AI-based multi-mode water level monitoring and well lid alarming method

Abstract

The invention discloses a multi-modal water level monitoring and well lid alarming method based on AI, and the method comprises the steps: arranging a water immersion and triggering alarm terminal of an integrated electrode sensor and a Hall sensor in an open bin of a well lid base, integrating terminal data, image data and environment data, and carrying out the standardization processing, thereby generating a multi-modal fusion feature. And by taking the characteristics as core input, constructing an AI model, and outputting an accumulated water abnormal probability and a well lid jacking probability. In combination with the pipe network topology, historical data and a dynamic threshold value of a digital twin system, the road surface ponding grade and the well lid abnormity reason are accurately judged, and the alarm effectiveness is verified through multi-source credibility fusion. Meanwhile, an alarm result and disposal feedback are synchronized to the digital twin system, and a training data set is updated to iteratively optimize model parameters. According to the method, the hysteresis quality and the false alarm rate of traditional monitoring are effectively reduced, the operation and maintenance efficiency and the emergency response capability of a municipal pipe network are remarkably improved, and reliable technical support is provided for urban inland inundation prevention and control and road safety guarantee.

Description

Technical Field

[0001] This invention relates to the field of pipeline network monitoring information, and in particular to a method for AI-based multimodal water level monitoring and manhole cover alarm. Background Technology

[0002] In the process of deepening new urbanization and digital transformation of cities, manhole covers, as key nodes of urban infrastructure networks, are widely distributed at the opening ends of various pipe wells such as power, telecommunications, gas, water supply and drainage. Real-time monitoring of their own status (such as whether they are abnormally opened or closed, or whether they are under pressure from inside the pipe) and the status of the surrounding environment (such as the depth of water accumulation on the road and the rate of water accumulation) is directly related to the safety of urban pipe network operation and maintenance, road traffic safety and the efficiency of urban flood disaster prevention and control. It is one of the core links to ensure the stable operation of the city's lifeline system.

[0003] However, current mainstream manhole cover and surrounding area condition monitoring solutions still have many technical bottlenecks and application limitations, making it difficult to adapt to the large-scale, high-precision, and low-cost urban-level monitoring needs.

[0004] Firstly, in terms of manhole cover condition monitoring, traditional solutions often employ a deployment method of "independent terminal + strong adhesive bonding / drilling and locking," whereby the monitoring terminal is fixed to the underside of the manhole cover using high-strength adhesive bonding or drilling. These terminals typically use a three-axis gyroscope as the core sensing element to detect the displacement and tilt of the manhole cover. However, because three-axis gyroscopes are extremely sensitive to environmental interference such as vibration and impact, in scenarios with heavy traffic and frequent road bumps in urban areas, they are prone to triggering false alarms due to instantaneous vibrations caused by vehicles running over them. This results in "ineffective inspections" by maintenance personnel, wasting manpower and reducing the efficiency of responding to real risks.

[0005] Secondly, in terms of monitoring road surface water and underground water levels, existing solutions rely on high-precision hydrostatic sensors for data acquisition. While these sensors offer high measurement accuracy, they have significant drawbacks when deployed on a large scale in scenarios such as storm and sewage pipe wells on urban main roads: From a cost perspective, the hardware procurement and calibration costs of a single hydrostatic sensor are high, making it difficult to promote their widespread use in the vast number of urban pipe wells; from an environmental adaptability perspective, hydrostatic sensors have stringent requirements for their working environment, and long-term exposure to harsh conditions such as rainwater immersion, siltation, high temperature and humidity can easily lead to problems such as sensor element aging and measurement accuracy degradation; from a maintenance difficulty perspective, main roads have dense traffic flow, and replacing and maintaining faulty sensors requires temporary lane closures, which not only affects traffic flow but also poses safety risks to maintenance personnel; in addition, the measurement results of hydrostatic sensors are easily affected by changes in air pressure and temperature fluctuations, especially during heavy rain, where sudden changes in air pressure often lead to distorted water level data, making it impossible to provide a reliable basis for urban flooding early warning.

