A tunnel waterlogging multi-event first threshold-crossing probability early warning method and system

Abstract

The present application relates to a kind of tunnel in flood multi-event first time over-limit probability early warning method and system, comprising: establishing directed event graph;For the facility node associated in directed event graph, determine the basic condition over-limit probability in preset prediction time window;The basic condition over-limit probability of facility node is multiplied with the non-over-limit probability in the previous time interval, and the first over-limit probability quality is obtained;Determine the target directed edge associated to facility node, and determine the corresponding dynamic hydraulic time lag kernel and propagation gating coefficient based on the hydraulic characteristic parameter of target directed edge;The basic condition over-limit probability of corresponding downstream facility node is obtained, and the basic condition over-limit probability is revised based on dynamic hydraulic time lag kernel and propagation gating coefficient, to obtain the condition over-limit probability revised by hydraulic propagation;According to the condition over-limit probability revised by hydraulic propagation of each facility node, determine the first over-limit facility, first over-limit time interval and corresponding early warning level.

Description

Technical Field

[0001] This invention relates to the field of smart urban flood control and tunnel safety monitoring technology, and in particular to a method and system for early warning of the probability of first exceedance of limits in multiple events of tunnel flooding. Background Technology

[0002] Urban tunnels, underpasses, and underground space entrances are prone to surface runoff inflow and rainwater backflow under heavy rainfall conditions. Backflowing rainwater typically develops tiered along the tunnel entrance, through hydraulic structures, and then through pumping stations, potentially crossing upstream and downstream cross-cutting ditches first, ultimately pushing rainwater pumping stations or wastewater wells to their warning levels. Quickly and reliably determining which facility will first exceed its limits and when is crucial for tunnel traffic control and drainage management.

[0003] To address the issue of early prediction of urban tunnel and underpass flooding risks during heavy rainfall and minimize traffic safety accidents caused by rainwater backflow, existing technologies primarily employ the following methods: 1) Rainfall, water level, and machine learning-based early warning for underpasses: This method outputs flood warning results based on data such as tunnel water level, street rainfall, and spatial distance; 2) Drainage network graph, time-series network, and multi-step state prediction: This method abstracts facilities along the tunnel, such as cross ditches, drainage networks, and pumping stations, into topological nodes, trains a time-series prediction model based on historical monitoring data, and extrapolates water level change trends for multiple future time steps node by node, outputting the overall water accumulation state sequence of the entire infrastructure; 3) Physical information graph network and water conservation constraints: This method embeds the water conservation rules of hydrodynamics into the training process of a graph neural network, combines real-time monitoring of inflow runoff and drainage data from the pipeline network, and constrains the water level extrapolation results of the model to avoid prediction deviations that violate fluid physics laws and improve the rationality of the simulation of water changes across the entire infrastructure.

[0004] While these methods can roughly predict the trend of tunnel flooding and provide overall water level warnings, they generally overlook the sequential correlation of multiple facilities exceeding their limits, failing to accurately pinpoint the specific facility that first triggers the limit exceedance and providing a precise time window for the first exceedance. In extreme scenarios where heavy rainfall runoff rapidly converges, this can easily lead to delayed warnings and missed key nodes, failing to meet the precise and timely requirements of tunnel traffic control and drainage. Therefore, to fill the technological gap in accurate early warning of the first exceedance of multiple nodes in tunnel flooding and improve the reliability and timeliness of tunnel flooding risk prediction under heavy rainfall, it is necessary to propose a probability-based early warning method for the first exceedance of multiple events in tunnel flooding. This method would accurately pinpoint the first exceedance facility and pre-determine the exceedance time interval, providing precise decision-making basis for emergency response actions such as temporary tunnel traffic control and early activation of drainage pumping stations, significantly reducing the risk of safety accidents caused by tunnel flooding. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for early warning of the probability of first exceedance of limit in multiple events of tunnel flooding, in order to address the shortcomings of the prior art.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for early warning of the probability of the first exceedance of limits in multiple events of tunnel flooding, the method comprising:

[0007] S1. Based on the topological relationships of hydraulic facilities along the tunnel backflow path, establish a directed event graph;

[0008] S2. For the facility nodes associated in the directed event graph, the state representation under the real-time rainfall monitoring characteristics will be input into the target basic condition limit exceedance probability generation model to obtain the basic condition limit exceedance probability within the preset prediction time window.

[0009] S3. Multiply the basic condition limit-breaking probability of the facility node with the non-limit-breaking probability in the previous time interval to obtain the first limit-breaking probability mass.

[0010] S4. Determine the target directed edge associated with the facility node, and input the hydraulic characteristic parameters of the target directed edge into the kernel parameter generator and the gate parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gate coefficient of the target directed edge.

[0011] S5. Obtain the basic condition limit violation probability of the corresponding downstream facility node, and correct the basic condition limit violation probability based on the dynamic hydraulic time delay kernel and the propagation gating coefficient to obtain the hydraulically propagated limit violation probability.

