A municipal building underground drainage pipe network operation risk assessment method and system

Abstract

The application discloses a kind of municipal building underground drainage pipe network operation risk assessment method and system, it is related to data processing technical field, the application constructs real-time hydraulic situation three-dimensional tensor by collecting the liquid level and flow rate data of monitoring point, input hydraulic topology expansion gate network obtains future hydraulic situation prediction tensor;Target analysis pipe section is spliced in time and space grid, and the internal physical field characteristics are solved in combination with prediction tensor and physical constraint, and the pipe wall mechanical response characteristic matrix is obtained;Based on the matrix, the pipe wall structure fatigue time-varying state transition chain is constructed, and the transient structure damage risk probability is obtained by Markov chain iterative calculation;Combined with real-time physical field characteristics, construct cross-modal cross attention operator to obtain weight square matrix;Finally, the comprehensive risk score is calculated in combination with the two. The application breaks through the blind area of traditional physical monitoring, realizes the continuous high-precision analysis of microscopic stress state and the dynamic quantification of transient fatigue fracture risk.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for risk assessment of the operation of underground drainage pipe networks in municipal buildings. Background Technology

[0002] Risk assessment of underground drainage pipe networks in municipal buildings is a key technical link to ensure the normal operation of cities and prevent urban flooding and secondary geological disasters. With the acceleration of urbanization, the scale and complexity of underground drainage systems are growing exponentially. Their operation status is directly related to the flood control and drainage safety of cities and the stability of road infrastructure. During rainstorms or under the condition of pumping station strong drainage, the pipe network not only has to withstand the severe impact of internal water flow, but may also face the pressure caused by uneven settlement of external soil. Currently, Chinese invention patent application number 202511093820.4 discloses an intelligent drainage pipe network management system. This technical solution improves the accuracy of data integrity assessment by comprehensively monitoring and time series analysis of water flow velocity, pipe pressure, fluid temperature, pipe wall vibration frequency and data transmission status. At the same time, it extracts key parameters of local pipes based on data integrity index, obtains local pipe health assessment values ​​by analyzing the trend change rate of data in the target time interval, and calls the water flow status, pressure and vibration characteristics of abnormal area pipes to analyze the abnormality category and match known fault mode data. Current technologies primarily rely on simple volatility and trend change rate analysis of discrete macroscopic data at sensor nodes. Limited by traditional analytical dimensions, they struggle to accurately capture the time delay characteristics and spatial propagation patterns of water hammer shock waves caused by upstream pumping stations in long-distance main pipelines across multiple time steps. This results in insufficient dynamic prediction capabilities for future hydraulic conditions. Existing technologies can only acquire discrete monitoring data at node locations or perform macroscopic structural stability assessments based on external pipe wall vibration frequencies; the continuous stress state inside the pipeline remains a blind spot in physical monitoring. The inability to accurately analyze the deep-level physical field characteristics of the pipe wall, such as circumferential stress and longitudinal strain, under the combined pressure of internal transient hydraulic compression and uneven external soil settlement makes it difficult to accurately analyze these characteristics. Traditional assessments of pipe network health or anomalies often lack dynamic mechanical evolution mechanisms and cannot quantify the instantaneous fatigue damage caused by water hammer impact on aging pipe walls. Although existing technologies can detect blockages or abnormal fluctuations, it is difficult to analyze the root physical causes when pipe network congestion leads to abnormal increases in liquid level. This lack of cause differentiation limits the accuracy of comprehensive risk assessment of pipe network operation and the scientific nature of operation and maintenance scheduling. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies are unable to accurately capture the time delay characteristics and spatial propagation law of water hammer shock waves caused by upstream pumping stations in long-distance main pipes across multiple time steps, resulting in insufficient dynamic prediction ability of future hydraulic conditions. Existing technologies still have blind spots in the physical detection of continuous stress state inside the pipeline. Existing technologies cannot accurately analyze the deep-level physical field characteristics such as circumferential stress and longitudinal strain of the pipe wall under the combined squeezing of internal transient hydraulic pressure and uneven settlement of external soil. They cannot quantitatively assess the instantaneous fatigue damage caused by water hammer impact on old pipe walls. When the pipeline is congested and the liquid level rises abnormally, existing technologies are unable to analyze the root physical cause.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for risk assessment of the operation of underground drainage pipe networks in municipal buildings, comprising the following steps: Step S1: Collect monitoring data at monitoring points in the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, input the real-time hydraulic situation three-dimensional tensor into a hydraulic topology expansion gated network, and obtain a future hydraulic situation prediction tensor. Step S2: Divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response characteristic matrix. Step S3: Based on the mechanical response feature matrix of the pipe wall, construct the fatigue time-varying state transition chain of the pipe wall structure, and obtain the transient structural failure risk probability through Markov chain iterative calculation. Step S4: Combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct a cross-modal cross attention operator, and obtain the cross attention weight matrix. Step S5: Based on the transient structural damage risk probability and the cross-attention weight matrix, calculate the comprehensive risk score and obtain the operational risk assessment results of the municipal building underground drainage network.

[0005] Preferably, the process of obtaining the future hydraulic situation prediction tensor in step S1 specifically includes: Collect monitoring data, including liquid level and flow rate, and regard the monitoring points as nodes in the drainage pipe network; Construct a real-time hydraulic situation three-dimensional tensor, which represents the liquid level and flow velocity of the monitoring node at the time step; Construct a directed adjacency matrix. The rows and columns of the directed adjacency matrix correspond to the nodes in the drainage network. The elements in the directed adjacency matrix represent the flow direction relationship between the corresponding two nodes. The real-time hydraulic situation three-dimensional tensor is input into a pre-trained hydraulic topology expansion gating network for computation to obtain the future hydraulic situation prediction tensor. The hydraulic topology expansion gated network includes a spatial topology transmission layer, parallel main information extraction branches and gated signal extraction branches, a gated fusion layer and a residual connection layer.

[0006] Preferably, the process of obtaining the pipe wall mechanical response feature matrix in step S2 specifically includes: The underground drainage pipe section between the two nodes is taken as the target analysis pipe section; The length of the target parsing pipe segment is collected, and the length of the target parsing pipe segment is discretized according to the preset length interval. The preset scheduling cycle is discretized according to the preset time interval. All length segmentation points and scheduling cycle segmentation points are combined in pairs to generate a spatiotemporal grid sampling point set; Construct a spatiotemporal coordinate matrix based on a set of spatiotemporal grid sampling points; The spatiotemporal coordinate matrix is ​​input into the pre-fixed Chebyshev polynomial equations of various orders for algebraic calculations to obtain the basis function matrix; Extract the hydraulic spatiotemporal feature slices corresponding to the upstream and downstream monitoring nodes of the target analytical pipe section from the future hydraulic situation prediction tensor. Flatten and stitch the data of the hydraulic spatiotemporal feature slices of the upstream and downstream monitoring nodes to generate boundary condition feature vectors. The boundary condition feature vector is input into the spectral domain physical evolution feedforward network, which outputs an extended coefficient matrix. This network performs data dimensionality reduction and nonlinear mapping through several hidden layers. The activation function in all hidden layers is a hydraulic catastrophe adaptive activation function, the mathematical expression of which is: ; in, Let be the adaptive scaling parameters to be solved. These are the linearly mapped feature values ​​of the input features received by the hidden layer neurons from the previous layer. The basis function matrix and the expansion coefficient matrix are multiplied to obtain the physical field feature matrix, which includes the predicted water depth, the predicted flow velocity, the predicted circumferential stress of the pipe wall, and the predicted longitudinal strain of the pipe wall.

[0007] Preferably, in step S2, the process of obtaining the pipe wall mechanical response characteristic matrix further includes: Using an automatic differentiation mechanism, analytical partial derivatives of the values ​​in each column of the physical field feature matrix with respect to the input spatiotemporal coordinate matrix are obtained, and the mass conservation residuals, momentum conservation residuals, and pipe-soil coupling mechanical residuals are calculated using the Saint-Venant equation and the Mohr-Coulomb model. A physical control loss function with volume weights is constructed. An adaptive moment estimation optimization algorithm is adopted to minimize the physical control loss function. All weight matrices, bias vectors and adaptive scaling parameter vectors inside the spectral domain physical evolution feedforward network are continuously updated through backpropagation. When the calculated value of the physical control loss function is continuously iterated to a preset number of times or the change is less than a preset minimum threshold, convergence is determined and the iteration is stopped. Extract the column vectors representing circumferential stress and longitudinal strain from the physical field feature matrix of the last iteration, and concatenate them according to the column vectors to generate the mechanical response feature matrix of the pipe wall.