[0006] In summary, traditional monitoring solutions have significant shortcomings in meeting the three core requirements of "accuracy, stability, and economy." They are unable to efficiently identify manhole cover anomalies and road surface water accumulation, and they struggle to support the refined and intelligent development of urban pipeline network operation and maintenance. Therefore, developing a new monitoring method that can integrate multi-source data, adapt to complex urban environments, and balance cost and accuracy has become a critical issue that urgently needs to be addressed in the field of municipal pipeline network monitoring. Summary of the Invention

[0007] This invention aims to provide a method for accurately determining road surface flooding and abnormal manhole cover opening by relying on existing topographic GIS data and pipeline network data from a digital drainage unit twin system, integrating real-time data collected from drainage unit terminals and edge sides, and optimizing the data structure through storage, analysis, and training. It utilizes an electrode-triggered alarm mechanism and water level data, combined with Hall effect sensors. The goal is to achieve low-cost, stable, and easy-to-maintain intelligent terminals, ensuring accurate and reliable monitoring and judgment, controllable costs, and a complete process. It integrates multi-source data and sensor characteristics through AI multimodal technology to form a complete and efficient closed loop.

[0008] This invention provides a method for AI-based multimodal water level monitoring and manhole cover alarm, comprising: S1. Deploy the terminal and collect real-time data, then integrate the real-time data into a dataset; S2. Standardize the dataset to generate multimodal fusion features; S3. Training a multimodal AI model based on multimodal fusion features; S4. Standardize the real-time multi-source data and input it into the trained multimodal AI model to output the anomaly probability; S5. Determine the cause of the anomaly based on the anomaly probability and trigger a tiered alarm. S6. Optimize the multimodal AI model based on the alarm results feedback.

[0009] In one embodiment of the present invention, step S1 includes: A water immersion and trigger alarm terminal is built into the open compartment of the manhole cover base. The terminal integrates an electrode sensor and a Hall sensor. The electrode sensor outputs a simulated water level h, and the Hall sensor detects the distance Δs between the well cover and the base. If Δs > 3cm, the Hall sensor outputs an opening / closing state signal S = 1; otherwise, it outputs S = 0. The h and S values ​​are uploaded synchronously to the edge node, and the digital twin system is called to associate the location coordinates of the manhole cover with the topology data of the pipeline network to which it belongs; A dataset is generated by integrating real-time terminal data, environmental data, digital twin system data, and image monitoring data.

[0010] In one embodiment of the present invention, step S2 includes: Standardize the dataset; The standardized multi-source data is unified into a time series with a sampling interval of 10 seconds. Relying on the timestamps and spatial coordinates of the digital twin system, the spatiotemporal consistency of terminal data, image data and environmental data is ensured. Cross-modal weights are calculated based on an attention mechanism, and the priority of multi-source data is adjusted according to data credibility. When the data confidence level changes, the sensor weight is increased by using the formula Ws=α×Cs+(1-α)×(1-CI), where α is the adjustment coefficient, Cs is the sensor confidence level, and CI is the image confidence level. When in a rain-affected environment, the sensor weights are enhanced using the formula Ws=0.7×(1-R')+0.3×h', at which point the image weight WI=1-Ws; Where h' = h / hmax, hmax is the highest historical water level stored in the digital twin system, R' = R / Rmax, Rmax is the extreme rainfall value stored in the digital twin system, and R is the rainfall amount; Multimodal fusion features are generated by weighted summation, using the following formula: F = Ws × Fs + WI × FI; Where Fs=[h',Δs',P'], h', Δs', and P' are the standardized water level, the distance between the manhole cover and the base, and the pressure; FI is the image feature vector.

[0011] In one embodiment of the present invention, step S3 includes: Based on the pipeline topology data of the digital twin system, the distance D between the manhole cover location and the main pipeline is calculated, and a spatial correlation model is established by combining the pressure P inside the pipe. A time-series correlation parameter, the rainfall lag coefficient τ, is introduced to learn the dynamic relationship between rainfall and water level: τ = Δh / (Δt × tr); Where tr represents the duration of rainfall, and Δh / Δt represents the rise in water level.