[0012] S6. Based on the hydraulically propagated conditional exceedance probability of each facility node, generate the first exceedance time probability distribution and cumulative exceedance probability of each facility node, and determine the facility that exceeds the limit for the first time, the time interval of the first exceedance, and the corresponding warning level based on the cumulative exceedance probability.

[0013] Furthermore, in step S2, the target basic condition exceedance probability generation model is specifically trained based on the following steps:

[0014] (1) For each pre-determined associated facility node, generate an over-limit label based on the over-limit judgment result of the actual state sequence of the facility, wherein the over-limit label includes the first over-limit time interval label and the right censoring label;

[0015] (2) Obtain the target state representation vector of each associated facility node, and input the limit violation label and the target state representation vector into the initial basic condition limit violation probability generation model. After multiple rounds of model training and parameter optimization, the target basic condition limit violation probability generation model is obtained. The basic condition limit violation probability generation model includes a node basic risk head, and the basic condition limit violation probability of each associated facility node in the future prediction time window is output through the node basic risk head.

[0016] Furthermore, in step S4, the hydraulic characteristic parameters include at least one static hydraulic parameter among the hydraulic distance between nodes, longitudinal slope or elevation difference, and at least two dynamic hydraulic state parameters among the real-time rainfall intensity, the remaining storage capacity of the node, the start-stop status of the pump group and the effective drainage capacity.

[0017] Furthermore, in step S4, the calculation formula for the dynamic hydraulic time delay kernel includes:

[0018] ;

[0019] in, This represents the unnormalized kernel score output by the kernel parameter generator for a directed edge (u,v), a time interval k, and a time delay step d. Let d represent the time-delay feasibility mask, d represent the discrete propagation time-delay steps, u represent the upstream facility node, and v represent the downstream facility node. Let represent the minimum allowable discrete propagation delay steps for a directed edge (u,v). Let represent the maximum allowed number of discrete propagation delay steps for a directed edge (u,v), and D represent the maximum allowed number of discrete propagation delay steps to consider.

[0020] Furthermore, in step S4, the calculation formula for the dynamic hydraulic time delay kernel includes:

[0021] ;

[0022] Where d represents the current discrete propagation time delay step, Indicates the time step size of adjacent prediction time intervals. Indicates the output center of the kernel parameter generator. This indicates the corresponding propagation delay dispersion or uncertainty width. This represents the discrete time delay index in truncated normalized summation.

[0023] Furthermore, in step S5, the correction of the basic condition violation probability based on the dynamic hydraulic time delay kernel and propagation gating coefficient to obtain the hydraulically propagated condition violation probability includes:

[0024] S51. Discretely convolve the probability mass with the dynamic hydraulic time delay kernel to obtain the propagation event mass of the current first over-limit event in the corresponding time interval to reach the corresponding downstream facility node.

[0025] S52. Normalize the quality of the propagation event according to the remaining unpropagated probability to obtain the propagation condition risk which is on the same probability scale as the basic condition exceedance probability of the downstream facility node.

[0026] S53. Based on the propagation condition risk and the propagation gating coefficient, the probability of exceeding the basic condition limit for the corresponding downstream facility node is gating to obtain the hydraulically propagated probability of exceeding the condition limit.

[0027] Furthermore, in step S53, if the downstream facility node v corresponds to multiple upstream parent nodes u, then the basic condition exceedance probability is gating using the following formula:

[0028] ;

[0029] If a downstream facility node v corresponds to a single upstream parent node u, then the gating of the probability of exceeding the basic condition limit is performed using the following formula:

[0030] ;

[0031] in, This represents the probability that downstream facility node v exceeds the limit under basic conditions at time k. This represents the set of upstream parent nodes corresponding to downstream facility node v. Represents the propagation gating coefficient. Indicates the risk of transmission conditions. This represents the probability of exceeding the condition limit after hydraulic propagation correction.

[0032] Furthermore, the target basic condition limit violation probability generation model, kernel parameter generator, and gating parameter generator are trained based on a joint loss function, which includes the first limit violation time and right censoring negative log-likelihood loss, hydraulic propagation supervision loss, and late reporting asymmetric loss, wherein:

[0033] The hydraulic propagation monitoring loss includes propagation interval loss, dynamic core monitoring loss, and gated monitoring loss;

[0034] The propagation interval loss is calculated based on the actual first time difference of adjacent facility nodes and the allowable propagation time delay interval determined by the hydraulic distance between nodes, longitudinal slope or elevation difference, real-time rainfall intensity, remaining storage capacity of nodes, pump start-up and shutdown status and effective drainage capacity.

[0035] The dynamic nuclear surveillance loss is calculated based on the observation propagation delay when both upstream and downstream facility nodes experience their first limit exceedance.

[0036] The gated monitoring loss is calculated based on the propagation result label of whether the downstream facility node has its first violation within the allowed propagation time delay interval after the upstream facility node has its first violation.

