Tunnel disaster prevention safety risk prediction system and method and storage medium

Abstract

The invention discloses a tunnel disaster prevention safety risk prediction system and method and a storage medium, and the system comprises a space-time analysis module which is used for analyzing the spatial correlation of spatial feature data, obtaining spatial state data, and constructing an optimized tunnel three-dimensional grid, obtaining a first risk mode from the spatial feature data by optimizing the combination of a tunnel three-dimensional grid and a graph neural network; the time analysis module is used for analyzing the time sequence periodicity of the time characteristic data to obtain time sequence state data, constructing an improved long short-term memory network and obtaining a second risk mode from the time characteristic data through the improved long short-term memory network; and the risk prediction module is used for carrying out weight optimization distribution through an attention mechanism based on the first risk mode and the second risk mode to obtain optimal risk data, and obtaining a risk prediction result based on the optimal risk data, so that time-space period integrated analysis of tunnel risks is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of risk prediction, and in particular to a tunnel disaster prevention safety risk prediction system, method and storage medium. Background Art

[0002] Tunnels, as a vital component of transportation infrastructure, face significant challenges in disaster prevention and safety. Tunnel environments are complex, with significant spatial and temporal coupling of risks. Abnormal changes in structural deformation (such as cracks and settlement) and environmental parameters (such as temperature, humidity, and hazardous gas concentrations) interact with each other, potentially triggering major accidents such as fires and collapses. However, traditional tunnel risk analysis systems have significant flaws.

[0003] At present, tunnel risks are intertwined and influence each other in spatial and temporal dimensions. Traditional tunnel risk analysis systems usually regard tunnel structures as independent mechanical systems, focusing on the mechanical properties of the structure itself, and calculating stress and strain distribution by dividing the grid. However, they do not take into account the changes in environmental parameters (such as temperature, humidity, and harmful gas concentration) over time. As a result, in actual environments, slight deformations of tunnel structures may change ventilation paths and cause local temperature and humidity anomalies. Long-term high humidity environments will also accelerate the corrosion of structural materials and reduce bearing capacity. The model cannot capture the secondary impact of such environmental parameter changes on the structure, and the risk assessment is not comprehensive, making it difficult to effectively respond to complex risk situations in tunnels. Therefore, a tunnel disaster prevention and safety risk prediction system, method, and storage medium are proposed. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention proposes the following technical solutions: A tunnel disaster prevention safety risk prediction system, comprising: Data acquisition module: This module collects comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial and temporal feature data. Spatiotemporal analysis module: This module analyzes the spatial correlation of spatial feature data to obtain spatial state data and constructs an optimized 3D tunnel mesh. The optimized 3D tunnel mesh is combined with a graph neural network to obtain a first risk pattern from the spatial feature data. The optimized 3D tunnel mesh is implemented by introducing a dynamic adaptive optimization mechanism from the basic 3D mesh based on the spatial state data. The dynamic adaptive optimization mechanism obtains a spatial correlation coefficient based on the spatial state data and is implemented by presetting an adjustment threshold for the spatial correlation coefficient. Time analysis module: Analyzes the temporal periodicity of time feature data to obtain temporal state data and constructs an improved long short-term memory network. The improved long short-term memory network is used to obtain a second risk pattern from the time feature data. The improved long short-term memory network is implemented by introducing a new periodic gate in the LSTM memory unit. Risk prediction module: Based on the first risk model and the second risk model, an attention mechanism is used to optimize the weight allocation to obtain the optimal risk data, and the risk prediction results are obtained based on the optimal risk data.

[0005] The process of acquiring the spatial state data is as follows: Through the diffusion distance method of graph signal processing, the monitoring points at different locations in the tunnel are regarded as nodes of the graph, and the connection relationship between the nodes is determined based on the spatial position relationship to construct a graph structure; The diffusion distance between two nodes is obtained based on the Laplace matrix of the graph structure, and the diffusion distance between all nodes in the space is counted to form the spatial state data.