[0012] In one embodiment of the present invention, the anomaly probability in step S4 includes: The probability of abnormal water accumulation is Pw = β × (Δh / Δt / H0) + (1-β) × (h / hmax); The probability of a manhole cover tipping over is Pj = γ × (P / P0) + (1-γ) × (Δs / smax); Where H0 is the dynamic threshold, smax is the maximum allowable displacement, P0 is the safe pressure of the pipeline network, and β and γ are empirical coefficients.

[0013] In one embodiment of the present invention, in step S5: If Pj > 0.6 and P > P0, it is determined that the water level pressure inside the pipe is causing the problem, and the cause label T = 1 is output; otherwise, it is determined that the problem is caused by external force intervention, and the cause label T = 0 is output. When S=1 and P>P0, the probability of the manhole cover being pushed up is Q=0.8×(P / P0)+0.2×d / 10, where d is the real-time distance between the manhole cover and the base; When Q≥0.7, it is determined that the abnormal opening is caused by the top support.

[0014] In one embodiment of the present invention, step S5 includes triggering a graded alarm for road surface water accumulation: Let the electrode spacing be d. When the water level h ≥ n × d, an n-level alarm is triggered, where n = 1, 2, 3, and the alarm level L = ceil (h / d). When h≥5cm and Pw≥0.5, the output signal A1=1, triggering a first-level warning and pushing the water accumulation rate to the management platform; When h≥10cm and Pw≥0.7, the output signal A2=1, triggering a level 2 warning and linking with surrounding warning signs; When h≥15cm and Pw≥0.8, the output signal A3=1, triggering a level 3 alarm and sending a drainage dispatch command to the pumping station.

[0015] In one embodiment of the present invention, step S5 includes a graded alarm for manhole cover top support: Calculate the top support coefficient K=P 实时 / P 安全 P 安全 The design pressure for manhole covers to resist jacking is stored for the digital twin system; When K≥0.8, d>3mm and Q≥0.6, the output signal B1=1, triggering a pre-alarm and pushing a reinforcement reminder to the inspection terminal; When K≥1.0, d>5mm and Q≥0.7, the output signal B2=1, triggering an emergency alarm, activating the terminal buzzer and marking the coordinates of the danger zone; The alarm is confirmed to be valid if the historical pressure curve of the pipeline network in the associated digital twin system is obtained and the conditions are met for three consecutive samplings.

[0016] In one embodiment of the present invention, it further includes: When the rainfall reaches the rainstorm standard, the alarm threshold is corrected by the indicator function I; At this point, the water level threshold correction value is H' = H × (1 - 0.2 × I (R ≥ 50)); The threshold correction value for the top support coefficient is K'=K×(1-0.1×I(R≥50)); Where I is an indicator function, which takes the value 1 if the condition is met, and 0 otherwise.

[0017] In one embodiment of the present invention, it further includes: The alarm confidence level is calculated using the following formula: C = 0.6 × sensor data confidence level + 0.4 × image / pipeline correlation; The confidence level of the sensor data is based on the historical stability records of the sensors in the digital twin system; Image / pipeline correlation is the degree of overlap between the water accumulation area in the image and the pipeline coverage area of ​​the digital twin system; The alarm is confirmed to be valid when C ≥ 0.7.

[0018] The present invention has the following beneficial effects: (1) This invention collects data in real time by using built-in dual sensors (electrode-graded water level measurement and Hall effect displacement monitoring) and associates it with digital twin system information. It also dynamically allocates weights by combining the attention mechanism of multimodal data fusion to achieve accurate alarm based on graded thresholds and state judgment, thus solving the problem of heterogeneous data fusion.