[0037] In the evening report asymmetric loss, the penalty weight for predicting the first violation time later than the actual first violation time is greater than the penalty weight for predicting the first violation time earlier than the actual first violation time.

[0038] Secondly, this application discloses a tunnel flooding multi-event first-time limit exceedance probability early warning system. The system includes a directed event graph establishment module, a basic condition limit exceedance probability calculation module, a first-time limit exceedance probability quality calculation module, a dynamic hydraulic parameter processing module, a hydraulic propagation correction module, and a limit exceedance probability analysis and early warning module, wherein:

[0039] The directed event graph building module is used to build a directed event graph based on the topological association of hydraulic facilities on the tunnel backflow path;

[0040] The basic condition exceedance probability calculation module is used to input the state representation under real-time rainfall monitoring characteristics of the facility nodes associated in the directed event graph into the target basic condition exceedance probability generation model to obtain the basic condition exceedance probability within the preset prediction time window.

[0041] The first-time limit exceedance probability quality calculation module is used to multiply the basic condition limit exceedance probability of the facility node with the non-limit exceedance probability in the preceding time interval to obtain the first-time limit exceedance probability quality.

[0042] The dynamic hydraulic parameter processing module is used to determine the target directed edge associated with the facility node, and input the hydraulic feature parameters of the target directed edge into the kernel parameter generator and the gate parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gate coefficient of the target directed edge.

[0043] The hydraulic propagation correction module is used to obtain the probability of the basic condition exceeding the limit for the corresponding downstream facility node, and to correct the probability of the basic condition exceeding the limit based on the dynamic hydraulic time delay kernel and the propagation gating coefficient, so as to obtain the probability of the condition exceeding the limit after hydraulic propagation correction.

[0044] The over-limit probability analysis and early warning module is used to generate the first over-limit time probability distribution and cumulative over-limit probability of each facility node based on the conditional over-limit probability after hydraulic propagation correction, and to determine the facility that first over-limits, the time interval of the first over-limit, and the corresponding early warning level based on the cumulative over-limit probability.

[0045] Thirdly, this application discloses a readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for early warning of the probability of first exceedance of limits in multiple events of tunnel flooding as described in any of the preceding claims.

[0046] The beneficial effects of this invention are: 1) It directly predicts the probability distribution of the first time limit exceedance of multiple facilities, making the training target correspond to the facility that first exceeds the limit, the time of the first exceedance, and the remaining handling time, which is different from the processing method of only predicting continuous water levels and then comparing thresholds; 2) This invention propagates the time probability distribution of the first exceedance event upstream. This distribution is convolved by a dynamic hydraulic time delay kernel and converted into propagation conditional risk, and then directly participates in the generation of downstream conditional exceedance probability, instead of just using the propagation time delay as an input lag, post-hoc explanation, or ordinary loss term; 3) The dynamic hydraulic time delay kernel is jointly determined by static parameters such as facility distance and slope, as well as dynamic states such as real-time rainfall, remaining storage capacity, pump status, and drainage capacity. The expected time delay can be supervised by the observed difference between the first exceedance time of upstream and downstream facilities, thereby establishing an interpretable event propagation time distribution; 4) The gating coefficient is trained based on the real-time hydraulic state and the propagation results of whether the downstream exceeds the limit within the prediction window after the upstream exceeds the limit, so that the system can suppress the unreasonable amplification of downstream risks under conditions of high remaining capacity, strong drainage capacity, or obstructed propagation. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a method for early warning of the probability of first exceedance of limits in multiple events of tunnel flooding disclosed in this invention.

[0048] Figure 2 This is a schematic diagram of the structure of a tunnel flooding multi-event first-limit-exceeding probability early warning system disclosed in this invention;

[0049] Figure 3 This is a schematic diagram of the structure of a readable storage medium disclosed in this invention. Detailed Implementation

[0050] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0051] like Figure 1 As shown, this application discloses a method for early warning of the probability of the first exceedance of limits in multiple events of tunnel flooding, the method comprising:

[0052] Step S1: Based on the topological relationships of hydraulic facilities along the tunnel backflow path, establish a directed event graph.

[0053] Step S2: For the facility nodes associated in the directed event graph, the state representation under the real-time rainfall monitoring characteristics is input into the target basic condition limit exceedance probability generation model to obtain the basic condition limit exceedance probability within the preset prediction time window.

[0054] Specifically, this application first collects and organizes data related to real-time rainfall monitoring characteristics and state characterization data of facility nodes under real-time rainfall monitoring characteristics; then, it organizes these data into the model according to the format and dimensions required by the target basic condition exceedance probability generation model; finally, the model performs feature extraction, fusion and other operations through correlation analysis, using its own algorithms and parameters, and continuously updates the calculation results according to the changes in input data within the preset prediction time window, and finally outputs the basic condition exceedance probability.