[0008] Preferably, step S3, the process of obtaining the probability of transient structural failure risk, specifically includes: According to the pre-defined discrete time steps, the preset scheduling cycle is divided into multiple time windows. The different length cutting points and row vectors corresponding to different time steps in each time window are extracted from the pipe wall mechanical response feature matrix. All absolute values ​​of circumferential stress in the same time window are traversed. The row vector containing the largest absolute value of circumferential stress in the time window is selected as the force row vector under that time step. The force row vectors under each time step are spliced ​​together in chronological order to obtain the force time sequence matrix. Extract the circumferential stress and longitudinal strain of the pipe wall from each row of the stress time series matrix, perform mechanical equivalent conversion, and construct a stress intensity column vector.

[0009] Preferably, step S3, the process of obtaining the probability of transient structural failure risk, further includes: A blank state transition probability matrix is ​​initialized for each time step within a preset scheduling period. The state transition probability matrix is ​​used to record the state transition probability of the pipeline state transitioning from a preset first state to a preset second state and from a preset second state to a preset third state at each time step. The state transition probability is calculated by combining the logistic activation function with the equivalent force intensity factor.

[0010] Preferably, step S3, the process of obtaining the probability of transient structural failure risk, further includes: Construct an initial health state row vector, wherein the initial health state row vector is a preset row vector; The initial health state row vector is multiplied continuously with the state transition probability matrix within the preset scheduling period, and finally the termination state row vector at the end of the last time step is output. Extract the last element of the terminal state row vector as the transient structure failure risk probability.

[0011] Preferably, step S4, the process of obtaining the cross-attention weight matrix, specifically includes: Extract the liquid level data of the downstream monitoring node of the target analysis pipe section from the monitoring data, and construct the actual liquid level column vector; Extract the corresponding real-time hydraulic situation slices of the upstream and downstream monitoring nodes of the target analytical pipe section from the real-time hydraulic situation three-dimensional tensor, flatten and stitch the data of the real-time hydraulic situation slices of the upstream and downstream monitoring nodes to generate real-time boundary condition feature vectors. Each time the target parsing pipe segment is parsed, a copy of the current spectral domain physical evolution feedforward network parameters is made as an independent copy. The real-time boundary condition feature vector is input into the network parameter copy, and the reverse optimization iteration update is continuously performed on the network parameter copy a preset number of times according to the physical control loss function and the adaptive moment estimation optimization algorithm. After the network parameter replicas converge, the real-time physical field feature matrix of the last iteration is obtained. For the spatial dimension expansion of the real-time physical field feature matrix caused by grid discretization, the liquid level feature column corresponding to the actual liquid level column vector on the spatial topology node is extracted as the theoretical liquid level column vector composed of the theoretical liquid level values ​​at each time step in the past preset time period. Subtract the theoretical liquid level column vector from the actual liquid level column vector to generate the hydraulic situation residual column vector. The column representing circumferential stress is extracted from the real-time physical field feature matrix. All spatial division points at each time step are traversed. The maximum circumferential stress values ​​corresponding to each time step are selected and concatenated into an extreme value column vector. The maximum circumferential stress difference between adjacent time steps is calculated to generate a pipe-soil coupling stress gradient column vector. Perform matrix multiplication on the hydraulic situation residual column vector and the preset first attention mapping matrix to output the query matrix; Perform matrix multiplication on the column vector of soil-tube coupling stress gradient and the preset second attention mapping matrix to output the key matrix; The dot product attention mechanism is used to calculate the cross-association degree between the query matrix and the key matrix.

[0012] Preferably, step S5, the process of obtaining the operational risk assessment results of the municipal building underground drainage network, specifically includes: Calculate the arithmetic mean of all elements in the column vector of hydraulic state residuals, and denote it as the residual mean; If the mean residual is less than the preset abnormal liquid level threshold, the pipeline network is considered to be operating normally. If the mean residual value is greater than or equal to the preset abnormal liquid level threshold, the pipeline network is determined to be in abnormal operation, and a comprehensive risk score is calculated. The comprehensive risk score is the product of the transient structural damage risk probability and the preset hard penalty amplification coefficient. The process of obtaining the hard penalty amplification factor specifically includes: Extract the main diagonal elements from the cross-attention weight matrix and calculate their arithmetic mean, which is denoted as the diagonal mean. If the mean value of the diagonal is less than the preset stress mutation threshold, the target analysis section is determined to have a functional siltation risk; if the mean value of the diagonal is greater than or equal to the preset stress mutation threshold, the target analysis section is determined to have a structural settlement and backwater risk.

[0013] A risk assessment system for the operation of underground drainage pipe networks in municipal buildings includes a data acquisition module, a physical field analysis module, a fatigue state evolution module, a cross-modal attention module, and an assessment module. The acquisition module is used to collect monitoring data at monitoring points of the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, and input the real-time hydraulic situation three-dimensional tensor into a hydraulic topology expansion gated network to obtain a future hydraulic situation prediction tensor. The physical field analysis module is used to divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response feature matrix. The fatigue state evolution module is used to construct a time-varying fatigue state transition chain of the pipe wall structure based on the mechanical response feature matrix of the pipe wall, and obtain the probability of transient structural failure risk through Markov chain iterative calculation. The cross-modal attention module is used to combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct a cross-modal cross attention operator, and obtain the cross attention weight matrix. The assessment module is used to calculate a comprehensive risk score based on the transient structural damage risk probability and the cross-attention weight matrix, and to obtain the operational risk assessment results of the municipal building underground drainage network.

[0014] The beneficial effects of this invention are as follows: By constructing a hydraulic topology-gated network and utilizing dilated causal convolution operations, this invention expands the receptive field of the convolution kernel in the time series without increasing the number of network training parameters. This allows for the accurate capture of the long-sequence time lag correlation between sudden changes in the forced drainage of upstream pumping stations and the surge in liquid level in distant downstream pipe sections. Based on this, this invention uses the macroscopic hydraulic characteristics at the nodes as dynamic boundary conditions, introduces a spectral domain physical evolution feedforward network, and constructs a volume-weighted physical control loss function by combining the Saint-Venant equation and the Mohr-Coulomb model. This not only considers the surge in water pressure caused by the transient impact of internal water flow but also takes into account the pressure caused by uneven settlement of external soil. As a result, it can analyze the physical field characteristics such as the circumferential stress and longitudinal strain of the pipe wall inside the target analytical pipe section, significantly improving the spatial resolution and accuracy of the monitoring of the stress state of the pipeline network. This invention constructs a time-varying state transition chain for pipe wall fatigue based on the pipe wall mechanical response characteristic matrix and introduces Markov chain iterative calculation. It can dynamically update the probability of the pipeline transitioning from an intact state to a microcrack yielding state or even a through-crack instability state. This time-varying state evolution model based on mechanical equivalent conversion and logistic activation function can measure the instantaneous fatigue damage caused by the water hammer shock wave to the old pipe wall at the moment the pump station starts forced drainage. It provides a reliable absolute mathematical expectation value and transient structural damage risk probability for the preventive maintenance of the pipeline network structure, filling the gap in the existing technology in transient pressure fatigue evolution analysis. When pipe network congestion leads to poor water flow, this invention constructs a cross-modal cross-attention operator to calculate the resonance probability between the hydraulic state residual caused by abnormal rise in liquid level and the pipe-soil coupled stress gradient. This allows for precise differentiation between functional siltation caused by internal sediment and foreign matter accumulation and structural settlement caused by external soil settlement that flattens and deforms the pipe. Based on this causal classification, a differentiated hard penalty amplification coefficient is matched when calculating the comprehensive risk score. A lower penalty weight is applied to functional siltation that does not damage the overall structural pressure-bearing skeleton, while a very high penalty weight is applied to structural deformation that leads to a precipitous drop in the mechanical yield limit and is prone to brittle fracture. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a method for assessing the operational risk of underground drainage pipe networks in municipal buildings, as provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 This paper provides a method for risk assessment of the operation of underground drainage pipe networks in municipal buildings, including the following steps: Step S1: Collect monitoring data at monitoring points in the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, input the real-time hydraulic situation three-dimensional tensor into the hydraulic topology expansion gated network, and obtain the future hydraulic situation prediction tensor. Step S2: Divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response characteristic matrix. Step S3: Based on the mechanical response feature matrix of the pipe wall, construct the fatigue time-varying state transition chain of the pipe wall structure, and obtain the transient structural failure risk probability through Markov chain iterative calculation. Step S4: Combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct a cross-modal cross attention operator, and obtain the cross attention weight matrix. Step S5: Based on the transient structural failure risk probability and the cross-attention weight matrix, calculate the comprehensive risk score and obtain the operational risk assessment results of the municipal building underground drainage network.