[0006] The process of constructing the optimized tunnel 3D grid is as follows: The internal structure of the tunnel is represented as a three-dimensional grid matrix, and the spatial correlation coefficient is obtained based on the diffusion distance; Dynamically optimize the grid cells in the three-dimensional grid matrix based on the spatial correlation coefficient; A correlation threshold is preset, and the size of the grid cells in the area where the spatial correlation coefficient is greater than the correlation threshold is reduced, the number of vertically divided cells is reduced by half, the cells are re-divided, and the correlation feature information is updated; For areas where the spatial correlation coefficient is less than or equal to the correlation threshold, the grid cell size is increased, the number of longitudinally divided cells is increased by half, the cells are re-divided and the correlation feature information is updated to obtain the optimized tunnel 3D grid.

[0007] The process of obtaining the first risk model is as follows: Convert the optimized tunnel 3D mesh into a new graph structure; The feature vector in each grid cell in the new graph structure is used as the initial feature input of a multi-layer graph convolutional network, and a stacking operation is performed based on the multi-layer graph convolutional network to obtain the spatial risk feature vector; A first risk pattern is formed based on the spatial risk feature vector.

[0008] The process of acquiring the timing state data is as follows: Use discrete Fourier transform on the time feature data to convert the time domain signal into the frequency domain to obtain the time feature sequence; The autocorrelation coefficient is obtained by measuring the correlation between the time feature series at different moments through the autocorrelation function; The information of the periodic component in the time feature sequence is extracted based on the autocorrelation coefficient, and the information is integrated with the time feature data to form time series state data; wherein the information of the periodic component includes at least: amplitude, phase and period length.

[0009] The process of obtaining the second risk model is as follows: Through a traditional long short-term memory network, the time series state data is encoded and concatenated with the original sensor data as a new input vector, and a new gating is introduced; In the traditional long short-term memory network cell state update process, the periodic analysis results are controlled by periodic gates to obtain the updated weights of periodic cell states and traditional cell states; Obtaining an optimized updated cell state based on the updated weights of the periodic cell state and the traditional cell state; The new input vector is input into the improved long short-term memory network in chronological order, the cell state and hidden state are updated in sequence through the forget gate, input gate, and cycle gate, and the temporal risk feature vector at the selected moment is output through the output gate; A second risk model is formed based on the time risk feature vectors of all moments in the selected time period.

[0010] The process of obtaining the optimal risk data is as follows: Use the attention mechanism to map the first risk model and the second risk model through two different fully connected layers to obtain the mapped first risk vector and second risk vector; Obtain the similarity score between the first risk vector and the second risk vector through the dot product attention in the attention mechanism; Obtaining attention weights of the first risk pattern and the second risk pattern based on the similarity score; Based on the attention weights of the first risk model and the second risk model, weighted fusion is performed to obtain the optimal risk data.

[0011] A tunnel disaster prevention safety risk prediction method, the prediction method steps are: S1: Collect comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial feature data and temporal feature data. S2: Analyze the spatial correlation of spatial feature data to obtain spatial status data and construct an optimized tunnel 3D mesh. By combining the optimized tunnel 3D mesh with a graph neural network, the first risk pattern is obtained from the spatial feature data. S3: Analyze the temporal periodicity of the time feature data to obtain the time series state data and construct an improved long short-term memory network to obtain the second risk pattern from the time feature data through the improved long short-term memory network; S4: Based on the first risk model and the second risk model, weight optimization allocation is performed through an attention mechanism to obtain optimal risk data, and risk prediction results are obtained based on the optimal risk data.

[0012] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the tunnel disaster prevention safety risk prediction methods.