[0019] (2) This invention integrates digital twin pipeline network data with real-time monitoring by multiple sensors to construct a spatiotemporal correlation model. Combined with multimodal technology, it achieves accurate quantification, graded early warning, and dynamic response to the risks of water accumulation and manhole cover bulging. The lightweight model ensures real-time edge computing, and federated learning enhances cross-regional generalization capabilities, effectively improving the accuracy and timeliness of risk assessment, reducing the risk of urban flooding and safety accidents, and optimizing the efficiency of pipeline network operation and maintenance and emergency response.

[0020] (3) This invention integrates terminal monitoring, edge environment and digital twin pipeline / historical data, combined with dynamic threshold adjustment, multimodal model cross-validation and standardization processing, and relies on the characteristics of digital twin system to accurately calculate the probability of water accumulation and manhole cover anomalies, effectively improving the accuracy of road water accumulation judgment and manhole cover opening and closing status identification, and providing a reliable basis for graded early warning. Attached Figure Description

[0021] Figure 1 A flowchart of a method for AI-based multimodal water level monitoring and manhole cover alarm is shown in one embodiment of the present invention. Detailed Implementation

[0022] In the following description, the invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be practiced without one or more specific details or with other alternatives and / or additional methods, materials, or components. In other instances, well-known structures, materials, or operations are not shown or described in detail so as not to obscure the inventive points of the invention. Similarly, for illustrative purposes, specific quantities, materials, and configurations are set forth to provide a comprehensive understanding of embodiments of the invention. However, the invention is not limited to these specific details.

[0023] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0024] In this specification, references to "an embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. The phrase "in one embodiment" appearing throughout this specification does not necessarily refer to the same embodiment in all instances.

[0025] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] Figure 1 A flowchart of a method for AI-based multimodal water level monitoring and manhole cover alarm is shown in one embodiment of the present invention.

[0028] like Figure 1 As shown, in one embodiment of the present invention, the method for AI-based multimodal water level monitoring and manhole cover alarm includes: S1. Deploy the terminal and collect real-time data, integrating the real-time data into a dataset, including: The terminal hardware deployment and real-time data acquisition are housed within the open compartment of the adjustable five-proof ball iron manhole cover base. This compartment integrates an electrode sensor for graded water level monitoring and a Hall sensor for monitoring the relative displacement between the manhole cover and base, triggering a water immersion alarm terminal. It is equipped with a nylon plastic structure to eliminate communication signal shielding, magnetic contacts, and an external charging module, establishing a data transmission architecture of "terminal + NB-IoT / 4G + digital twin system". The electrode sensor outputs an analog water level signal h; the Hall sensor detects the distance Δs between the manhole cover and base. If Δs > 3cm, it outputs an open / closed status signal S=1; otherwise, it outputs S=0.

[0029] The h and S values ​​are uploaded synchronously to the edge node, and the basic data of the digital twin system is called to associate the location coordinates (x, y) of the manhole cover and the topology data of the pipeline network to which it belongs (such as the distance D from the main road), so as to complete the binding of real-time data with the spatiotemporal information of the digital twin.

[0030] Multi-source data integration: Four types of core data are integrated at the edge to form the basic dataset for model training, including: Terminal data: h (water level), S (manhole cover status), Δs (displacement); Environmental data: Rainfall R (mm / h), pipe pressure P (kPa); Digital twin system data: historical water level h0, historical alarm cases, pipeline topology distance D, manhole cover anti-backflow design pressure P (kPa); and Image data: Visual features of water accumulation I (area of ​​water accumulation, edge contour).

[0031] S2. Standardize the dataset to generate multimodal fusion features, including: Data standardization eliminates differences in units, providing a unified format for model input: The water level normalization is h' = h / hmax, where hmax is the highest historical water level stored in the digital twin system. In this embodiment, hmax = 20cm, so h' = h / 20. The rainfall is standardized as R' = R / Rmax, where Rmax is the extreme rainfall value stored in the digital twin system. In this embodiment, Rmax = 100 mm / h, so R' = R / 100. Parameters such as pressure and displacement are synchronized and standardized according to the safety thresholds in the digital twin system, such as P'=P / P0, where P0 is the safe pressure of the pipeline network.

