An emergency drainage risk assessment method and system based on machine learning

Abstract

The application discloses an emergency drainage risk assessment method and system based on machine learning, comprising the following steps: collecting and preprocessing multi-source state data of emergency drainage, generating a synchronous state sequence; matching single-pump performance curves, series performance curves and parallel performance curves to generate a drainage capacity benchmark sequence; extracting disaster load, inflow blockage, pump group working condition deviation and equipment health characteristics to construct a risk feature set; performing deviation calculation on the risk feature set and the drainage capacity benchmark sequence to generate a risk trigger residual sequence; constructing a risk correlation feature map; inputting the risk correlation feature map into an improved space-time graph neural controlled differential model to generate multi-class risk probability; calculating a comprehensive risk assessment value and outputting a risk level, a dominant risk type and a risk assessment result. The application can improve the accuracy, continuity and reliability of emergency drainage risk assessment.

Description

Technical Field

[0001] This invention relates to the field of emergency drainage risk assessment technology, and in particular to an emergency drainage risk assessment method and system based on machine learning. Background Technology

[0002] In recent years, with the acceleration of urbanization and the increase in extreme rainfall events, emergency drainage operations such as urban flooding, flood control and disaster relief, and underground space drainage have placed higher demands on the operational reliability and risk prediction capabilities of drainage equipment. Existing technologies typically acquire drainage operation status through methods such as rainfall monitoring, water level monitoring, pump pressure monitoring, flow rate monitoring, motor current monitoring, temperature monitoring, and inlet image monitoring. They also utilize human experience, fixed threshold judgments, single-device alarms, or conventional machine learning models to identify pump blockages, equipment malfunctions, decreased drainage efficiency, and operational risks. Some solutions also combine single-pump performance curves, series performance curves, and parallel performance curves to schedule pump operation modes, ensuring that the pumps operate within the working range corresponding to the target flow rate and target head.

[0003] However, existing technologies still suffer from insufficient expression of risk coupling relationships and low accuracy in dynamic risk identification in emergency drainage risk assessment scenarios. On the one hand, existing solutions often treat disaster status, inlet blockage, pump condition, and equipment health as independent monitoring objects, lacking a processing method to conduct unified deviation analysis between multi-source status data and drainage capacity benchmarks, making it difficult to accurately reflect the impact of actual risks on drainage capacity. On the other hand, existing solutions typically rely on fixed thresholds or single risk probability outputs, lacking graph structure modeling of the node correlation relationships between load changes, blockage attenuation, operating condition deviations, and equipment attenuation. They also struggle to adapt to unstable sampling intervals and continuous risk changes in emergency operations, making it difficult to comprehensively assess risks such as disaster expansion, inlet blockage, pump instability, equipment overload, and insufficient drainage capacity in a timely and accurate manner. Summary of the Invention

[0004] One objective of this invention is to propose an emergency drainage risk assessment method and system based on machine learning. This invention achieves emergency drainage risk assessment based on machine learning and risk graph modeling, and has the advantages of accuracy, timeliness and reliability.

[0005] An emergency drainage risk assessment method based on machine learning according to an embodiment of the present invention includes the following steps:

[0006] Collect multi-source status data during emergency drainage operations, preprocess the multi-source status data, and generate a synchronized status sequence;

[0007] Based on the synchronous state sequence matching of single pump performance curves, series performance curves and parallel performance curves, a drainage capacity benchmark sequence is generated.

[0008] Based on the synchronous state sequence, disaster load characteristics, water inlet blockage characteristics, pump set operating condition deviation characteristics, and equipment health characteristics are extracted to construct a risk characteristic set;

[0009] The deviation between the risk feature set and the drainage capacity benchmark sequence is calculated to generate a risk trigger residual sequence.

[0010] Risk nodes are constructed and node associations are calculated based on risk feature sets and risk trigger residual sequences to build a risk association feature map.

[0011] The risk association feature map is input into the improved spatiotemporal graph neural controlled differential model. The node state, edge weight and time interval in the risk association feature map are jointly encoded in continuous time to generate the probability of disaster expansion risk, water ingress blockage risk, pump set instability risk, equipment overload risk and insufficient drainage capacity risk.

[0012] The probabilities of various risks are integrated to generate a comprehensive risk assessment value. Based on the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the risk assessment result are output.

[0013] Optionally, the multi-source status data includes rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate and target drainage head in the emergency drainage operation area, pump operation mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed and power, inlet image, bar screen image, pump body vibration, motor current and motor temperature; the preprocessing includes data acquisition time calibration, abnormal sampling point removal, missing sampling point completion, duplicate data merging and data dimension unification.

[0014] Optionally, the generation of the drainage capacity benchmark sequence includes:

[0015] Read the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, power, target drainage flow rate, and target drainage head from the synchronization status sequence;

[0016] The operating mode of the pump set determines the single pump operating state, series operating state, or parallel operating state. The single pump operating state matches the single pump performance curve, the series operating state matches the series performance curve, and the parallel operating state matches the parallel performance curve.

[0017] Based on inlet pressure, outlet pressure, drainage flow rate, engine speed and power, locate the current operating point in the corresponding performance curve;

[0018] Based on the target drainage flow rate and target drainage head, locate the target operating point in the corresponding performance curve;

[0019] Calculate the flow rate deviation, head deviation, and power deviation between the current operating point and the target operating point, and combine them with the valve opening to generate a baseline value for drainage capacity at a single moment;

[0020] The drainage capacity benchmark values ​​at each single moment are arranged in chronological order according to the synchronization state sequence to generate the drainage capacity benchmark sequence.

[0021] Optionally, the construction of the risk feature set includes:

[0022] Rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate, and target drainage head are extracted from the synchronous state sequence, and disaster load characteristics are generated according to the time window.

[0023] Extract inlet images and grid area images from the synchronous state sequence, and perform feature processing on debris coverage area, debris accumulation density, debris distribution location and grid occlusion range to generate inlet blockage features;

[0024] Calculate the flow rate deviation, head deviation, power deviation, and efficiency deviation between the current operating point and the target operating point. Combine the flow rate deviation, head deviation, power deviation, and efficiency deviation with the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, and power according to the time window to generate the pump set operating condition deviation characteristics.

[0025] Pump body vibration, motor current and motor temperature are extracted from the synchronous state sequence. The vibration change amplitude, current fluctuation amplitude and temperature rise amplitude are processed to generate equipment health characteristics.

[0026] Based on the time window correspondence, the characteristics of disaster load, water inlet blockage, pump set operating condition deviation, and equipment health are combined to construct a risk characteristic set.

[0027] Optionally, the generation of the risk-triggered residual sequence includes:

[0028] According to the time window correspondence, the various risk characteristics in the risk characteristic set are matched with the single-moment drainage capacity benchmark values ​​in the drainage capacity benchmark sequence;

[0029] Based on the characteristics of the disaster load, such as water level, water level rise rate and drainage target flow, a drainage demand index value is generated. The difference between the drainage demand index value and the single-moment drainage capacity benchmark value is calculated to generate the load trigger residual.

[0030] Based on the debris coverage area, debris accumulation density and grid obstruction range in the characteristics of water inlet blockage, the water inlet attenuation ratio is generated. The baseline value of drainage capacity at a single moment is multiplied by the water inlet attenuation ratio to generate the drainage capacity attenuation value caused by blockage, and the blockage trigger residual is determined.

[0031] Based on the flow deviation, head deviation, power deviation and efficiency deviation values ​​in the pump set operating condition deviation characteristics, the operating condition deviation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the operating condition deviation ratio to generate the drainage capacity attenuation value caused by the operating condition deviation, and the operating condition trigger residual is determined.