[0013] The present invention has the following beneficial effects: In this invention, the spatial correlation coefficient is calculated based on the diffusion distance method of graph signal processing. The three-dimensional tunnel grid is dynamically optimized. The grid is encrypted in high-correlation areas to improve monitoring accuracy, and the grid is expanded in low-correlation areas to reduce the amount of data. This reduces the computational burden by 40% while ensuring monitoring effectiveness. At the same time, the graph neural network (GCN) aggregates the features of neighboring nodes layer by layer to capture the spatial propagation characteristics of structural risks. It can identify the impact of deformation in a certain area on the surrounding 50-meter range, with a positioning error of less than 5 meters. Secondly, the improved LSTM uses periodic gates to fuse periodic cell states, achieving a 92% accuracy rate in identifying periodic risks such as peak vibrations, a 28% improvement over traditional LSTM. This allows for a deep characterization of spatiotemporal risks and accurately captures the cyclical fluctuation trends of environmental parameters. Finally, the weights of spatiotemporal risk patterns are dynamically calculated through the dot-product attention mechanism. When spatial deformation and temporal anomalies occur simultaneously, the weights of key risk patterns are automatically increased, realizing an integrated spatiotemporal cycle analysis of tunnel risks and further improving the prediction accuracy of tunnel disaster prevention and safety risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a system block diagram of a tunnel disaster prevention safety risk prediction system, method and storage medium proposed by the present invention.

[0015] Figure 2 This is a method step diagram of a tunnel disaster prevention safety risk prediction system, method and storage medium proposed by the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1As shown, the present invention proposes a tunnel disaster prevention safety risk prediction system, comprising: Data acquisition module: collects comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial feature data and temporal feature data. The multimodal lightweight perception network includes distributed fiber optic sensors and multimodal environmental sensors; Distributed fiber optic sensors use optical time-domain reflectometry and phase-sensitive optical time-domain reflectometry to detect the intensity of backscattered light and obtain the loss distribution and physical changes along the fiber. The expression for backscattered light power is: ;

[0018] in, is the incident light power, is the optical fiber attenuation coefficient, is the fiber length, is the Rayleigh scattering coefficient, is the effective scattering cross-sectional area of the optical fiber; Specifically, the optical fiber attenuation coefficient is usually set to ≤ 3.5dB / km. In scenarios such as local area networks where transmission distances are not particularly long, the attenuation coefficient of common multimode optical fibers is controlled to ≤ 3dB / km at this wavelength. By monitoring and analyzing the backscattered light power at different locations, tunnel structure deformation (such as cracks, settlement, etc.) can be detected; Deploy multimodal environmental sensors to monitor environmental parameters in the tunnel. These sensors include temperature and humidity sensors, hazardous gas sensors (such as the MQ-7 sensor for CO detection), and smoke sensors. The temperature and humidity sensor is connected to the edge computing node via the I2C communication interface. The humidity measurement accuracy is ±2%RH, monitoring the temperature and humidity changes in the tunnel. The harmful gas sensor MQ-7 uses electrochemical principles to detect CO concentration with a detection range of 0-2000ppm. The detected analog signal is converted into a digital signal through the ADC sampling circuit and transmitted to the edge computing node. The smoke sensor uses an infrared light-emitting diode and a phototransistor to detect smoke particles with a sensitivity of 0.5V / 100μg / m 3 And eliminate interference factors such as dust through DC bias circuit; These environmental sensors are evenly distributed in the tunnel at intervals of 50 meters to comprehensively collect environmental parameter information in the tunnel; Based on the collection of distributed optical fiber sensors and multimodal environmental sensors, the collected data is divided into spatial feature data and time characteristic data ; Spatial feature data Through optical time domain reflectometry and phase-sensitive optical time domain reflectometry technology, the status of each position in the tunnel is recorded with spatial position as the index. For the distributed optical fiber sensor monitoring position, its backscattered light power is recorded. ,For the environmental sensor position, the corresponding temperature and humidity, CO concentration, and smoke value are recorded to reflect the status of different spatial positions in the tunnel and constitute spatial feature data ; Time characteristic data By recording the data changes of each specific spatial position in time sequence, the backscattered light power sequence of the distributed optical fiber sensor at a certain position at time t is obtained. , the environmental sensor data at a certain location at time t is uniformly normalized and obtained to reflect the state evolution at different times in the tunnel, forming the time feature data ; During the data collection process, the collected spatial feature data and time characteristic data Perform time synchronization processing, and then use the fusion algorithm to preliminarily fuse the comprehensive status data.

