arranged in the tunnel according to a preset spatial interval, and the Internet of Things sensor nodes specifically include a temperature sensor, a humidity sensor, a carbon monoxide sensor, a methane sensor, a smoke sensor, a wind speed sensor, and a vibration sensor to ensure multidimensional monitoring of the environmental state in the tunnel.

[0093] Each Internet of Things sensor node collects data of the tunnel environment in a set sampling period, and the collected data includes temperature data obtained by the temperature sensor, humidity data obtained by the humidity sensor, carbon monoxide concentration data obtained by the carbon monoxide sensor, methane concentration data obtained by the methane sensor, smoke concentration data obtained by the smoke sensor, wind speed data obtained by the wind speed sensor, and vibration amplitude data obtained by the vibration sensor.

[0094] The data output by all sensors at their respective sampling times are marked according to the sensor node number and the sensor category to form an original sensor data set, and the original sensor data set covers multi-source environmental monitoring data collected by all sensor nodes at different times in the entire tunnel space.

[0095] The original sensor data set is transmitted to an edge computing device or a remote server through a wireless communication module, and the wireless communication module supports LoRa, NB-IoT, or 4G / 5G communication modes to ensure real-time uploading and centralized storage of the original sensor data set.

[0096] In the embodiment, the original sensor data set is divided into a partition sub-data set, which specifically includes:

[0097] The original sensor data set is divided according to the spatial position of the sensor node, and the tunnel is equally divided into a front section region, a middle section region and a rear section region along the length direction according to the actual total length of the tunnel, wherein the data of the sensor node located between the starting point of the tunnel and the first division boundary is divided into the front section region, the data located between the first division boundary and the second division boundary is divided into the middle section region, and the data located between the second division boundary and the ending point of the tunnel is divided into the rear section region, and finally a spatial division result region set is obtained;

[0098] The original sensor data set is classified according to the functional type of the sensor, the data collected by the temperature sensor, the humidity sensor and the wind speed sensor is classified as environmental data, the data collected by the carbon monoxide sensor, the methane sensor and the smoke sensor is classified as gas data, and the data collected by the vibration sensor is classified as structural data, and finally a category division result category set is obtained;

[0099] The spatial division result region set and the category division result category set are combined to construct a plurality of region-category pairs;

[0100] According to each region-category pair, the corresponding data in the original sensor data set is classified and arranged to generate a partitioned sub-data set, which is used as an input data set for subsequent fuzzification processing.

[0101] In the embodiment, the fuzzification processing is performed on each data in the partitioned sub-data set to generate a region-level fuzzy feature set, which specifically includes:

[0102] The original monitoring values in the partitioned sub-data set are normalized to map data of different dimensions into the interval [0, 1] to obtain a normalized data set;

[0103] On the basis of the normalized data set, a preset fuzzy membership function is selected for each type of sensor data, the fuzzy membership function includes a Gaussian function, a triangular function or a trapezoidal function, and according to the definition of the fuzzy membership function, the membership degree value corresponding to each piece of sensor data under different fuzzy sets is calculated, and finally an initial membership matrix is generated.

[0104] The membership value calculation specifically includes inputting the normalized value of each piece of sensor data into the fuzzy membership function, and calculating the membership value under different fuzzy sets according to the curve shape of the function;

[0105] When the triangular function is used, the membership value increases in proportion in the interval from the left edge to the vertex of the triangle, and decreases in proportion in the interval from the vertex to the right edge;

[0106] When a trapezoidal function is adopted, the membership value of the normalized value of the strip sensor data remains one in the flat-top interval, and increases or decreases proportionally in the two side slope intervals;

[0107] When a Gaussian function is adopted, the membership value continuously changes according to the distance between the normalized value of the strip sensor data and the center position in a Gaussian mode;

[0108] The initial membership matrix is sorted and aggregated according to the division results of the spatial regions and the functional categories, the membership values of the same functional category belonging to the same spatial region are collected together to form a corresponding regional-level membership matrix, wherein each regional-level membership matrix is used to represent the membership distribution of the sensor data of a functional category in each fuzzy set in a spatial region, and finally a set of regional-level membership matrices covering all spatial regions and functional categories is obtained;

[0109] On the basis of the regional-level membership matrix, a preset fuzzy rule is introduced, each membership value is substituted into the corresponding fuzzy rule condition, the activation strength of each fuzzy rule is calculated, and the result is inferred according to the combination relationship of different rules to obtain a regional-level fuzzy activation result corresponding to each spatial region and functional category;