[0032] The standardized multi-source data is unified into a time series with a sampling interval of 10 seconds. Relying on the timestamps and spatial coordinates of the digital twin system, the spatiotemporal consistency of terminal data, image data, and environmental data is ensured.

[0033] Cross-modal weights are calculated based on an attention mechanism, and priorities are adjusted according to data reliability. When the confidence level of the data changes, such as when the confidence level CI of the nighttime image is less than 0.7, the threshold is referenced from the historical nighttime data of the digital twin system, and the sensor weight is increased by the formula: Ws=α×Cs+(1-α)×(1-CI), α=0.6-0.8, and Cs is the sensor confidence level; In scenarios with rainfall interference, such as when R≥50mm / h, the sensor weights are further enhanced by the formula Ws=0.7×(1-R')+0.3×h', and the image weight WI=1-Ws.

[0034] The multimodal fusion feature F is generated by weighted summation, and the formula is F=Ws×Fs+WI×FI; Where Fs is the sensor feature vector [h', Δs', P']; FI is the image feature vector [water accumulation area, edge contour matching degree], where the edge contour matching degree is the comparison result between the image and the digital twin road contour.

[0035] S3. Train a multimodal AI model based on multimodal fusion features. In this embodiment, a multimodal Transformer architecture is used to process the temporal and spatial dual-dimensional features of the data.

[0036] The temporal features are learned by the dynamic relationship between R' and h', which is achieved by introducing the rainfall lag coefficient τ=Δh / (Δt×tr), where tr is the duration of rainfall and Δh / Δt is the rise in water level. The spatial characteristics are related to the distance D in the learning pipeline topology and the pressure P in the pipeline. The closer the manhole cover is to the main road, the faster the pressure is transmitted.

[0037] By combining knowledge distillation and reinforcement learning techniques, complex models are compressed into lightweight versions adapted for edge computing, ensuring real-time inference latency of ≤50ms. Furthermore, a federated learning framework is used to jointly train models using cross-regional data (different cities, different network types) from digital twin systems, avoiding model bias caused by data from a single region.

[0038] Finally, the model parameters are fine-tuned periodically using newly collected terminal data and digital twin historical data to adapt to seasonal changes and pipeline renovations.

[0039] S4. Standardize the real-time multi-source data and input it into the trained multimodal AI model, outputting the anomaly probability, which includes: The probability of abnormal water accumulation is Pw = β × (Δh / Δt / H0) + (1-β) × (h / hmax); The probability of a manhole cover tipping over is Pj = γ × (P / P0) + (1-γ) × (Δs / smax); Where H0 is the dynamic threshold, smax is the maximum allowable displacement, P0 is the safe pressure of the pipeline network, and β and γ are empirical coefficients.

[0040] S5. Determine the cause of the anomaly based on the anomaly probability and trigger a tiered alarm, including: If Pj > 0.6 and P > P0, it is determined that the water level pressure inside the pipe is causing the problem, and the cause label T = 1 is output; otherwise, it is determined that the problem is caused by external force intervention, and the cause label T = 0 is output. When S=1 and P>P0, the probability of the manhole cover being pushed up is Q=0.8×(P / P0)+0.2×d / 10, where d is the real-time distance between the manhole cover and the base; When Q≥0.7, it is determined that the abnormal opening is caused by the top support.

[0041] By calling upon scenario-based data from the digital twin system and associating it with the edge elevation of the manhole cover and the critical water level for passage on surrounding roads (sidewalks / vehicle lanes), a classification threshold can be determined. Let the electrode spacing be d. When the water level h ≥ n × d, an n-level alarm is triggered, where n = 1, 2, 3, and the alarm level L = ceil (h / d).