[0018] This invention integrates data-driven hydraulic situation prediction with physical constraint-based pipe wall mechanical analysis, transforming macroscopic liquid level and flow velocity monitoring data into microscopic continuous pipe wall mechanical response characteristics. Furthermore, it utilizes a Markov chain iterative model and a cross-modal cross-attention mechanism to calculate transient failure risks and comprehensive scores, achieving a leap from static, empirical assessment of pipeline network operation risks to dynamic, high-resolution physical evolution deduction, significantly improving the scientific rigor and comprehensiveness of urban drainage system safety early warning under extreme conditions.

[0019] Step S1, the process of obtaining the future hydraulic situation prediction tensor, specifically includes: Collect continuous monitoring data at various monitoring points in the drainage pipe network, including liquid level and flow velocity, and regard the monitoring points as nodes in the drainage pipe network; Construct a real-time hydraulic situation three-dimensional tensor, which represents the liquid level and flow velocity of the monitoring node at the time step; Construct a directed adjacency matrix. The rows and columns of the directed adjacency matrix correspond to the nodes in the drainage network. The elements in the directed adjacency matrix represent the flow direction relationship between the corresponding two nodes. The real-time hydraulic situation three-dimensional tensor is input into a pre-trained hydraulic topology expansion gating network for computation. The hydraulic topology expansion gating network includes a spatial topology transmission layer, parallel main information extraction branches and gating signal extraction branches, a gating fusion layer and a residual connection layer. The calculation process for the spatial topology transmission layer includes: The real-time hydraulic situation three-dimensional tensor is divided into independent two-dimensional feature slices along the time step dimension. The feature slices represent the monitoring data of all nodes of the drainage network in one time step. For any feature slice, use it as the multiplicand matrix. Multiply the directed adjacency matrix with the preset spatial weight matrix as the left multiplication matrix. Perform matrix multiplication operations in sequence to obtain intermediate feature slices. Reassemble all intermediate feature slices in the third dimension according to the original time sequence to obtain the three-dimensional tensor of spatial hydraulic features. The three-dimensional tensor of spatial hydraulic features is synchronously split into parallel main information extraction branch and gated signal extraction branch to obtain the three-dimensional tensor of main information and the three-dimensional tensor of gated signal, respectively. The master information 3D tensor and the gated signal 3D tensor are input into the gated fusion layer for Hadamard product operation, and the intermediate hydraulic tensor is output. The intermediate hydraulic tensor is input into the residual connection layer, and the predicted hydraulic spatiotemporal features at the next time step are output. The hydraulic topology expansion gated network also generates predicted hydraulic characteristics at multiple time steps within a preset scheduling period through an autoregressive rolling prediction mechanism; The predicted hydraulic features at multiple time steps are spliced ​​together to generate a tensor for predicting the future hydraulic situation.

[0020] In one specific embodiment of the present invention, the IoT sensors at each monitoring point of the drainage pipe network collect monitoring data from the past hour at a preset sampling frequency of 5 minutes. The first dimension of the real-time hydraulic situation three-dimensional tensor is the number of detection points, the second dimension is fixed at 2, representing the two extracted features of liquid level and flow velocity; the third dimension represents the time step within the time window of continuous monitoring in the past. In this embodiment, monitoring data within the past 60 minutes are continuously collected at a sampling frequency of 5 minutes, and the third dimension has a length of 12 time steps. Each element in the real-time hydraulic situation three-dimensional tensor represents the liquid level and flow velocity values ​​of the corresponding node at a specific time step. A directed adjacency matrix is ​​constructed based on the direction of underground drainage pipes. The directed adjacency matrix is ​​a two-dimensional square matrix, and the number of rows and columns are the number of monitoring points. When there is a real underground drainage pipe that flows unidirectionally from node i to node j, the number of elements in the i-th row and j-th column of the directed adjacency matrix is ​​set to 1. When there is no real underground drainage pipe that flows unidirectionally from node i to node j, the number of elements in the i-th row and j-th column of the directed adjacency matrix is ​​set to 0. The hydraulic situation tensor is input into the hydraulic topology expansion gated network for calculation. Since the water hammer shock wave caused by the pump station’s forced discharge has a very obvious time delay characteristic in the propagation of the long-distance main pipe, the ordinary convolutional network is limited by a small receptive field and cannot extract the propagation law of the hydraulic wave peak across multiple time steps. Therefore, in this embodiment, conventional convolution is excluded in the hydraulic topology expansion gated network and the expansion causal convolution operation is adopted. Both the main information extraction branch and the gated signal extraction branch adopted the dilated causal convolution operation with a dilation factor of 2. This operation expands the receptive field of the convolution kernel in the time series exponentially without increasing the number of network training parameters, ensuring that the network processes data strictly in chronological order and accurately captures the long-sequence time lag correlation between the sudden change in the upstream pumping station's forced drainage and the surge in liquid level in the downstream remote pipe section. In the main information extraction branch, the real-time hydraulic situation three-dimensional tensor is processed by the expansion causal convolution operation of the first weight matrix and the first bias vector and hyperbolic tangent activation function to output the main information three-dimensional tensor. In the main information extraction branch, the real-time hydraulic situation three-dimensional tensor is processed by the expansion causal convolution operation of the second weight matrix, and by the second bias vector and the sigmoid activation function to output the gated signal three-dimensional tensor. In the gated fusion layer, the main information three-dimensional tensor and the gated signal three-dimensional tensor are subjected to Hadamard product operation to output the intermediate hydraulic tensor. Using the zero-to-one weight coefficients in the gated signal three-dimensional tensor, the invalid redundant features generated by daily stable drainage are automatically suppressed and eliminated, while retaining the sudden spatial surge and time-dependent information caused by water hammer wave peaks. In the residual connection layer, the real-time hydraulic situation three-dimensional tensor is added element by element to the above-filtered intermediate hydraulic tensor to output the predicted hydraulic spatiotemporal characteristics. This residual connection structure forcibly preserves the original bottom-level water flow physical information, solving the problem of water flow characteristic attenuation and gradient vanishing caused by the deepening of the hydraulic transmission layer in the pipeline network. The autoregressive rolling prediction mechanism specifically includes: The hydraulic topology expansion gated network is based on the monitoring data of the past 12 time steps. In a single forward calculation, it outputs the predicted hydraulic spatiotemporal features of the first time step in the future, i.e. the 13th time step. The predicted hydraulic spatiotemporal features are a two-dimensional matrix, which is used to represent the predicted liquid level and flow velocity at each node in the next time step. The data from the 2nd to the 12th time steps in the real-time hydraulic situation 3D tensor are concatenated with the predicted hydraulic spatiotemporal features from the 13th time step to form a new hydraulic situation 3D tensor, which is used as the output of the hydraulic topology expansion gated network. The predicted hydraulic spatiotemporal features of the 13th time step are used as known real data to supplement the predicted hydraulic spatiotemporal features of the real-time hydraulic situation three-dimensional tensor in the future 2nd time step, i.e. the 14th time step. This process is repeated 12 times to output the predicted hydraulic spatiotemporal features for 12 consecutive time steps within a preset scheduling period (60 minutes in this example). The predicted hydraulic spatiotemporal features of all nodes in the network are then stitched together to generate a three-dimensional future hydraulic situation prediction tensor. The future hydraulic situation prediction tensor includes: the first dimension is the number of nodes, the second dimension is fixed at 2, representing the two extracted features of liquid level and flow velocity; the third dimension represents the time steps within the time window of future continuous monitoring. In this embodiment, the third dimension is 12 time steps. Each element in the real-time hydraulic situation three-dimensional tensor represents the liquid level and flow velocity values ​​of the corresponding node at a specific time step. The pre-training process of hydraulic topology-extended gated networks specifically includes: The target city's drainage network real IoT monitoring historical dataset from the past twelve months is retrieved. From this dataset, continuous monitoring data including rainstorm periods, pumping station flood discharge periods, and stable operation periods are extracted to construct a large amount of training data. In this embodiment, the input data in the training data is a real-time hydraulic situation three-dimensional tensor of continuous time steps. In this embodiment, there are 12 time steps. The label data corresponding to the input data is the 13th time step, that is, the liquid level and flow rate in the 5th minute in the future. The training data is batch-input into an uninitialized hydraulic topology expansion gated network. The network forward computes and outputs the predicted liquid level and flow velocity at the 13th time step. The mean square error function is used to calculate the error values ​​between the predicted liquid level and flow velocity and the actual label data matrix. An adaptive moment estimation optimization algorithm is used to calculate the error derivative layer by layer through backpropagation along the gradient direction where the mean square error decreases. The spatial weight matrix, the first weight matrix, the second weight matrix, the first bias vector, and the second bias vector are iteratively updated repeatedly. When the mean square error value calculated on the validation set is less than the set convergence threshold of 0.001 for ten consecutive iterations, the network training is completed and the iteration stops. The first weight matrix, the second weight matrix, the first bias vector, and the second bias vector in the network are then fixed.