[0110] The calculation of the activation strength of the fuzzy rule specifically includes that when the antecedent condition of the rule contains multiple membership values, the minimum value among all the membership values is taken as the activation strength in the minimum mode, which is used to represent the degree of simultaneous establishment of multiple conditions, and in the product mode, all the membership values are multiplied in order, which is used to represent the strength and weakness of the common enhancement or weakening of multiple conditions on the triggering of the rule, when the rule contains only a single antecedent condition, the membership value of the condition is directly taken as the activation strength;

[0111] The preset fuzzy rule is established based on the tunnel operating environment safety specification, historical accident statistical data and the change characteristics of various sensor monitoring indicators, and is used to represent the corresponding relationship between different environmental states and risk levels;

[0112] The fuzzy rule adopts the IF-THEN form, takes the fuzzy membership of the sensor input as the condition variable, and takes the risk level as the output variable, specifically, the fuzzy rule can include the following examples:

[0113] When the membership of temperature is high and the membership of smoke is high, the output risk level is high;

[0114] When the membership of carbon monoxide concentration is medium and the membership of wind speed is low, the output risk level is medium;

[0115] When the membership of vibration amplitude is high and the membership of methane concentration is high, the output risk level is extremely high;

[0116] The region-level fuzzy activation results are sorted and combined in chronological order at each sampling time, and the fuzzy activation results of the same spatial region and the same functional category are summarized at the current time step to form a corresponding region-level fuzzy feature vector, which is used to represent the comprehensive risk characteristics of the region at the current time;

[0117] The region-level fuzzy feature vectors corresponding to each spatial region and each functional category at the same time step are uniformly collected and processed to obtain a complete region-level fuzzy feature set, which is used as input data for the subsequent hierarchical spatio-temporal fuzzy neural network model to support comprehensive modeling and risk assessment of the overall tunnel environment state.

[0118] In this embodiment, the input of the region-level fuzzy feature set into the hierarchical spatio-temporal fuzzy neural network model to generate a region-level fuzzy risk output set specifically includes:

[0119] The region-level fuzzy feature set is input as input data into the hierarchical spatio-temporal fuzzy neural network model, which includes an input access module, a local fuzzy rule processing layer, a time sequence fuzzy memory unit, a graph rule propagation layer, and an output aggregation module. The input access module is used to receive and format the region-level fuzzy feature set, and pass the formatted input data to the local fuzzy rule processing layer;

[0120] In the local fuzzy rule processing layer, based on the input data, a preset fuzzy rule mapping matrix is called to combine and reason the fuzzy feature vectors of the same spatial region and the same functional category, and an initial risk activation value is calculated. The activation value is summarized as an initial risk representation tensor, which is used to represent the local fuzzy risk response of each region at the current time step;

[0121] The fuzzy rule mapping matrix is used to express the preset fuzzy rules in matrix form to realize the corresponding relationship between the input fuzzy features and the risk output. The rows of the matrix represent fuzzy conditions composed of sensor input membership degrees, the columns represent risk level output categories, and the elements in the matrix are weight values between conditions and outputs. The weight values can be determined by expert experience, tunnel safety specifications, or accident case statistics, or can be adaptively updated through a backpropagation algorithm during training;

[0122] In actual operation, the input region-level fuzzy feature vector and the fuzzy rule mapping matrix are operated to obtain the initial activation of each fuzzy rule, and finally output the initial risk representation tensor. The matrix structure can realize parallel computing of multi-source sensor inputs and improve the efficiency of fuzzy rule reasoning, while supporting dynamic optimization of rule weights during model operation to balance the explainability of rules and the adaptive ability of the model;

[0123] inputting the initial risk representation tensor into a time sequence fuzzy memory unit and fusing the time sequence risk representation tensor of the previous time step to output a time sequence risk representation tensor;

[0124] In the graph rule propagation layer, a graph structure is constructed based on the spatial division result area set, and a graph propagation calculation is performed on the time sequence risk representation tensor to generate a spatial propagated risk representation tensor;

[0125] The spatial propagated risk representation tensor is input into an output aggregation module, and the risk representations of all areas are merged for processing to generate an area-level fuzzy risk output set, which is used for defuzzification processing and risk level determination.