[0055] Step S3: Multiply the basic condition exceedance probability of the facility node with the non-exceedance probability in the previous time interval to obtain the first exceedance probability quality.

[0056] Specifically, this application first calculates the probability of facility node i not exceeding the limit at each time point within the preceding time interval. (where j ranges from 1 to k-1), and then these probabilities of not exceeding the limit are multiplied together to obtain... Finally, the probability of exceeding the limit under the basic conditions at the current time k is calculated. The result of multiplication Multiply to obtain the first out-of-limit probability mass. .

[0057] Step S4: Determine the target directed edge associated with the facility node, and input the hydraulic characteristic parameters of the target directed edge into the kernel parameter generator and the gating parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gating coefficient of the target directed edge.

[0058] Specifically, this application generates discrete hydraulic time delay kernels and gating coefficients for each directed edge based on the static hydraulic parameters and real-time hydraulic state of the directed edge.

[0059] Step S5: Obtain the basic condition limit violation probability of the corresponding downstream facility node, and correct the basic condition limit violation probability based on the dynamic hydraulic time delay kernel and propagation gating coefficient to obtain the hydraulically propagated limit violation probability.

[0060] Specifically, this application first performs a discrete convolution between the probability quality and the dynamic hydraulic time delay kernel to obtain the propagation event quality of the first violation event reaching the corresponding downstream facility node within the corresponding time interval. Then, based on the remaining unpropagated probability, the propagation event quality is normalized so that the obtained propagation conditional risk is on the same probability scale as the basic condition violation probability of the downstream facility node.

[0061] Finally, based on the propagation condition risk and propagation gating coefficient, the probability of exceeding the basic condition limit for the corresponding downstream facility node is gating, resulting in the hydraulically propagated probability of exceeding the condition limit. This allows the influence of the initial upstream limit exceedance probability propagated through a time-delay kernel to exert a differentiable gating effect on the downstream condition exceedance probability.

[0062] Step S6: Based on the hydraulically propagated conditional exceedance probability of each facility node, generate the first exceedance time probability distribution and cumulative exceedance probability of each facility node, and determine the facility that exceeds the limit for the first time, the time interval of the first exceedance, and the corresponding warning level based on the cumulative exceedance probability.

[0063] Specifically, this application will calculate the cumulative out-of-limit probability F of facility node i. i,k First time reaching the warning threshold α i The time interval is determined as the predicted first over-limit time interval. Based on the predicted first over-limit time interval of each node, the earliest over-limit node in the entire map is determined, and the earliest over-limit node in the entire map, the first over-limit time interval, the over-limit confidence, the remaining handling time, the warning level, etc. are output to provide timely and accurate risk warning information to relevant management departments and personnel, so that they can rationally allocate resources, take effective measures to deal with the upcoming over-limit events, reduce the potential risks and losses caused by over-limits, and ensure the safe and stable operation of the facilities.

[0064] As can be seen from the above, the tunnel flooding multi-event first-limit probability early warning method disclosed in this application 1) directly predicts the time probability distribution of the first limit exceeding multiple facilities, so that the training target corresponds to the first limit exceeding facility, the first limit exceeding time, and the remaining handling time, which is different from the processing method of only predicting continuous water levels and then comparing thresholds; 2) the present invention propagates the time probability distribution of the upstream first limit exceeding event, which is converted into propagation conditional risk after being convolved by dynamic hydraulic time delay kernel, and directly participates in the generation of downstream conditional limit exceeding probability, rather than only using the propagation time delay as the input lag quantity, event 3) The dynamic hydraulic time delay kernel is jointly determined by static parameters such as facility distance and slope, as well as dynamic states such as real-time rainfall, remaining storage capacity, pump status and drainage capacity. It can also be monitored by the observed first over-limit time difference between upstream and downstream, thereby establishing an interpretable event propagation time distribution; 4) The gating coefficient is trained based on the real-time hydraulic state and the propagation results of whether the downstream over-limit is within the prediction window after the upstream over-limit, so that the system can suppress the unreasonable amplification of downstream risks under conditions of high remaining capacity, strong drainage capacity or obstructed propagation.

[0065] In one embodiment, in step S2, the target basic condition exceedance probability generation model is specifically trained based on the following steps:

[0066] (1) For each pre-determined associated facility node, generate an over-limit label based on the over-limit judgment result of the actual state sequence of the facility. The over-limit label includes the first over-limit time interval label and the right censoring label.

[0067] Specifically, this application will generate a first over-limit time interval label based on the actual state sequence of the facility within the future prediction time window, when it is determined that there is a moment when the state sequence first exceeds the preset over-limit threshold and the moment when it first falls back below the threshold, and generate a right censoring label when it is determined that the facility has not experienced an over-limit event by the end of the observation period.

[0068] (2) Obtain the target state representation vector of each associated facility node, and input the limit violation label and the target state representation vector into the initial basic condition limit violation probability generation model. After multiple rounds of model training and parameter optimization, the target basic condition limit violation probability generation model is obtained. The basic condition limit violation probability generation model includes a node basic risk head, and the basic condition limit violation probability of each associated facility node in the future prediction time window is output through the node basic risk head.