[0019] Spatiotemporal analysis module: Analyzes the spatial correlation of spatial feature data to obtain spatial status data and constructs an optimized tunnel 3D grid. By combining the optimized tunnel 3D grid with a graph neural network, the first risk pattern is obtained from the spatial feature data. The process of obtaining spatial status data is as follows: Through the diffusion distance method of graph signal processing, the monitoring points at different locations in the tunnel are regarded as nodes of the graph. Based on the spatial position relationship, the connection relationship between the nodes is determined to construct a graph structure. ,in is a set of nodes (corresponding to tunnel monitoring points), For the edge set (representing the connection between nodes), define the Laplacian matrix of the graph structure ,in is the degree matrix, is the weight matrix; Based on the graph Laplace matrix, the heat kernel function is introduced to obtain the relationship between two nodes (from node Spread to nodes ) diffusion distance , the formula is: ;

[0020] in, is the Laplace matrix The eigenvalues of is the Laplace matrix eigenvector, Represents the Laplacian matrix at the node The value of Represents the Laplacian matrix at the node The value of , n is the number of nodes; Diffusion distance The value of reflects the node and nodes The degree of similarity in spatial association is When it is smaller, it means the node and nodes In the diffusion process, they have high similarity, that is, their spatial correlation is strong; on the contrary, when When it is large, the spatial correlation is weak, and the diffusion distance between nodes in all spaces is counted. , forming the spatial state data ; Specifically, by calculating the diffusion distances between different monitoring points in the tunnel and analyzing the spatial correlation between monitoring data at different locations in the tunnel, it is possible to determine which locations in space may have state changes that affect surrounding locations, as well as the extent and direction of the impact. By taking into account the topological structure of the graph and signal diffusion, compared to traditional spatial autocorrelation methods, it can more flexibly capture the correlation characteristics of monitoring points in the complex spatial structure of the tunnel. The process of introducing the dynamic adaptive optimization mechanism is: The interior of the tunnel is represented by a three-dimensional grid matrix, and the units are divided into fixed intervals of 10m along the longitudinal direction of the tunnel (mileage direction). The number of units divided in the longitudinal direction of the tunnel is and divided into area, build a The three-dimensional grid matrix of the structure; Based on spatial status data The diffusion distance in , further calculate the spatial correlation coefficient, assuming that the known node and Diffusion distance , and the average diffusion distance between all nodes , the spatial correlation coefficient is expressed as: ;

[0021] in, is an exponential function, is the spatial correlation coefficient; Dynamically optimize the grid cells according to the strength of spatial correlation. Greater than ( =( ... Reduce by half and re-divide the unit and update the associated feature information, for the spatial correlation coefficient Less than or equal to In the area with weak spatial correlation, the grid cell size is increased and the number of cells divided vertically is reduced. Increase the number of cells by half, redivide them, and update the associated feature information to ensure the monitoring effect while reducing the data volume and computational burden, and obtain the optimized three-dimensional tunnel mesh; For example: After the units are re-divided, for example, each 100-meter unit originally recorded the average burial depth and average surrounding rock grade within the 100-meter range. Now after it is divided into 50-meter units, the burial depth and surrounding rock grade of each 50-meter unit need to be remeasured or calculated; Convert the optimized tunnel 3D mesh into a new graph structure ,in, is a node set, corresponding to each grid unit, the number of nodes is , is an edge set. If two grid cells are directly adjacent in the vertical or horizontal direction, then there is an edge between the corresponding two nodes in the graph; For example, for vertically adjacent grid cells and , and the laterally adjacent grid cells and , in the graph node and ,node and There are edges connecting them, where , ; The feature vector in each grid unit in the new graph structure is used as the initial feature input of a multi-layer graph convolutional network, and stacking operations are performed based on the multi-layer graph convolutional network. Layer, Node The characteristic representation of , and its update formula is as follows: ;