[0032] Based on the vibration change amplitude, current fluctuation amplitude, and temperature rise amplitude in the equipment health characteristics, the equipment attenuation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the equipment attenuation ratio to generate the drainage capacity attenuation value caused by the equipment state, and the health trigger residual is determined.

[0033] By combining load-triggered residuals, congestion-triggered residuals, operating condition-triggered residuals, and health-triggered residuals in chronological order, a risk-triggered residual sequence is generated.

[0034] Optionally, the construction of the risk association feature map includes:

[0035] The disaster load characteristics and load trigger residuals are combined to form the disaster load node state; the water inlet blockage characteristics and blockage trigger residuals are combined to form the water inlet blockage node state; the pump set operating condition deviation characteristics and operating condition trigger residuals are combined to form the pump set operating condition deviation node state; and the equipment health characteristics and health trigger residuals are combined to form the equipment health node state.

[0036] Based on the degree of synchronization of numerical changes, the degree of similarity of residual strength, and the degree of influence of drainage capacity among disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, and equipment health node status, the node correlation value is calculated.

[0037] Establish risk transmission edges between nodes whose node association values ​​satisfy the node connection conditions, and write the node association values ​​into the corresponding risk transmission edges to form edge weights;

[0038] By combining the disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, equipment health node status, risk transmission edge and edge weight in the order of time window, a risk association feature graph is constructed.

[0039] Optionally, the generation of the risk probability includes:

[0040] The risk-related feature map is input into the improved spatiotemporal graph neural controlled differential model, which includes a graph embedding layer, a time-controlled path generation layer, an edge weight graph propagation layer, a controlled differential evolution layer, and a multi-task risk prediction layer.

[0041] In the graph embedding layer, the node states in the risk association feature graph are mapped to the initial hidden states of the nodes, and the edge weights are organized into an edge weight matrix.

[0042] In the time-controlled path generation layer, the initial hidden state of the node is continuously interpolated based on the time interval between adjacent time windows to generate the node state control path.

[0043] In the edge-weighted graph propagation layer, the initial hidden state of the node is propagated using the edge weight matrix with neighborhood weighting to generate the graph propagation hidden state.

[0044] In the controlled differential evolution layer, the hidden state derivative is calculated based on the node state control path and the graph propagation hidden state, and the hidden state is updated along the time interval to generate the risk evolution hidden state.

[0045] In the multi-task risk prediction layer, the hidden states of risk evolution are mapped to the probability of disaster expansion risk, the probability of water blockage risk, the probability of pump set instability risk, the probability of equipment overload risk, and the probability of insufficient drainage capacity risk.

[0046] Optionally, the generation of the emergency drainage risk level, dominant risk type, and risk assessment results includes:

[0047] The probability of disaster spread, water blockage, pump instability, equipment overload, and insufficient drainage capacity are normalized to generate a set of normalized risk probabilities.

[0048] The corresponding fusion ratio is determined based on the magnitude of each risk probability in the normalized risk probability set;

[0049] The normalized risk probabilities are weighted and summed according to the integration ratio to generate a comprehensive risk assessment value.

[0050] The emergency drainage risk level is determined based on the comprehensive risk assessment value, the maximum risk probability, and the number of risk probabilities higher than the comprehensive risk assessment value.

[0051] The risk type corresponding to the risk probability with the largest value in the normalized risk probability set is determined as the dominant risk type.

[0052] The risk assessment result is generated by combining the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the probability of each risk.

[0053] An emergency drainage risk assessment system based on machine learning according to an embodiment of the present invention includes:

[0054] The synchronization processing module is used to collect and process multi-source status data during emergency drainage operations and generate a synchronization status sequence.

[0055] The benchmark generation module is used to generate a drainage capacity benchmark sequence by matching the performance curves of a single pump, series performance curves, and parallel performance curves based on the synchronous state sequence.

[0056] The feature construction module is used to construct a set of risk features based on the synchronization state sequence;

[0057] The residual generation module is used to calculate the deviation between the risk feature set and the drainage capacity benchmark sequence to generate a risk trigger residual sequence.

[0058] The graph construction module is used to construct a risk association feature graph based on the risk feature set and the risk trigger residual sequence.

[0059] The risk prediction module is used to input the risk association feature map into the improved spatiotemporal graph neural controlled differential model to generate the probability of each risk.

[0060] The assessment output module is used to integrate the probabilities of various risks, generate a comprehensive risk assessment value, and output the emergency drainage risk level, dominant risk type, and risk assessment results.

[0061] The beneficial effects of this invention are:

[0062] This invention collects multi-source state data during emergency drainage operations and preprocesses this data to generate a synchronized state sequence. This allows for the correlation analysis of data such as rainfall intensity, water level, pump operation mode, pressure, flow rate, engine speed, power, inlet image, pump vibration, motor current, and motor temperature on a simultaneous basis, avoiding the problem of incomplete information caused by single sensor monitoring or human experience judgment. By matching the synchronized state sequence with single pump performance curves, series performance curves, and parallel performance curves, a drainage capacity benchmark sequence is generated. This provides the risk assessment process with a drainage capacity reference adapted to the current pump operation mode, accurately reflecting the drainage capacity level that should be achieved under different operating conditions.

[0063] This invention unifies disaster load characteristics, inlet blockage characteristics, pump set operating condition deviation characteristics, and equipment health characteristics into a risk feature set. It further calculates the deviation between this set and the drainage capacity benchmark sequence to generate a risk trigger residual sequence. This allows risk assessment to move beyond simply identifying abnormal conditions and quantify the impact of various risk factors on drainage capacity. Through load trigger residuals, blockage trigger residuals, operating condition trigger residuals, and health trigger residuals, the impact of increased disaster severity, decreased inlet capacity, pump set operating point deviation, and equipment condition degradation on drainage tasks can be characterized, improving the ability to identify latent and early-stage risks.

[0064] This invention constructs a risk association feature graph based on a risk feature set and a risk trigger residual sequence, enabling the correlation between disaster load, water inlet blockage, pump set operating condition deviation, and equipment health to be expressed in the form of node states, risk propagation edges, and edge weights. The improved spatiotemporal graph neural controlled differential model performs continuous-time joint encoding of node states, edge weights, and time intervals, adapting to the characteristics of unstable sampling intervals, continuous changes in risk states, and dynamic evolution of risk relationships in emergency drainage operations. It generates multiple risk probabilities, including disaster expansion, water inlet blockage, pump set operating condition instability, equipment overload, and insufficient drainage capacity. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of an emergency drainage risk assessment method based on machine learning proposed in this invention;

[0067] Figure 2 This is a schematic diagram illustrating the generation of risk-triggered residual sequences in an emergency drainage risk assessment method based on machine learning proposed in this invention.

[0068] Figure 3 This diagram illustrates the generation of risk probability for an emergency drainage risk assessment method based on machine learning proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figures 1-3 An emergency drainage risk assessment method based on machine learning includes the following steps:

[0071] Collect multi-source status data during emergency drainage operations, preprocess the multi-source status data, and generate a synchronized status sequence;

[0072] Based on the synchronous state sequence matching of single pump performance curves, series performance curves and parallel performance curves, a drainage capacity benchmark sequence is generated.

[0073] Based on the synchronous state sequence, disaster load characteristics, water inlet blockage characteristics, pump set operating condition deviation characteristics, and equipment health characteristics are extracted to construct a risk characteristic set;

[0074] The deviation between the risk feature set and the drainage capacity benchmark sequence is calculated to generate a risk trigger residual sequence.

