Drainage pipe network real-time fullness degree evaluation method based on machine learning

Abstract

The invention discloses a drainage pipe network real-time fullness degree evaluation method based on machine learning, particularly relates to the technical field of urban drainage pipe network operation monitoring and intelligent evaluation, and aims to solve the problem that an existing drainage pipe network is not complete in observation, unstable in sensor quality and complex in pipe network structure. And the fullness degree of each pipe section is difficult to accurately assess in time and carry out risk early warning. The method comprises the following steps: constructing a drainage network time sequence evidence graph fusing node evidence vectors and hydraulic connection attributes under unified time service and operation slice constraints, and inputting the drainage network time sequence evidence graph into a graph time sequence neural network which is subjected to offline training and version locking; physical consistency correction is carried out on the intermediate fullness degree representation under the constraint of mass conservation, and a fullness degree estimation value with an uncertainty index is output; therefore, a drainage pipe network fullness degree evaluation result which is consistent in physical constraint, quantifiable in credibility and traceable and reproducible in process is provided for each pipe section under the conditions of incomplete observation, uneven sensor quality and complex pipe network structure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban drainage network operation monitoring and intelligent evaluation, in particular to a drainage network real-time fullness evaluation method based on machine learning. BACKGROUND

[0002] In the urban drainage system, the scale of the pipe network gradually expands and the structure becomes increasingly complex. In order to prevent waterlogging and overflow pollution, the operation unit usually relies on rain gauges, water level gauges, flow meters, and pump station and gate monitoring systems to monitor key nodes, and conducts working condition analysis in combination with traditional hydrodynamic models. In the prior art, one approach focuses on hydrodynamic simulation based on design records and boundary conditions, and offline deduction is performed for typical rainfall processes. Another approach relies on single-point water level or flow threshold, empirical rules or simple statistical indicators to determine whether there is a full pipe, jacking or overflow risk locally. However, these methods generally assume that the monitoring data is complete and the sensor is in a stable working state. Sensor missing, delay and quality fluctuations are often only handled by simple rejection or interpolation, and the analysis object is mostly a local section or a small number of key sections, making it difficult to uniformly depict the real-time fullness of each pipe segment on a network scale.

[0003] With the development of the Internet of Things and data-driven methods, some schemes have attempted to use machine learning or time series models to predict and evaluate drainage conditions. However, existing methods mostly treat each monitoring point as a mutually independent time series, or only introduce a small number of topological features in a local range, and still cannot fully reflect the hydraulic connection relationship along the pipe network topology. At the same time, there is a lack of systematic modeling and coding mechanisms for unified timekeeping of monitoring data, division of operation time slices, sensor missing and delay status, and long-term health degree. Monitoring anomalies are often roughly classified or directly discarded, making the model sensitive to incomplete observations and uneven sensor quality. In addition, existing fullness evaluation results mostly only give a certain "full pipe degree" or risk level, lack explicit combination with physical constraints such as mass conservation, and still do not generally provide a physical consistency correction process for each pipe segment and corresponding uncertainty quantification. The evaluation process and basis are difficult to be completely traced and reproduced afterwards.

[0004] Therefore, under the engineering conditions of missing, delay and anomaly of monitoring data, uneven sensor quality, and complex drainage pipe network topology, the existing technology still cannot provide real-time fullness evaluation results that are physically consistent, have quantifiable reliability, and have traceable evaluation processes for each pipe segment in a unified time and space reference framework. How to construct time series data identifiers that reflect the hydraulic connection relationship for the entire drainage pipe network under the premise of incomplete observations and uneven sensor quality, and obtain real-time fullness evaluation results that can be used for subsequent risk identification and scheduling decisions, has become a technical problem that needs to be solved in the field. SUMMARY

[0005] To address the shortcomings of existing technologies, this invention provides a machine learning-based method for real-time assessment of the fullness of drainage pipe networks, thereby solving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for real-time assessment of the fullness of drainage pipe networks, comprising: S1. Establish a drainage network topology model, configure unified time synchronization and operation slicing rules, and generate operation slice identifiers for each node and pipe segment; S2. Collect rainfall, water level, flow rate, and status of pumping stations and gates in each operating slice. Combine the pipe section structural parameters and sensor missing, delay, and health information to construct node evidence vectors and assign them mask features. S3. Construct a drainage network evidence graph using node evidence vectors and hydraulic connection attributes, and combine the evidence graphs corresponding to several consecutive running slices into a time-series evidence graph sequence. S4. Input the time-series evidence graph sequence into the pre-trained graph time-series neural network, perform weighted aggregation and time-series encoding between nodes to obtain the intermediate fill degree representation of each pipe segment. S5. Based on the mass conservation constraint, perform physical consistency correction on the intermediate filling degree characterization, and output the real-time filling degree estimate and uncertainty index of each pipe segment; S6. Calculate the fullness risk index based on the fullness estimate, uncertainty index and fullness change trend, and classify and output the risk of drainage pipe network sections and areas.

[0007] Furthermore, S1 includes: A topology model of the drainage network was established based on pipeline ledgers and geographic information. The water flow direction of each pipe section is determined by comparing the elevation of the starting structure and the elevation of the ending structure, and combined with the design slope, and the manually specified type is recorded. Configure a unified time synchronization reference and time synchronization tolerance threshold, and verify the timestamps of the observed measurements based on the time synchronization tolerance threshold; Based on the baseline start time and the running slice rules, the running slices are divided, and the timestamps are mapped to the running slices. A running slice identifier is generated for each node unit and each pipe segment unit, which consists of a topology version number, an object type tag, an object code, and a running slice sequence number, and has a unique index in the running management library.

[0008] Furthermore, S2 includes: The control center obtains observations from the rainfall monitoring system, water level monitoring system, flow monitoring system, pump station monitoring system, and gate monitoring system based on the operational slice identifiers. The timestamps of the observations are assigned to a unique running slice according to the running slice rules. Based on the allowed value range and the reasonable change range threshold, the representative value of the current running slice is classified into one of the normal observation category and the abnormal observation category; Missing information is generated when there are no valid observations in the running slice, and delayed information is generated when the difference between the observation reception time and the end time of the running slice is greater than the delay threshold.

[0009] Furthermore, the control center maintains sensor health information for each monitoring point. When the observation is normal, the missing information is normal, and the delay information is normal within the running slice, the sensor health information is increased. When the observation is marked as a missing observation, a delayed observation, or an abnormal observation, the sensor health information is decreased and limited to a value range of zero to one. Based on rainfall, water level, flow rate, pump station status, gate status, pipe section structural parameters, missing information, delay information, and sensor health information, node evidence vectors are formed by encoding them in a fixed field order. The field order and field definition are locked by the node evidence vector version number.

[0010] Furthermore, S3 includes: Within each operational slice, the control center constructs a drainage network evidence map based on the drainage network topology model and operational slice rules. Node units and pipe segment units are treated as points in the map, and the connection relationship between node units and pipe segment units is treated as a line with attached node evidence vectors and hydraulic connection attributes. The control center generates a time-series evidence map sequence from the drainage network evidence maps sorted by the running slice identifier according to the observation window length and sliding step size, and registers the time-series evidence map sequence together with the topology version number, running slice rule version number, hydraulic connection attribute configuration version number and window configuration version number in the operation management library.

[0011] Furthermore, S4 includes: The control center inputs the drainage network evidence map into the graph-time neural network in the order of operation slices within the observation window; In the spatial aggregation layer, adjacency weights are calculated based on hydraulic connectivity attributes, sensor mask features, and sensor health information. The node evidence vectors are then weighted and aggregated according to the adjacency weights, and a nonlinear transformation is performed on the weighted results to obtain node representations. In the temporal coding layer, the node representation is recursively updated according to the running slice order using the temporal state vector; At the end of the observation window, the temporal representations of the nodes at both ends of the pipe segment are combined with the structural parameters of the pipe segment into an extended feature vector, which is then transformed linearly to generate the intermediate fill degree representation.

[0012] Furthermore, S5 includes: The control center configures mass conservation constraint rules based on the drainage network topology model and hydraulic connection properties; The pipe section diameter and slope are converted into water conveyance capacity coefficients according to the calculation rules for water conveyance capacity coefficients; The total inflow and total outflow of the upstream pipe section of each node are calculated based on the filling degree component in the intermediate filling degree characterization and the water conveyance capacity coefficient to obtain the flow loss measurement. When the absolute value of the flow rate error exceeds the allowable range of node balance error, the relative water flow indicators of the upstream and downstream pipe sections are scaled by a uniform ratio, and the scaled filling degree component is truncated in intervals. The truncated filling degree component is used as the real-time filling degree estimate.

[0013] Furthermore, after obtaining the real-time filling estimate, the control center sets an initial confidence value for each pipe segment; The initial confidence value is multiplied by the missing sensor information recorded in the node evidence vector, the delay scaling factor is multiplied by the sensor delay information, the health reference value is multiplied by the sensor health information, and the adjustment range recorded in the mass conservation constraint step is multiplied by the corresponding adjustment scaling factor to obtain the uncertainty index. The real-time fill rate estimate, uncertainty index, as well as the running slice identifier, topology version number, model version number, and mass conservation constraint version number are registered in the fill rate assessment database to form evidence chain record entries.

[0014] Furthermore, S6 includes: After each operating segment is completed, the control center reads the real-time estimated value of the fullness and uncertainty index of each pipe segment from the fullness assessment database, filters the historical fullness according to the confidence threshold within the preset observation period, and obtains the fullness change trend based on the normalization of the difference in fullness of the start and end operating segments. Based on the risk assessment rules, the current filling level, the filling level change trend and uncertainty are weighted and superimposed to obtain a filling level risk index between zero and one. The risk level of the pipe section and area is determined according to the graded threshold and an early warning output is triggered.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By establishing a time-series evidence graph of the drainage network containing node evidence vectors and hydraulic connection attributes under unified time synchronization and running slice rule constraints, and inputting it into a graph time-series neural network trained offline and version-locked, the intermediate fullness representation is physically consistent under mass conservation constraints, and a real-time fullness estimate with uncertainty index is output. This achieves the effect of providing real-time fullness assessment results of the drainage network with consistent physical constraints, quantifiable credibility, and traceable and reproducible assessment process for each pipe segment, even under conditions of incomplete observation, inconsistent sensor quality, and complex network structure.

[0016] 2. By tracking real-time fill level estimates and their uncertainties on continuously running slices, the fill level change trend is calculated, and a fill level risk index that takes into account the current fill level, change trend, and confidence penalty is constructed. Pipe sections and areas are classified into risk levels, and early warnings are output to external scheduling and monitoring systems through an interface with idempotent keys and error codes. This achieves the effect of accurately identifying high-risk pipe sections and areas under complex rainfall and abnormal monitoring conditions, reducing false alarms and missed alarms, and providing a unified and replayable risk basis for drainage scheduling and operation and maintenance decisions. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a machine learning-based method for real-time filling degree assessment of drainage pipe networks according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figure 1 A flowchart illustrating a machine learning-based real-time filling degree assessment method for drainage pipe networks is provided. The method includes: S1. Establish a drainage pipe network topology model, configure unified time synchronization and operation slicing rules, and generate operation slice identifiers for each node and pipe segment. The specific implementation is as follows: In the control center of the urban drainage area, a drainage network modeling and timing system for that area is first configured. This system is used to establish the drainage network topology model, unify the timing reference, and define operational slicing rules. The drainage network topology model refers to a directed structure model that uses manholes, pumping stations, gates, and overflow outlets as node units, and the pipelines connecting these structures as pipe segment units, organizing the connection relationships between node units and pipe segment units according to the direction of water flow.