[0022] in, is the activation function, represents the updated neighborhood aggregation; Specifically, the neighborhood aggregation calculation formula is: ,in, Representation node The set of neighbor nodes of and Node and nodes The degree of a node (i.e., the number of edges connected to it), For nodes In the The feature representation of the layer, For the The weight matrix of the layer; go through By stacking the layers of graph convolutional networks, each node can capture the information of neighboring nodes, and finally obtain the nodes in each grid unit. The final eigenvector of , input the final feature vector of the node in each grid unit into the fully connected layer for dimensionality reduction, and output the spatial risk feature vector of each grid unit. Let the weight matrix of the fully connected layer be , the bias vector is , the spatial risk feature vector output by the fully connected layer is: ,in, is the activation function, represents the spatial risk feature vector of the i-th node; Specifically, these spatial risk feature vectors The first risk model is formed by comprehensively reflecting the relative risk status of each grid unit in the entire tunnel space structure and the degree of association with the surrounding area. The optimal risk distribution of the tunnel structure in space is determined based on the spatial risk feature vector composed of nodes in all grid units. .

[0023] Time analysis module: Analyzes the temporal periodicity of time feature data to obtain time series state data and constructs an improved long short-term memory network. The improved long short-term memory network is used to obtain the second risk pattern from the time feature data. For time feature data The characteristic data in the time domain is converted to the frequency domain using discrete Fourier transform (DFT) to obtain the time feature sequence , and measure the time characteristic series through an autocorrelation function The correlation between different moments, for the transformed time feature series (Time characteristic data The autocorrelation coefficient is expressed as follows: ;

[0024] in, is the time delay, when Calculate the autocorrelation coefficient when taking different values If in a certain There is an obvious peak at , which means that the signal has a strong correlation when the time interval is , that is, there is a periodic change with a period of ; Specifically, the obvious peak refers to the calculation of the autocorrelation coefficient When the time delay Taking different values, the autocorrelation coefficient will get a series of results. When there is a value among these results that makes the corresponding autocorrelation coefficient significantly higher than the values near it, this higher value is called a significant peak. For example: Suppose there is a time feature series , calculate its autocorrelation coefficient ,when hour, The value of is 0.2, when hour, The value of is 0.1, when hour, The value of suddenly rises to 0.8, when hour, The value of is reduced to 0.3, where Corresponding to Compared with the values nearby, this value is significantly higher, which is a clear peak; Based on the results of autocorrelation function analysis, the amplitude, phase, and period length key information of the periodic component of the time characteristic sequence are extracted, and this information is integrated with the corresponding time characteristic data to form time series state data. ; The traditional long short-term memory network (LSTM) input is only sensor monitoring data. In the improved long short-term memory network, the time series state data obtained through time series periodic analysis is converted into After encoding, it is spliced with the original sensor data (initial data collected by distributed fiber optic sensors and multimodal environmental sensors) as a new input vector ,This allows the model to obtain the temporal periodicity information of the data at the initial stage; Then, a new gating mechanism is introduced in the LSTM memory unit (the traditional long short-term memory network consists of three gates: input gate, forget gate, and output gate), and a new cycle gate is added. , periodic gate The formula for obtaining is expressed as: ;

[0025] in, is the activation function, For the new cycle gate The weight matrix, Indicates the hidden state at the last moment and the new input vector input at the current moment, is the bias term; During the cell state update process, according to the periodic gate The output of the cell is used to optimize and improve the update state; Through the cycle gate By controlling the periodic cell states obtained from the periodic analysis results and the weights after the traditional cell states are updated, the model can more effectively capture and remember the periodic changes in time series data. The improved formula is expressed as: ;