[0075] Risk nodes are constructed and node associations are calculated based on risk feature sets and risk trigger residual sequences to build a risk association feature map.

[0076] The risk association feature map is input into the improved spatiotemporal graph neural controlled differential model. The node state, edge weight and time interval in the risk association feature map are jointly encoded in continuous time to generate the probability of disaster expansion risk, water ingress blockage risk, pump set instability risk, equipment overload risk and insufficient drainage capacity risk.

[0077] The probabilities of various risks are integrated to generate a comprehensive risk assessment value. Based on the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the risk assessment result are output.

[0078] In this embodiment, the multi-source status data includes rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate and target drainage head in the emergency drainage operation area, pump operation mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed and power, inlet image, bar screen image, pump body vibration, motor current and motor temperature; preprocessing includes data acquisition time calibration, abnormal sampling point removal, missing sampling point completion, duplicate data merging and data dimension unification.

[0079] In this embodiment, the generation of the drainage capacity reference sequence includes:

[0080] Read the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, power, target drainage flow rate, and target drainage head from the synchronization status sequence;

[0081] The operating mode of the pump set determines the single pump operating state, series operating state, or parallel operating state. The single pump operating state matches the single pump performance curve, the series operating state matches the series performance curve, and the parallel operating state matches the parallel performance curve.

[0082] The determination of the operating status specifically includes: pre-establishing a drainage capacity curve library, which includes pump performance curves, series performance curves, and parallel performance curves; using the pump group operating mode in the synchronous state sequence as the basis for state discrimination, extracting the mode marker, operating scheme number, activated pump number, pipeline connection marker, and valve combination marker corresponding to the pump group operating mode; when the mode marker is a single pump mode and the operating scheme number corresponds to an activated pump number, the current sampling time is determined as a single pump operating state; when the mode marker is a series mode and the pipeline connection marker indicates that the outlet of the upstream pump is connected to the inlet of the downstream pump, the activated pump numbers are arranged according to the water flow direction to generate a series stage sequence, and the current sampling time is determined as a series operating state; when the mode marker is a parallel mode and the pipeline connection marker indicates that multiple activated pump inlets are connected to the same inlet area and outlets are connected to the same outlet pipeline, the activated pump numbers are arranged according to the branch number to generate a parallel branch sequence, and the current sampling time is determined as a parallel operating state.

[0083] Based on inlet pressure, outlet pressure, drainage flow rate, engine speed and power, locate the current operating point in the corresponding performance curve;

[0084] Determining the current operating point specifically includes: using the performance curve corresponding to the current sampling time as the positioning range, determining the current speed curve segment based on the engine speed within the corresponding performance curve; if no curve in the current speed curve segment is exactly the same as the engine speed, selecting two adjacent speed curve segments for speed interpolation to generate the current speed matching curve segment; calculating the pressure difference between the outlet pressure and the inlet pressure, and converting the pressure difference into the current head characterization value; determining the drainage flow rate as the current flow rate characterization value, and combining the current flow rate characterization value and the current head characterization value to form the flow-head positioning coordinates; finding the curve position corresponding to the current flow rate characterization value within the current speed matching curve segment, with the curve position located between two... When there are multiple adjacent flow sampling points, interpolation is performed on the head, power, and efficiency corresponding to the adjacent flow sampling points to generate candidate curve positions. The difference between the curve head corresponding to the candidate curve position and the current head value is calculated, and the difference between the curve power corresponding to the candidate curve position and the power at the current sampling time is calculated. The candidate curve positions are filtered according to the head difference and power difference, and the candidate curve position that meets the positioning conditions for both head difference and power difference is determined as the current operating point. When there are multiple candidate curve positions that meet the positioning conditions, they are sorted from smallest to largest according to the comprehensive deviation of the head difference and power difference, and the candidate curve position with the highest deviation is determined as the current operating point.

[0085] Based on the target drainage flow rate and target drainage head, locate the target operating point in the corresponding performance curve;

[0086] The determination of the target operating point specifically includes: using the performance curve corresponding to the current sampling time as the positioning range, determining the target drainage flow rate as the target flow rate characterization value, and the target drainage head as the target head characterization value; based on the engine speed at the current sampling time, determining the target speed curve segment in the corresponding performance curve; if there is no curve in the target speed curve segment that is completely consistent with the engine speed, selecting two adjacent speed curve segments for speed interpolation to generate a target speed matching curve segment; using the target flow rate characterization value and the target head characterization value to form the target positioning coordinates, and finding the curve position corresponding to the target flow rate characterization value in the target speed matching curve segment; the curve position is located between two adjacent... When sampling flow points are adjacent, interpolation is performed on the head, power, and efficiency corresponding to the adjacent flow sampling points to generate target candidate curve positions. The difference between the curve head corresponding to the target candidate curve position and the target head characterization value is calculated to generate the target head deviation. The curve power and curve efficiency corresponding to the target candidate curve position are correlated and extracted to generate target power and target efficiency values. When the target head deviation meets the target positioning conditions, the target candidate curve position is determined as the target operating point. When multiple target candidate curve positions meet the target positioning conditions, they are sorted in ascending order of target head deviation, and the target candidate curve position with the largest deviation is determined as the target operating point.

[0087] Calculate the flow rate deviation, head deviation, and power deviation between the current operating point and the target operating point, and combine them with the valve opening to generate a baseline value for drainage capacity at a single moment;

[0088] The calculation of the baseline value of drainage capacity at a single moment is as follows: Extract the current flow rate, current head, and current power value corresponding to the current operating point; extract the target flow rate, target head, and target power value corresponding to the target operating point; calculate the difference between the target flow rate and the current flow rate to generate a flow deviation value; calculate the difference between the target head and the current head to generate a head deviation value; calculate the difference between the target power and the current power to generate a power deviation value; perform absolute value processing on the flow deviation value, head deviation value, and power deviation value respectively, and then proportionalize them according to the target flow rate, target head, and target power values ​​respectively to generate the flow deviation ratio, head deviation ratio, and power deviation ratio. The system calculates the valve deviation ratio and power deviation ratio; converts the valve opening at the current sampling time into the effective valve opening ratio, and generates a valve correction coefficient based on the effective valve opening ratio; it then merges the flow deviation ratio, head deviation ratio, and power deviation ratio according to their respective weights to generate the operating point deviation value; it multiplies the operating point deviation value with the valve correction coefficient to generate the drainage capacity deviation correction value; it normalizes and merges the target flow value, target head value, and target efficiency value corresponding to the target operating point to generate the target drainage capacity value; and it subtracts the drainage capacity deviation correction value from the target drainage capacity value to generate the single-moment drainage capacity benchmark value at the current sampling time.

[0089] The drainage capacity benchmark values ​​at each single moment are arranged in chronological order according to the synchronization state sequence to generate the drainage capacity benchmark sequence.

[0090] In this embodiment, the construction of the risk feature set includes:

[0091] Rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate, and target drainage head are extracted from the synchronous state sequence, and disaster load characteristics are generated according to the time window.

[0092] Extract inlet images and grid area images from the synchronous state sequence, and perform feature processing on debris coverage area, debris accumulation density, debris distribution location and grid occlusion range to generate inlet blockage features;

[0093] Calculate the flow rate deviation, head deviation, power deviation, and efficiency deviation between the current operating point and the target operating point. Combine the flow rate deviation, head deviation, power deviation, and efficiency deviation with the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, and power according to the time window to generate the pump set operating condition deviation characteristics.