[0126] In the embodiment, the inputting the initial risk representation tensor into a time sequence fuzzy memory unit and fusing the time sequence risk representation tensor of the previous time step to output a time sequence risk representation tensor specifically includes:

[0127] The initial risk representation tensor is taken as the input of the current time step, and the time sequence risk representation tensor of the previous time step is called as the historical input;

[0128] A fuzzy update gate and a fuzzy forgetting gate are set in the time sequence fuzzy memory unit, the fuzzy update gate is generated according to the fuzzy membership degree distribution in the initial risk representation tensor of the current time step, and is used to adjust the influence degree of the current input in the output result, the fuzzy forgetting gate is generated according to the fuzzy activation degree in the time sequence risk representation tensor of the previous time step, and is used to control the reservation proportion of the historical state in the output result, thereby realizing the dynamic balance of the current fuzzy input and the historical fuzzy state;

[0129] The fuzzy membership degree is calculated by a preset membership function from the numerical data collected by the sensor, and the sensor numerical input is input into the membership function to obtain the membership degree value in the interval [0, 1];

[0130] The fuzzy activation degree is calculated by combining the membership degree vector and the fuzzy rule mapping matrix, each fuzzy rule includes the antecedent condition and the consequent conclusion, the antecedent condition is the combination of multi-dimensional membership degrees, and the consequent conclusion is the corresponding risk level label;

[0131] When the sensor membership degree satisfies the rule antecedent condition, the activation strength is calculated according to the preset operation operator, for example, when the minimum operator is used, the fuzzy activation degree takes the minimum membership degree value in the antecedent condition, and when the product operator is used, the fuzzy activation degree is the product of the membership degrees of the antecedent conditions;

[0132] Based on the initial risk representation tensor of the current time step and the time sequence risk representation tensor of the previous time step, a weighted operation and a nonlinear transformation are performed to obtain a candidate state tensor:

[0133] ;

[0134] wherein, is a candidate state tensor, is an initial risk representation tensor, is a time-series risk representation tensor of a previous time step, is an input weight matrix, used for linear transformation of the initial risk representation tensor to adjust the influence of each input fuzzy feature in the candidate state calculation, is a state weight matrix, used for linear transformation of the time-series risk representation tensor of the previous time step to adjust the influence of the historical fuzzy state in the candidate state calculation, is a bias vector, used to provide a translation term in the candidate state calculation process to correct the calculation result of the overall risk representation, is a hyperbolic tangent activation function;

[0135] Based on the fuzzy update gate, the fuzzy forget gate and the candidate state tensor, a time-series risk representation tensor is updated:

[0136] ;

[0137] wherein, is a time-series risk representation tensor, is a fuzzy update gate, is a fuzzy forget gate, is an element-wise multiplication operator, used for element-wise calculation in the tensor dimension to realize one-to-one correspondence between the gating weight and the state value.

[0138] In the embodiment, the graph propagation layer based on the spatial division result region set constructs a graph structure, performs graph propagation calculation on the time-series risk representation tensor, and generates a spatial propagation risk representation tensor, which specifically includes:

[0139] The graph structure is constructed based on the spatial division result region set, wherein the node set is composed of each region in the spatial division result region set, and the edge set is determined according to the physical topological relationship and airflow channel relationship between the regions, forming a graph structure for representing the association between regions;

[0140] An adjacency matrix is obtained according to the graph structure, and a unit matrix is introduced based on the adjacency matrix to ensure that each region retains its own characteristics during the propagation process, then a degree matrix is calculated and the adjacency matrix is normalized according to the degree matrix, thereby forming a normalized adjacency matrix, which is used as a weight matrix for spatial propagation;

[0141] The degree matrix comprises: summing all elements of each row in the adjacency matrix, taking the sum as the degree value of the region corresponding to the row, filling the degree value into the diagonal position of the degree matrix, and setting the elements not in the diagonal position to zero, so that the degree matrix records the connection number or connection strength of each region in a diagonal form;

[0142] The normalization processing of the adjacency matrix according to the degree matrix comprises scaling each non-zero element in the adjacency matrix according to the degree values corresponding to the row and column where the element is located, so that the regions with larger connection numbers will not cause the propagation result to be too large due to value accumulation in the propagation process, and the regions with smaller connection numbers will not be weakened in the propagation due to insufficient connection;

[0143] The time sequence risk representation tensor is input as an initial node feature matrix to the graph rule propagation layer, and the graph propagation calculation is performed, wherein each row corresponds to a spatial division result region, and each column corresponds to a risk feature dimension. The graph propagation calculation comprises performing operation on the normalized adjacency matrix and the first layer trainable weight matrix, and processing through a nonlinear activation function to obtain a first layer spatial propagation representation, which is used to represent the risk feature distribution of each region after introducing the spatial adjacency relationship.