[0020] To build the model, the control center reads records one by one from the pipeline ledger and geographic information of the drainage management department. Each record in the pipeline ledger includes at least the starting structure number, the ending structure number, the pipe diameter, material, burial depth, length, design slope, and whether it is connected to an overflow outlet or a storage structure. The geographic information includes at least the plane coordinates and elevation of the structure.

[0021] The system registers all structure numbers as node codes and assigns them as node units. The starting and ending structure numbers in each pipeline record are registered as the two ends of a pipe segment unit, and static attributes such as pipe diameter, material, length, and design slope are attached to this segment unit. For determining the water flow direction, the system prioritizes comparing the elevation values ​​of the starting and ending points. When the starting elevation is greater than the ending elevation, the water flow direction is defined as from the starting point to the ending point; when the starting elevation is less than the ending elevation, the water flow direction is defined as from the ending point to the starting point. When the starting and ending elevations are equal within the allowable error range, the system reads the design slope field from the pipeline ledger. If the design slope is positive, the water flow direction is defined according to the positive direction of the design slope. If the design slope is zero or missing, the design flow direction mark in the ledger is used, or the water flow direction is manually specified by engineering technicians, and this specification is recorded as a manually specified type. This ensures that each pipe segment unit is assigned a unique and clear water flow direction attribute.

[0022] After completing the registration of node units and pipe segment units, the system generates internal codes for all node units and pipe segment units according to unified rules. It also organizes the pipeline ledger version number, geographic information collection date, modeling time, and records of manual confirmation used in this modeling to generate a unique topology version number. The topology version number can be set as a string composed of date, area number, and serial number, and is stored in the operation management library to identify the physical state corresponding to the current drainage network topology model. In subsequent topology adjustments or modifications, a new topology version number is generated while the old version number is retained to avoid confusion between different versions.

[0023] Subsequently, a unified time synchronization reference is configured in the control center. This unified time synchronization reference refers to a set of time reference sources and their usage rules provided uniformly by the control center. In this embodiment, a time synchronization server connected to a BeiDou time synchronization device is preferred to obtain the standard time and periodically transmit it to rainfall monitoring terminals, water level monitoring terminals, flow monitoring terminals, and pump station and gate monitoring terminals via wired or wireless networks. Upon receiving the time synchronization information, each monitoring terminal adjusts its local clock to match the time synchronization server. To constrain time synchronization accuracy, the system registers a time synchronization tolerance threshold in the operation management database. This threshold can be set to several seconds and is reported by each monitoring terminal. When monitoring values, the control center calculates the absolute time difference between the timestamp attached to the monitored value and the current time of the time synchronization server. When the absolute value is less than or equal to the time synchronization tolerance threshold, the terminal is marked as having normal time synchronization. When the absolute value is greater than the time synchronization tolerance threshold, the terminal is marked as having abnormal time synchronization. The terminal number, the time when the abnormality was discovered, and the magnitude of the deviation are recorded. The value of the time synchronization tolerance threshold can be set to be significantly smaller than the length of the running slice time. Preferably, during the trial operation phase, the distribution of time deviations of each terminal is statistically analyzed, and a value that can cover the upper limit of the time deviation of most normal terminals is selected so as to detect time synchronization faults in a timely manner without misjudging normal terminals.

[0024] When the connection between the time synchronization server and the time synchronization source is abnormal or the self-test determines that the time synchronization source is unreliable, the time synchronization server can mark the current time synchronization status as failed and broadcast it to each monitoring terminal. During the period when the time synchronization status is failed, each monitoring terminal will not adjust its local clock, but will use the local time reference when it was last confirmed to be normal. The control center will record the start time of the time synchronization failure, the recovery time, and the estimated time deviation range during this period in the operation management library, which will be reflected in the subsequent construction of node evidence vectors and evaluation of uncertainty indicators.

[0025] After the unified time reference is determined, the control center sets up operation slicing rules, dividing the continuous operation time into a series of operation slices with fixed time lengths. The operation slicing rules include at least the reference start time, the time length of a single operation slice, and the slice numbering method. In this embodiment, it can be set to divide operation slices according to a fixed rhythm starting from midnight of a certain day. The time length of a single operation slice can be set to several minutes, for example, ten minutes per slice. The slice number can be incremented sequentially from zero. When the system receives data reported by the monitoring terminal, it reads the timestamp attached to the data, compares it with the reference start time, calculates the total time from the timestamp to the reference start time, divides the total time by the time length of a single operation slice, and takes the integer part of the quotient as the operation slice number, thereby realizing the use of integer... The division method combined with rounding down uniquely maps any timestamp to a specific operational slice. When a timestamp is exactly equal to the end time of a certain slice, it can be preferentially assigned to the next operational slice, and this assignment rule is recorded in the operation management library. This allows those skilled in the art to reproduce the operational slice division according to the same logic in any implementation environment. The specific value of the operational slice duration can be selected based on the drainage system's requirements for timeliness and stability. Generally, ten minutes per slice can balance the capture of rainfall process changes and noise smoothing. If it is necessary to improve timeliness in key areas or during the response phase of heavy rainfall, the operational slice duration can be shortened to five minutes or even less. When it is necessary to change the operational slice duration, the control center registers a new version number for the operational slice rule and records the effective time.

[0026] After configuring the topology model and running slice rules, the control center generates a unique running slice identifier for each node unit and pipe segment unit combined with each running slice. The running slice identifier is a code used to uniquely identify the status of a node unit or pipe segment unit within a specific running slice under a given topology version number. In this implementation, it can be generated in the form of "topology version number + separator + object type marker + separator + object code + separator + running slice sequence number". The object type marker distinguishes between node units and pipe segment units, and the separator is a fixed character that will not appear in any part of the content. The system initiates a database transaction when generating the identifier, writing the newly generated running slice identifier into the running slice. In the row management database, fields with unique indexes are used to roll back transactions when the existence of the identifier is detected or a write failure occurs. One or more check characters are appended to the end of the original identifier, and the write is retried until successful. This ensures that the run slice identifier in the run management database is unique and searchable globally. The run slice identifier, along with the corresponding topology version number, run slice rule version, unified timing reference version, and timing tolerance threshold, are stored in the run management database to form a complete record entry containing the topology context and time partitioning context. This allows subsequent steps to uniquely locate all observations and calculation results of a node or segment within a specific time slice and backtrack its context using the run slice identifier.

[0027] This step can be deployed on local industrial servers, edge computing nodes, or cloud platforms. For smaller drainage areas, topology model maintenance, unified time synchronization, and operation slice identifier generation can be achieved on a single server. For large drainage systems covering multiple areas, an edge node plus cloud platform approach is preferred. The edge nodes establish sub-topology version numbers and sub-operation management libraries in their respective areas, execute all the above modeling and identifier generation steps, and synchronize the sub-topology version numbers, operation slice rule versions, and operation slice identifier summaries to the cloud platform. A total topology version number and area mapping relationship are established on the cloud platform, thereby ensuring independent operation of each area while achieving consistent modeling and unified referencing across the entire network. This allows those skilled in the art to reproduce all the behaviors of this step under different system architectures according to the above disclosure.

[0028] S2. Within each operational slice, collect rainfall, water level, flow rate, and pump station and gate status. Combine this with pipe segment structural parameters and sensor missing, delay, and health information to construct node evidence vectors and assign them masked features. The specific implementation is as follows: Within each operational slice, the control center, based on the operational slice identifier generated in the previous step, retrieves various on-site observations within the time slice corresponding to that operational slice identifier from the rainfall monitoring system, water level monitoring system, flow monitoring system, pump station monitoring system, and gate monitoring system. Based on this, it constructs the node evidence vector required for subsequent evaluation. The rainfall monitoring system refers to multiple rain gauges and their acquisition devices deployed in the drainage area to record rainfall within their respective service areas under a unified timing reference constraint. Rainfall refers to the cumulative rainfall height at a specific rainfall monitoring point within an operational slice, uniformly measured in millimeters. The observation rhythm is consistent with the length of the operational slice, meaning that each operational slice outputs a cumulative value once.

[0029] A water level monitoring system refers to water level gauges and their acquisition devices installed in inspection wells or channels to record the liquid level height at corresponding locations. Water level refers to the representative water depth at a specific monitoring point within an operational segment, uniformly expressed in meters relative to a reference elevation. When multiple instantaneous water level samples are taken within the operational segment, the control center sorts all sampled values ​​chronologically, removes those marked as abnormal, and calculates the arithmetic mean of the remaining sampled values. This average value is taken as the representative water level of the operational segment. A flow rate monitoring system refers to flow meters and their acquisition devices installed at key pipe sections or outlets to record the volumetric flow rate passing through a cross-section under a unified timing reference. Flow rate refers to the average volumetric flow rate at a specific flow rate monitoring point within an operational segment, uniformly expressed in cubic meters per hour. When multiple instantaneous flow rate samples are taken within the operational segment, the control center calculates the arithmetic mean of all instantaneous flow rate values ​​not marked as abnormal, converts it to cubic meters per hour according to the sampling period, and takes this average value as the representative flow rate of the operational segment.

[0030] A pump station monitoring system is a system that collects the start / stop status, speed signals, and operating modes of each pump in a pump station. Pump station status refers to the start / stop combination status and representative operating parameters of each pump within a specific operating slice. The control center statistically analyzes the percentage of time each pump is in the on state and its average speed within that operating slice, encoding this data into a fixed-format status description. A gate monitoring system is a system that collects the opening and closing positions or opening degrees of gates. Gate status refers to the representative opening degree of a gate within a specific operating slice, uniformly expressed as a percentage representing the average opening degree of that gate within the operating slice. The control center calculates the arithmetic mean of the sampled opening degrees within that slice after removing anomalies, using this as the representative value.

[0031] When the control center obtains these observations, it first checks the timestamps attached to each observation according to the aforementioned operating slice rules. For each observation, it calculates the relationship between its timestamp and the start and end times of the operating slice. When the timestamp is strictly later than the start time and earlier than the end time, or when the slice boundary is equal to the end time and is assigned to the next operating slice as agreed, the observation is assigned to the corresponding operating slice, thereby avoiding an observation being assigned to two slices or not being assigned to any slice.

[0032] To ensure the physical rationality of the observed quantities, the system configures allowable value ranges and reasonable variation amplitudes for each type of monitoring quantity, and records them in the operation management library with a separate configuration version number. The allowable value range refers to the minimum and maximum values ​​that the monitoring quantity may occur in engineering. For example, the lower limit of rainfall is set to zero, and the upper limit is set based on the local historical extreme rainfall statistics multiplied by a safety factor; the lower limit of water level is set to zero, and the upper limit is set based on the difference between the elevation of the manhole or ditch bottom and the ground elevation plus a safety margin; the lower limit of flow rate is set to zero, and the upper limit is set as a certain multiple of the design flow rate.

[0033] The reasonable variation range refers to the maximum allowable change in the monitored quantity between two adjacent operating slices. Specific values ​​can be configured according to different drainage systems and fixed through configuration version numbers. For example, the water level change between two adjacent operating slices should not exceed a certain height, and the flow rate change should not exceed a certain flow rate difference. Specific values ​​can be determined by combining the maximum variation difference from historical continuous observations with a safety factor. After the control center completes the representative value calculation within an operating slice, it calculates the difference between the current representative value and the representative value of the previous operating slice. When the absolute value of the difference does not exceed the reasonable variation range threshold and the representative value is within the allowable range, the current representative value is marked as a normal observation. When the representative value exceeds the allowable range or the absolute value of the difference between it and the representative value of the previous operating slice is greater than the reasonable variation range threshold, the current representative value is marked as an abnormal observation, and an abnormal label is generated for this monitored quantity during node evidence vector construction. It will no longer be considered a reliable observation for subsequent learning and inference.