[0094] The difference between the target flow rate and the current flow rate is calculated to generate a flow rate deviation value; the difference between the target head and the current head is calculated to generate a head deviation value; the difference between the target power and the current power is calculated to generate a power deviation value; and the difference between the target efficiency and the current efficiency is calculated to generate an efficiency deviation value.

[0095] Pump body vibration, motor current and motor temperature are extracted from the synchronous state sequence. The vibration change amplitude, current fluctuation amplitude and temperature rise amplitude are processed to generate equipment health characteristics.

[0096] Based on the time window correspondence, the characteristics of disaster load, water inlet blockage, pump set operating condition deviation, and equipment health are combined to construct a risk characteristic set.

[0097] In this embodiment, the generation of the risk-triggered residual sequence includes:

[0098] According to the time window correspondence, the various risk characteristics in the risk characteristic set are matched with the single-moment drainage capacity benchmark values ​​in the drainage capacity benchmark sequence;

[0099] Based on the characteristics of the disaster load, such as water level, water level rise rate and drainage target flow, a drainage demand index value is generated. The difference between the drainage demand index value and the single-moment drainage capacity benchmark value is calculated to generate the load trigger residual.

[0100] The calculation of the drainage demand index value is as follows: Extract the accumulated water level, water level rise rate, and drainage target flow rate for the current and previous time windows; calculate the difference between the accumulated water level in the current time window and the accumulated water level in the previous time window to generate the water level increase; calculate the difference between the water level rise rate in the current time window and the water level rise rate in the previous time window to generate the rise rate increase; set the values ​​less than zero in the water level increase and rise rate increase to zero to generate the effective water level increase and effective rise rate increase; multiply the drainage target flow rate by the effective water level increase and the effective rise rate increase respectively to obtain the water level load increment and rise rate load increment; add the drainage target flow rate, water level load increment, and rise rate load increment to obtain the drainage demand index value.

[0101] Based on the debris coverage area, debris accumulation density and grid obstruction range in the characteristics of water inlet blockage, the water inlet attenuation ratio is generated. The baseline value of drainage capacity at a single moment is multiplied by the water inlet attenuation ratio to generate the drainage capacity attenuation value caused by blockage, and the blockage trigger residual is determined.

[0102] The calculation of the influent attenuation ratio is as follows: Determine the effective influent area and shielding range of the bar screen corresponding to the current time window; extract the debris coverage area and debris accumulation density of each shielded area within the shielding range; sum the debris coverage areas of each shielded area to generate the total debris coverage area; calculate the shielding area corresponding to the bar screen shielding range to generate the bar screen shielding area; divide the total debris coverage area by the effective influent area of ​​the bar screen to generate the debris coverage ratio; divide the bar screen shielding area by the effective influent area of ​​the bar screen to generate the bar screen shielding ratio; multiply the debris coverage area of ​​each shielded area by the debris accumulation density and sum them to generate the debris accumulation amount; divide the debris accumulation amount by the effective influent area of ​​the bar screen to generate the debris accumulation ratio; sum the debris coverage ratio, bar screen shielding ratio, and debris accumulation ratio to generate the influent attenuation ratio.

[0103] Based on the flow deviation, head deviation, power deviation and efficiency deviation values ​​in the pump set operating condition deviation characteristics, the operating condition deviation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the operating condition deviation ratio to generate the drainage capacity attenuation value caused by the operating condition deviation, and the operating condition trigger residual is determined.

[0104] The calculation of the operating condition deviation ratio is as follows: the absolute value of the flow deviation value is divided by the target flow value to generate the flow deviation ratio; the absolute value of the head deviation value is divided by the target head value to generate the head deviation ratio; the absolute value of the power deviation value is divided by the target power value to generate the power deviation ratio; the absolute value of the efficiency deviation value is divided by the target efficiency value to generate the efficiency deviation ratio; the flow deviation ratio, head deviation ratio, power deviation ratio, and efficiency deviation ratio are added together and averaged to generate the operating condition deviation ratio.

[0105] Based on the vibration change amplitude, current fluctuation amplitude, and temperature rise amplitude in the equipment health characteristics, the equipment attenuation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the equipment attenuation ratio to generate the drainage capacity attenuation value caused by the equipment state, and the health trigger residual is determined.

[0106] The calculation of the equipment attenuation ratio is as follows: Extract the vibration change amplitude, current fluctuation amplitude, and temperature rise amplitude corresponding to the current time window and the previous time window from the equipment health characteristics; divide the vibration change amplitude by the pump body vibration amplitude corresponding to the previous time window to generate the vibration change ratio; divide the current fluctuation amplitude by the motor current value corresponding to the previous time window to generate the current fluctuation ratio; divide the temperature rise amplitude by the motor temperature value corresponding to the previous time window to generate the temperature rise ratio; add the vibration change ratio, current fluctuation ratio, and temperature rise ratio together and average them to generate the equipment attenuation ratio.

[0107] By combining load-triggered residuals, congestion-triggered residuals, operating condition-triggered residuals, and health-triggered residuals in chronological order, a risk-triggered residual sequence is generated.

[0108] In this embodiment, the construction of the risk association feature map includes:

[0109] The disaster load characteristics and load trigger residuals are combined to form the disaster load node state; the water inlet blockage characteristics and blockage trigger residuals are combined to form the water inlet blockage node state; the pump set operating condition deviation characteristics and operating condition trigger residuals are combined to form the pump set operating condition deviation node state; and the equipment health characteristics and health trigger residuals are combined to form the equipment health node state.

[0110] Based on the degree of synchronization of numerical changes, the degree of similarity of residual strength, and the degree of influence of drainage capacity among disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, and equipment health node status, the node correlation value is calculated.

[0111] The calculation of node correlation values ​​is as follows: Within the same time window, any two risk nodes are selected as the node pair to be calculated. The node pair to be calculated includes any two nodes from the following: disaster load node, water inlet blockage node, pump set operating condition deviation node, and equipment health node. The node states of the node pair to be calculated are extracted in the current time window and the previous time window. The node states in the current time window and the previous time window are subtracted item by item, and the absolute values ​​are taken to generate the node state change amounts corresponding to the two risk nodes. The smaller value of the two node state change amounts is divided by the larger value to generate the numerical change synchronization value. The node state change values ​​to be calculated are then extracted. The corresponding trigger residuals of the node pairs in the current time window are calculated. The smaller of the two corresponding trigger residuals is divided by the larger value to generate a residual strength approximation value. The two corresponding trigger residuals are divided by the single-moment drainage capacity benchmark value of the current time window to generate two drainage capacity influence ratios. The smaller of the two drainage capacity influence ratios is divided by the larger value to generate the drainage capacity influence value. The numerical change synchronization value, the residual strength approximation value, and the drainage capacity influence value are added together and averaged to generate the node association value of the node pairs to be calculated. All node pairs to be calculated are traversed in the above manner to obtain the node association value between each risk node.