[0021] This invention addresses the severe time delay characteristic of water hammer shock waves caused by forced drainage from pumping stations propagating in long-distance main pipelines. By introducing an expanded causal convolution operation with an expansion factor of 2 into the main information extraction branch and the gated signal extraction branch, the network can expand the receptive field of the convolution kernel in the time series without increasing the number of model training parameters. Combined with a gated fusion layer using Hadamard product operation, the network can automatically suppress the ineffective redundant features generated by daily stable drainage and retain the sudden spatial surge and time-dependent information caused by water hammer wave peaks, thereby ensuring extremely high accuracy and computational efficiency in long-sequence hydraulic situation prediction.

[0022] Step S2, the process of obtaining the characteristic matrix of the pipe wall mechanical response, specifically includes: The underground drainage pipe section between the two nodes is taken as the target analysis pipe section; The length of the target analysis pipe segment is collected, and the length of the target analysis pipe segment is discretized and divided according to a preset length interval of 1 meter. The preset scheduling cycle of 60 minutes is discretized according to a preset time interval of 5 minutes. All length segmentation points and scheduling cycle segmentation points are combined in pairs to generate a set of spatiotemporal grid sampling points covering all spatiotemporal locations within the target parsing pipe segment; A spatiotemporal coordinate matrix is ​​constructed based on the set of spatiotemporal grid sampling points. The number of rows in the spatiotemporal coordinate matrix is ​​the total number of spatiotemporal grid sampling points, and the number of columns is 2. The first column represents the distance of the spatiotemporal grid sampling point from the starting point of the target analytical pipe segment, and the element values ​​in the second column represent the time of the sampling point. The spatiotemporal coordinate matrix is ​​input into the pre-fixed Chebyshev polynomial equations of various orders for algebraic calculations to obtain the basis function matrix; Extract the hydraulic spatiotemporal feature slices corresponding to the upstream and downstream monitoring nodes of the target analytical pipe section from the future hydraulic situation prediction tensor. Flatten and stitch the data of the hydraulic spatiotemporal feature slices of the upstream and downstream monitoring nodes to generate boundary condition feature vectors. The boundary condition feature vector is input into the spectral domain physical evolution feedforward network, which outputs an extended coefficient matrix. The spectral domain physical evolution feedforward network performs data dimensionality reduction and nonlinear mapping through several hidden layers (four in this embodiment). The activation function in all hidden layers is the hydraulic catastrophe adaptive activation function, and the mathematical expression of the hydraulic catastrophe adaptive activation function is: ; in, The adaptive scaling parameters to be solved are defined as follows: each hidden layer neuron has an independent adaptive scaling parameter, and the adaptive scaling parameters of all neurons in the network together constitute the adaptive scaling parameter vector. During backpropagation, the adaptive scaling parameter vector is optimized. These are the linearly mapped feature values ​​of the input features received by the hidden layer neurons from the previous layer. The basis function matrix and the expansion coefficient matrix are multiplied to obtain the physical field feature matrix, which includes the predicted water depth, the predicted flow velocity, the predicted circumferential stress of the pipe wall, and the predicted longitudinal strain of the pipe wall.

[0023] In a specific embodiment of the present invention, the monitoring points are used as topological nodes in step S1, and the underground drainage pipe sections connecting these monitoring points are used as topological edges. Since IoT sensors can only acquire discrete data of node locations, the continuous stress state inside the pipe is a blind spot of physical monitoring. Therefore, the present invention uses the macroscopic hydraulic characteristics at the nodes as dynamic boundary conditions to drive the physical equations to analyze the microscopic continuous stress field inside the pipe. In this embodiment, the highest order of the Chebyshev polynomial is 19, the number of rows of the basis function matrix is ​​the total number of spatiotemporal grid sampling points, and the number of columns is 20. The elements in the basis function matrix represent the basic algebraic mapping values ​​of the spatial coordinates and time coordinates inside the pipe at different polynomial frequencies. The hydraulic spatiotemporal feature slice is a two-dimensional matrix with the number of rows representing the number of time steps and the number of columns being 2, representing the liquid level value and the flow velocity value respectively. The boundary condition feature vector condenses the transient water pressure information caused by the hydraulic shock wave of the entire network to both ends of this section of the pipeline. The output of the spectral domain physical evolution feedforward network is a two-dimensional expansion coefficient matrix with 20 rows and 4 columns. The elements inside the expansion coefficient matrix represent the physical field expansion coefficients predicted by the network based on the hydraulic pressure situation at both ends of the target analytical pipe segment. The physical field feature matrix has 4 columns, with the number of rows equal to the total number of spatiotemporal grid sampling points. The elements in the first column of the physical field feature matrix represent the predicted water depth, the elements in the second column represent the predicted flow velocity, the elements in the third column represent the predicted pipe wall circumferential stress, and the elements in the fourth column represent the predicted pipe wall longitudinal strain.

[0024] This invention further defines the specific process of obtaining the pipe wall mechanical response feature matrix through a spectral domain physical evolution feedforward network, eliminating the physical monitoring blind spot where traditional IoT sensors cannot detect the continuous stress state inside the pipe. This invention uses the macroscopic hydraulic features at the nodes as dynamic boundary conditions, constructs the basis function matrix using Chebyshev polynomials, and constructs a hydraulic catastrophe adaptive activation function with adaptive scaling parameters. This function can drive the physical equations to solve for the microscopic continuous circumferential stress and longitudinal strain of the pipe wall inside the pipe, realizing the nonlinear mapping and reconstruction of discrete macroscopic hydraulic data into a continuous microscopic mechanical field.

[0025] Step S2, the process of obtaining the characteristic matrix of the pipe wall mechanical response, also includes: Using the automatic differentiation mechanism in the deep learning framework, the analytical partial derivatives of each column of the physical field feature matrix with respect to the input spatiotemporal coordinate matrix are obtained, and the mass conservation residuals, momentum conservation residuals, and pipe-soil coupling mechanical residuals are calculated through the Saint-Venant equation and the Moore-Coulomb model. Construct a physical control loss function with volume weights. The mathematical expression of the physical control loss function is as follows: ; in, This represents the final calculated value of the physical control loss function. The total number of spatiotemporal grid sampling points. For the first The unit volume value of the finite volume element corresponding to each sampling point is obtained by calculating the cross-sectional area corresponding to the inner diameter of the pipe and multiplying it by the length interval of 1 meter set by the spatial discretization. The multiplication weighting of this value forcibly amplifies the error penalty for large diameter pipe areas. , , Representing the first Normalized mass conservation residuals, momentum conservation residuals, and pipe-soil coupling mechanics residuals at each sampling point; An adaptive moment estimation optimization algorithm is adopted to minimize the physical control loss function. All weight matrices, bias vectors and adaptive scaling parameter vectors inside the spectral domain physical evolution feedforward network are continuously updated through backpropagation. When the calculated value of the physical control loss function is continuously iterated to a preset number of 50 times or the change is less than a preset minimum threshold of 0.001, convergence is determined and the iteration stops. Extract the column vectors representing circumferential stress and longitudinal strain from the physical field feature matrix of the last iteration, and concatenate them according to the column vectors to generate the mechanical response feature matrix of the pipe wall.