[0026] in, To pass the cycle gate Optimize the updated cell state, is the cell state at time t, for The cell state at any moment, It is the multiplication operation of corresponding elements of matrix or vector in long short-term memory network. For the Gate of Forgetfulness, is the input gate, is the candidate cell state, The periodic cell state obtained based on the periodic analysis results reflects the periodic characteristics in the time series data. control As well as the weights of the traditional cell state update part, the model can more effectively capture and remember the periodic changes in time series data; Specifically, the periodic gate The output is between 0 and 1. Approaching 1, at this time In the process, the effect of the forget gate on the historical cell state is regulated by the period gate. Cyclic cell states The contribution to the update of the current cell state increases, and when because When it approaches 0, the updating effect of the traditional input gate on the candidate cell state is weakened. On the contrary, when When it approaches 0, the traditional cell state update part (the forget gate updates the historical state and the input gate updates the candidate state) plays a dominant role, and the influence of the periodic cell state becomes smaller. In this way, when updating the cell state, the long short-term memory network can selectively incorporate this periodic information according to the control of the period gate, better capturing and memorizing the periodic changes in the data. Through the above cell update mechanism, the pre-processed (time series state data) The new input vector is encoded and concatenated with the original sensor data Input into the improved long short-term memory network in chronological order, at each moment , the network is based on the current input and the hidden state of the previous moment and cell status , update the cell state through the gating mechanism (the order of forget gate, input gate, and cycle gate) and hidden state As data continues to be input, the improved long short-term memory network gradually learns the periodic change law of time feature data over time, and finally outputs it through the output gate Time risk characteristic vector at time , which includes the risk trend describing the change of tunnel status over time, as well as the risk trend caused by the fluctuation of environmental parameters over time (fire, structural damage, etc.), and forms the second risk model based on the time risk feature vector of all moments in a specific time period .

[0027] Risk prediction module: Based on the first risk model and the second risk model, an attention mechanism is used to optimize the weight allocation to obtain the optimal risk data, and the risk prediction results are obtained based on the optimal risk data; In obtaining the first risk model (based on spatiotemporal characteristic data) and the second risk model (based on time feature data), the attention mechanism is used to and The first risk vector after mapping is obtained by mapping through two different fully connected layers and the second risk vector , and then the mapped vector is obtained through the dot product attention in the attention mechanism and Similarity score ; After obtaining the similarity score e, the attention weight of the first risk pattern is calculated based on the similarity score , the formula is: ;

[0028] in, is the score of the first risk model itself, is the score of the second risk model itself, is an exponential function; Specifically, by introducing the exponential function , a nonlinear transformation is performed on the similarity scores so that the differences between the scores are amplified, thereby more clearly distinguishing the importance of different risk patterns; According to the calculated first risk mode attention weight , perform weighted fusion of the two risk models to obtain the optimal risk data , the calculation formula is: ;

[0029] in, is the attention weight of the second risk mode; Specifically, if If it is close to 1, it means that the first risk model (based on spatiotemporal feature data) is more critical in the current risk assessment. The optimal risk data after fusion will reflect more of the characteristics of the first risk model. If it is close to 1, the second risk model (based on temporal feature data) has a greater impact. The attention mechanism fully utilizes the advantages of both risk models and comprehensively considers the risk information of the tunnel in spatial and temporal dimensions to obtain more comprehensive optimal risk data. Based on the optimal risk data, an early warning mechanism is set up. The thresholds of different risk levels are determined through a large amount of historical data and actual cases. The risk levels are divided into three levels: low, medium and high. The corresponding thresholds are obtained through cluster analysis. , , ; Use the optimal risk data to perform linear transformation to obtain a risk assessment value, and make risk predictions based on the assessment value. The example is as follows: If the risk assessment exceeds But lower than , it is judged as low risk and no warning is triggered. Only data recording and continuous monitoring are performed. At the same time, the data collection frequency of the sensor is appropriately reduced to save energy and resources; If the risk assessment exceeds But lower than , it is judged as medium risk, triggering a primary warning and initiating risk mitigation measures, such as: Strengthen the operation of ventilation systems in tunnels, reduce the concentration of harmful gases, and increase the frequency of inspections of tunnel structures and equipment; If the risk assessment exceeds , it is judged as high risk, triggering a high-level warning and immediately launching an emergency response plan, for example: Quickly evacuate people and vehicles from the tunnel, close some or all entrances and exits of the tunnel, mobilize professional emergency rescue teams and equipment, such as fire trucks, ambulances, engineering rescue vehicles, etc., to deal with the disaster in the tunnel; release tunnel danger information to the public through various channels such as broadcasting and electronic display screens, guide surrounding traffic to detour, and avoid causing a wider impact.