[0144] In the spatial propagation process, the first layer propagation representation is sequentially subjected to iterative operation, each layer taking the propagation result of the previous layer as input, combining the normalized adjacency matrix and the corresponding trainable weight matrix to perform feature transmission and aggregation, and obtaining a new propagation representation through a nonlinear activation function. When the number of iterations reaches the preset number of layers, the risk representation tensor after spatial propagation is formed.

[0145] In the embodiment, the defuzzification processing of the region-level fuzzy risk output set to form the partition risk score set and the grading judgment to obtain the corresponding risk level result specifically comprises:

[0146] The defuzzification operation is performed on the region-level fuzzy risk output set, and a preset weighted average method or a centroid method is used to combine the membership values corresponding to the region-level fuzzy risk output set and the risk output scalar to obtain an initial risk score value of each region at the current time step. The initial risk score value is used to represent the comprehensive risk level of the region at the current time.

[0147] The initial risk score value is subjected to normalization processing to uniformly map the risk score results of different regions into a standard interval of zero to one, so that the risk score values of the regions have comparability in the same dimension range, and a standardized risk score value is obtained.

[0148] The standardized risk score values of each region are sequentially arranged according to the order of the spatial division result region set, to form a partition risk score set at the same time step, and the partition risk score set is used to collectively represent the risk levels of different regions at the time step;

[0149] The partition risk score set is compared with a preset risk level mapping table, a corresponding risk level label is determined according to a threshold interval in which each risk score value is located, and risk level label sets of all regions are collected to form a risk level result set.

[0150] In the embodiment, triggering a disaster early warning response mechanism according to the risk level result, generating an early warning control instruction, and issuing the early warning control instruction to a linkage device specifically includes:

[0151] The risk level result set is taken as an input, and a corresponding early warning rule table is called according to different risk level labels;

[0152] The early warning rule table is established according to a tunnel operation safety specification, historical accident disposal experience, and an emergency plan of a management party, and is used to establish a mapping relationship between the risk level result set and specific early warning actions;

[0153] The early warning rule table takes a risk level label as an index, and stores an early warning action set corresponding to each risk level, wherein the early warning action includes operations such as starting and stopping of a ventilation device, starting of a fire extinguishing system, tunnel entrance control, and traffic signal adjustment;

[0154] When the risk level is low, the corresponding action is only an alarm prompt without triggering a device action, when the risk level is medium, the corresponding action is to start a ventilation device to dilute harmful gas, when the risk level is high, the corresponding action is to close part of a tunnel entrance and start a fire extinguishing system in preparation, and when the risk level is extremely high, the corresponding action is to completely close the tunnel and simultaneously start all emergency linkage devices;

[0155] By calling the early warning rule table, the risk level result set can be quickly converted into a standardized early warning action set, to ensure automation and standardization of a process from risk identification to emergency disposal, and to support adjustment and expansion according to different scales of tunnels, device configurations, and management strategies;

[0156] The risk level labels of each region are matched with the early warning rule table one by one, to determine an early warning action instruction set that needs to be executed by each region at the current time step, and the early warning action instruction sets corresponding to all regions are collected and integrated to form a global early warning instruction set;

[0157] The global early warning instruction set is matched according to the binding relationship between the region and the device, and the abstract early warning action in the instruction set is mapped into a numerical or symbolic control parameter executable by each linkage device, and then the control vector is arranged in order according to the device number or category to form a control vector composed of multiple elements, each element corresponding to the control instruction of each linkage device at the current time step, so as to ensure that the execution results of different devices can be uniformly managed and synchronously issued;

[0158] The early warning control instruction is generated according to the control vector, and is issued to the corresponding linkage device through the communication module, wherein the linkage device includes a ventilation device, a fire fighting system, a tunnel entrance control system and a traffic signal control system.

[0159] In the embodiment, the collection of risk level results and the disposal feedback information after the execution of the early warning control instruction are used to construct an online feedback data set, and the hierarchical spatio-temporal fuzzy neural network model is updated.