[0112] Establish risk transmission edges between nodes whose node association values ​​satisfy the node connection conditions, and write the node association values ​​into the corresponding risk transmission edges to form edge weights;

[0113] The generation of risk transmission edges and edge weights specifically includes: within the current time window, pairing disaster load nodes, water inlet blockage nodes, pump set operating condition deviation nodes, and equipment health nodes to generate multiple candidate node pairs; extracting the node correlation values ​​corresponding to each candidate node pair and calculating the average of all node correlation values ​​within the current time window to generate a node connection judgment value; determining candidate node pairs whose node correlation values ​​are not lower than the node connection judgment value as valid node pairs; for each valid node pair, extracting the trigger residuals corresponding to the two risk nodes respectively, and dividing each trigger residual by the single-moment drainage capacity benchmark value to generate the drainage capacity influence ratio of the two risk nodes; determining the risk node with a larger drainage capacity influence ratio as the edge start point and the risk node with a smaller drainage capacity influence ratio as the edge end point, and establishing a risk transmission edge between the edge start point and the edge end point; when the drainage capacity influence ratios of the two risk nodes are the same, establishing a bidirectional risk transmission edge between the two risk nodes; and writing the node correlation values ​​corresponding to the valid node pairs into the risk transmission edge to form the edge weight of the risk transmission edge.

[0114] By combining the disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, equipment health node status, risk transmission edge and edge weight in the order of time window, a risk association feature graph is constructed.

[0115] In this embodiment, the generation of risk probability includes:

[0116] The risk-related feature map is input into the improved spatiotemporal graph neural controlled differential model, which includes a graph embedding layer, a time-controlled path generation layer, an edge weight graph propagation layer, a controlled differential evolution layer, and a multi-task risk prediction layer.

[0117] The improvements to the spatiotemporal graph neural controlled differential model are specifically reflected in the following aspects: The basic spatiotemporal graph neural controlled differential model typically completes node propagation based on fixed adjacency relationships. This implementation introduces an edge weight matrix from the risk association feature graph, causing the propagation strength between disaster load nodes, water ingress blockage nodes, pump set operating condition deviation nodes, and equipment health nodes to change with the degree of risk association; the hidden state of a node is no longer constructed solely from node observation features, but rather integrates the risk feature set and the risk trigger residual sequence, allowing the hidden state to simultaneously represent risk performance information and drainage capacity deviation information; for situations where sampling intervals are inconsistent in emergency drainage operations, a node state control path is generated based on the actual time interval between adjacent time windows, ensuring that the continuous time update process corresponds to the actual sampling process; during the calculation of the hidden state derivative, the node state control path, edge weight matrix, and time interval are simultaneously introduced, causing the state evolution of risk nodes to be constrained by the node's own changes, the strength of inter-node association, and the time interval; the model output is set with multi-task risk prediction branches corresponding to disaster expansion, water ingress blockage, pump set operating condition instability, equipment overload, and insufficient drainage capacity, generating probabilities for each risk respectively;

[0118] The training data comes from historical emergency drainage operation records, on-site drill operation records, and pump unit operation test records. The training data includes multi-source state data, drainage capacity benchmark sequences, risk feature sets, risk trigger residual sequences, risk association feature maps, and corresponding risk labeling results. The risk labeling results include disaster expansion labels, water ingress blockage labels, pump unit instability labels, equipment overload labels, and insufficient drainage capacity labels. The training data is processed by correcting the acquisition time, completing missing data, removing anomalies, unifying units, and dividing time windows. According to the above processing, training synchronization state sequences, training drainage capacity benchmark sequences, training risk feature sets, training risk trigger residual sequences, and training risk association feature maps are generated. The node states, edge weights, and time intervals in the training risk association feature maps are organized into training samples, and the corresponding risk labeling results are organized into training labels. The training set, validation set, and test set are divided according to the operation batch.

[0119] The model training parameters are set, including the node hidden state dimension, the order of the time control path, the number of graph propagation layers, the learning rate, the batch size, the number of training epochs, the gradient pruning threshold, and the number of early stopping epochs. The Adam optimizer is used to update the model parameters. In each training epoch, the probabilities of disaster spread risk, water ingress blockage risk, pump set instability risk, equipment overload risk, and insufficient drainage capacity risk are calculated. The binary classification cross-entropy loss is calculated with the corresponding risk labeling results to generate disaster spread loss, water ingress blockage loss, pump set instability loss, equipment overload loss, and insufficient drainage capacity loss. Five types of losses are weighted and fused according to risk category weights to generate a multi-task risk prediction loss. Edge weight sparse constraint loss and hidden state smoothing constraint loss are added to the multi-task risk prediction loss to generate the total model loss. The parameters of the graph embedding layer, time-controlled path generation layer, edge weight graph propagation layer, controlled differential evolution layer, and multi-task risk prediction layer are updated in reverse according to the total model loss. When the total model loss on the validation set continuously reaches the early stopping condition or the number of training rounds reaches the set upper limit, training is stopped, and the model parameters with the highest comprehensive risk assessment accuracy on the validation set are saved to obtain the improved spatiotemporal graph neural controlled differential model after training is completed.

[0120] In the graph embedding layer, the node states in the risk association feature graph are mapped to the initial hidden states of the nodes, and the edge weights are organized into an edge weight matrix.

[0121] The generation of the initial hidden state of a node specifically includes: determining the node state of each risk node as its own state vector, which includes the corresponding risk feature and the corresponding triggering residual; taking the current risk node as the target node, extracting the node states of adjacent risk nodes based on the risk propagation edges connected to the target node in the risk association feature graph, and weighting and summarizing the node states of adjacent risk nodes according to the corresponding edge weights to generate the adjacency association state vector of the target node; concatenating the target node's own state vector with the adjacency association state vector to generate the target node's graph association state vector; and performing linear mapping, nonlinear activation, and layer normalization on the target node's graph association state vector to generate the initial hidden state of the node corresponding to the target node.

[0122] In the time-controlled path generation layer, the initial hidden state of the node is continuously interpolated based on the time interval between adjacent time windows to generate the node state control path.

[0123] The generation of the node state control path specifically includes: configuring a sampling time marker for each time window; subtracting the sampling times of two adjacent time windows to generate the time interval between adjacent time windows; for the same risk node, extracting the node's initial hidden state in the current time window, the node's initial hidden state in the next time window, and the corresponding time interval; calculating the amount of hidden state change between the current time window and the next time window, and distributing the amount of hidden state change according to the time interval to generate the state change slope of the risk node between adjacent time windows; setting continuous time positions between the current time window and the next time window, dividing the time difference between the continuous time position and the sampling time of the current time window by the time interval to generate the normalized time position; performing piecewise interpolation based on the node's initial hidden state in the current time window, the node's initial hidden state in the next time window, the state change slope, and the normalized time position to obtain the continuous hidden state of the risk node at any continuous time position between adjacent time windows; and connecting the continuous hidden states of each risk node according to the time window order to generate the node state control path.

[0124] In the edge-weighted graph propagation layer, the initial hidden state of the node is propagated using the edge weight matrix with neighborhood weighting to generate the graph propagation hidden state.

[0125] The generation of the graph propagation hidden state specifically includes: normalizing the edge weights of each row in the edge weight matrix to generate a normalized edge weight matrix; for any target risk node, determining the adjacent risk nodes connected to the target risk node based on the normalized edge weight matrix, and extracting the initial hidden state of the adjacent risk nodes; multiplying the initial hidden state of each adjacent risk node by its corresponding normalized edge weight to generate the weighted hidden state of the adjacent nodes; summing the weighted hidden states of all adjacent nodes corresponding to the target risk node to generate the neighborhood propagation hidden state of the target risk node; concatenating the initial hidden state of the target risk node with the neighborhood propagation hidden state, and performing linear transformation, nonlinear activation, and normalization to generate the graph propagation hidden state of the target risk node; performing neighborhood weighted propagation on each risk node in the same way, arranging the graph propagation hidden states corresponding to disaster load nodes, water inlet blockage nodes, pump set operating condition deviation nodes, and equipment health nodes in node order to generate the graph propagation hidden state corresponding to the current time window;

[0126] In the controlled differential evolution layer, the hidden state derivative is calculated based on the node state control path and the graph propagation hidden state, and the hidden state is updated along the time interval to generate the risk evolution hidden state.