[0042] The triggering condition for a Level 1 warning is: H(t)≥5cmH(t) (i.e., h) and Pw≥0.5, output signal A1=1, push the water accumulation rate V_water accumulation=(H(t)-H(t-1)) / 1min to the management platform; The trigger conditions for a Level 2 warning are: H(t)≥10cm and Pw≥0.7, output signal A2=1 (synchronization A1=1), and linkage with surrounding traffic lights / warning signs; The trigger conditions for a Level 3 alarm are: H(t)≥15cm and Pw≥0.8, output signal A3=1 (synchronous A1=1, A2=1), and send drainage dispatch instructions to the pumping station. The location of the pumping station and its connection to the pipeline network are taken from the digital twin system.

[0043] The support coefficient is K=P 实时 / P 安全 P 安全 The design pressure for manhole cover resistance to top pressure is stored for the digital twin system.

[0044] When K≥0.8, d>3mm, and Q≥0.6, a pre-alarm is triggered, output signal B1=1, and a reinforcement reminder is pushed to the inspection terminal; When K≥1.0, D>5mm, and Q≥0.7, an emergency alarm is triggered, output signal B2=1 (synchronous B1=1), activate the terminal buzzer, and mark the coordinates of the danger zone. These coordinates are from the digital twin system.

[0045] The alarm is confirmed to be valid if the historical pressure curve of the pipeline network is correlated with the digital twin system and the conditions are met for three consecutive samplings, thus avoiding false alarms due to single data fluctuations.

[0046] S6. Optimize the multimodal AI model based on alarm results feedback, including: Dynamic threshold adjustment: When rainfall R ≥ 50 mm / h, the alarm threshold is corrected through an indicator function I, thereby improving the sensitivity of early warnings under extreme weather conditions. Water level threshold correction value H' = H × (1 - 0.2 × I (R≥50)); The threshold correction value for the top support coefficient K' = K × (1 - 0.1 × I (R≥50)); Where R is taken from environmental data, the threshold is referenced from the rainstorm standard of the digital twin system, which is 50 mm / h in this embodiment, and I is an indicator function, which is 1 if the condition is met, and 0 otherwise.

[0047] Alarm confidence verification involves calculating the alarm confidence level using a formula: C = 0.6 × sensor data confidence level + 0.4 × image / pipeline correlation; Where C∈[0,1], the sensor data confidence is based on the historical stability record of the sensor in the digital twin system, and the image / pipeline correlation is the overlap between the water accumulation area in the image and the pipeline coverage area in the digital twin system.

[0048] The alarm is confirmed to be valid when C≥0.7; otherwise, it is marked as "suspected false alarm" and requires manual verification.

[0049] Alarm results and on-site handling feedback (such as pump station scheduling effectiveness and inspection confirmation results) are synchronized to the digital twin system, updating the historical database and real-time pipeline network status. Based on the updated digital twin data, the multimodal AI model is retrained, and the dynamic weight coefficients (such as α and β), probability calculation parameters (such as 0.6 and 0.4), and threshold adjustment strategies are optimized.

[0050] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A method for AI-based multi-modal water level monitoring and manhole cover alarming, characterized in that, Comprise: S1, deploy terminal and collect real-time data, integrate real-time data into dataset; S2, standardize dataset, generate multi-modal fusion features; S3, train multi-modal AI model based on multi-modal fusion features; S4, input standardized real-time multi-source data into trained multi-modal AI model, output abnormal probability; S5, determine abnormal cause according to abnormal probability and trigger hierarchical alarm; S6, feedback optimization of multi-modal AI model according to alarm result.

2. The method of claim 1, wherein, The step S1 comprises: Water immersion and trigger alarm terminal are placed in the open bin of the manhole cover base, the terminal integrates electrode sensor and hall sensor; The electrode sensor outputs water level analog quantity h, and the hall sensor detects the distance Δs between the manhole cover and the base; If Δs>3cm, the hall sensor outputs open / close state signal S=1, otherwise S=0; h and S are synchronously uploaded to the edge node, and the position coordinates and the associated pipe network topology data of the manhole cover are called by the digital twin system; Integrate terminal real-time data, environmental data, digital twin system data and image monitoring data to generate dataset.