[0160] The collection of risk level results and the execution of the early warning control instruction are used to obtain disposal feedback information, and the online feedback data set is constructed by associating the region-level fuzzy feature set, the risk representation tensor after spatial propagation, the partition risk score set and the control vector.

[0161] A target label vector is generated based on the disposal feedback information, and a corresponding confidence weight is assigned to the feedback result of each region to form a sample confidence weight vector, wherein the target label vector is used to represent the actual risk intensity of each region at the current time step, and the sample confidence weight is used to reflect the credibility of the feedback data in the model optimization process.

[0162] The target label vector is generated based on the disposal feedback information; the disposal feedback information records the actual execution after the risk level result and the early warning control instruction are issued, including the return result of the environmental monitoring data, the execution state of the linkage device and the manual intervention record.

[0163] The environmental monitoring data is compared with the preset safety threshold interval, and when the value exceeds the threshold, the risk label of the corresponding region is assigned as high risk, otherwise it is assigned as low risk.

[0164] The event disposal situation recorded in the feedback is compared with the preset risk level standard to determine the actual risk level of the region at the current time step.

[0165] The continuous indicators (such as gas concentration and smoke intensity) in the feedback data are normalized to form the risk intensity label.

[0166] In the above manner, each spatial division result region at the current time step can obtain an actual risk label, and the risk labels of all regions are combined to form a target label vector.

[0167] In the model adaptive optimization process, a target function is defined, the partition risk score set is compared with the target label vector region by region, the difference value between the predicted risk score and the actual risk label is calculated, and the difference value is weighted by combining the sample confidence weight vector to obtain a comprehensive loss result, which is used to measure the overall prediction deviation of the model in a given time window and as the basis for parameter updating;

[0168] Based on the online feedback data set and the comprehensive loss result, the gradient direction of the hierarchical spatio-temporal fuzzy neural network model parameter set is calculated, and the fuzzy rule mapping matrix, the input weight matrix and the state weight matrix of the time sequence fuzzy memory unit and the bias vector thereof, and the trainable weight matrix of each layer of the graph rule propagation layer are adaptively adjusted according to the set learning rate, so as to obtain the updated parameter set;

[0169] The calculation of the gradient direction of the hierarchical spatio-temporal fuzzy neural network model parameter set includes, after obtaining the comprehensive loss result, judging the influence change of each type of parameter in the model on the comprehensive loss result when the value is slightly increased and slightly reduced, respectively, when the increase of the parameter causes the comprehensive loss result to rise, the change is regarded as an adverse direction, when the increase of the parameter causes the comprehensive loss result to decrease, the change is regarded as a beneficial direction, and the direction in which the parameter should be adjusted to make the comprehensive loss result decrease is determined, thereby obtaining the gradient direction of each parameter for updating;

[0170] The updated parameter set is written into the hierarchical spatio-temporal fuzzy neural network model to generate an updated hierarchical spatio-temporal fuzzy neural network model.

[0171] Embodiment 1:

[0172] In order to verify the feasibility of the application in implementation, the application is applied to the environmental monitoring and disaster warning scene of a mountainous highway tunnel. The tunnel is about 3 kilometers long, located in a humid area, with poor internal air flow, and there is a risk of vehicle exhaust accumulation, high humidity caused by water seepage, and increasing concentration of harmful gases when ventilation is poor. In the existing monitoring mode, the tunnel operator mainly relies on traditional threshold alarms, which trigger an alarm when the monitoring value of a single-point sensor exceeds the preset threshold. However, there are often problems such as high false alarm rate, delayed response, and inability to reflect regional risk diffusion. For example, when the smoke concentration in a certain area increases, the system may issue multiple redundant alarms, but it cannot predict the diffusion path of the smoke in the tunnel, resulting in delayed deployment of emergency personnel.