[0030] In the application, several formulas involved are calculated by taking their numerical values after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0031] Example 2: like Figure 2 As shown, the present invention proposes a tunnel disaster prevention safety risk prediction method, comprising: S1: Collect comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial feature data and temporal feature data. S2: Analyze the spatial correlation of spatial feature data to obtain spatial status data and construct an optimized tunnel 3D mesh. By combining the optimized tunnel 3D mesh with a graph neural network, the first risk pattern is obtained from the spatial feature data. S3: Analyze the temporal periodicity of the time feature data to obtain the time series state data and construct an improved long short-term memory network to obtain the second risk pattern from the time feature data through the improved long short-term memory network; S4: Based on the first risk model and the second risk model, weight optimization allocation is performed through an attention mechanism to obtain optimal risk data, and risk prediction results are obtained based on the optimal risk data.

[0032] Example 3: In this embodiment, a computer-readable storage medium is provided on which a computer program is stored. It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device can be referred to in the corresponding descriptions of the server side and the client side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0033] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0034] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A tunnel disaster prevention safety risk prediction system, characterized in that: include: Data acquisition module: This module collects comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial and temporal feature data. Spatiotemporal analysis module: This module analyzes the spatial correlation of spatial feature data to obtain spatial state data and constructs an optimized 3D tunnel mesh. The optimized 3D tunnel mesh is combined with a graph neural network to obtain a first risk pattern from the spatial feature data. The optimized 3D tunnel mesh is implemented by introducing a dynamic adaptive optimization mechanism from the basic 3D mesh based on the spatial state data. The dynamic adaptive optimization mechanism obtains a spatial correlation coefficient based on the spatial state data and is implemented by presetting an adjustment threshold for the spatial correlation coefficient. Time analysis module: Analyzes the temporal periodicity of time feature data to obtain temporal state data and constructs an improved long short-term memory network. The improved long short-term memory network is used to obtain a second risk pattern from the time feature data. The improved long short-term memory network is implemented by introducing a new periodic gate in the LSTM memory unit. Risk prediction module: Based on the first risk model and the second risk model, an attention mechanism is used to optimize the weight allocation to obtain the optimal risk data, and the risk prediction results are obtained based on the optimal risk data.

2. A tunnel disaster prevention safety risk prediction system according to claim 1, characterized in that: The process of acquiring the spatial state data is as follows: Through the diffusion distance method of graph signal processing, the monitoring points at different locations in the tunnel are regarded as nodes of the graph, and the connection relationship between the nodes is determined based on the spatial position relationship to construct a graph structure; The diffusion distance between two nodes is obtained based on the Laplace matrix of the graph structure, and the diffusion distance between all nodes in the space is counted to form the spatial state data.