[0026] In a specific embodiment of the present invention, since the underground pipeline is not suspended but is encased in the soil, under the condition of heavy rain and strong drainage, the pipeline network not only bears the surge in internal water pressure, but also faces the pressure from uneven settlement of the external soil due to local soil erosion or changes in the upper load. The Saint-Venant equation is used to represent the surge in internal water pressure borne by the target analytical pipe section, and the Mohr-Coulomb model is used to represent the pressure from uneven settlement of the external soil. The actual circumferential stress and longitudinal strain of the pipe wall of the target analytical pipe section are determined by the combined compression of the internal water hammer impact force and the external Mohr-Coulomb soil pressure. The inputs to the Saint-Venant equation include: water depth values, flow velocity values, spatial partial derivatives of water depth, spatial partial derivatives of flow velocity, time partial derivatives of water depth, time partial derivatives of flow velocity, pipe inner diameter, pipe inner wall roughness coefficient, and pipe bottom slope values ​​extracted from the physical field feature matrix. The data on pipe inner diameter and pipe inner wall roughness coefficient are derived from pipe specification data. The value of pipe bottom slope is obtained by dividing the difference in elevation data of the upstream and downstream nodes by the physical length of the pipe. The inputs to the Mohr-Coulomb model include: the pipe wall circumferential stress value, pipe wall longitudinal strain value, circumferential stress spatial partial derivative value, soil cohesion value, soil internal friction angle value, overburden unit weight value, pipe burial depth value, and pipe wall thickness value extracted from the physical field feature matrix. The soil cohesion, soil internal friction angle, and overburden unit weight data are derived from geological survey data; the pipe burial depth data are derived from drainage network design drawings; and the pipe wall thickness data are derived from the pipe specifications provided by the manufacturer.

[0027] This invention deeply couples the Saint-Venant equation with the Mohr-Coulomb model, recreating the real and complex physical environment under heavy rain and strong drainage conditions where the pipe network simultaneously bears the pressure of a surge in internal water pressure and uneven settlement of external soil. By using an automatic differentiation mechanism to obtain partial derivatives and calculate the mass, momentum, and pipe-soil coupling mechanical residuals, and by introducing the volume weight of the pipe diameter cross-sectional area element, the error penalty intensity in large-diameter areas is forcibly amplified. This not only ensures that the neural network prediction results follow the conservation laws of fluid mechanics and solid mechanics.

[0028] Step S3, the process of obtaining the probability of transient structural failure risk, specifically includes: According to the pre-defined discrete time steps, the preset scheduling cycle is divided into multiple time windows. The different length cutting points and row vectors corresponding to different time steps in each time window are extracted from the pipe wall mechanical response feature matrix. All absolute values ​​of circumferential stress in the same time window are traversed. The row vector containing the largest absolute value of circumferential stress in the time window is selected as the force row vector under that time step. The force row vectors under each time step are spliced ​​together in chronological order to obtain the force time sequence matrix. Extract the circumferential stress and longitudinal strain of the pipe wall from each row of the stress time series matrix, perform mechanical equivalent conversion, and construct a stress intensity column vector. The mathematical expression for the mechanical equivalent conversion is: ; in, The equivalent stress strength factor is calculated as follows. For the circumferential stress of the pipe wall, For the longitudinal strain of the pipe wall, This represents the elastic modulus constant of the pipe material, with data sourced from pipe specification data.

[0029] In a specific embodiment of the present invention, since traditional pipeline risk assessments mostly rely on static data such as the laying age of pipe sections, they cannot measure the instantaneous damage caused to the aging pipe walls by the water hammer shock wave at the moment the pump station starts forced drainage, this embodiment constructs a time-varying state transition chain for pipe wall structure fatigue. Using the high-precision physical field data calculated in step S2, the probability of pipeline rupture is dynamically updated second by second, specifically including: After performing mechanical equivalent conversion on each row of the force time series matrix, a one-dimensional force intensity column vector is obtained. Each element inside the column vector is the equivalent force intensity factor at the corresponding time step.

[0030] This invention abandons the crude evaluation mode that relies solely on a single stress value. By integrating the circumferential stress of the pipe wall, longitudinal strain, and the elastic modulus constant of the pipe material, an equivalent stress strength factor is constructed. This factor can condense the true structural stress strength state of the pipe under extreme hydraulic impact, laying a mechanical data foundation for the subsequent accurate calculation of the yield and rupture probability of the pipe.

[0031] Step S3, the process of obtaining the probability of transient structural failure risk, also includes: A blank state transition probability matrix is ​​initialized for each time step within the preset scheduling period. The state transition probability matrix is ​​used to record the state transition probability of the pipeline state transitioning from the preset first state to the preset second state and from the preset second state to the preset third state at each time step. The state transition probability is calculated by combining the logistic activation function with the equivalent stress intensity factor. The mathematical expression for the state transition probability is: ; ; in, Let be the first probability of transitioning from the first state to the second state at the i-th time step. Let be the second probability of transitioning from the second state to the third state at the i-th time step. This is the mutation sensitivity amplification factor, which is set to 10 in this embodiment. Let be the equivalent stress intensity factor at the i-th time step. The first-level yield strength threshold is set to 2.5 MPa in this embodiment. The second-level fracture strength threshold is set to 5.0 MPa in this embodiment.

[0032] The present invention presupposes three progressive states of irreversible pipeline conditions: the first state represents an intact pipeline structure, the second state represents the yielding of micro-cracks in the inner wall of the pipeline, and the third state represents the pipeline experiencing a through-through rupture and instability. The state transition probability matrix has a dimension of 3 rows and 3 columns. The element in the first row and first column is 1, which is the difference between the first probability and the first probability. The element in the first row and second column is the first probability, and the element in the first row and third column is 0. The element in the second row and first column is 0, the element in the second row and second column is 1, which is the difference between the second probability and the second probability, and the element in the second row and third column is the second probability. The elements in the third row and first column and the second column are both 0, and the element in the third row and third column is 1. There is a corresponding state transition probability matrix for each time step.

[0033] This invention, by pre-setting three irreversible progressive states—intact, microcrack yielding, and through-fracture instability—and introducing a mutation sensitivity amplification factor and multi-level strength thresholds, can characterize the nonlinear damage evolution law caused by material fatigue when old pipe walls are subjected to hydraulic impact. It can also dynamically update the theoretical probability of pipeline condition deterioration based on the ever-changing equivalent stress intensity.

[0034] Step S3, the process of obtaining the probability of transient structural failure risk, also includes: Construct an initial health state row vector, which is a preset row vector; Multiply the initial health state row vector by the state transition probability matrix of the first time step within the preset scheduling period to obtain the state row vector at the end of the first time step. Multiply the output state row vector by the state transition probability matrix of the second time step, and perform matrix multiplication continuously to finally output the termination state row vector at the end of the last time step. Extract the last element of the terminal state row vector as the transient structure failure risk probability.

[0035] In a specific embodiment of the present invention, the initial health state row vector is: This means that the target pipe section is in the first state of 100% before the pump station's forced drainage is started. Perform Markov chain iterative computation, which specifically includes: Multiply the initial healthy state row vector by the state transition probability matrix of the first time step to obtain the state row vector at the end of the first time step. Then multiply the output state row vector by the state transition probability matrix of the second time step. Perform 12 consecutive matrix multiplication operations in the order of time progression to finally output the termination state row vector at the end of the last time step. The third element in the termination state row vector represents the absolute mathematical expectation of the target analytical pipe segment finally falling into the third state after experiencing a hydraulic impact evolution of up to 60 minutes. This single absolute mathematical expectation of the final falling into the third state is used as the probability of transient structural damage risk.