[0173] In this scenario, the Internet of Things tunnel environment monitoring and disaster early warning method of the application forms an original sensor data set by arranging temperature sensors, humidity sensors, carbon monoxide sensors, methane sensors, smoke sensors, wind speed sensors and vibration sensors, a total of 120 nodes, in the front, middle and rear sections of the tunnel. The data is first divided into three areas according to the spatial position of the tunnel, and is divided into environment, gas and structure classes according to function, to obtain a partitioned sub-data set. In the data processing process, the original values collected by the sensors in each area are converted into membership vectors through fuzzy processing, and the area-level fuzzy risk representation results are calculated through a fuzzy rule mapping matrix and input into a hierarchical spatio-temporal fuzzy neural network model for processing. This model not only reflects the change law of risk with time through the time sequence fuzzy memory unit, but also simulates the diffusion trend of smoke and harmful gases in space through the graph rule propagation layer combined with the relationship of the tunnel airflow channel, thereby realizing dynamic modeling of the overall risk.

[0174] In actual operation, when the smoke concentration in the middle section gradually increases to 2.3 mg per cubic meter, and the carbon monoxide concentration rises to 85 ppm per cubic meter, the system of the application completes data acquisition, fuzzy processing and model calculation within 60 seconds, and obtains that the risk level of the middle section is high, the risk level of the front section is medium, and the risk level of the rear section is low. The system automatically generates a global warning instruction set and sends it to the ventilation system and traffic signal control system in the tunnel, starts forced ventilation and switches the tunnel entrance signal light to red, prohibiting new vehicles from entering. Compared with the traditional threshold alarm, the false alarm rate of the application under similar conditions is reduced to 2.7%, which is lower than the 8.5% of the traditional system; the response time is shortened to within 60 seconds, while the average response time of the traditional system is 180 seconds; the risk diffusion prediction accuracy rate reaches more than 94%, providing reliable support for emergency disposal.

[0175] Table 1 Comparison data of the application method and the traditional method in tunnel environment monitoring and early warning

[0176] Test conditions Indicators Conventional method values Invention method values Mid-section smoke concentration increased to 2.3 mg / m³ Response time (seconds) 180 60 Mid-section carbon monoxide concentration increased to 85 ppm False alarm rate (%) 8.5 2.7 Multi-source data comprehensive analysis Risk diffusion prediction accuracy (%) 71.2 94.3 Regional risk level identification (early / mid / late section) Output accuracy (%) 75.6 96.8 Emergency disposal scheme issuance Instruction coverage rate (%) 68.9 95.1 Model adaptability (corrected within 1 hour after environmental change) Parameter update success rate (%) Not available 92.5

[0177] As can be seen from the above table, the present application shows significant improvement in multiple key performance indicators of tunnel environment monitoring and disaster early warning compared with the traditional method. First, in terms of response time, when the middle section smoke concentration rises to 2.3 mg / m³, the traditional method needs an average of 180 seconds to complete data collection, judgment and alarm, while the present application shortens the response time to within 60 seconds through the rapid calculation of the hierarchical spatio-temporal fuzzy neural network, and the response efficiency is improved by more than 2 times, enhancing the timeliness of disposal in emergency. Secondly, in terms of false alarm rate, when the carbon monoxide concentration rises to 85 ppm, the false alarm rate of the traditional method is 8.5%, while that of the present application is only 2.7%, indicating that the introduction of fuzzy logic and multi-source data fusion in complex environment can more accurately distinguish between risk signals and environmental fluctuations, effectively reducing invalid alarms.

[0178] In terms of risk diffusion prediction ability, the accuracy of the present application method reaches 94.3%, while that of the traditional method is only 71.2%. This result shows that the present application, by combining the tunnel space topology relationship and the airflow channel modeling, enables the system to more accurately simulate the diffusion law of smoke and harmful gases in the tunnel, thereby providing more reliable risk area distribution reference for emergency personnel. In terms of regional risk level identification accuracy, the present application reaches 96.8%, which is much higher than the 75.6% of the traditional method, ensuring the scientificity and rationality of risk level division and avoiding the misallocation of emergency resources due to inaccurate judgment.

[0179] In addition, in terms of the coverage rate of emergency disposal scheme, the effective coverage rate of the traditional method is 68.9%, while that of the present application reaches 95.1%, which can ensure that the linkage devices such as ventilation system, fire fighting system, entrance control and traffic signal control device are executed in coordination under the condition of early warning, thereby realizing systematic linkage. Finally, in terms of model adaptability, the traditional method lacks parameter updating mechanism and cannot be adjusted according to real-time environmental changes, while the parameter updating success rate of the present application reaches 92.5% within 1 hour after the environment changes, which reflects the advantages of self-adaptive optimization mechanism. In summary, the present application not only surpasses the traditional method in response speed and accuracy, but also shows systematic improvement in risk prediction, linkage control and self-learning ability, providing more efficient and reliable technical support for disaster early warning in tunnel environment.