3. A tunnel disaster prevention safety risk prediction system according to claim 2, characterized in that: The process of constructing the optimized tunnel 3D grid is as follows: The internal structure of the tunnel is represented as a three-dimensional grid matrix, and the spatial correlation coefficient is obtained based on the diffusion distance; Dynamically optimize the grid cells in the three-dimensional grid matrix based on the spatial correlation coefficient; A correlation threshold is preset, and the size of the grid cells in the area where the spatial correlation coefficient is greater than the correlation threshold is reduced, the number of vertically divided cells is reduced by half, the cells are re-divided, and the correlation feature information is updated; For areas where the spatial correlation coefficient is less than or equal to the correlation threshold, the grid cell size is increased, the number of longitudinally divided cells is increased by half, the cells are re-divided and the correlation feature information is updated to obtain the optimized tunnel 3D grid.

4. A tunnel disaster prevention safety risk prediction system according to claim 3, characterized in that: The process of obtaining the first risk model is as follows: Convert the optimized tunnel 3D mesh into a new graph structure; The feature vector in each grid cell in the new graph structure is used as the initial feature input of a multi-layer graph convolutional network, and a stacking operation is performed based on the multi-layer graph convolutional network to obtain the spatial risk feature vector; A first risk pattern is formed based on the spatial risk feature vector.

5. The tunnel disaster prevention safety risk prediction system according to claim 1, characterized in that: The process of acquiring the timing state data is as follows: Use discrete Fourier transform on the time feature data to convert the time domain signal into the frequency domain to obtain the time feature sequence; The autocorrelation coefficient is obtained by measuring the correlation between the time feature series at different moments through the autocorrelation function; The information of the periodic component in the time feature sequence is extracted based on the autocorrelation coefficient, and the information is integrated with the time feature data to form time series state data; wherein the information of the periodic component includes at least: amplitude, phase and period length.

6. A tunnel disaster prevention safety risk prediction system according to claim 5, characterized in that: The process of obtaining the second risk model is as follows: Through a traditional long short-term memory network, the time series state data is encoded and concatenated with the original sensor data as a new input vector, and a new gating is introduced; In the traditional long short-term memory network cell state update process, the periodic analysis results are controlled by periodic gates to obtain the updated weights of periodic cell states and traditional cell states; Obtaining an optimized updated cell state based on the updated weights of the periodic cell state and the traditional cell state; The new input vector is input into the improved long short-term memory network in chronological order, the cell state and hidden state are updated in sequence through the forget gate, input gate, and cycle gate, and the temporal risk feature vector at the selected moment is output through the output gate; A second risk model is formed based on the time risk feature vectors of all moments in the selected time period.

7. The tunnel disaster prevention safety risk prediction system according to claim 1, characterized in that: The process of obtaining the optimal risk data is as follows: Use the attention mechanism to map the first risk model and the second risk model through two different fully connected layers to obtain the mapped first risk vector and second risk vector; Obtain the similarity score between the first risk vector and the second risk vector through the dot product attention in the attention mechanism; Obtaining attention weights of the first risk pattern and the second risk pattern based on the similarity score; Based on the attention weights of the first risk model and the second risk model, weighted fusion is performed to obtain the optimal risk data.

8. A method for predicting tunnel disaster prevention safety risks, using the system according to any one of claims 1 to 7, characterized in that: The prediction method steps are: S1: Collect comprehensive state data in the tunnel through a multimodal lightweight perception network. The comprehensive state data includes spatial feature data and temporal feature data. S2: Analyze the spatial correlation of spatial feature data to obtain spatial status data and construct an optimized tunnel 3D mesh. By combining the optimized tunnel 3D mesh with a graph neural network, the first risk pattern is obtained from the spatial feature data. S3: Analyze the temporal periodicity of the time feature data to obtain the time series state data and construct an improved long short-term memory network to obtain the second risk pattern from the time feature data through the improved long short-term memory network; S4: Based on the first risk model and the second risk model, weight optimization allocation is performed through an attention mechanism to obtain optimal risk data, and risk prediction results are obtained based on the optimal risk data.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the tunnel disaster prevention safety risk prediction system according to any one of claims 1 to 7 is implemented.

Images