[0034] Within a running slice, if a monitoring point has no original sampled values ​​whose timestamps fall within the time range of that slice, or if all sampled values ​​are marked as abnormal in the above checks, then the monitoring point is considered to lack effective observations within that running slice. The system generates a sensor missing information field for that monitoring point and sets the missing flag to one. If the original sampling timestamp of the monitoring point falls within the time range of that running slice, but the control center actually receives the sampled value later than the end time of the running slice, and the time difference between the receiving time and the end time of the running slice is greater than a pre-configured delay threshold, then the system records the observation as a delayed observation.

[0035] The delay threshold is a single value configured in the operation management library for each type of monitoring quantity. It can preferably be set as a certain proportion of the length of the operation slice, such as half or one-third, and remains unchanged during operation. For each observation, the control center calculates the difference between its reception time and the end time of its operation slice. When the difference is less than or equal to the delay threshold, the observation is considered to be acceptable in terms of time, and the delay information field can be set to zero or "normal". When the difference is greater than the delay threshold, the observation is considered to be too delayed, and the delay information field is set to record the difference or the pre-divided delay level. At the same time, the observation is not included in the representative value calculation in this operation slice. It is only reflected in the node evidence vector in the form of delay information that the monitoring point has a time delay in the slice.

[0036] Sensor health information is a reliability index derived from a long observation period of a specific monitoring point, considering its missing data, latency, and the number of anomaly markers. The value ranges from zero to one, with one indicating complete reliability and zero indicating unreliability. The control center maintains a current health value and two sets of update step size parameters for each monitoring point: an increase step size and a decrease step size. The increase step size is a fixed positive number, and the decrease step size is a fixed positive number greater than the increase step size. All three are recorded in the operation management database as health configuration version numbers.

[0037] When the system is first launched or after the sensors are manually calibrated, the health information is initialized to a relatively high intermediate value, such as 0.8. During operation, whenever the representative value of a monitoring point within a certain operating slice is marked as a normal observation with a missing value of zero and a normal delay value, the control center increases the health information of that monitoring point by one increment. If the summed value is greater than one, it is truncated to one. Conversely, whenever the monitoring point within a certain operating slice is marked as a missing observation, a delayed observation, or an abnormal observation, the control center decreases the health information of that monitoring point by one decrement. If the decremented value is less than zero, it is truncated to zero. Through this monotonic update and truncation rule, under the cumulative effect of multiple operating slices, the long-term stable sensor health will gradually approach one, while the long-term frequently missing or abnormal sensor health will gradually decrease to near zero. This health information is used in subsequent mask feature construction to reflect the difference in the long-term reliability of the monitoring point.

[0038] Meanwhile, according to the pipeline ledger, pipe segment structural parameters are attached to each pipe segment. These pipe segment structural parameters are static parameters that describe the geometric and functional attributes of the pipe segment, including pipe diameter, slope, length, material, and whether it is connected to an overflow outlet or a storage structure. These parameters are registered as static attributes when the topology model is established. Within each operating slice, they are retrieved from the operation management database through the pipe segment code and attached to the corresponding node or pipe segment, and do not change over time.

[0039] The aforementioned rainfall, water level, flow rate, pump station status, gate status, pipe section structural parameters, missing sensor information, sensor delay information, and sensor health information constitute the organized observation and prior information of a node unit within a given operational slice. The control center arranges and concatenates this information into a vector combining numerical values ​​and labels according to a pre-agreed field order; this vector is the node evidence vector. The node evidence vector is a set of data encoding various observations, structural information, and quality information of a node unit within an operational slice, under the constraints of a specific topology version number and operational slice rule version. The control center implements version locking for the field order, field meaning, physical units, value range, and position within the vector of the node evidence vector, recording this information in the operational management database using the node evidence vector version number to ensure the uniqueness and traceability of the parsing method during model training and online evaluation.

[0040] Mask features are additional markers in the node evidence vector used to summarize the credibility status of various observations. To avoid ambiguity, this implementation assigns a mask value to each type of observation. The mask value takes only three states, representing high credibility, low credibility, and unavailable. The control center determines the credibility status of this type of observation based on the sensor missing information, sensor delay information, and sensor health information within the current running slice: when the missing flag is zero, the delay information field indicates no delay or a delay time less than or equal to the delay threshold, and the health information is greater than or equal to a preset high credibility threshold, the mask value for this type of observation is set to high credibility; when the missing flag is one, or the health information is less than or equal to a preset unavailable threshold, the mask value is set to unavailable; and when neither of the above two conditions applies, the mask value is set to low credibility.

[0041] The high-confidence threshold and the unavailability threshold are two fixed values, which are recorded in the runtime management library in the form of a mask configuration version number. The judgment order is fixed as follows: first judge the unavailability condition, then judge the high-confidence condition, and the remaining cases are classified as low-confidence. This ensures that the same set of missing information, delay information and health information will have a uniquely determined mask value in any engineering implementation.

[0042] This step preferably covers several operational slices within one hour. With a setting of one slice every ten minutes, six operational slices will be formed within one hour. Within each operational slice, node evidence vectors for each node are constructed in the order described above. For critical nodes located near the main pipeline or key flood-prone areas, the sampling rhythm of the monitoring terminal can be set to be shorter than the length of the operational slice, for example, once per minute. When constructing node evidence vectors, the control center still generates representative values ​​according to the rules of "by slice affiliation, elimination of anomalies, and arithmetic mean", thereby improving the ability to distinguish critical nodes without changing the operational slice division logic. In drainage scenarios of different scales and underlying systems, those skilled in the art can adjust the length of the running slice, the allowable range of values, the reasonable range of changes, the delay threshold, the health update step size, and the mask threshold. However, the method for constructing node evidence vectors always follows the underlying logic of "unified unit and time attribution, identification of missing and abnormalities according to fixed rules, updating health according to a clear step size, generating a mask according to fixed conditions, and encoding according to a fixed field order." This ensures that this step can be reproduced in different engineering environments and consistent node evidence vectors can be obtained, providing a stable and reliable foundation for the subsequent construction of drainage network evidence graphs and graph time-series neural network evaluation.

[0043] S3. Construct a drainage network evidence graph using node evidence vectors and hydraulic connection attributes. Combine the evidence graphs corresponding to several consecutive operating slices into a time-series evidence graph sequence. The specific implementation is as follows: After the node evidence vectors are constructed, the control center constructs the drainage network evidence graph for the current time slice within each operational slice, based on the aforementioned drainage network topology model and under the constraints of a unified topology version number and operational slice rule version. The drainage network evidence graph is a graphical representation within a specific operational slice, using node units and pipe segment units as points, and the connections between node units and between node units and pipe segment units as lines. Node evidence vectors and hydraulic connection attributes are attached to each point and each line. When constructing the drainage network evidence graph, the control center first reads all node unit codes and pipe segment unit codes, as well as pipe segment structure parameters, corresponding to the current topology version number from the operational management database, and then reads all node evidence vectors corresponding to the current operational slice identifier. Subsequently, for each node evidence vector, it is attached to the node unit with the same name in the topology model according to the node code recorded internally, ensuring that each node unit is associated with a unique node evidence vector within the operational slice.

[0044] Hydraulic connection attributes are a set of attributes used to describe the hydraulic transmission relationship between node units and pipe segment units. They include at least flow direction markers, trunk or branch pipe markers, whether a pipe crosses a pumping station marker, and whether it is connected to an overflow outlet or a storage structure marker. The flow direction marker directly adopts the water flow direction determined in the topology model. The specific division rules for trunk or branch pipe markers are locked in a versioned manner in the hydraulic connection attribute configuration: preferably, when a pipe segment has a "function type" field in the pipeline ledger and the field is marked as "trunk", the trunk or branch pipe marker of the pipe segment is set as trunk, and when the function type field is marked as branch, it is set as branch. For pipe segments without explicitly marked function types, the control center divides them according to the comparison result between the pipe diameter of the pipe segment and the preset trunk pipe diameter threshold. If the pipe diameter is greater than or equal to the trunk pipe diameter threshold, it is marked as trunk; otherwise, it is marked as branch. The trunk pipe diameter threshold is recorded in the operation management library in the form of a single value and a version number.

[0045] The rule for determining whether a line crosses a pump station mark is as follows: when a connection represented by a line contains a pump station unit in the topology model, that is, when one end or the middle of the line passes through a node unit marked as a pump station, the rule for determining whether the line crosses a pump station mark is set to yes; otherwise, it is set to no. The rule for determining whether a line is connected to an overflow outlet or a storage structure mark is as follows: when a pipe segment's structural parameters record that one end or the middle of it is connected to an overflow outlet or a storage structure, the mark for that pipe segment and its related lines is set to yes; otherwise, it is set to no.

[0046] When constructing the drainage network evidence map, the control center generates a connection record for the connection relationship between each node or between a node and a pipe segment. This record stores the starting node code, ending node code, flow direction mark, main or branch pipe mark, whether it crosses a pumping station mark, and whether it is connected to an overflow outlet or a storage structure mark in a fixed order. These records, together with the node evidence vector record, form the drainage network evidence map data set for the current operating slice. This data set, along with the current operating slice identifier, topology version number, operating slice rule version number, and hydraulic connection attribute configuration version number, is registered in the operation management database, enabling those skilled in the art to uniquely reconstruct the drainage network evidence map at any subsequent time using the operating slice identifier and topology version number.

[0047] To reflect the evolution of the drainage network's fullness over time, the control center constructs a time-series evidence map sequence based on the drainage network evidence map. A time-series evidence map sequence is an ordered set of several time-series consecutive drainage network evidence maps arranged in order from earliest to latest according to the operational slice identifiers. It is used to describe the temporal changes in the drainage network's state within a certain observation window.

[0048] The observation window is a time range consisting of a fixed number of adjacent running slices. This number is recorded in the runtime management library as a window length parameter, and the window length is a positive integer. The sliding step size refers to the number of slices that the running slice window moves forward on the time axis each time a new time series evidence graph sequence is generated. It is also a positive integer and is recorded in the runtime management library along with the window length in the form of a window configuration version number.

[0049] In a preferred configuration, the run slice length is ten minutes, the window length can be set to six run slices, and the sliding step size can be set to one run slice. In this case, each observation window covers one hour, and the observation window is scrolled once every time a new run slice is generated. When constructing the time-series evidence map sequence, the control center follows a fixed algorithm: First, a currently running slice is selected, and its running slice number is denoted as k. With a window length of L and a sliding step size of S, when k is greater than or equal to L minus one and the drainage network evidence maps of the previous L running slices have been registered, the control center sequentially reads the drainage network evidence maps corresponding to the L running slices with sequence numbers from k minus L plus one to k from the operation management library, forming an ordered sequence according to the order of the running slice identifiers. Subsequently, a time-series sequence number is assigned to this ordered sequence. This number can be formed by concatenating the window configuration version number, the window's starting running slice identifier, and the window's ending running slice identifier in a fixed format. The correspondence between this time-series sequence number and the window length, sliding step size, window's starting running slice identifier, window's ending running slice identifier, and the list of drainage network evidence map identifiers included is registered in the operation management library. When the sliding step size is one running slice, the control center automatically triggers the above process to generate a new time-series evidence map sequence every time a new drainage network evidence map for a running slice is generated. When the sliding step size is greater than one running slice, the construction operation can only be triggered when the current running slice number k satisfies the condition that "k minus the initial starting number is divisible by the sliding step size", thereby reducing the number of time-series evidence map sequences. The above triggering conditions and divisibility judgment rules are recorded together in the window configuration record of the operation management library. If, when constructing an observation window, it is found that the drainage network evidence map of a certain running slice in the window is missing or marked as unavailable, the control center can set it not to construct the time-series evidence map sequence for that window, and record the starting running slice identifier, ending running slice identifier, and the reason for not constructing the window in the operation management library to avoid inconsistencies in the handling of missing windows in different implementations.