[0069] Specifically, the target state representation vector of each associated facility node is determined through the following steps: extracting facility state features of each associated facility node under historical rainfall events based on a shared temporal encoder, and performing standardization processing based on the facility state features to obtain the target state representation vector of each associated facility node. The shared temporal encoder, as a temporal feature extraction network, can simultaneously adapt to two different types of temporal data: rainfall and facility state.

[0070] In one embodiment, in step S4, the hydraulic characteristic parameters include at least one static hydraulic parameter among the hydraulic distance between nodes, longitudinal slope or elevation difference, and at least two dynamic hydraulic state parameters among the real-time rainfall intensity, the remaining storage capacity of the nodes, the start-stop status of the pump group and the effective drainage capacity.

[0071] In one embodiment, step S4 includes the following calculation formula for the dynamic hydraulic time delay kernel:

[0072] ;

[0073] in, This represents the unnormalized kernel score output by the kernel parameter generator for a directed edge (u,v), a time interval k, and a time delay step d. Let d represent the time-delay feasibility mask, d represent the discrete propagation time-delay steps, u represent the upstream facility node, and v represent the downstream facility node. Let represent the minimum allowable discrete propagation delay steps for a directed edge (u,v). Let represent the maximum allowed number of discrete propagation delay steps for a directed edge (u,v), and D represent the maximum allowed number of discrete propagation delay steps to consider.

[0074] In one embodiment, step S4 includes the following calculation formula for the dynamic hydraulic time delay kernel:

[0075] ;

[0076] Where d represents the current discrete propagation time delay step, Indicates the time step size of adjacent prediction time intervals. Indicates the output center of the kernel parameter generator. This indicates the corresponding propagation delay dispersion or uncertainty width. This represents the discrete time delay index in truncated normalized summation.

[0077] In one embodiment, step S5, which involves correcting the basic condition violation probability based on the dynamic hydraulic time delay kernel and the propagation gating coefficient to obtain the hydraulically propagated condition violation probability, includes:

[0078] Step S51: Discretely convolve the probability mass with the dynamic hydraulic time delay kernel to obtain the propagation event mass of the current first limit violation event in the corresponding time interval to reach the corresponding downstream facility node.

[0079] Step S52: Normalize the quality of the propagation event according to the remaining unpropagated probability to obtain the propagation condition risk at the same probability scale as the basic condition exceedance probability of the downstream facility node.

[0080] Specifically, this application uses the following formula to measure the quality of the dissemination event. Transformation into transmission condition risk : ,in, This indicates the quality of the propagated event at time step j. This represents the numerical stability constant to prevent the denominator from being zero.

[0081] In one embodiment, in actual calculations... Bounded truncation is performed according to the following formula:

[0082] It should be noted that, It is a pre-defined small positive number used to determine the truncation boundary. This formula can be used to... The value is truncated in the interval [ ,1- Within this range, the quality of the propagation event and the probability of downstream conditions exceeding limits are kept at the same conditional risk scale.

[0083] Step S53: Based on the propagation condition risk and the propagation gating coefficient, the basic condition exceedance probability of the corresponding downstream facility node is gating to obtain the hydraulically propagated condition exceedance probability.

[0084] In one embodiment, in step S53, if the downstream facility node v corresponds to multiple upstream parent nodes u, then the basic condition exceedance probability is gated using the following formula:

[0085] .

[0086] Specifically, when When equal to 0, it is stipulated that The above equation satisfies And it has the following quantitative properties: when When, the propagation contribution of edge (u,v) is zero; when When, upstream events do not increase downstream risk in the k-th interval; when and When both increase, Monotonically increasing.

[0087] If a downstream facility node v corresponds to a single upstream parent node u, then the gating of the probability of exceeding the basic condition limit is performed using the following formula:

[0088] ;

[0089] in, This represents the probability that downstream facility node v exceeds the limit under basic conditions at time k. This represents the set of upstream parent nodes corresponding to downstream facility node v. Represents the propagation gating coefficient. Indicates the risk of transmission conditions. This represents the probability of exceeding the condition limit after hydraulic propagation correction.

[0090] It should be noted that, This setting is to prevent the probability from exceeding 1 due to simple addition, thereby balancing the probability of exceeding the limit under the basic conditions. Propagation gating coefficient Risks related to transmission conditions The relationship between them makes the calculated hydraulic propagation correction conditional failure probability... It can remain within a reasonable probability range of 0 to 1.

[0091] In one embodiment, the target baseline condition exceedance probability generation model, kernel parameter generator, and gating parameter generator are trained based on a joint loss function, which includes the first exceedance time and right censoring negative log-likelihood loss, hydraulic propagation supervision loss, and late reporting asymmetric loss, wherein:

[0092] The hydraulic propagation monitoring loss includes propagation interval loss, dynamic core monitoring loss, and gated monitoring loss.