[0036] This invention, by continuously performing matrix multiplication operations within multiple time steps of a preset scheduling cycle, propagates and accumulates the transient hydraulic impact impact of each second step by step. The absolute mathematical expectation value output by the time-series Markov chain deduction objectively reflects the ultimate risk probability of the target pipe section undergoing a complete hydraulic impact evolution cycle, thus making up for the shortcomings of traditional static risk assessment in dealing with transient high-pressure impact fatigue.

[0037] Step S4, the process of obtaining the cross-attention weight matrix, specifically includes: Extract the liquid level data of the downstream monitoring node of the target analysis pipe section from the monitoring data, and construct the actual liquid level column vector; Extract the corresponding real-time hydraulic situation slices of the upstream and downstream monitoring nodes of the target analytical pipe section from the real-time hydraulic situation three-dimensional tensor, flatten and stitch the data of the real-time hydraulic situation slices of the upstream and downstream monitoring nodes to generate real-time boundary condition feature vectors. Each time the target parsing pipe segment is parsed, a copy of the current spectral domain physical evolution feedforward network parameters is made as an independent copy. The real-time boundary condition feature vector is input into the copy of the network parameters, and the reverse optimization iteration update is continuously performed on the copy of the network parameters a preset number of times (50 times in this embodiment) according to the physical control loss function and the adaptive moment estimation optimization algorithm. After the network parameter replicas converge, the real-time physical field feature matrix of the last iteration is obtained. In view of the spatial dimension expansion of the real-time physical field feature matrix caused by grid discretization, the liquid level feature column corresponding to the actual liquid level column vector on the spatial topology node is extracted as the theoretical liquid level column vector composed of the theoretical liquid level values ​​at each time step in the past preset time period. Subtract the theoretical liquid level column vector from the actual liquid level column vector to generate the hydraulic situation residual column vector; The column representing circumferential stress is extracted from the real-time physical field feature matrix. All spatial division points at each time step are traversed. The maximum circumferential stress values ​​corresponding to each time step are selected and concatenated into an extreme value column vector. The maximum circumferential stress difference between adjacent time steps is calculated to generate a pipe-soil coupling stress gradient column vector. The hydraulic state residual column vector is compared with a preset size. Perform matrix multiplication on the first attention mapping matrix and output the query matrix; The column vector of the soil-tube coupling stress gradient is coupled with a preset size. The second attention mapping matrix is ​​multiplied to output the key matrix; The dot product attention mechanism is used to calculate the cross-association degree between the query matrix and the key matrix. The mathematical expression for the cross-association degree is: ; in, To calculate the cross-attention weight matrix of the output; For query matrix; This represents the transpose operation on the key matrix; The dimension representing the feature mapping is 16 in this embodiment.

[0038] In a specific embodiment of the present invention, when the pipe network is congested and water flow is obstructed, the causes are divided into two categories: the first category is functional siltation caused by foreign objects or silt entering the pipe; the second category is structural settlement caused by the subsidence of the soil outside the pipe, which flattens and deforms the pipe. By accurately distinguishing these two causes, the present invention constructs a cross-modal cross-attention operator. Real-time hydraulic situation slices corresponding to the upstream and downstream monitoring nodes of the target analytical pipe segment are extracted from the three-dimensional tensor of the real-time hydraulic situation. The real-time hydraulic situation slices are two-dimensional matrices with the number of rows representing the number of time steps and two columns representing the liquid level and flow velocity values, respectively. The data of the real-time hydraulic situation slices of the upstream and downstream monitoring nodes are flattened and stitched together to generate a one-dimensional real-time boundary condition feature vector. When analyzing the target analytical pipe segment, in order to avoid interfering with the global model, a copy of the parameters of the spectral domain physical evolution feedforward network in step S2 is made, and the real-time boundary condition feature vector is input into this parameter copy. Following the physical control loss function and adaptive moment estimation optimization algorithm described in step S2, the network forward mapping and backward optimization iteration process is executed 50 times on the parameter copy. When the set convergence condition is met, the real-time physical field feature matrix of the last iteration output is obtained. Since this matrix contains data from many spatiotemporal grid sampling points inside the pipe section, the spatial dimension has been expanded. Therefore, it is necessary to extract feature columns that match the actual sensor spatial positions at both ends of the pipe section and represent the water depth values, as theoretical liquid level column vectors composed of theoretical liquid level values ​​in 12 time steps over the past hour. Extract the third column representing circumferential stress, calculate the difference in circumferential stress between adjacent time steps, and generate a column vector of pipe-soil coupling stress gradient. The length of the column vector of pipe-soil coupling stress gradient is 12. The value of the first element adopts a zero-value padding strategy, that is, the change in circumferential stress at the initial time step is set to 0, and the remaining elements are the difference in circumferential stress between the subsequent 11 time steps and the previous time step. In the generated cross-attention weight matrix, the main diagonal elements Representing the past The abnormal rise in liquid level at time step 1 is related to the first time step 2. The probability of resonance between the sudden changes in pipe wall stress occurring at each time step; The process of obtaining the first attention mapping matrix and the second attention mapping matrix specifically includes: Historical operation and maintenance records of the municipal underground drainage network were retrieved, and pipe sections that had experienced abnormal water level backlogs were identified. For each historical backlog event, the true causal qualitative conclusions were extracted from the accident detection reports in the operation and maintenance records and converted into labeled data for supervision and training. Specifically, when the detection report confirmed that there was a large amount of sludge and garbage accumulation inside the pipe and the pipe wall was intact, the true label for this event was strictly set as functional siltation, and the label value was set to 0. When the monitoring report confirmed that the pipe was disconnected, misaligned, or the pipe bed subsided, resulting in a reduction in the water flow cross-section, the true label for this event was set as structural settlement and deformation, and the label value was set to 1. The monitoring data recorded by the IoT device at the time of the historical backlog event was extracted. According to the method of the present invention, the historical hydraulic state residual column vector and the historical pipe-soil coupling stress gradient column vector were obtained. These two sets of vectors were combined with the above real labels to form the sea-span modal genesis supervised training dataset. Extract the input historical hydraulic state residual column vector, perform matrix multiplication with the first attention mapping matrix, and output the query matrix; simultaneously, extract the input historical pipe-soil coupled stress gradient column vector, perform matrix multiplication with the second attention mapping matrix, and output the key matrix; perform dot product operation between the query matrix and the transpose of the key matrix, and perform numerical smoothing through a scaling factor; finally, input the result into the Softmax normalization function to calculate the cross-attention weight coefficients assigned to the hydraulic state residual and the pipe-soil stress gradient; based on these weight coefficients, output the predicted probability distribution array of whether the current event belongs to functional siltation or structural settlement deformation. The cross-entropy loss function is used to calculate the classification error between the predicted probability distribution array and the corresponding real operation and maintenance label values. The adaptive moment estimation optimization algorithm is used to minimize the cross-entropy classification error as the absolute objective. Along the direction of gradient descent, the values ​​inside the first attention mapping matrix and the second attention mapping matrix are continuously updated through the backpropagation mechanism. When the classification error calculated on the validation set data converges, the iterative update stops and the first attention mapping matrix and the second attention mapping matrix are fixed. After the first and second attention mapping matrices have completed supervised training and stopped iterating, the complete historical data containing real labels for functional siltation and structural settlement deformation in the validation set are re-input into the network. The cross-attention weight matrices generated by each are extracted, and the arithmetic mean of their main diagonal elements is calculated. The optimal split point for distinguishing between functional siltation and structural settlement deformation is determined by using the maximum inter-class variance method, and this point is set as the stress mutation threshold.

[0039] This invention extracts the hydraulic state residual column vector and the pipe-soil coupled stress gradient column vector, and uses a dot product attention mechanism to calculate the cross-attention weight matrix. This allows it to capture the resonance probability between abnormal rises in liquid level and sudden changes in pipe wall stress. This cross-modal data fusion analysis breaks away from the single dependence on apparent water level and can distinguish between functional siltation caused by sediment and foreign matter and structural settlement deformation caused by external soil compression, thus improving the transparency and intelligence level of pipeline fault diagnosis.