[0050] For large-scale drainage systems, to avoid excessive computation time or resource consumption in subsequent graph-based time-series neural network calculations due to the large size of a single drainage network evidence map, the entire network can be divided into several sub-regions according to administrative or hydraulic regions. The sub-region division results are fixed in the operation management library through a region configuration record. This record assigns a unique region identifier to each node unit and each pipe segment unit, ensuring that any node unit and pipe segment unit belongs to only one region under the same topology version. A subgraph refers to a drainage network evidence map within a specific region that only contains the node units and pipe segment units of that region and their internal hydraulic connection attributes. Its construction method is the same as that of the entire network drainage network evidence map, except that the input scope is limited to this region. A subgraph time-series evidence map sequence refers to a set of ordered drainage network evidence map sequences constructed within that region according to the same window length and sliding step rules as the entire network.

[0051] When constructing subgraphs and subgraph time-series evidence graph sequences, the control center first groups the node units and pipe segment units under the current topology version according to the area configuration records. Within each running slice, it selects the node evidence vectors belonging to a certain area and the corresponding hydraulic connection attributes to form the drainage network evidence graph of that area. Then, it constructs the time-series evidence graph sequence of that area according to the same window length and sliding rules as the entire network, and assigns a subgraph time-series sequence number to each time-series evidence graph sequence of each area.

[0052] After construction, the control center registers each subgraph and its corresponding time-series evidence graph sequence, along with the subgraph identifier, the network-wide topology version number, and the observation window configuration version number, into the operation management library. Simultaneously, it establishes a mapping relationship between the subgraph identifier and the network-wide topology node and pipe segment codes for each subgraph. This mapping relationship records the node and pipe segment unit codes in the network-wide topology model corresponding to each node and pipe segment code in the subgraph. This allows for the reading of drainage network evidence graphs and time-series evidence graph sequences from each subgraph, and the recombination of multiple subgraphs or their time-series evidence graph sequences into a single logical network-wide drainage network evidence graph or network-wide time-series evidence graph sequence, thus maintaining a consistent topology and time reference between parallel subgraph computation and unified network-wide analysis.

[0053] In practical engineering applications, this step can be implemented centrally on a single industrial server or in parallel on multiple edge computing nodes. In edge computing scenarios, it is preferable that each edge node locally constructs the subgraph and subgraph temporal evidence graph sequence of its respective area, and periodically synchronizes the subgraph identifier, subgraph temporal sequence number, and mapping relationship increment to the cloud operation management library. Based on the above definitions and construction order of drainage network evidence graph, hydraulic connection attributes, observation window, sliding step size, temporal evidence graph sequence, and subgraph, those skilled in the art can reproduce the complete process of this step on any common monitoring and computing platform along a unique path, obtaining a spatial topology and temporal evolution representation consistent with this method, providing a stable and traceable data foundation for subsequent fullness assessment and risk identification based on graph temporal neural networks.

[0054] S4. Input the time-series evidence graph sequence into a pre-trained graph-time neural network, perform weighted aggregation and time-series encoding between nodes to obtain the intermediate fill degree representation of each pipe segment. The specific implementation is as follows: When a fullness assessment is required, after the current operating slice ends, the control center retrieves a sequence of time-series evidence maps covering the observation window from the operation management library, based on the aforementioned observation window configuration. The observation window consists of several consecutive operating slices, and the window length and sliding step size are recorded in the operation management library as window configuration version numbers. The control center reads the drainage network evidence map corresponding to each operating slice within the window, and forms an ordered drainage network evidence map in ascending order of operating slice identifiers. This ordered map and its window configuration parameters are used as the basic information for a fullness assessment session and input into the graph temporal neural network. The graph temporal neural network is a type of model that runs on the drainage network evidence map. Its characteristic is that it updates node representations through multiple rounds of neighborhood aggregation in the spatial dimension and updates the temporal state vectors of nodes through recursion in the temporal dimension, ultimately generating an intermediate fullness representation for each pipe segment at the end of the observation window. The model consists of several spatial aggregation layers and temporal encoding layers. The operation rules and parameters of each layer are determined during the training phase and locked in the model version record, remaining unchanged during online operation.

[0055] In the spatial aggregation layer, the control center processes the corresponding drainage network evidence map for each running slice within the observation window. For each node unit in the map, it takes the node's own node evidence vector and the node evidence vectors of all adjacent nodes with hydraulic connections to that node. Adjacency weights are assigned to each connection relationship according to the hydraulic connection attributes and observation quality information. To avoid ambiguity, this implementation method fixes the calculation rules for adjacency weights as follows: First, a basic coefficient is calculated for each connection relationship. The basic coefficient is determined by the trunk or branch pipe label and the flow direction label. If the connection relationship belongs to the trunk and the flow direction is consistent with the water flow direction in the topology model, the basic coefficient is one. If the connection relationship belongs to the trunk and the flow direction is opposite to the water flow direction, the basic coefficient is half. If the connection relationship belongs to the branch pipe and the flow direction is consistent with the water flow direction, the basic coefficient is half. If the connection relationship belongs to the branch pipe and the flow direction is opposite to the water flow direction, the basic coefficient is one-quarter. Then, the foundation coefficient is adjusted according to whether it crosses a pumping station and whether it is connected to an overflow outlet or a storage structure: when the connection crosses a pumping station, the foundation coefficient is multiplied by the pumping station adjustment coefficient, which is a fixed value greater than zero, preferably set to a value greater than one to increase the impact of the lifting section; when the connection is connected to an overflow outlet or a storage structure, the foundation coefficient is multiplied by the overflow adjustment coefficient, which is a fixed value greater than zero, preferably set to a value less than one to reduce the direct impact on the main trunk under high overflow conditions. The specific values ​​of the pumping station adjustment coefficient and the overflow adjustment coefficient, along with the version number, are recorded in the operation management library.

[0056] Subsequently, the control center reads the sensor health information and mask features of neighboring nodes within the current running slice. When the mask features indicate that the node's observation is unavailable, the adjacency weight of that connection is directly set to zero. When the mask features indicate that the node's observation is low confidence, the base coefficient is multiplied by a low confidence scaling factor (a fixed value between zero and one), and then multiplied by the node's sensor health information. When the mask features indicate that the node's observation is high confidence, the base coefficient is directly multiplied by the node's sensor health information as the initial adjacency weight of that connection. For all connections to the current node, the control center sums their initial adjacency weights to obtain the total weight. When the total weight is greater than zero, the initial adjacency weight of each connection is divided by the total weight to obtain a set of standardized adjacency weights that sum to one. When the total weight is equal to zero, it indicates that the current node has no available neighbors in this round of aggregation, and the system will not perform neighborhood weighting on that node, but will only use the node's own node evidence vector for updating.

[0057] During the specific aggregation process, the control center performs a linear transformation on the node's own evidence vector using a set of fixed coefficients to obtain the node's self-mapping vector. For each neighboring node's evidence vector, a linear transformation is performed using another set or the same set of fixed coefficients to obtain a neighbor mapping vector. Then, all neighbor mapping vectors are weighted and summed according to the standardized adjacency weights mentioned above to obtain the neighborhood aggregation vector. Subsequently, the self-mapping vector and the neighborhood aggregation vector are added at a fixed ratio, for example, by adding them one-to-one, to obtain the node's intermediate vector before this round of aggregation. For each component of this intermediate vector, the control center applies a fixed nonlinear transformation rule: if the component is less than zero, it is replaced with zero; if the component is greater than or equal to zero, it remains unchanged, thus obtaining the node representation vector after this round of spatial aggregation. The linear transformation coefficients, the mixing ratio of neighborhood and self-information, and the nonlinear transformation rule are all determined during the model training phase, and their values ​​are recorded along with the model version number in the model version record. The control center applies the same spatial aggregation rule to all nodes in the same drainage network evidence map. After completing one round of spatial aggregation, the new representation vectors of all nodes are used as the input for the next round of spatial aggregation. The preset number of rounds is repeated. The number of rounds is fixed in the model configuration and is used to control the range of neighborhoods from which nodes absorb information.

[0058] After completing the spatial aggregation of each running slice within the observation window, the control center performs temporal encoding on the representation of each node in the time dimension. To this end, for each node within the observation window, the control center processes the spatial aggregation representation of that node in each running slice sequentially according to the time order of the running slice identifier, and maintains a temporal state vector for that node. The dimension of the temporal state vector is fixed in the model configuration, and the initial value is uniformly set to an all-zero vector. This setting method is recorded in the model version record along with the model version number.

[0059] When processing the first running slice within the window, the control center applies a set of fixed linear transformations to the current time state vector to obtain the time propagation vector. It then applies another set of fixed linear transformations to the spatial aggregation representation vector of the node within that running slice to obtain the input mapping vector. The time propagation vector and the input mapping vector are added to obtain the vector before update. Then, a nonlinear transformation is performed on each component of this vector. If a component is greater than a certain upper threshold, it is truncated to that upper threshold; if a component is less than a certain lower threshold, it is truncated to that lower threshold. The transformation remains unchanged between the upper and lower thresholds to prevent the time state from diverging numerically. When processing subsequent running slices within the window, the above steps are repeated for each running slice. Each time, the time state vector obtained at the end of the previous running slice and the node spatial aggregation representation vector of the current running slice are used for updating, until the last running slice within the observation window is processed. The time state vector obtained at the end of the window is the temporal representation of that node within the entire observation window. The time propagation transformation coefficients, input mapping transformation coefficients, and the upper and lower thresholds are all determined during the training phase and locked in the model version record.

[0060] After the node temporal representation calculation is completed, the control center generates an intermediate fill degree representation for each pipe segment. The intermediate fill degree representation is a numerical vector derived by a graph temporal neural network based on graph structure and temporal information, before mass conservation constraint correction, used to characterize the water filling degree of the pipe segment at the end of the observation window. For each pipe segment, the control center reads the temporal representation vectors of the nodes at both ends of the segment at the end of the observation window, as well as the pipe segment's diameter, length, slope, material, whether it crosses a pumping station, and whether it connects to an overflow outlet or a storage structure. The temporal representation vectors of the two ends are concatenated in a fixed order and then combined with the structural parameters in a fixed order to form an extended feature vector. A set of fixed linear transformations is applied to this extended feature vector to obtain a lower-dimensional intermediate vector. Then, the same nonlinear transformation rules as spatial aggregation are applied to each component of the intermediate vector, replacing components less than zero with zero and keeping components greater than or equal to zero unchanged, thus obtaining the intermediate fill degree representation of the pipe segment.

[0061] The intermediate fill level representation can contain several components, at least one of which is used to approximate the fill level of the pipe segment. Its numerical range can be limited to zero to one, where zero represents a completely empty pipe and one represents a completely full pipe. The remaining components can be used to characterize the fluctuation of the pipe segment's fill level within the observation window or its sensitivity to upstream and downstream conditions. The linear transformation coefficients used to map from the extended feature vector to the intermediate fill level representation, and their output dimensions, are also determined during the training phase and recorded in the model version record.

[0062] To ensure the traceability and reproducibility of the fill degree evaluation process, the control center implements version locking for the structure and parameters of the graph temporal neural network. The model version record includes at least the model version number, the number of spatial aggregation layers, the number of temporal encoding layers, the node representation dimension, the temporal state vector dimension, the intermediate fill degree representation dimension, the observation window length and sliding step size used, the time range covered by the training data, the set of topology version numbers used during training, and a summary of training rules. The training rule summary includes the error measurement method and the parameter update method. The error measurement method can be set to calculate the square of the difference between the component representing the fill degree in the intermediate fill degree representation output by the model and the on-site measured fill degree or the high-confidence simulation fill degree on the training sample set, and sum them to obtain the total error. The parameter update method can be set to iteratively adjust the model parameters in each round of training based on the direction of the influence of the total error on the model parameters; that is, decrease the parameter values ​​in directions with larger total errors, and maintain or fine-tune the parameter values ​​in directions with smaller total errors or those that have already decreased, thereby gradually reducing the total error after multiple rounds of training.