[0180] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A tunnel environment monitoring and disaster warning method based on Internet of Things, characterized in that, The method comprises the following steps: a plurality of types of Internet of Things sensor nodes are arranged in the tunnel to collect environmental monitoring data in the tunnel and form an original sensor data set; the original sensor data set is spatially and categorically divided to form a partitioned sub-data set; each item of data in the partitioned sub-data set is subjected to fuzzification processing to generate a regional fuzzy feature set; the regional fuzzy feature set is input into a hierarchical spatio-temporal fuzzy neural network model to generate a regional fuzzy risk output set; the regional fuzzy risk output set is subjected to defuzzification processing to form a partitioned risk score set, and grading is performed to obtain a corresponding risk level result; a disaster early warning response mechanism is triggered according to the risk level result, a warning control instruction is generated, and the warning control instruction is issued to a linkage device; disposal feedback information after the risk level result and the warning control instruction are executed is collected to build an online feedback data set, and the hierarchical spatio-temporal fuzzy neural network model is updated; the original sensor data set is spatially divided according to the spatial positions of the sensor nodes to obtain a spatial division result area set; the original sensor data set is categorized according to the functional types of the sensors to obtain a category division result category set; the spatial division result area set and the category division result category set are combined to construct a plurality of region-category pairs; the corresponding data in the original sensor data set is classified and arranged according to each region-category pair to generate the partitioned sub-data set; the regional fuzzy feature set is input into the hierarchical spatio-temporal fuzzy neural network model, and the hierarchical spatio-temporal fuzzy neural network model comprises an input access module, a local fuzzy rule processing layer, a time sequence fuzzy memory unit, a graph rule propagation layer, and an output aggregation module; in the local fuzzy rule processing layer, a preset fuzzy rule mapping matrix is called based on the input data, the fuzzy feature vectors under the same spatial region and the same functional category are combined and reasoned to obtain an initial risk representation tensor; the initial risk representation tensor is input into the time sequence fuzzy memory unit and fused with the time sequence risk representation tensor at the previous time step to output a time sequence risk representation tensor; in the graph rule propagation layer, a graph structure is constructed based on the spatial division result area set, and the time sequence risk representation tensor is subjected to graph propagation calculation to generate a spatially propagated risk representation tensor; the spatially propagated risk representation tensor is input into the output aggregation module, and the risk representations of all regions are merged to generate the regional fuzzy risk output set. The original sensor data set comprises temperature data collected by a temperature sensor, humidity data collected by a humidity sensor, carbon monoxide concentration data collected by a carbon monoxide sensor, methane concentration data collected by a methane sensor, smoke concentration data collected by a smoke sensor, wind speed data collected by a wind speed sensor, and vibration amplitude data collected by a vibration sensor. ​ 2.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, ​ 3.The tunnel environment monitoring and disaster warning method based on the Internet of Things according to claim 1, characterized in that, The fuzzy processing is performed on each item of data in the partitioned sub-data set to generate a region-level fuzzy feature set, specifically comprising: The original monitoring values in the partitioned sub-data set are normalized to obtain a normalized data set; On the basis of the normalized data set, a preset fuzzy membership function is selected for each type of sensor data, the membership degree value corresponding to each piece of sensor data under different fuzzy sets is calculated, and an initial membership matrix is generated; The initial membership matrix is arranged and aggregated according to the division results of the spatial regions and the functional categories to form a corresponding region-level membership matrix; On the basis of the region-level membership matrix, a preset fuzzy rule is introduced, each membership degree value is substituted into the corresponding fuzzy rule condition, the activation strength of each fuzzy rule is calculated, and a region-level fuzzy activation result is obtained; The region-level fuzzy activation result is arranged and combined in time sequence at each sampling time to form a region-level fuzzy feature vector; The region-level fuzzy feature vectors corresponding to each spatial region and each functional category at the same time step are uniformly collected and set to obtain a region-level fuzzy feature set. 4.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, The initial risk representation tensor is input into the time sequence fuzzy memory unit, and the time sequence risk representation tensor of the last time step is fused to output the time sequence risk representation tensor, specifically comprising: The initial risk representation tensor is input as the input of the current time step, and the time sequence risk representation tensor of the last time step is called as the historical input; In the time sequence fuzzy memory unit, a fuzzy update gate and a fuzzy forgetting gate are set, the fuzzy update gate is generated according to the fuzzy membership degree distribution in the initial risk representation tensor of the current time step, and the fuzzy forgetting gate is generated according to the fuzzy activation degree in the time sequence risk representation tensor of the last time step; Based on the initial risk representation tensor of the current time step and the time sequence risk representation tensor of the last time step, weighted operation and nonlinear transformation are performed to obtain a candidate state tensor; Based on the fuzzy update gate, the fuzzy forgetting gate and the candidate state tensor, the time sequence risk representation tensor is updated. 