[0093] The propagation interval loss is calculated based on the actual first time difference of adjacent facility nodes and the allowable propagation time delay interval determined by the hydraulic distance between nodes, longitudinal slope or elevation difference, real-time rainfall intensity, remaining storage capacity of nodes, pump start-up and shutdown status, and effective drainage capacity.

[0094] Specifically, the propagation interval loss is used to constrain the first time the downstream facility node exceeds the limit to be no earlier than the upstream facility node, and the time difference between the two falls within the allowable propagation delay interval.

[0095] The dynamic nuclear surveillance loss is calculated based on the observation propagation delay when both upstream and downstream facility nodes experience their first over-limit events.

[0096] Specifically, the expected propagation delay of the dynamic hydraulic time delay kernel is compared with the observed propagation delay, or the discrete time delay step corresponding to the observed propagation delay is taken as the target step, and a negative log-likelihood constraint is applied to the dynamic hydraulic time delay kernel so that the dynamic hydraulic time delay kernel is concentrated near the actual propagation delay.

[0097] The gated monitoring loss is calculated based on the propagation result label of whether the downstream facility node has its first violation within the allowed propagation time delay interval after the upstream facility node has its first violation.

[0098] Specifically, when a downstream facility node experiences its first violation within the allowed propagation time lag interval, the propagation result is labeled as effective propagation; when a downstream facility node does not experience its first violation within the allowed propagation time lag interval, or when propagation is hindered due to a large remaining storage capacity or strong effective drainage capacity of the pumping unit, the propagation result is labeled as invalid propagation. It should be noted that the gated supervision loss is used to train the propagation gating coefficients, enhancing the impact of upstream violation events on the probability of downstream violation conditions under effective propagation conditions, and suppressing unreasonable risk amplification under propagation hindered conditions.

[0099] In the evening report asymmetric loss, the penalty weight for predicting the first violation time later than the actual first violation time is greater than the penalty weight for predicting the first violation time earlier than the actual first violation time.

[0100] Please refer to Figure 2This application discloses a tunnel flooding multi-event first-limit-exceedance probability early warning system. The system includes a directed event graph establishment module, a basic condition limit-exceedance probability calculation module, a first-limit-exceedance probability quality calculation module, a dynamic hydraulic parameter processing module, a hydraulic propagation correction module, and a limit-exceedance probability analysis and early warning module, wherein:

[0101] The directed event graph building module is used to build a directed event graph based on the topological association of hydraulic facilities on the tunnel backflow path.

[0102] The basic condition exceedance probability calculation module is used to input the state representation under real-time rainfall monitoring characteristics of the facility nodes associated in the directed event graph into the target basic condition exceedance probability generation model to obtain the basic condition exceedance probability within a preset prediction time window.

[0103] The first-time exceedance probability quality calculation module is used to multiply the basic condition exceedance probability of the facility node with the non-exceedance probability in the preceding time interval to obtain the first-time exceedance probability quality.

[0104] The dynamic hydraulic parameter processing module is used to determine the target directed edge associated with the facility node, and input the hydraulic feature parameters of the target directed edge into the kernel parameter generator and the gating parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gating coefficient of the target directed edge.

[0105] The hydraulic propagation correction module is used to obtain the probability of exceeding the basic conditions of the corresponding downstream facility node, and to correct the probability of exceeding the basic conditions based on the dynamic hydraulic time delay kernel and the propagation gating coefficient to obtain the probability of exceeding the conditions after hydraulic propagation correction.

[0106] The over-limit probability analysis and early warning module is used to generate the first over-limit time probability distribution and cumulative over-limit probability of each facility node based on the conditional over-limit probability after hydraulic propagation correction, and to determine the facility that first over-limits, the time interval of the first over-limit, and the corresponding early warning level based on the cumulative over-limit probability.

[0107] In one embodiment, the above modules are also used to implement a method for early warning of the probability of first exceedance of limit in multiple events of tunnel flooding as described in any of the preceding claims, and this application does not limit this.

[0108] As can be seen from the above, the tunnel flooding multi-event first-limit probability early warning system disclosed in this application 1) directly predicts the time probability distribution of the first limit exceeding multiple facilities, so that the training target corresponds to the first-limit exceeding facility, the first-limit exceeding time, and the remaining handling time, which is different from the processing method of only predicting continuous water levels and then comparing thresholds; 2) the present invention propagates the time probability distribution of the upstream first-limit exceeding event. This distribution is convolved by a dynamic hydraulic time delay kernel and converted into propagation conditional risk, and then directly participates in the generation of downstream conditional limit exceeding probability, rather than only using the propagation time delay as the input lag quantity. 3) The dynamic hydraulic time delay kernel is jointly determined by static parameters such as facility distance and slope, as well as dynamic states such as real-time rainfall, remaining storage capacity, pump status and drainage capacity. It can also be monitored by the observed first over-limit time difference between upstream and downstream, thereby establishing an interpretable event propagation time distribution; 4) The gating coefficient is trained based on the real-time hydraulic state and the propagation results of whether the downstream over-limit is within the prediction window after the upstream over-limit, so that the system can suppress the unreasonable amplification of downstream risks under conditions of high remaining capacity, strong drainage capacity or obstructed propagation.