[0063] During the offline phase, the control center selects a representative historical time interval from the operation management library, pairs the corresponding time-series evidence map sequence with reference fullness data within the same time period, and adjusts the model parameters multiple times using the aforementioned error measurement and parameter update methods until the total error drops below a preset threshold or no longer significantly decreases in several training rounds. After training, the obtained set of parameters, along with the aforementioned structural information, is registered as a new model version number, while the old version number record is retained for comparison or playback. During online operation, before each intermediate fullness representation calculation, the control center selects a compatible model version number from the model version record based on the current topology version number and the observation window configuration version number, preferably the most recently marked as usable version. At the start of this calculation, the model version number, the observation window start and end slice identifiers, and the time-series evidence map sequence identifier used are all registered in the operation log, enabling those skilled in the art to completely reconstruct the calculation process of a particular intermediate fullness representation afterward based on the operation log.

[0064] In terms of real-time performance, the control center can set the maximum allowable delay for each fullness assessment. For example, the intermediate fullness representation calculation of the entire network or each subgraph can be completed within a few seconds or tens of seconds after the slice is finished. In large-scale drainage systems, the overall calculation time can be shortened by distributing the forward computation of the graph temporal neural network to multiple edge nodes or multiple industrial servers and running in parallel on a subgraph basis without changing the above spatial aggregation rules and time coding rules.

[0065] Based on the above-disclosed rules for spatial aggregation weight calculation, node representation update, temporal state update, intermediate fullness representation construction, and model training and version locking, those skilled in the art can implement the complete process of this step in a unique path within a common graph computing framework or custom program environment. Under the same temporal evidence graph sequence and model version number, consistent intermediate fullness representation results can be obtained, thereby ensuring the feasibility and reproducibility of this step.

[0066] S5. Based on the mass conservation constraint, perform physical consistency correction on the intermediate filling degree characterization, and output the real-time filling degree estimate and uncertainty index of each pipe segment. The specific implementation is as follows: After obtaining the intermediate fill degree characterization, the control center performs physical consistency correction on the intermediate fill degree characterization within the current operating slice based on the mass conservation constraint rules configured according to the drainage network topology model and hydraulic connection attributes. The mass conservation constraint rules refer to checking whether the relative water delivery of the upstream and downstream pipe segments of each node is within the allowable balance error range at a given operating slice time scale. When there is a significant imbalance, the intermediate fill degree characterization of the relevant pipe segments is adjusted in groups according to a fixed strategy to bring the water volume near the node close to balance within the allowable error range. To this end, the control center first reads the water delivery capacity coefficient calculation rules, node balance error coefficients, pipe segment adjustment parameters, and node category labels from the mass conservation constraint configuration record. The node category labels are used to distinguish different categories such as nodes located on the main trunk, branch pipe end nodes, and nodes in sparsely monitored areas, so that different node balance error coefficients can be configured for different categories of nodes.

[0067] The water conveyance capacity coefficient is a non-negative number pre-calculated and fixed for each pipe segment, reflecting the relative water conveyance capacity of that segment under a unit degree of fullness. This implementation uses a unified calculation rule: for each pipe segment, the inner diameter is expressed in meters. The square of the inner diameter is multiplied by the square root of the absolute value of the slope, and then multiplied by a fixed proportionality coefficient. The result is the water conveyance capacity coefficient for that pipe segment. This fixed proportionality coefficient has the same value across the entire network, and its value, along with the mass conservation constraint version number, is recorded in the operation management library and does not change during operation. Within the same topology version, the larger the pipe diameter and the greater the slope, the larger the water conveyance capacity coefficient, and the more adjustment the pipe segment can bear in the mass conservation correction.

[0068] For a given node, the control center, based on the hydraulic connection attributes, considers all pipe segments with water flowing towards that node as upstream pipe segments and all pipe segments flowing out of that node as downstream pipe segments. It reads the component representing the degree of fullness in the intermediate fullness representation at the end of the current operating slice for each relevant pipe segment, restricts this component to the range between zero and one, and reads the corresponding water conveyance capacity coefficient. Multiplying the fullness component by the water conveyance capacity coefficient, the relative water conveyance index of that pipe segment on the current operating slice time scale is obtained.

[0069] The control center sums the relative water delivery indicators of all upstream pipe sections to obtain the total inflow of the node, and sums the relative water delivery indicators of all downstream pipe sections to obtain the total outflow of the node. The total inflow is then subtracted from the total outflow to obtain the flow error of the node. When both the total inflow and total outflow are zero, the flow error is considered zero. The allowable range of node balance error is calculated according to a unified rule in this implementation: for each node, the water delivery capacity coefficients of all its related pipe sections are summed to obtain the total water delivery capacity. This total water delivery capacity is multiplied by a node balance error coefficient to obtain the allowable range of balance error for that node. The node balance error coefficient is a fixed value between zero and one. A smaller value can be configured for nodes located on the main trunk, and a larger value can be configured for nodes located at the end of branch pipes or in sparsely monitored areas. The error coefficient values ​​corresponding to various types of nodes are registered in the operation management database along with the mass conservation constraint version number.

[0070] During the mass conservation correction process for each operating segment, the control center performs balance checks and adjustments on each node in a fixed sequence. For a given node, if the absolute value of its flow loss is less than or equal to the allowable balance error range for that node, the system considers the water volume balance of that node within the current operating segment to be within an acceptable range, and does not adjust the intermediate fill degree representation of the relevant pipe section. The system only records the flow loss of that node when calculating the uncertainty index later. When the absolute value of the flow loss is greater than the allowable balance error range, the system determines that the intermediate fill degree representation near that node deviates from the mass conservation constraint as a whole, and the relevant pipe sections need to be corrected in groups according to the preset balance adjustment strategy.

[0071] The balancing adjustment strategy uniformly adopts a method of scaling the relative water conveyance index of upstream or downstream pipe segments proportionally to make the total inflow and total outflow of a node as close to equal as possible. The specific rules are as follows: when the total inflow of a node is greater than the total outflow of a node and the flow imbalance is positive, the imbalance is first attempted to be eliminated in the downstream pipe segment. The target total outflow is set as the current total inflow, and then the scaling factor is calculated. This factor is equal to the ratio of the target total outflow to the current total outflow. If the current total outflow is greater than zero, the relative water conveyance index of each downstream pipe segment is multiplied by the scaling factor, and the fullness component of the pipe segment is calculated accordingly. The new fullness component is the smaller of the original fullness component multiplied by the scaling factor and one, ensuring that the fullness component does not exceed one. After scaling all downstream pipe segments, the control center recalculates the total outflow of the node to obtain the adjusted total outflow and the new flow imbalance.

[0072] If scaling down the downstream pipe segment reduces the absolute value of the flow imbalance to within the allowable range of balance error, the current adjustment cycle ends. If the fullness component of all downstream pipe segments has reached one and the total outflow of the node is still less than the target total outflow, it indicates that adjusting the downstream pipe segment alone is insufficient to eliminate the imbalance. In this case, the control center reduces the relative water delivery index in the upstream pipe segment by a uniform ratio: the target total inflow is set as the total outflow of the node plus the allowable range of balance error, the new scaling factor is set as the ratio of the target total inflow to the current total inflow, and the fullness component of each upstream pipe segment is multiplied by this scaling factor. The portions exceeding zero and one are truncated to zero and one, respectively. After completion, the total inflow and flow imbalance are recalculated. If the absolute value of the flow imbalance is still greater than the allowable range of balance error after the above two adjustments, the system will not make further adjustments to the relevant pipe segment, but will retain the remaining imbalance and record it in the imbalance statistics of that node.

[0073] When the total inflow to a node is less than the total outflow and the flow imbalance is negative, the control center adopts a symmetrical strategy: first, the relative flow rate index is proportionally increased in the upstream pipe section to increase the total inflow; if this does not completely eliminate the imbalance, the relative flow rate index is proportionally decreased in the downstream pipe section to reduce the total outflow. The scaling factor and cutoff rule are set symmetrically with the aforementioned case. Since each round of scaling uses a uniform scaling factor within the node, and the fullness component is limited to between zero and one after each adjustment, the entire balance adjustment process is a deterministic mapping, not relying on human experience judgment or random selection.

[0074] After completing the balancing adjustment at the node level, the control center updates the component representing the degree of fullness in the intermediate fullness representation for each pipe segment, obtaining the corrected fullness value under the mass conservation constraint. This value is then used as the real-time fullness estimate for the current operating slice. The real-time fullness estimate is represented by linearly mapping the corrected fullness between zero and one to a percentage value between zero and one hundred. That is, zero corresponds to zero percent, one corresponds to one hundred percent, and values ​​between zero and one are proportionally mapped to the corresponding percentages. When a pipe segment is not scaled during the mass conservation constraint adjustment, its corrected fullness is the same as the fullness component in the intermediate fullness representation, and there is a one-to-one linear relationship between the real-time fullness estimate and the intermediate fullness representation.

[0075] After obtaining the real-time fill level estimate for each pipe segment, the control center calculates an uncertainty index for that segment based on the sensor missing information, sensor delay information, sensor health information, and the adjustment range of that segment in the mass conservation constraint step, all recorded in the evidence vector of the nodes associated with that segment. The uncertainty index is a value between zero and one that reflects the reliability of the real-time fill level estimate; the closer the value is to one, the more reliable the estimate; the closer the value is to zero, the lower the reliability.

[0076] In this embodiment, the uncertainty index is calculated using a fixed multi-penalty rule. The calculation process is as follows: First, an initial confidence value is set for each pipe segment, and this initial value is uniformly set to one. Then, a first penalty is applied based on sensor missing information. When any node connected to the pipe segment has an observation with a missing flag of one in the current running slice, the confidence value of the pipe segment is multiplied by a missing scaling factor. The missing scaling factor is a fixed value between zero and one, and is recorded in the mass conservation constraint configuration record along with the version number. Second, a second penalty is applied based on sensor delay information. When a sensor of a node connected to the pipe segment has an observation with a delay time greater than a delay threshold, the confidence value of the pipe segment is multiplied by a delay scaling factor. The delay scaling factor is a fixed value between zero and one. Third, a third penalty is applied based on sensor health information. The health information of each node connected to the pipe segment in the current running slice is taken as the highest value. The small value is used as a health reference value for the pipe section. The current confidence value is multiplied by this health reference value to weaken the uncertainty index of monitoring points with long-term unstable performance. Finally, a fourth penalty is applied based on the adjustment range of the pipe section in the quality conservation constraint. The adjustment range is defined as the absolute difference between the fullness component of the pipe section before and after the correction. The control center configures two thresholds for the adjustment range. If the adjustment range is less than the first threshold, it is considered a small adjustment, and the confidence value is not changed. If the adjustment range is between the first and second thresholds, it is considered a medium adjustment, and the confidence value is multiplied by the medium adjustment scaling factor. If the adjustment range is greater than the second threshold, it is considered a large adjustment, and the confidence value is multiplied by the large adjustment scaling factor. The medium adjustment scaling factor and the large adjustment scaling factor are both fixed values ​​between zero and one. The two thresholds and the three scaling factors are all registered in the quality conservation constraint configuration record along with the version number.

[0077] The confidence value obtained after the above multiple scaling is the uncertainty index of the pipe section. When the value is less than zero, it is truncated to zero; when the value is greater than one, it is truncated to one, thus ensuring that the uncertainty index always falls between zero and one. Under the conditions of complete sensor observations, high health, small node imbalance, and small adjustment range of mass conservation constraints, the uncertainty index of the pipe section is close to one. Under the conditions of severe observation gaps, significant delays, or large scaling of the filling degree to eliminate severe node imbalance, the uncertainty index of the pipe section is significantly reduced.