[0040] Step S5, the process of obtaining the operational risk assessment results of the municipal building underground drainage network, specifically includes: The arithmetic mean of all elements in the hydraulic state residual column vector is calculated and denoted as the residual mean. In this embodiment, the abnormal liquid level threshold is set to 0.15 meters. If the mean residual value is less than the preset abnormal liquid level threshold of 0.15 meters, the pipeline network is considered to be operating normally. If the mean residual value is greater than or equal to the preset abnormal liquid level threshold, the pipeline network is judged to be in abnormal operation, and a comprehensive risk score is calculated. The comprehensive risk score is the product of the transient structural damage risk probability and the preset hard penalty amplification coefficient. The process of obtaining the hard penalty amplification factor specifically includes: Extract the main diagonal elements from the cross-attention weight matrix and calculate their arithmetic mean, which is denoted as the diagonal mean. If the mean of the diagonal is less than the preset stress mutation threshold, it indicates that the liquid level in the pipeline network is abnormally high, but the stress on the pipe wall is stable and has not been subjected to abnormal compression by the external soil. The pipe section is judged to be at risk of functional siltation. In this embodiment, the value of the hard penalty amplification factor is set to 1.2, which means that although simple silt blockage will raise the local liquid level, the pressure-bearing skeleton of the overall pipeline structure has not been damaged. Therefore, only a slight penalty weight of 1.2 is applied. If the mean of the diagonal is greater than or equal to the preset stress mutation threshold, it indicates that while the liquid level in the pipeline is rising abnormally, the pipe wall is simultaneously subjected to severe pressure from the external soil, leading to deformation. This pipe section is determined to be at risk of structural settlement and backwater. In this embodiment, the value of the hard penalty amplification factor is set to 5.0. Once the pipe wall undergoes structural deformation, its mechanical yield limit will drop precipitously, making it extremely prone to brittle fracture under water hammer impact. Therefore, a larger penalty weight needs to be applied to increase the risk score.

[0041] This invention analyzes the diagonal mean of the cross-attention weight matrix, applying only a slight penalty weight to functional sedimentation that does not damage the overall pressure-bearing framework, while applying a high penalty weight to structural settlement risks that are highly susceptible to brittle bursting under water hammer impact. This differentiated penalty mechanism ensures that the final comprehensive risk score accurately reflects the urgency and severity of pipeline safety hazards, thereby guiding municipal maintenance departments to allocate emergency repair resources and avoid ineffective waste of manpower and resources.

[0042] A risk assessment system for the operation of underground drainage pipe networks in municipal buildings includes a data acquisition module, a physical field analysis module, a fatigue state evolution module, a cross-modal attention module, and an assessment module. The data acquisition module is used to collect monitoring data at monitoring points in the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, input the real-time hydraulic situation three-dimensional tensor into the hydraulic topology expansion gated network, and obtain the future hydraulic situation prediction tensor. The physical field analysis module is used to divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response characteristic matrix. The fatigue state evolution module is used to construct a time-varying fatigue state transition chain of the pipe wall structure based on the mechanical response feature matrix of the pipe wall, and obtain the probability of transient structural failure risk through Markov chain iterative calculation. The cross-modal attention module is used to combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct the cross-modal cross attention operator, and obtain the cross attention weight matrix; The assessment module is used to calculate a comprehensive risk score based on the transient structural failure risk probability and the cross-attention weight matrix, and obtain the operational risk assessment results of the underground drainage network of municipal buildings.

[0043] This invention constructs a hydraulically topologically expanded gated network and utilizes expanded causal convolution operations to increase the receptive field of the convolution kernel in the time series without increasing the number of network training parameters. This allows for the accurate capture of long-sequence time-lag correlations between sudden changes in forced drainage at upstream pumping stations and sudden surges in liquid levels in distant downstream pipe sections. Based on this, the invention introduces macroscopic hydraulic features at nodes as dynamic boundary conditions into a spectral domain physical evolution feedforward network. It also constructs a volume-weighted physical control loss function by combining the Saint-Venant equation and the Mohr-Coulomb model. This not only considers the surge in water pressure caused by transient impacts of internal water flow but also takes into account the pressure caused by uneven settlement of external soil. As a result, it can analyze the physical field characteristics such as the circumferential stress and longitudinal strain of the pipe wall within the target pipe section, significantly improving the spatial resolution and accuracy of monitoring the stress state of the pipeline network. This invention constructs a time-varying state transition chain for pipe wall fatigue based on the pipe wall mechanical response characteristic matrix and introduces Markov chain iterative calculation. It can dynamically update the probability of the pipeline transitioning from an intact state to a microcrack yielding state or even a through-crack instability state. This time-varying state evolution model based on mechanical equivalent conversion and logistic activation function can measure the instantaneous fatigue damage caused by the water hammer shock wave to the old pipe wall at the moment the pump station starts forced drainage. It provides a reliable absolute mathematical expectation value and transient structural damage risk probability for the preventive maintenance of the pipeline network structure, filling the gap in the existing technology in transient pressure fatigue evolution analysis. When pipe network congestion leads to poor water flow, this invention constructs a cross-modal cross-attention operator to calculate the resonance probability between the hydraulic state residual caused by abnormal rise in liquid level and the pipe-soil coupled stress gradient. This allows for precise differentiation between functional siltation caused by internal sediment and foreign matter accumulation and structural settlement caused by external soil settlement that flattens and deforms the pipe. Based on this causal classification, a differentiated hard penalty amplification coefficient is matched when calculating the comprehensive risk score. A lower penalty weight is applied to functional siltation that does not damage the overall structural pressure-bearing skeleton, while a very high penalty weight is applied to structural deformation that leads to a precipitous drop in the mechanical yield limit and is prone to brittle fracture.