[0109] Please refer to Figure 3 This application discloses a readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for early warning of the probability of first exceedance of limits in multiple events of tunnel flooding as described in any of the preceding claims.

[0110] As can be seen from the above, the readable storage medium disclosed in this application 1) directly predicts the probability distribution of the first time limit exceeding the limit for multiple facilities, so that the training target corresponds to the facility that first exceeds the limit, the time of the first exceedance, and the remaining handling time, which is different from the processing method of only predicting continuous water levels and then comparing thresholds; 2) the present invention propagates the time probability distribution of the first exceedance event upstream. This distribution is convolved by a dynamic hydraulic time delay kernel and converted into propagation conditional risk, and then directly participates in the generation of downstream conditional exceedance probability, instead of just using the propagation time delay as an input lag, post-hoc explanation, or ordinary loss term; 3) the dynamic hydraulic time delay kernel is jointly determined by static parameters such as facility distance and slope, as well as dynamic states such as real-time rainfall, remaining storage capacity, pump status, and drainage capacity, and its expected time delay can be supervised by the observed first exceedance time difference between upstream and downstream, thereby establishing an interpretable event propagation time distribution; 4) the gating coefficient is trained based on the real-time hydraulic state and the propagation results of whether the downstream exceeds the limit within the prediction window after the upstream exceeds the limit, so that the system can suppress the unreasonable amplification of downstream risks under conditions of high remaining capacity, strong drainage capacity, or obstructed propagation.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning of the probability of the first exceedance of limits in multiple events of tunnel flooding, characterized in that, The method includes: S1. Based on the topological relationships of hydraulic facilities along the tunnel backflow path, establish a directed event graph; S2. For the facility nodes associated in the directed event graph, the state representation under the real-time rainfall monitoring characteristics will be input into the target basic condition limit exceedance probability generation model to obtain the basic condition limit exceedance probability within the preset prediction time window. S3. Multiply the basic condition limit-breaking probability of the facility node with the non-limit-breaking probability in the previous time interval to obtain the first limit-breaking probability mass. S4. Determine the target directed edge associated with the facility node, and input the hydraulic characteristic parameters of the target directed edge into the kernel parameter generator and the gate parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gate coefficient of the target directed edge. S5. Obtain the basic condition limit violation probability of the corresponding downstream facility node, and correct the basic condition limit violation probability based on the dynamic hydraulic time delay kernel and the propagation gating coefficient to obtain the hydraulically propagated limit violation probability. S6. Based on the hydraulically propagated conditional exceedance probability of each facility node, generate the first exceedance time probability distribution and cumulative exceedance probability of each facility node, and determine the facility that exceeds the limit for the first time, the time interval of the first exceedance, and the corresponding warning level based on the cumulative exceedance probability.

2. The method according to claim 1, characterized in that, In step S2, the target basic condition exceedance probability generation model is specifically trained based on the following steps: (1) For each pre-determined associated facility node, generate an over-limit label based on the over-limit judgment result of the actual state sequence of the facility, wherein the over-limit label includes the first over-limit time interval label and the right censoring label; (2) Obtain the target state representation vector of each associated facility node, and input the limit violation label and the target state representation vector into the initial basic condition limit violation probability generation model. After multiple rounds of model training and parameter optimization, the target basic condition limit violation probability generation model is obtained. The basic condition limit violation probability generation model includes a node basic risk head, and the basic condition limit violation probability of each associated facility node in the future prediction time window is output through the node basic risk head.

3. The method according to claim 1, characterized in that, In step S4, the hydraulic characteristic parameters include at least one static hydraulic parameter among the hydraulic distance between nodes, longitudinal slope or elevation difference, and at least two dynamic hydraulic state parameters among the real-time rainfall intensity, the remaining storage capacity of the node, the start-stop status of the pump group and the effective drainage capacity.

4. The method according to claim 1, characterized in that, In step S4, the calculation formula for the dynamic hydraulic time delay kernel includes: ； in, This represents the unnormalized kernel score output by the kernel parameter generator for a directed edge (u,v), a time interval k, and a time delay step d. This represents the time-delay feasibility mask, where d represents the number of discrete propagation time delay steps, u represents the upstream facility node, and v represents the downstream facility node. Let represent the minimum allowable discrete propagation delay steps for a directed edge (u,v). Let represent the maximum allowed number of discrete propagation delay steps for a directed edge (u,v), and D represent the maximum allowed number of discrete propagation delay steps to consider.