[0078] The control center registers the real-time fill rate estimate, uncertainty index, and corresponding operating slice identifier, topology version number, model version number, and mass conservation constraint version number for each pipe segment under each operating slice into the fill rate assessment database, forming a record entry with a complete chain of evidence. The chain of evidence refers to all version information and contextual information associated with generating the real-time fill rate estimate within the same record, including the time-series evidence graph sequence identifier, node evidence vector version number, mass conservation constraint configuration version number, and model version number. This allows those skilled in the art to subsequently reconstruct the assessment process of a specific pipe segment under a given operating slice based on this information.

[0079] In practical applications, for critical trunk pipeline segments located on the main line, operators can set smaller node balance error coefficients, as well as smaller missing scaling coefficients, delay scaling coefficients, and adjustment scaling coefficients for the nodes in the mass conservation constraint configuration, in order to improve the physical consistency and credibility threshold of the fullness estimate of such pipeline segments. For a large number of sparsely monitored branch pipes, the node balance error coefficients can be appropriately increased, and a more moderate scaling coefficient can be used, so that the system can still provide reference-worthy estimation results under the premise of certain incomplete observations.

[0080] During on-site operation, after multiple rainfall events, operators can compare actual overflow records and surface water accumulation to adjust the node balance error coefficient, the fixed proportional coefficient used to calculate the water conveyance capacity coefficient, and various scaling coefficients in groups. Each time an adjustment is made, the control center generates a new version number of the mass conservation constraint and retains the old version number and its parameter values ​​so that the constraint rules used at that time can be restored when analyzing historical events.

[0081] Based on the above-mentioned disclosures regarding the calculation method of water conveyance capacity coefficient, the definition of allowable range of node balance error, the order of balance adjustment, the calculation rules of uncertainty index, and the evidence chain registration method, those skilled in the art can achieve physical consistency correction and real-time fullness estimation on any common monitoring and calculation platform along a unique path, obtaining consistent results under the same input conditions and constraint version numbers, thereby ensuring the feasibility and sufficiency of this step.

[0082] S6. Calculate the fullness risk index based on the estimated fullness, uncertainty indicators, and fullness change trends, and classify and issue early warnings for drainage pipe network sections and areas. The specific implementation is as follows: Based on the continuous generation of real-time fill level estimates and uncertainty indicators, the control center calculates the fill level change trend and fill level risk index for each pipe segment after each operational slice, according to pre-configured risk assessment rules. The fill level change trend refers to the direction and magnitude of the change in the fill level of the pipe segment over time within a preset observation period, used to characterize the rising, falling, or basically stable state of the pipe segment over a period of time. The observation period consists of several consecutive operational slices, the length of which is represented by the number of slices and locked in the risk assessment configuration by a version number. With an operational slice length of ten minutes, the observation period length can be set to six operational slices, corresponding to a one-hour time range.

[0083] For a given pipe segment and a specific operational slice, the control center first determines the slice's sequence number on the timeline. An observation window is created by combining the current operational slice with several preceding operational slices. When the number of historical operational slices is insufficient to fill the observation window, only existing historical operational slices are used for calculation. The control center retrieves the real-time fullness estimate and uncertainty index of each operational slice within the observation window from the fullness assessment database. The real-time fullness estimate is converted from a percentage to a fullness value between zero and one, and arranged chronologically. During the calculation process, the control center filters out historical points with uncertainty indices below a certain confidence threshold, excluding these points from trend calculations. The confidence threshold is a fixed value between zero and one, and its value, along with the risk assessment rule version number, is recorded in the operation management database.

[0084] To calculate the filling degree change trend, the control center uses a fixed difference weighting rule: within the filtered observation window, the earliest operating slice and the current operating slice are selected as the start and end points. The filling degree value of the current operating slice is subtracted from the filling degree value of the earliest operating slice to obtain the filling degree change. This change is then divided by the number of operating slices between the two to obtain the average filling degree change amplitude for each operating slice. When there is only one point in the observation window, such as the current operating slice, or when all historical points other than the current operating slice are filtered out due to uncertainty indicators being below the confidence threshold, the control center directly considers the filling degree change trend of that pipe section under the current operating slice to be zero and records in the operation log that this trend calculation has been processed as a "no valid historical point" branch.

[0085] After obtaining the average change range, the control center can normalize it according to the pre-configured upper limit of the change range, mapping cases with absolute values ​​less than or equal to zero to zero, and cases with absolute values ​​greater than or equal to the upper limit to one, with the mapping in between proportionally, thereby obtaining the trend strength between zero and one. At the same time, the direction of change is retained for subsequent risk assessment. The upper limit of the change range is recorded as a configuration parameter in the risk assessment configuration. Its value is selected by the operators in combination with historical operating conditions and is fixed with the risk assessment version number.

[0086] After obtaining the real-time fill level estimate, uncertainty index, and trend for the current operating segment, the control center synthesizes these three factors into a fill level risk index according to the risk assessment rules. The fill level risk index is a value between zero and one that comprehensively reflects the probability of overflow, backflow, or siltation occurring in a certain pipe section within a subsequent period. Its calculation uses a fixed weighted summation rule. The risk assessment rules assign a set of weighting coefficients to each factor, including at least the current operating segment fill level weighting coefficient, the trend weighting coefficient, and the uncertainty penalty coefficient. These coefficients are all values ​​between zero and one and are recorded in the operation management database as risk assessment version numbers.

[0087] Under a specific operational slice, the control center first scales the real-time fill level estimate of the pipe segment back to a fill level value between zero and one by a percentage, obtaining the current fill level component. Then, the aforementioned normalized trend strength and direction are used as the trend component. Different weighting adjustment rules can be applied to upward and downward trends, but these adjustment rules are fixed as parameters in the risk assessment configuration. Next, the uncertainty index is converted into a risk penalty component. The larger the uncertainty index, the smaller the risk penalty component. This component can be obtained by subtracting the uncertainty index from one, so that the lower the confidence level of the estimate, the greater its contribution to the risk index.

[0088] Subsequently, the control center multiplies the current fullness component by the current fullness weight coefficient, the trend component by the changing trend weight coefficient, and the risk penalty component by the uncertainty penalty coefficient, and sums these three to obtain the initial risk value of the pipe segment under the current operating slice. If the initial risk value is less than zero, it is set to zero; if the initial risk value is greater than one, it is truncated to one. The truncated value is the fullness risk index. In actual engineering settings, the current fullness weight coefficient can be set slightly larger than the changing trend weight coefficient, and the uncertainty penalty coefficient can be set to a medium level. Based on this, the weights can be adjusted to adapt to different operating conditions. The specific values ​​of these weights are recorded in the operation management library along with the risk assessment version number.

[0089] For scenarios requiring risk assessment at the area level, the control center aggregates the risk indices of all pipe segments within the same area using a fixed statistical method to form an area risk index. The aggregation method for the area risk index is fixed as a single method in the risk assessment configuration by the aggregation rule type. The aggregation rule type is limited to either "maximum value aggregation" or "over-threshold ratio aggregation," and only one type is specified and registered in the runtime management library under the same risk assessment version. When the aggregation rule type is "maximum value aggregation," the control center uses the maximum value among all pipe segment risk indices within the area as the area's risk index. When the aggregation rule type is "over-threshold ratio aggregation," the control center counts the number of pipe segments within the area whose risk index exceeds a certain risk threshold, divides this number by the total number of pipe segments in the area to obtain the proportion of pipe segments in a high-risk or relatively high-risk state, and uses this proportion as the area risk index. Both the risk threshold and aggregation rule type of the area risk index are locked with the risk assessment version number, ensuring that the calculation path of the area risk index is uniquely determined under any version.

[0090] After obtaining the fullness risk index for each pipe segment and each area, the control center classifies the pipe segments and areas into different risk levels according to the grading thresholds in the risk assessment rules. The grading thresholds refer to several pre-selected boundary values ​​between zero and one, dividing the entire risk index range into several level intervals. Each level corresponds to a closed or semi-open interval. The name, color, and prompt of each level are determined in the risk assessment configuration along with the grading thresholds and are bound to the risk assessment version number. During operation, the control center compares the risk index of each pipe segment or each area with these grading thresholds. When the risk index falls within a certain segment, it is marked as the corresponding risk level.

[0091] The monitoring interface uses preset colors and prompts to display information based on the risk level. For example, a high-risk level is displayed in red with a prompt that it needs immediate attention, a medium-risk level is displayed in orange with a prompt that it is recommended to arrange an inspection, and a low-risk level is displayed in green with a prompt that it is basically operating normally. The specific colors and text descriptions are fixed in the configuration file of the monitoring interface and correspond to the risk assessment version number.

[0092] When the risk level of a certain pipe section or area reaches or exceeds the preset warning trigger level, the control center sends a warning output to the drainage dispatch system. The warning content includes at least the pipe section or area identifier, the currently operating slice identifier, the real-time filling degree estimate, the uncertainty index, the filling degree risk index, and the risk level. It can also be accompanied by a suggested dispatch action type label. The definitions of these fields are uniformly defined in the interface documentation and risk assessment configuration.

[0093] To ensure the traceability and replayability of the risk assessment process, the control center records the risk assessment version number and key parameters used in the operation log for each risk grading and early warning output action. It also records the corresponding topology version number, model version number, mass conservation constraint version number, and fullness assessment record identifier at that time. The fullness assessment record identifier indicates the set of database entries containing the set of real-time fullness estimates and uncertainty indicators used in this risk assessment. This allows those skilled in the art to retrospectively trace back to the topology, model configuration, mass conservation constraint parameters, and risk assessment rules used at that time based on the records in the operation log. They can then recalculate the risk index using the same time-series evidence graph sequence, fullness estimation results, and configuration version number to confirm whether the risk grading and early warning output at that time conformed to the set rules.

[0094] To facilitate integration with external monitoring systems, drainage scheduling systems, or historical analysis systems, the control center provides a set of text-based network interfaces. These interfaces use fixed field names and structures, and the meaning, unit, rhythm, and allowed value range of each field are clearly specified in the interface documentation. At least two types of interfaces are included: one type is an interface for submitting the operating slice identifier and node evidence vector set, used by external systems or edge nodes to send the node evidence vector set within a certain operating slice to the control center. The request content of this interface includes at least the operating slice identifier, topology version number, node code list and corresponding node evidence vector sequence, and the interface agreement clearly states that the field order of the node evidence vectors must be consistent with the definition corresponding to the node evidence vector version number; the other type is an interface for querying the real-time fullness estimate, uncertainty index and fullness risk index of a specified operating slice identifier and pipe segment. The request content of this interface includes at least the operating slice identifier and pipe segment code, and the response content includes at least the real-time fullness estimate, uncertainty index, fullness risk index and risk level of the pipe segment under the operating slice. When an external system needs to obtain risk information at the area level, it can query the area risk index and classification results by area identifier through the area query interface.

[0095] All API calls must carry a unique request identifier, defined as an idempotent key in the API documentation, used to identify a business request. If multiple request messages with identical idempotent keys and identical request content appear in the same API, the control center will perform the calculation only once and record the result in the runtime log. Subsequent repeated calls will be treated as duplicate queries of existing results, and the already recorded result will be returned directly without recalculation, thus preventing duplicate recording and order disorder. When the idempotent keys are the same but the request content is different, the control center will handle them differently according to the request content and add this situation to the runtime log. The specific behavior is defined by fixed rules in the API documentation.