[0127] The generation of risk evolution hidden states specifically includes: determining the continuous time solution interval corresponding to the node state control path according to the time window sequence, and determining the graph propagation hidden state of the current time window as the initial evolution hidden state of each risk node; for disaster load nodes, water inlet blockage nodes, pump set operating condition deviation nodes, and equipment health nodes, extracting the control path state corresponding to the current continuous time position along the node state control path, and determining the control path change based on the control path state difference between adjacent continuous time positions; concatenating the current evolution hidden state, control path state, control path change, and corresponding graph propagation hidden state, and generating the current continuous time hidden state through linear mapping, nonlinear activation, and gating filtering. The hidden state derivatives corresponding to the time positions are calculated; based on the time interval between adjacent consecutive time positions, the hidden state derivatives are integrated and updated to generate the updated evolutionary hidden state corresponding to the next consecutive time position; when the sampling time of the next time window is reached, the updated evolutionary hidden state is gated and fused with the graph propagation hidden state corresponding to the next time window to generate the evolutionary hidden state of the next time window; the hidden state derivative calculation and integration update are continuously performed in the order of the time windows to obtain the evolutionary hidden state of each risk node in the complete operation period; the evolutionary hidden states corresponding to disaster load nodes, water inlet blockage nodes, pump set operating condition deviation nodes, and equipment health nodes are combined in the order of nodes to generate the risk evolutionary hidden state;

[0128] In the multi-task risk prediction layer, the hidden states of risk evolution are mapped to the probability of disaster expansion risk, the probability of water blockage risk, the probability of pump set instability risk, the probability of equipment overload risk, and the probability of insufficient drainage capacity risk.

[0129] The calculation of each risk probability is as follows: Extract the node evolution hidden states corresponding to each risk node in the risk evolution hidden states; aggregate the node evolution hidden states to generate a global risk evolution vector; concatenate the disaster load node evolution hidden states with the global risk evolution vector to generate a disaster expansion discriminant vector; concatenate the water inlet blockage node evolution hidden states with the global risk evolution vector to generate a water inlet blockage discriminant vector; concatenate the pump unit operating condition deviation node evolution hidden states with the global risk evolution vector to generate a pump unit operating condition instability discriminant vector; and concatenate the equipment... The hidden state of healthy nodes is concatenated with the global risk evolution vector to generate a vector to be judged for equipment overload; the hidden state of each node is concatenated with the global risk evolution vector to generate a vector to be judged for insufficient drainage capacity; each vector to be judged is subjected to linear transformation, nonlinear activation and single-output mapping to generate corresponding risk judgment values; each risk judgment value is normalized by Sigmoid to obtain the probability of disaster expansion risk, water inlet blockage risk, pump set instability risk, equipment overload risk, and insufficient drainage capacity risk.

[0130] In this embodiment, the generation of emergency drainage risk level, dominant risk type, and risk assessment results includes:

[0131] The probability of disaster spread, water blockage, pump instability, equipment overload, and insufficient drainage capacity are normalized to generate a set of normalized risk probabilities.

[0132] The corresponding fusion ratio is determined based on the magnitude of each risk probability in the normalized risk probability set;

[0133] The normalized risk probabilities are weighted and summed according to the integration ratio to generate a comprehensive risk assessment value.

[0134] The emergency drainage risk level is determined based on the comprehensive risk assessment value, the maximum risk probability, and the number of risk probabilities higher than the comprehensive risk assessment value.

[0135] The emergency drainage risk level classification is as follows: The average risk probability is calculated by averaging the risk probabilities in the normalized risk probability set; the deviation of each risk probability from the average risk probability is calculated to generate a risk probability dispersion value; a low-risk level is defined as a case where the comprehensive risk assessment value is not higher than the average risk probability, and the number of risk probabilities higher than the comprehensive risk assessment value does not exceed one; a medium-risk level is defined as a case where the comprehensive risk assessment value is higher than the average risk probability, but not higher than the sum of the average risk probability and the risk probability dispersion value; a high-risk level is defined as a case where the comprehensive risk assessment value is higher than the sum of the average risk probability and the risk probability dispersion value, and the number of risk probabilities higher than the comprehensive risk assessment value does not exceed two; and an extremely high-risk level is defined as a case where the maximum risk probability is higher than the sum of the average risk probability and the risk probability dispersion value, and the number of risk probabilities higher than the comprehensive risk assessment value is greater than two.

[0136] The risk type corresponding to the risk probability with the largest value in the normalized risk probability set is determined as the dominant risk type.

[0137] The risk assessment result is generated by combining the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the probability of each risk.

[0138] An emergency drainage risk assessment system based on machine learning includes:

[0139] The synchronization processing module is used to collect and process multi-source status data during emergency drainage operations and generate a synchronization status sequence.

[0140] The benchmark generation module is used to generate a drainage capacity benchmark sequence by matching the performance curves of a single pump, series performance curves, and parallel performance curves based on the synchronous state sequence.

[0141] The feature construction module is used to construct a set of risk features based on the synchronization state sequence;

[0142] The residual generation module is used to calculate the deviation between the risk feature set and the drainage capacity benchmark sequence to generate a risk trigger residual sequence.

[0143] The graph construction module is used to construct a risk association feature graph based on the risk feature set and the risk trigger residual sequence.

[0144] The risk prediction module is used to input the risk association feature map into the improved spatiotemporal graph neural controlled differential model to generate the probability of each risk.

[0145] The assessment output module is used to integrate the probabilities of various risks, generate a comprehensive risk assessment value, and output the emergency drainage risk level, dominant risk type, and risk assessment results.

[0146] Example 1: To verify the feasibility of this invention in practice, it was applied to an emergency drainage operation scenario during the rainy season in a riverside city. During the flood season, the city experienced continuous heavy rainfall at night, leading to water accumulation in low-lying roads, underpasses, and surrounding residential areas. After dispatching vehicles carrying mobile pump sets to the drainage points, it was necessary to continuously assess changes in water level, drainage capacity, inlet blockage, pump set operating conditions, and equipment load. Traditional methods primarily rely on on-site personnel observing water levels, checking pressure gauges, listening to equipment operation sounds, and waiting for fixed alarm signals to assess risk. However, in environments with heavy rainfall, limited visibility at night, complex floating debris at the inlet, and frequent adjustments to pump set operating modes, human experience is insufficient to distinguish the source of risk. This can easily lead to problems such as decreased drainage capacity going unnoticed, increased equipment load without a clear cause, and partial blockage of the grating failing to trigger an obvious alarm while continuously affecting drainage efficiency.

[0147] In this scenario, the on-site drainage task is issued by the dispatch terminal, and the target drainage flow rate and target drainage head are recorded synchronously with the task. The system collects multi-source status data such as rainfall, water level, pump operation mode, valve opening, pressure, flow rate, engine speed, power, inlet image, bar screen area image, pump vibration, motor current and motor temperature, and performs preprocessing to form a synchronous status sequence. The system matches the performance curves of a single pump, series pump, or parallel pump based on the pump set's operating mode to generate a baseline sequence of drainage capacity. It then extracts disaster load characteristics, inlet blockage characteristics, pump set operating condition deviation characteristics, and equipment health characteristics from the synchronous state sequence to construct a risk feature set. The system calculates the deviation between the risk feature set and the baseline sequence of drainage capacity to form a risk triggering residual sequence. Based on the risk feature set and the risk triggering residual sequence, a risk correlation feature graph is constructed. This graph is then fed into an improved spatiotemporal graph neural controlled differential model to generate probabilities of disaster expansion, inlet blockage, pump set operating condition instability, equipment overload, and insufficient drainage capacity. Finally, the system integrates these probabilities and outputs a comprehensive risk assessment value, an emergency drainage risk level, the dominant risk type, and the risk assessment result.