5.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, In the graph rule propagation layer, a graph structure is constructed based on the spatial division result region set, the graph propagation calculation is performed on the time sequence risk representation tensor, and the spatial propagation risk representation tensor is generated, specifically comprising: A graph structure is constructed based on the spatial division result region set, wherein the node set is composed of each region in the spatial division result region set, and the edge set is determined according to the physical topological relationship and airflow channel relationship between the regions; An adjacency matrix is obtained according to the graph structure, a unit matrix is introduced based on the adjacency matrix, a degree matrix is calculated, and the adjacency matrix is normalized according to the degree matrix to form a normalized adjacency matrix; The time sequence risk representation tensor is input as an initial node feature matrix into the graph rule propagation layer, the normalized adjacency matrix and the first layer trainable weight matrix are operated, and the first layer spatial propagation representation is obtained through a nonlinear activation function. In the spatial propagation process, the first layer propagation representation is sequentially subjected to iterative operation, each layer takes the propagation result of the last layer as input, and when the number of iterations reaches the preset number of layers, the spatial propagation risk representation tensor is formed. 6.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, The de-fuzzing processing is performed on the regional fuzzy risk output set to form a partition risk score set, and a hierarchical judgment is performed to obtain a corresponding risk level result, and the specific process includes: The de-fuzzing operation is performed on the regional fuzzy risk output set, the membership value corresponding to each regional fuzzy risk output set is weighted and combined with the risk output scalar to obtain an initial risk score value of each region; The initial risk score value is normalized to uniformly map the risk score results of different regions into a standard interval to obtain a standardized risk score value; The standardized risk score values of each region are sequentially arranged according to the order of the spatial division result region set to form a partition risk score set; The partition risk score set is compared with a preset risk level mapping table, the corresponding risk level label is determined according to the threshold interval in which each risk score value is located, and the risk level label of all regions is collected to form a risk level result set. 7.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, The disaster early warning response mechanism is triggered according to the risk level result, the early warning control instruction is generated, and the early warning control instruction is sent to the linkage device, and the specific process includes: The risk level result set is taken as an input, and a corresponding early warning rule table is called according to different risk level labels; The risk level labels of each region are matched with the early warning rule table one by one to determine the early warning action instruction set of each region, and the early warning action instruction sets corresponding to all regions are collected and integrated to form a global early warning instruction set; The global early warning instruction set is matched according to the binding relationship between the region and the device to form a control vector; The early warning control instruction is generated according to the control vector, and is sent to the corresponding linkage device through the communication module. 8.The tunnel environment monitoring and disaster early warning method based on the Internet of Things according to claim 1, characterized in that, The risk level result and the disposal feedback information after the execution of the early warning control instruction are collected, an online feedback data set is constructed, and the hierarchical spatio-temporal fuzzy neural network model is updated, and the specific process includes: The execution of the risk level result set and the early warning control instruction is collected to obtain disposal feedback information, which is associated with the regional fuzzy feature set, the risk representation tensor after spatial propagation, the partition risk score set, and the control vector to construct an online feedback data set; A target label vector is generated based on the disposal feedback information, and a corresponding confidence weight is assigned to the feedback result of each region to form a sample confidence weight vector; A target function is defined, the partition risk score set is compared with the target label vector region by region, the difference value between the predicted risk score and the actual risk label is calculated, and the difference value is weighted combined with the sample confidence weight vector to obtain a comprehensive loss result; The gradient direction of the hierarchical spatio-temporal fuzzy neural network model parameter set is calculated based on the online feedback data set and the comprehensive loss result to obtain an updated parameter set; The updated parameter set is written into the hierarchical spatio-temporal fuzzy neural network model to generate an updated hierarchical spatio-temporal fuzzy neural network model.

Images