[0096] The interface design retains a fixed set of error codes to represent common interface call failure scenarios. These error codes and their meanings are fixed in the interface documentation and do not change with the runtime status. Error codes include at least the following categories: parameter mismatch, time synchronization mismatch, unavailable model version, and insufficient system resources. Parameter mismatch indicates missing fields, incorrect units, or values ​​exceeding the allowed range in the request message. Time synchronization mismatch indicates that the running slice identifier in the request message is inconsistent with the control center's current running slice rules or that the timestamp exceeds the allowed deviation. Unavailable model version indicates that the current topology version or observation window configuration cannot match any model version marked as available. Insufficient system resources indicate that the evaluation calculation cannot be completed on time at the current moment. When returning an error code, the control center simultaneously records the error code, a summary of the request content, and the current topology version number and risk assessment version number in the runtime log for subsequent troubleshooting.

[0097] In abnormal operating conditions such as prolonged disconnection of the monitoring subsystem, failure of the unified time synchronization source, and significant deviation between rainfall and water level, the control center automatically adjusts the risk level or marks the current assessment result as unsuitable for direct use in dispatching during the risk index calculation phase, referring to uncertainty indicators and anomaly judgment rules. In the risk assessment rules, a handling strategy flag is configured for each type of abnormal operating condition. The value of the handling strategy flag is limited to one of two options: "force escalation and marking" or "marking without escalation." Under the same risk assessment version, only one handling strategy flag is configured for each type of abnormal operating condition, and it is registered together with the risk assessment version number. When the monitoring subsystem is out of contact for an extended period, resulting in incomplete observations with a missing flag of 1 in a large number of node evidence vectors or a significant decrease in sensor health, the uncertainty index will decrease significantly. When the uncertainty index of a certain section or area is lower than the preset threshold for unsuitable scheduling, the control center will perform corresponding operations according to the handling strategy: when the handling strategy is marked as mandatory upgrade and marking, the risk level of the section or area will be upgraded by at least one level, and a flag indicating that it is not suitable for direct scheduling will be added to the result; when the handling strategy is marked as only marking without upgrade, the risk level will not be adjusted, and only a note will be added to the result indicating that the current assessment result is not suitable for direct scheduling.

[0098] When a unified time synchronization source failure causes the timestamps of multiple monitoring systems to deviate from the unified time synchronization benchmark and exceed the allowable tolerance, the control center can mark the entire operational slice as unreliable in terms of time synchronization, uniformly mark the risk index results of all pipe sections and areas under that operational slice as unsuitable for direct use in scheduling, and decide whether to uniformly upgrade the risk level based on the handling strategy marking. When the relationship between rainfall and water level is significantly divergent, such as water level not rising for a long time under high rainfall intensity or water level rising rapidly under no rainfall conditions, the control center will impose additional penalties on the uncertainty indicators of the affected area based on the predefined rainfall and water level consistency judgment conditions in the risk assessment rules, and can decide whether to upgrade the risk level of the corresponding pipe section or area based on the handling strategy marking. These judgment conditions, unsuitable scheduling thresholds, and handling strategy marking combinations are registered together with the risk assessment version number.

[0099] During field operation, operators can statistically analyze a large number of overflow and water accumulation samples under heavy rain conditions, compare the relationship between the fullness risk index and actual results under different versions, and gradually adjust the weighting coefficients, classification thresholds, unsuitable scheduling thresholds, aggregation rule types, and abnormal condition handling strategies. This allows the method to achieve a reasonable balance between early warning sensitivity and false alarm rate while meeting safety and compliance requirements. Based on the aforementioned publicly available information regarding the fullness change trend calculation rules, risk index weighting rules, regional risk aggregation methods, risk classification and early warning logic, interface design and idempotent strategies, and decision-making adjustment methods under abnormal conditions, those skilled in the art can implement this step in a unique path within common monitoring and scheduling system environments. Under given parameter configurations and version numbers, the same risk assessment results can be reconstructed, thus ensuring the feasibility and sufficiency of disclosure of this step.

[0100] In the operating scenario shown in this embodiment: The drainage area in an old urban district of a coastal city covers approximately 30 square kilometers. Upstream are mostly old residential areas and back streets, while downstream it connects to a booster pumping station and a culvert leading to a river. Before the renovation, this area frequently experienced road flooding and overflow complaints during short-duration heavy rainfalls. After completing the pipeline survey and detection, the drainage management department deployed the aforementioned drainage network modeling and evaluation system at the control center. First, during the project preparation phase, step S1 was completed, identifying approximately two thousand manholes, four pumping stations, three gates, and dozens of overflow outlets and regulating reservoirs forming the drainage network. Each structure was registered as a node unit, and approximately three thousand five hundred pipelines as pipe segment units. The unique flow direction of each pipe segment was determined using elevation and design slope, and a unified internal code was generated for all nodes and pipe segments. Based on this, the ledger version, geographic information collection date, and manual confirmation records used in this process were summarized into a topology version number, "V Urban Rainwater 202505". A unified time reference was established using a BeiDou time server, and the operating slice length was set to ten minutes, with rolling numbering starting from midnight each day. The topology version number, unified time reference version, and operating slice rule version were recorded in the operation management database, providing a stable structural and time reference for subsequent steps.

[0101] After the system went live, on the afternoon of a typical summer day with heavy rainfall, the weather forecast predicted short-duration heavy rainfall within the next two hours. Following step S2, more than ten rain gauges in the area began uploading cumulative rainfall according to the rhythm of the operating slices. Water level gauges in each major inspection well and culvert sampled at a rhythm of one to two minutes and summarized them into representative water levels within each operating slice. Flow meters at each key outlet and main pipeline section provided representative flow rates in a similar manner. At the same time, two pumping stations and three key gates located downstream of the area continuously uploaded pumping station status (including the percentage of operating time and average speed of each pump within a ten-minute slice) and gate status (including the average opening degree within the slice). At the end of each operational slice, the control center assigns all observations to the corresponding slice based on the timestamp. It then checks the allowable range and reasonable variation of the observed values, marking sudden increases in single-point values ​​far exceeding the design water level as abnormal, marking observations not uploaded due to communication interruptions as missing, and marking data received only after significantly exceeding the delay threshold as delayed. During long-term operation, the health information of each monitoring point is updated based on the frequency of missing, delayed, and abnormal occurrences. This ensures that the health of rain gauges, water level gauges, and flow meters that have been operating stably for a long time gradually approaches one, while the health of monitoring points with frequent distortions or interruptions gradually decreases. During this rainstorm, the water level gauge at the outlet of an upstream residential area experienced multiple unreasonable jumps within an hour due to inadequate initial maintenance. The system marked its observation as abnormal in several consecutive operational slices, and its health gradually decreased from 0.8 to around 0.3. The corresponding mask feature also changed to a low-confidence or unavailable state.

[0102] As rainfall intensified and entered its peak intensity phase, the average cumulative rainfall in the area approached 50 millimeters within approximately one hour, causing a rapid rise in water levels in the upstream branch pipes. Following the S2 rule, at the end of each ten-minute slice, the control center constructs a node evidence vector for each node unit. This vector encodes the node's representative rainfall, water level, flow rate, connected pump station status and gate status within the current slice, as well as structural parameters of the corresponding pipe section, including pipe diameter, slope, length, material, and whether it is connected to an overflow outlet or storage structure. These parameters, along with missing sensor information, delay information, and health information, are encoded in a predetermined field order, generating a mask feature representing the reliability of each observation. At this point, for key inspection well nodes located on the main pipe, because the surrounding rain gauges and water level gauges provide complete observations and have a health level close to one, the masks are mostly judged as high-reliability. However, for the aforementioned upstream sub-area outlet nodes, due to frequent abnormal water level observations, the masks are judged as low-reliability or even unusable.

[0103] Following step S3, within each operational slice, the control center attaches the evidence vectors of each node to the corresponding node unit based on the topology version number. It then attaches markers indicating main or branch pipes (based on pipe diameter and function type), whether a pipe crosses a pumping station, and whether it connects to an overflow outlet or storage structure to the connection relationship between nodes and pipe segments, thus constructing the drainage network evidence map for the current time slice. Several sub-maps are simultaneously built on a regional basis. As the rainstorm progresses, the system uses an observation window configuration with a window length of six operational slices and a sliding step size of one operational slice. It reviews six drainage network evidence maps from the most recent hour, arranging them into a time-series evidence map sequence according to the operational slice identifiers. A time-series sequence number is assigned to each sub-map, and the sequence, along with the identifiers of each included map and the start and end slices of the window, are registered in the operational management database. Subsequently, every ten minutes when a new operational slice is generated, an update window covering the most recent hour is automatically created, continuously providing the latest time-series evidence map sequence to the graph time-series neural network in a sliding manner.

[0104] After the graph-time neural network is trained and registered as a specific model version, the system performs online evaluation according to step S4 during the several hours of heavy rain. At the end of the current running slice, the control center selects the time-series evidence map sequence for the corresponding area and inputs six consecutive drainage network evidence maps into the model in chronological order. The model performs multiple rounds of spatial aggregation on each map, assigning different adjacency weights to the connections between main nodes, between main and branch pipes, and across pumping stations according to the aforementioned weighting rules. Nodes with high confidence masks and health scores close to one contribute significantly to neighborhood aggregation, while low-confidence or unusable nodes are automatically suppressed or removed. Through this process, the representation vector of a key node on the main line not only includes its own water level, flow rate, and pumping station status information, but also integrates the water level changes of adjacent upstream and downstream nodes, the water inflow of branch pipes, and the opening status of overflow outlets in the past hour. Subsequently, the model recursively updates the representation vector of each node in the time dimension according to a ten-minute rhythm, using nonlinear truncation to prevent state divergence, and finally forms the time-series representation of each node at the end of the observation window. For each pipe segment, the model merges the temporal representations and structural parameters of its two endpoints into an extended feature, mapping it to obtain an intermediate fill degree representation. One major component is between zero and one, approximately representing the fill degree of the pipe segment at the end of the window. For example, during peak rainfall, this component of the downstream culvert inlet segment may be close to 0.9, while some branch pipes located in upstream areas with significant elevation differences may still be below 0.5. The model version number, window length, and sequence number of the temporal evidence map used are recorded in the operation log at each round of evaluation for subsequent traceability.

[0105] To avoid the model generating results that do not conform to physical conservation based solely on statistical correlation, the system performs mass conservation constraint correction on the intermediate fill degree representation within each operational slice according to step S5. At the end of the current rainstorm window, the control center calculates the pre-calculated water conveyance capacity coefficient for all pipe segments according to the water conveyance capacity coefficient calculation rules. The fill degree component in the intermediate fill degree representation of each pipe segment is multiplied by this coefficient to obtain the relative water conveyance index of each pipe segment at the current ten-minute slice scale. For a node located on the main line, there are two large-diameter box culvert sections and several branch pipes upstream, and a pumping station downstream. The system sums the relative water delivery volumes of each upstream and downstream pipe section to obtain the total inflow and total outflow, and calculates the flow imbalance. If the absolute value of the imbalance exceeds the allowable balance error range configured according to the node type, the system first amplifies the relative water delivery volume of the downstream pipe section of the pumping station by a uniform scaling ratio on the downstream pipe section side, and back-calculates the corrected fullness component of the downstream pipe section, and cuts off the part exceeding 1 to 1. If the imbalance cannot be eliminated even when the fullness of all downstream pipe sections has reached 1, the system then reduces the fullness component of the upstream pipe section by a uniform ratio until the imbalance is compressed to within the allowable balance error range or cannot be further adjusted. Through this process, the corrected fill levels of the three key pipe segments surrounding the aforementioned main nodes within the current slice may be adjusted to 0.85, 0.92, and 0.80, respectively. Nodes connected to upstream branches with very low health levels, due to insufficient neighborly monitoring, experience smaller adjustments under the mass conservation constraint, contributing only a limited amount to the overall water balance based on their water conveyance capacity coefficient and limited observational information. The system then linearly maps the corrected fill levels to real-time fill level estimates in percentage form. For example, the three key pipe segments may be recorded as 85%, 92%, and 80% respectively. Simultaneously, based on the missing, delayed, and health levels of the connected sensors, combined with their respective adjustments under the mass conservation correction, an uncertainty index is calculated for each pipe segment. When sensor observations are stable and the adjustment range is small, the uncertainty index for the main pipe segments can approach or exceed 0.9, while the uncertainty for monitoring sparsely populated branches may decrease to below 0.5. All real-time fill estimates, uncertainty metrics, and corresponding running slice identifiers, topology version numbers, model version numbers, mass conservation constraint version numbers, and the sequence numbers of the time-series evidence graphs used are written into the fill assessment database, forming a traceable chain of evidence.