[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for assessing the operational risk of underground drainage pipe networks in municipal buildings, characterized in that, Includes the following steps: Step S1: Collect monitoring data at monitoring points in the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, input the real-time hydraulic situation three-dimensional tensor into a hydraulic topology expansion gated network, and obtain a future hydraulic situation prediction tensor. Step S2: Divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response characteristic matrix. Step S3: Based on the mechanical response feature matrix of the pipe wall, construct the fatigue time-varying state transition chain of the pipe wall structure, and obtain the transient structural failure risk probability through Markov chain iterative calculation. Step S4: Combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct a cross-modal cross attention operator, and obtain the cross attention weight matrix. Step S5: Based on the transient structural damage risk probability and the cross-attention weight matrix, calculate the comprehensive risk score and obtain the operational risk assessment result of the municipal building underground drainage pipe network; The process of obtaining the pipe wall mechanical response feature matrix in step S2 specifically includes: The underground drainage pipe section between the two nodes is taken as the target analysis pipe section; The length of the target parsing pipe segment is collected, and the length of the target parsing pipe segment is discretized according to the preset length interval. The preset scheduling cycle is discretized according to the preset time interval. All length segmentation points and scheduling cycle segmentation points are combined in pairs to generate a spatiotemporal grid sampling point set; Construct a spatiotemporal coordinate matrix based on a set of spatiotemporal grid sampling points; The spatiotemporal coordinate matrix is ​​input into the pre-fixed Chebyshev polynomial equations of various orders for algebraic calculations to obtain the basis function matrix; Extract the hydraulic spatiotemporal feature slices corresponding to the upstream and downstream monitoring nodes of the target analytical pipe section from the future hydraulic situation prediction tensor. Flatten and stitch the data of the hydraulic spatiotemporal feature slices of the upstream and downstream monitoring nodes to generate boundary condition feature vectors. The boundary condition feature vector is input into the spectral domain physical evolution feedforward network, which outputs an extended coefficient matrix. This network performs data dimensionality reduction and nonlinear mapping through several hidden layers. The activation function in all hidden layers is a hydraulic catastrophe adaptive activation function, the mathematical expression of which is: ； in, Let be the adaptive scaling parameters to be solved. These are the linearly mapped feature values ​​of the input features received by the hidden layer neurons from the previous layer. The basis function matrix and the expansion coefficient matrix are multiplied to obtain the physical field feature matrix, which includes the predicted water depth, the predicted flow velocity, the predicted pipe wall circumferential stress, and the predicted pipe wall longitudinal strain. The process of obtaining the pipe wall mechanical response feature matrix further includes: Using an automatic differentiation mechanism, analytical partial derivatives of the values ​​in each column of the physical field feature matrix with respect to the input spatiotemporal coordinate matrix are obtained, and the mass conservation residuals, momentum conservation residuals, and pipe-soil coupling mechanical residuals are calculated using the Saint-Venant equation and the Mohr-Coulomb model. A physical control loss function with volume weights is constructed. An adaptive moment estimation optimization algorithm is adopted to minimize the physical control loss function. All weight matrices, bias vectors and adaptive scaling parameter vectors inside the spectral domain physical evolution feedforward network are continuously updated through backpropagation. When the calculated value of the physical control loss function is continuously iterated to a preset number of times or the change is less than a preset minimum threshold, convergence is determined and the iteration is stopped. Extract the column vectors representing circumferential stress and longitudinal strain from the physical field feature matrix of the last iteration, and concatenate them according to the column vectors to generate the mechanical response feature matrix of the pipe wall. Step S4, the process of obtaining the cross-attention weight matrix, specifically includes: Extract the liquid level data of the downstream monitoring node of the target analysis pipe section from the monitoring data, and construct the actual liquid level column vector; Extract the corresponding real-time hydraulic situation slices of the upstream and downstream monitoring nodes of the target analytical pipe section from the real-time hydraulic situation three-dimensional tensor, flatten and stitch the data of the real-time hydraulic situation slices of the upstream and downstream monitoring nodes to generate real-time boundary condition feature vectors. Each time the target parsing pipe segment is parsed, a copy of the current spectral domain physical evolution feedforward network parameters is made as an independent copy. The real-time boundary condition feature vector is input into the network parameter copy, and the reverse optimization iteration update is continuously performed on the network parameter copy a preset number of times according to the physical control loss function and the adaptive moment estimation optimization algorithm. After the network parameter replicas converge, the real-time physical field feature matrix of the last iteration is obtained. For the spatial dimension expansion of the real-time physical field feature matrix caused by grid discretization, the liquid level feature column corresponding to the actual liquid level column vector on the spatial topology node is extracted as the theoretical liquid level column vector composed of the theoretical liquid level values ​​at each time step in the past preset time period. Subtract the theoretical liquid level column vector from the actual liquid level column vector to generate the hydraulic situation residual column vector. The column representing circumferential stress is extracted from the real-time physical field feature matrix. All spatial division points at each time step are traversed. The maximum circumferential stress values ​​corresponding to each time step are selected and concatenated into an extreme value column vector. The maximum circumferential stress difference between adjacent time steps is calculated to generate a pipe-soil coupling stress gradient column vector. Perform matrix multiplication on the hydraulic situation residual column vector and the preset first attention mapping matrix to output the query matrix; Perform matrix multiplication on the column vector of soil-tube coupling stress gradient and the preset second attention mapping matrix to output the key matrix; The dot product attention mechanism is used to calculate the cross-association degree between the query matrix and the key matrix; Step S5, the process of obtaining the operational risk assessment results of the municipal building underground drainage pipe network, specifically includes: Calculate the arithmetic mean of all elements in the column vector of hydraulic state residuals, and denote it as the residual mean; If the mean residual is less than the preset abnormal liquid level threshold, the pipeline network is considered to be operating normally. If the mean residual value is greater than or equal to the preset abnormal liquid level threshold, the pipeline network is determined to be in abnormal operation, and a comprehensive risk score is calculated. The comprehensive risk score is the product of the transient structural damage risk probability and the preset hard penalty amplification coefficient. The process of obtaining the hard penalty amplification factor specifically includes: Extract the main diagonal elements from the cross-attention weight matrix and calculate their arithmetic mean, which is denoted as the diagonal mean. If the mean value of the diagonal is less than the preset stress mutation threshold, the target analysis section is determined to have a functional siltation risk; if the mean value of the diagonal is greater than or equal to the preset stress mutation threshold, the target analysis section is determined to have a structural settlement and backwater risk.

2. The method for risk assessment of underground drainage pipe network operation in municipal buildings as described in claim 1, characterized in that, The process of obtaining the future hydraulic situation prediction tensor in step S1 specifically includes: Collect monitoring data, including liquid level and flow rate, and regard the monitoring points as nodes in the drainage pipe network; Construct a real-time hydraulic situation three-dimensional tensor, which represents the liquid level and flow velocity of the monitoring node at the time step; Construct a directed adjacency matrix. The rows and columns of the directed adjacency matrix correspond to the nodes in the drainage network. The elements in the directed adjacency matrix represent the flow direction relationship between the corresponding two nodes. The real-time hydraulic situation three-dimensional tensor is input into a pre-trained hydraulic topology expansion gating network for computation to obtain the future hydraulic situation prediction tensor. The hydraulic topology expansion gated network includes a spatial topology transmission layer, parallel main information extraction branches and gated signal extraction branches, a gated fusion layer and a residual connection layer.

3. The method for assessing the operational risk of underground drainage pipe networks in municipal buildings as described in claim 2, characterized in that, The process of obtaining the probability of transient structural damage risk in step S3 specifically includes: According to the pre-defined discrete time steps, the preset scheduling cycle is divided into multiple time windows. The different length cutting points and row vectors corresponding to different time steps in each time window are extracted from the pipe wall mechanical response feature matrix. All absolute values ​​of circumferential stress in the same time window are traversed. The row vector containing the largest absolute value of circumferential stress in the time window is selected as the force row vector under that time step. The force row vectors under each time step are spliced ​​together in chronological order to obtain the force time sequence matrix. Extract the circumferential stress and longitudinal strain of the pipe wall from each row of the stress time series matrix, perform mechanical equivalent conversion, and construct a stress intensity column vector.

4. The method for assessing the operational risk of underground drainage pipe networks in municipal buildings as described in claim 3, characterized in that, The process of obtaining the transient structural failure risk probability in step S3 further includes: A blank state transition probability matrix is ​​initialized for each time step within a preset scheduling period. The state transition probability matrix is ​​used to record the state transition probability of the pipeline state transitioning from a preset first state to a preset second state and from a preset second state to a preset third state at each time step. The state transition probability is calculated by combining the logistic activation function with the equivalent force intensity factor.

5. The method for assessing the operational risk of underground drainage pipe networks in municipal buildings as described in claim 4, characterized in that, The process of obtaining the transient structural failure risk probability in step S3 further includes: Construct an initial health state row vector, wherein the initial health state row vector is a preset row vector; The initial health state row vector is multiplied continuously with the state transition probability matrix within the preset scheduling period, and finally the termination state row vector at the end of the last time step is output. Extract the last element of the terminal state row vector as the transient structure failure risk probability.

6. A risk assessment system for the operation of underground drainage pipe networks in municipal buildings, applied in a risk assessment method for the operation of underground drainage pipe networks in municipal buildings as described in any one of claims 1-5, characterized in that, It includes a data acquisition module, a physics field analysis module, a fatigue state evolution module, a cross-modal attention module, and an evaluation module; The acquisition module is used to collect monitoring data at monitoring points of the drainage pipe network, construct a real-time hydraulic situation three-dimensional tensor, and input the real-time hydraulic situation three-dimensional tensor into a hydraulic topology expansion gated network to obtain a future hydraulic situation prediction tensor. The physical field analysis module is used to divide the target analytical pipe segment between monitoring points, construct a spatiotemporal grid sampling point set, combine the future hydraulic situation prediction tensor, solve the physical field characteristics inside the target analytical pipe segment through physical constraints, and obtain the pipe wall mechanical response feature matrix. The fatigue state evolution module is used to construct a time-varying fatigue state transition chain of the pipe wall structure based on the mechanical response feature matrix of the pipe wall, and obtain the probability of transient structural failure risk through Markov chain iterative calculation. The cross-modal attention module is used to combine the spatiotemporal grid sampling point set with the real-time hydraulic situation three-dimensional tensor, obtain the real-time physical field characteristics inside the target analytical pipe section through the spectral domain physical evolution feedforward network, construct a cross-modal cross attention operator, and obtain the cross attention weight matrix. The assessment module is used to calculate a comprehensive risk score based on the transient structural damage risk probability and the cross-attention weight matrix, and to obtain the operational risk assessment results of the municipal building underground drainage network.

Images