5. The method according to claim 4, characterized in that, In step S4, the calculation formula for the dynamic hydraulic time delay kernel includes: ； Where d represents the current discrete propagation time delay step, Indicates the time step size of adjacent prediction time intervals. Indicates the output center of the kernel parameter generator. This indicates the corresponding propagation delay dispersion or uncertainty width. This represents the discrete time delay index in truncated normalized summation.

6. The method according to claim 1, characterized in that, In step S5, the correction of the basic condition violation probability based on the dynamic hydraulic time delay kernel and propagation gating coefficient to obtain the hydraulically propagated condition violation probability includes: S51. Discretely convolve the probability mass with the dynamic hydraulic time delay kernel to obtain the propagation event mass of the current first over-limit event in the corresponding time interval to reach the corresponding downstream facility node. S52. Normalize the quality of the propagation event according to the remaining unpropagated probability to obtain the propagation condition risk which is on the same probability scale as the basic condition exceedance probability of the downstream facility node. S53. Based on the propagation condition risk and the propagation gating coefficient, the probability of exceeding the basic condition limit for the corresponding downstream facility node is gating to obtain the hydraulically propagated probability of exceeding the condition limit.

7. The method according to claim 6, characterized in that, In step S53, if the downstream facility node v corresponds to multiple upstream parent nodes u, then the basic condition exceedance probability is gating using the following formula: ； If a downstream facility node v corresponds to a single upstream parent node u, then the gating of the probability of exceeding the basic condition limit is performed using the following formula: ； in, This represents the probability that downstream facility node v exceeds the limit under basic conditions at time k. This represents the set of upstream parent nodes corresponding to downstream facility node v. Represents the propagation gating coefficient. Indicates the risk of transmission conditions. This represents the probability of exceeding the condition limit after hydraulic propagation correction.

8. The method according to any one of claims 1 to 7, characterized in that, The target baseline condition exceedance probability generation model, kernel parameter generator, and gating parameter generator are trained based on a joint loss function, which includes the first exceedance time and right censoring negative log-likelihood loss, hydraulic propagation supervision loss, and late reporting asymmetric loss, wherein: The hydraulic propagation monitoring loss includes propagation interval loss, dynamic core monitoring loss, and gated monitoring loss; The propagation interval loss is calculated based on the actual first time difference of adjacent facility nodes and the allowable propagation time delay interval determined by the hydraulic distance between nodes, longitudinal slope or elevation difference, real-time rainfall intensity, remaining storage capacity of nodes, pump start-up and shutdown status and effective drainage capacity. The dynamic nuclear surveillance loss is calculated based on the observation propagation delay when both upstream and downstream facility nodes experience their first limit exceedance. The gated monitoring loss is calculated based on the propagation result label of whether the downstream facility node has its first violation within the allowed propagation time delay interval after the upstream facility node has its first violation. In the evening report asymmetric loss, the penalty weight for predicting the first violation time later than the actual first violation time is greater than the penalty weight for predicting the first violation time earlier than the actual first violation time.

9. A tunnel flooding multi-event first-limit-exceeding probability early warning system, characterized in that, The system includes a directed event graph construction module, a basic condition limit violation probability calculation module, an initial limit violation probability quality calculation module, a dynamic hydraulic parameter processing module, a hydraulic propagation correction module, and a limit violation probability analysis and early warning module, wherein: The directed event graph building module is used to build a directed event graph based on the topological association of hydraulic facilities on the tunnel backflow path; The basic condition exceedance probability calculation module is used to input the state representation under real-time rainfall monitoring characteristics of the facility nodes associated in the directed event graph into the target basic condition exceedance probability generation model to obtain the basic condition exceedance probability within the preset prediction time window. The first-time limit exceedance probability quality calculation module is used to multiply the basic condition limit exceedance probability of the facility node with the non-limit exceedance probability in the preceding time interval to obtain the first-time limit exceedance probability quality. The dynamic hydraulic parameter processing module is used to determine the target directed edge associated with the facility node, and input the hydraulic feature parameters of the target directed edge into the kernel parameter generator and the gate parameter generator respectively to obtain the dynamic hydraulic time delay kernel and propagation gate coefficient of the target directed edge. The hydraulic propagation correction module is used to obtain the probability of the basic condition exceeding the limit for the corresponding downstream facility node, and to correct the probability of the basic condition exceeding the limit based on the dynamic hydraulic time delay kernel and the propagation gating coefficient, so as to obtain the probability of the condition exceeding the limit after hydraulic propagation correction. The over-limit probability analysis and early warning module is used to generate the first over-limit time probability distribution and cumulative over-limit probability of each facility node based on the conditional over-limit probability after hydraulic propagation correction, and to determine the facility that first over-limits, the time interval of the first over-limit, and the corresponding early warning level based on the cumulative over-limit probability.

10. A readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a method for early warning of the probability of first exceedance of limits in multiple events of tunnel flooding as described in any one of claims 1-8.

Images