[0106] Based on this, the system performs a risk assessment according to step S6. Within one hour before and after the peak of the rainstorm, the control center calculates the filling rate change trend for each pipe segment after each operational slice. Assuming that when operating to a certain slice, the estimated real-time filling rate of the downstream culvert inlet pipe segment gradually increases from 55%, 68%, 76%, 83%, and 87% to 92% in the last six slices, with the uncertainty index consistently above 0.85, after restoring these values ​​to a filling rate between 0 and 1, the system filters out points with excessively low uncertainty, retaining all six records. Using the first and current records as the starting and ending points, it calculates the average filling rate change for each slice and normalizes it to a trend strength close to 1 with an upward direction based on a pre-configured upper limit for the change range, indicating that the filling rate of this pipe segment has been continuously and rapidly increasing in the last hour. The control center then calculates the fullness risk index for the pipe section according to the risk assessment rules. It superimposes the current fullness component (approximately 0.92), trend strength (close to 1), and the risk penalty component (obtained by subtracting the uncertainty index from 1) with fixed weights. For example, if the current fullness weight coefficient is 0.5, the trend weight coefficient is 0.3, and the uncertainty penalty coefficient is 0.2, the risk index for this pipe section under this operating slice might be close to 0.87. This is higher than the high-risk threshold of 0.8 set in the risk assessment configuration, and is therefore classified as high-risk. It is highlighted in red on the monitoring interface, accompanied by the message "High risk at the culvert inlet; attention should be paid to overflow and backflow." Meanwhile, an upstream branch pipe, due to its real-time fullness estimate fluctuating between 40% and 50% and its low uncertainty index, exhibits a weak trend within the same time window. Its risk index is between 0.3 and 0.4, and it is classified as low to medium-risk, displayed in green or yellow on the monitoring interface.

[0107] The system simultaneously assesses the risk level at the area level within the same operational slice. In this embodiment, the area risk aggregation rule type is set to "maximum value aggregation." Therefore, the high-risk index of key pipe sections such as the downstream culvert inlet directly determines the area risk index under this operational slice. The area risk level is classified as high risk, and the control center immediately sends an early warning message to the drainage dispatching system, including the area identifier, the current slice identifier, real-time fullness estimates of several key pipe sections, uncertainty indicators, risk indices, and corresponding risk levels. Based on the early warning, dispatchers prioritize increasing the operating speed of downstream pumping stations without changing the safety conditions of other pumping stations, and appropriately adjust the operating strategies of certain gates and storage tanks according to the gate status and upstream water inflow, releasing some downstream storage space in advance. As the dispatching action takes effect, the real-time fullness estimates of the culvert inlet pipe section in the subsequent operational slices are controlled at around 80, no longer rising continuously, and the fullness change trend gradually approaches zero. The corresponding risk index also falls from the high-risk range to the medium-risk level.

[0108] During this rainstorm, a rain gauge in an older residential area upstream experienced a prolonged power outage in the later stages due to a power failure. The system marked the relevant observations as missing based on sensor loss and health information in the node evidence vector. This resulted in a significant penalty for adjacent pipe sections during uncertainty index calculations. When the uncertainty index of these pipe sections fell below the pre-configured threshold for unsuitable scheduling, the system, according to the risk assessment rules for "prolonged disconnection of the monitoring subsystem," marked the risk result of the corresponding pipe section as "for reference only, not suitable for direct scheduling," without forcibly raising the risk level. This prompt was clearly visible to dispatchers on the interface, allowing them to rely more on nearby normal monitoring points and on-site inspection results during actual scheduling. After the entire rainstorm event concluded, the operating unit compared the three recorded short-term water accumulation points and one near-manhole overflow hazard with the system's output fullness risk index. They found a high degree of overlap between high-risk and medium-high-risk pipe sections and the actual problem points. Only a few pipe sections corresponding to high-risk warnings showed no obvious abnormalities on-site, indicating acceptable warning redundancy. Based on these statistical results, the operating unit adjusted individual risk weights and thresholds appropriately afterward and generated a new risk assessment version number. All assessment actions related to this rainstorm event can still be fully replayed and reproduced under the original version number through the operation logs and evidence chain records, enabling those skilled in the art to verify the effectiveness and feasibility of this method from an engineering practice perspective.

[0109] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0110] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0111] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center containing one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0112] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0114] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0116] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0118] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for real-time filling degree assessment of drainage pipe networks, characterized in that, include: S1. Establish a drainage network topology model, configure unified time synchronization and operation slicing rules, and generate operation slice identifiers for each node and pipe segment; S2. Collect rainfall, water level, flow rate, and status of pumping stations and gates in each operating slice. Combine the pipe section structural parameters and sensor missing, delay, and health information to construct node evidence vectors and assign them mask features. S3. Construct a drainage network evidence graph using node evidence vectors and hydraulic connection attributes, and combine the evidence graphs corresponding to several consecutive running slices into a time-series evidence graph sequence. S4. Input the time-series evidence graph sequence into the pre-trained graph time-series neural network, perform weighted aggregation and time-series encoding between nodes to obtain the intermediate fill degree representation of each pipe segment. S5. Based on the mass conservation constraint, perform physical consistency correction on the intermediate filling degree characterization, and output the real-time filling degree estimate and uncertainty index of each pipe segment; S6. Calculate the fullness risk index based on the fullness estimate, uncertainty index and fullness change trend, and classify and output the risk of drainage pipe network sections and areas.

2. The method for real-time filling degree assessment of drainage pipe networks based on machine learning according to claim 1, characterized in that, S1 includes: A topology model of the drainage network was established based on pipeline ledgers and geographic information. The water flow direction of each pipe section is determined by comparing the elevation of the starting structure and the elevation of the ending structure, and combined with the design slope, and the manually specified type is recorded. Configure a unified time synchronization reference and time synchronization tolerance threshold, and verify the timestamps of the observed measurements based on the time synchronization tolerance threshold; Based on the baseline start time and the running slice rules, the running slices are divided, and the timestamps are mapped to the running slices. A running slice identifier is generated for each node unit and each pipe segment unit, which consists of a topology version number, an object type tag, an object code, and a running slice sequence number, and has a unique index in the running management library.

3. The method for real-time filling degree assessment of drainage pipe network based on machine learning according to claim 1, characterized in that, S2 include: The control center obtains observations from the rainfall monitoring system, water level monitoring system, flow monitoring system, pump station monitoring system, and gate monitoring system based on the operational slice identifiers. The timestamps of the observations are assigned to a unique running slice according to the running slice rules. Based on the allowed value range and the reasonable change range threshold, the representative value of the current running slice is classified into one of the normal observation category and the abnormal observation category; Missing information is generated when there are no valid observations in the running slice, and delayed information is generated when the difference between the observation reception time and the end time of the running slice is greater than the delay threshold.

4. The method for real-time filling degree assessment of drainage pipe networks based on machine learning according to claim 3, characterized in that: The control center maintains sensor health information for each monitoring point. When the observation is normal, the missing information is normal, and the delay information is normal within the running slice, the sensor health information is increased. When the observation is marked as a missing observation, a delayed observation, or an abnormal observation, the sensor health information is decreased and limited to a value range of zero to one. Based on rainfall, water level, flow rate, pump station status, gate status, pipe section structural parameters, missing information, delay information, and sensor health information, node evidence vectors are formed by encoding them in a fixed field order. The field order and field definition are locked by the node evidence vector version number.

5. The method for real-time filling degree assessment of drainage pipe networks based on machine learning according to claim 1, characterized in that, S3 includes: Within each operational slice, the control center constructs a drainage network evidence map based on the drainage network topology model and operational slice rules. Node units and pipe segment units are treated as points in the map, and the connection relationship between node units and pipe segment units is treated as a line with attached node evidence vectors and hydraulic connection attributes. The control center generates a time-series evidence map sequence from the drainage network evidence maps sorted by the running slice identifier according to the observation window length and sliding step size, and registers the time-series evidence map sequence together with the topology version number, running slice rule version number, hydraulic connection attribute configuration version number and window configuration version number in the operation management library.

6. The method for real-time filling degree assessment of drainage pipe network based on machine learning according to claim 1, characterized in that, S4 include: The control center inputs the drainage network evidence map into the graph-time neural network in the order of operation slices within the observation window; In the spatial aggregation layer, adjacency weights are calculated based on hydraulic connectivity attributes, sensor mask features, and sensor health information. The node evidence vectors are then weighted and aggregated according to the adjacency weights, and a nonlinear transformation is performed on the weighted results to obtain node representations. In the temporal coding layer, the node representation is recursively updated according to the running slice order using the temporal state vector; At the end of the observation window, the temporal representations of the nodes at both ends of the pipe segment are combined with the structural parameters of the pipe segment into an extended feature vector, which is then transformed linearly to generate the intermediate fill degree representation.

7. The method for real-time filling degree assessment of drainage pipe network based on machine learning according to claim 1, characterized in that, S5 include: The control center configures mass conservation constraint rules based on the drainage network topology model and hydraulic connection properties; The pipe section diameter and slope are converted into water conveyance capacity coefficients according to the calculation rules for water conveyance capacity coefficients; The total inflow and total outflow of the upstream pipe section of each node are calculated based on the filling degree component in the intermediate filling degree characterization and the water conveyance capacity coefficient to obtain the flow loss measurement. When the absolute value of the flow rate error exceeds the allowable range of node balance error, the relative water flow indicators of the upstream and downstream pipe sections are scaled by a uniform ratio, and the scaled filling degree component is truncated in intervals. The truncated filling degree component is used as the real-time filling degree estimate.

8. The method for real-time filling degree assessment of drainage pipe networks based on machine learning according to claim 7, characterized in that: After obtaining the real-time filling level estimate, the control center sets an initial confidence value for each pipe segment; The initial confidence value is multiplied by the missing sensor information recorded in the node evidence vector, the delay scaling factor is multiplied by the sensor delay information, the health reference value is multiplied by the sensor health information, and the adjustment range recorded in the mass conservation constraint step is multiplied by the corresponding adjustment scaling factor to obtain the uncertainty index. The real-time fill rate estimate, uncertainty index, as well as the running slice identifier, topology version number, model version number, and mass conservation constraint version number are registered in the fill rate assessment database to form evidence chain record entries.

9. The method for real-time filling degree assessment of drainage pipe network based on machine learning according to claim 1, characterized in that, S6 include: After each operating segment is completed, the control center reads the real-time estimated value of the fullness and uncertainty index of each pipe segment from the fullness assessment database, filters the historical fullness according to the confidence threshold within the preset observation period, and obtains the fullness change trend based on the normalization of the difference in fullness of the start and end operating segments. Based on the risk assessment rules, the current filling level, the filling level change trend and uncertainty are weighted and superimposed to obtain a filling level risk index between zero and one. The risk level of the pipe section and area is determined according to the graded threshold and an early warning output is triggered.

Images