[0148] In practical applications, situations have arisen where rainfall intensifies, water level recedes only slightly, and pump power increases with limited improvement in drainage flow. Traditionally, field personnel would prioritize checking for pump shutdowns, abnormal pressure, or significant grid blockage. However, these phenomena are often caused by multiple factors, and a single parameter is insufficient to directly explain the problem. This invention provides a reference for the current pump operating status through a drainage capacity benchmark sequence, enabling field personnel to determine if the current drainage capacity deviates from task requirements. It also transforms increased disaster severity, obstructed water intake, operational deviations, and equipment degradation into impacts on drainage capacity through a risk-triggered residual sequence. Furthermore, it expresses the correlation between various risks through a risk association feature graph, allowing the system to go beyond simply outputting a single alarm and instead identify changes in risk level and the primary source of risk.

[0149] In this operation, the on-site central control unit, pump control unit, image acquisition unit, and dispatching unit jointly stored operation records. The records covered changes in rainfall, water accumulation, pump operation mode, pressure and flow, inlet image changes, vibration, current, and temperature changes, as well as the risk assessment results output by the system. Compared with the original method relying on manual inspection and fixed threshold alarms, this invention can provide risk change warnings before the decline in drainage capacity develops into a significant fault. It can identify the dominant risk type when partial obstruction of the inlet, pump operation point deviation, and equipment load fluctuation occur simultaneously. It can also assist on-site personnel in arranging bar screen cleaning, operation mode verification, and equipment inspection in advance.

[0150] Table 1. Comparison of the overall performance of the method of the present invention and traditional emergency drainage risk assessment methods.

[0151] Accuracy rate of risk level assessment (%) 82.6 88.7 Accuracy rate of identifying dominant risk types (%) 79.4 86.2 Accuracy rate of disaster spread risk identification (%) 81.3 87.5 Accuracy rate of water ingress blockage risk identification (%) 84.1 90.2 Accuracy rate of pump set instability identification (%) 80.7 87.1 Equipment overload risk identification accuracy (%) 81.5 87.8 Accuracy rate of identifying insufficient drainage capacity risk (%) 79.8 86.9 Time taken for a single risk assessment (min) 4.2 2.6

[0152] As shown in Table 1, the method of this invention outperforms traditional methods in all risk identification indicators. The accuracy of risk level judgment increased from 82.6% to 88.7%, indicating that the present invention, through joint processing of multi-source state data using synchronous state sequences, drainage capacity benchmark sequences, and risk trigger residual sequences, can more accurately reflect the current risk level of drainage operations than traditional methods that rely on fixed thresholds or human experience. This improvement does not come from single sensor data enhancement, but from comprehensive modeling of multiple factors such as disaster load, inlet blockage, pump set operating condition deviation, and equipment health.

[0153] The accuracy rate of identifying the dominant risk type increased from 79.4% to 86.2%, indicating that the present invention has a better ability to distinguish the source of risk. Traditional methods, when faced with simultaneous occurrences of decreased flow, pressure fluctuations, increased current, and inlet debris accumulation, easily categorize multiple phenomena as ordinary equipment anomalies or blockage anomalies. The present invention combines various risk characteristics with corresponding trigger residuals and constructs a risk correlation feature graph, enabling the formation of correlations between disaster load nodes, inlet blockage nodes, pump set operating condition deviation nodes, and equipment health nodes. Therefore, it can more accurately determine the dominant risk type.

[0154] From the perspective of individual risk identification results, the accuracy rate of identifying inlet blockage risk increased from 84.1% to 90.2%, and the accuracy rate of identifying insufficient drainage capacity risk increased from 79.8% to 86.9%, both of which showed significant improvements. This is because the present invention does not solely rely on inlet images or bar screen obstruction to determine blockage. Instead, it converts the area covered by debris, debris accumulation density, and bar screen obstruction range into an inlet attenuation ratio, and combines this with a single-moment drainage capacity benchmark value to generate a blockage trigger residual, thus establishing a correlation between blockage risk and the impact on drainage capacity. The improvement in the risk of insufficient drainage capacity stems from the matching process of single-pump performance curves, series performance curves, and parallel performance curves. This allows the system to form a drainage capacity reference based on the current pump unit operating status, rather than solely relying on real-time flow rate declines for judgment.

[0155] The accuracy rate for identifying pump set instability improved from 80.7% to 87.1%, and the accuracy rate for identifying equipment overload risk improved from 81.5% to 87.8%. This indicates that the present invention provides a more stable identification of pump set operating conditions and equipment health status. Traditional methods typically assess inlet pressure, outlet pressure, drainage flow rate, power, current, and temperature separately, making it difficult to reflect the combined impact of these data on drainage capacity. The present invention generates pump set operating condition deviation characteristics by measuring the flow rate deviation, head deviation, power deviation, and efficiency deviation between the current operating point and the target operating point, and generates equipment health characteristics by measuring vibration variation amplitude, current fluctuation amplitude, and temperature rise amplitude, enabling instability and equipment overload to be identified within the same risk assessment framework.

[0156] The time required for a single risk assessment was reduced from 4.2 minutes to 2.6 minutes. This is mainly because the present invention directly outputs the comprehensive risk assessment value, emergency drainage risk level, dominant risk type, and probability of various risks, reducing the time spent by on-site personnel on repeatedly comparing multiple monitoring data and making manual inferences. Overall, the data in Table 1 demonstrates that the present invention has achieved a stable improvement in the efficiency of risk level judgment, dominant risk identification, single risk identification, and on-site assessment. The reasons for this improvement are mainly reflected in the drainage capacity benchmark reference, risk trigger residual quantification, risk correlation feature map expression, and the improved spatiotemporal graph neural controlled differential model continuous-time joint encoding.

[0157] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based emergency drainage risk assessment method, characterized in that, Includes the following steps: Collect multi-source status data during emergency drainage operations, preprocess the multi-source status data, and generate a synchronized status sequence; Based on the synchronous state sequence matching of single pump performance curves, series performance curves and parallel performance curves, a drainage capacity benchmark sequence is generated. Based on the synchronous state sequence, disaster load characteristics, water inlet blockage characteristics, pump set operating condition deviation characteristics, and equipment health characteristics are extracted to construct a risk characteristic set; The deviation between the risk feature set and the drainage capacity benchmark sequence is calculated to generate a risk trigger residual sequence. Risk nodes are constructed and node associations are calculated based on risk feature sets and risk trigger residual sequences to build a risk association feature map. The risk association feature map is input into the improved spatiotemporal graph neural controlled differential model. The node state, edge weight and time interval in the risk association feature map are jointly encoded in continuous time to generate the probability of disaster expansion risk, water ingress blockage risk, pump set instability risk, equipment overload risk and insufficient drainage capacity risk. The probabilities of various risks are integrated to generate a comprehensive risk assessment value. Based on the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the risk assessment result are output.

2. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The multi-source status data includes rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate and target drainage head in the emergency drainage operation area; pump operation mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed and power; inlet image, bar screen image, pump body vibration, motor current and motor temperature; the preprocessing includes data acquisition time calibration, abnormal sampling point removal, missing sampling point completion, duplicate data merging and data dimension unification.

3. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The generation of the drainage capacity benchmark sequence includes: Read the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, power, target drainage flow rate, and target drainage head from the synchronization status sequence; The operating mode of the pump set determines the single pump operating state, series operating state, or parallel operating state. The single pump operating state matches the single pump performance curve, the series operating state matches the series performance curve, and the parallel operating state matches the parallel performance curve. Based on inlet pressure, outlet pressure, drainage flow rate, engine speed and power, locate the current operating point in the corresponding performance curve; Based on the target drainage flow rate and target drainage head, locate the target operating point in the corresponding performance curve; Calculate the flow rate deviation, head deviation, and power deviation between the current operating point and the target operating point, and combine them with the valve opening to generate a baseline value for drainage capacity at a single moment; The drainage capacity benchmark values ​​at each single moment are arranged in chronological order according to the synchronization state sequence to generate the drainage capacity benchmark sequence.

4. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The construction of the risk feature set includes: Rainfall intensity, cumulative rainfall, water level, water level rise rate, target drainage flow rate, and target drainage head are extracted from the synchronous state sequence, and disaster load characteristics are generated according to the time window. Extract inlet images and grid area images from the synchronous state sequence, and perform feature processing on debris coverage area, debris accumulation density, debris distribution location and grid occlusion range to generate inlet blockage features; Calculate the flow rate deviation, head deviation, power deviation, and efficiency deviation between the current operating point and the target operating point. Combine the flow rate deviation, head deviation, power deviation, and efficiency deviation with the pump set operating mode, valve opening, inlet pressure, outlet pressure, drainage flow rate, engine speed, and power according to the time window to generate the pump set operating condition deviation characteristics. Pump body vibration, motor current and motor temperature are extracted from the synchronous state sequence. The vibration change amplitude, current fluctuation amplitude and temperature rise amplitude are processed to generate equipment health characteristics. Based on the time window correspondence, the characteristics of disaster load, water inlet blockage, pump set operating condition deviation, and equipment health are combined to construct a risk characteristic set.

5. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The generation of the risk-triggered residual sequence includes: According to the time window correspondence, the various risk characteristics in the risk characteristic set are matched with the single-moment drainage capacity benchmark values ​​in the drainage capacity benchmark sequence; Based on the characteristics of the disaster load, such as water level, water level rise rate and drainage target flow, a drainage demand index value is generated. The difference between the drainage demand index value and the single-moment drainage capacity benchmark value is calculated to generate the load trigger residual. Based on the debris coverage area, debris accumulation density and grid obstruction range in the characteristics of water inlet blockage, the water inlet attenuation ratio is generated. The baseline value of drainage capacity at a single moment is multiplied by the water inlet attenuation ratio to generate the drainage capacity attenuation value caused by blockage, and the blockage trigger residual is determined. Based on the flow deviation, head deviation, power deviation and efficiency deviation values ​​in the pump set operating condition deviation characteristics, the operating condition deviation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the operating condition deviation ratio to generate the drainage capacity attenuation value caused by the operating condition deviation, and the operating condition trigger residual is determined. Based on the vibration change amplitude, current fluctuation amplitude, and temperature rise amplitude in the equipment health characteristics, the equipment attenuation ratio is generated. The single-moment drainage capacity benchmark value is multiplied by the equipment attenuation ratio to generate the drainage capacity attenuation value caused by the equipment state, and the health trigger residual is determined. By combining load-triggered residuals, congestion-triggered residuals, operating condition-triggered residuals, and health-triggered residuals in chronological order, a risk-triggered residual sequence is generated.

6. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The construction of the risk association feature map includes: The disaster load characteristics and load trigger residuals are combined to form the disaster load node state; the water inlet blockage characteristics and blockage trigger residuals are combined to form the water inlet blockage node state; the pump set operating condition deviation characteristics and operating condition trigger residuals are combined to form the pump set operating condition deviation node state; and the equipment health characteristics and health trigger residuals are combined to form the equipment health node state. Based on the degree of synchronization of numerical changes, the degree of similarity of residual strength, and the degree of influence of drainage capacity among disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, and equipment health node status, the node correlation value is calculated. Establish risk transmission edges between nodes whose node association values ​​satisfy the node connection conditions, and write the node association values ​​into the corresponding risk transmission edges to form edge weights; By combining the disaster load node status, water inlet blockage node status, pump set operating condition deviation node status, equipment health node status, risk transmission edge and edge weight in the order of time window, a risk association feature graph is constructed.

7. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The generation of the risk probability includes: The risk-related feature map is input into the improved spatiotemporal graph neural controlled differential model, which includes a graph embedding layer, a time-controlled path generation layer, an edge weight graph propagation layer, a controlled differential evolution layer, and a multi-task risk prediction layer. In the graph embedding layer, the node states in the risk association feature graph are mapped to the initial hidden states of the nodes, and the edge weights are organized into an edge weight matrix. In the time-controlled path generation layer, the initial hidden state of the node is continuously interpolated based on the time interval between adjacent time windows to generate the node state control path. In the edge-weighted graph propagation layer, the initial hidden state of the node is propagated using the edge weight matrix with neighborhood weighting to generate the graph propagation hidden state. In the controlled differential evolution layer, the hidden state derivative is calculated based on the node state control path and the graph propagation hidden state, and the hidden state is updated along the time interval to generate the risk evolution hidden state. In the multi-task risk prediction layer, the hidden states of risk evolution are mapped to the probability of disaster expansion risk, the probability of water blockage risk, the probability of pump set instability risk, the probability of equipment overload risk, and the probability of insufficient drainage capacity risk.

8. The emergency drainage risk assessment method based on machine learning according to claim 1, characterized in that, The generation of the emergency drainage risk level, dominant risk type, and risk assessment results includes: The probability of disaster spread, water blockage, pump instability, equipment overload, and insufficient drainage capacity are normalized to generate a set of normalized risk probabilities. The corresponding fusion ratio is determined based on the magnitude of each risk probability in the normalized risk probability set; The normalized risk probabilities are weighted and summed according to the integration ratio to generate a comprehensive risk assessment value. The emergency drainage risk level is determined based on the comprehensive risk assessment value, the maximum risk probability, and the number of risk probabilities higher than the comprehensive risk assessment value. The risk type corresponding to the risk probability with the largest value in the normalized risk probability set is determined as the dominant risk type. The risk assessment result is generated by combining the comprehensive risk assessment value, the emergency drainage risk level, the dominant risk type, and the probability of each risk.

9. A machine learning-based emergency drainage risk assessment system, comprising executing the machine learning-based emergency drainage risk assessment method according to any one of claims 1 to 8, characterized in that, include: The synchronization processing module is used to collect and process multi-source status data during emergency drainage operations and generate a synchronization status sequence. The benchmark generation module is used to generate a drainage capacity benchmark sequence by matching the performance curves of a single pump, series performance curves, and parallel performance curves based on the synchronous state sequence. The feature construction module is used to construct a set of risk features based on the synchronization state sequence; The residual generation module is used to calculate the deviation between the risk feature set and the drainage capacity benchmark sequence to generate a risk trigger residual sequence. The graph construction module is used to construct a risk association feature graph based on the risk feature set and the risk trigger residual sequence. The risk prediction module is used to input the risk association feature map into the improved spatiotemporal graph neural controlled differential model to generate the probability of each risk. The assessment output module is used to integrate the probabilities of various risks, generate a comprehensive risk assessment value, and output the emergency drainage risk level, dominant risk type, and risk assessment results.

Images