3. The method of claim 2, wherein, Step S2 comprises: Standardize dataset; Standardized multi-source data is unified into time series sequence with sampling interval of 10s, relying on the timestamp and spatial coordinates of the digital twin system to ensure the spatio-temporal consistency of terminal data, image data and environmental data; Based on attention mechanism, cross-modal weight is calculated, and multi-source data priority is adjusted according to data credibility; When data credibility changes, the sensor weight is improved through the formula Ws=α×Cs+(1-α)×(1-CI), where α is the adjustment coefficient, Cs is the sensor confidence, and CI is the image confidence; When in the rainfall interference scene, the sensor weight is strengthened through the formula Ws=0.7×(1-R')+0.3×h', and the image weight WI=1-Ws at this time; Where h'=h / hmax, hmax is the highest water level stored in the digital twin system, R'=R / Rmax, Rmax is the rainfall extreme value stored in the digital twin system, and R is the rainfall; Multi-modal fusion features are generated by weighted summation, and the formula is: F=Ws×Fs+WI×FI; Where Fs=[h', Δs', P'], h', Δs' and P' are the standardized water level, manhole cover and base distance and pressure; FI is the image feature vector.

4. The method of claim 3, wherein, Step S3 comprises: Based on the pipe network topology data of the digital twin system, the distance D between the manhole cover position and the pipe network trunk is calculated, and the spatial correlation model is established combined with the pipe pressure P; Rainfall lag coefficient τ is introduced as a time sequence correlation parameter to learn the dynamic correlation between rainfall and water level: τ=Δh / (Δt×tr); Where tr is the rainfall duration, and Δh / Δt is the water level rise.

5. The method of claim 4, wherein, In step S4, the abnormal probability comprises: Water accumulation abnormal probability Pw=β×(Δh / Δt / H0)+(1-β)×(h / hmax); Manhole cover jacking probability Pj=γ×(P / P0)+(1-γ)×(Δs / smax); Wherein H0 is a dynamic threshold, smax is the maximum allowable displacement, P0 is the safety pressure of the pipe network, β and γ are empirical coefficients.

6. The method of claim 5, wherein, In step S5: If Pj>0.6 and P>P0, it is determined that the water level pressure in the pipe is jacked, and the cause label T=1 is output; otherwise, it is determined that external force intervention occurs, and the cause label T=0 is output; When S=1 and P>P0, the manhole cover jacking cause probability Q=0.8×(P / P0)+0.2×d / 10, wherein d is the actual distance between the manhole cover and the base; When Q≥0.7, it is judged that jacking causes abnormal opening.

7. The method of claim 6, wherein, Step S5 includes road surface water grading alarm triggering: Let the electrode spacing be d, and when the water level h≥n×d, n-level alarm is triggered, wherein n=1, 2, 3, and the alarm level L=ceil(h / d); When h≥5cm and Pw≥0.5, output signal A1=1, triggering a first-level early warning, and pushing the water accumulation speed to the management platform; When h≥10cm and Pw≥0.7, output signal A2=1, triggering a second-level early warning and linking to the surrounding warning signs; When h≥15cm and Pw≥0.8, output signal A3=1, triggering a third-level alarm, and sending a drainage scheduling instruction to the pump station.

8. The method of claim 6, wherein, Step S5 includes manhole cover jacking grading alarm: The jacking coefficient K = P 实时 / P 安全 , P 安全 is the design pressure of the manhole cover against jacking stored by the digital twin system; When K≥0.8, d>3mm and Q≥0.6, output signal B1=1, triggering a pre-alarm, and pushing a reinforcement prompt to the inspection terminal; When K≥1.0, d>5mm and Q≥0.7, output signal B2=1, triggering an emergency alarm, activating the terminal buzzer and marking the coordinates of the dangerous area; Correlate the pipe network pressure history curve of the digital twin system, and confirm the alarm effective when the condition is met for 3 consecutive times.

9. The method of claim 8, wherein, Also includes: When the rainfall reaches the heavy rain standard, the alarm threshold is corrected by the indicator function I; At this time, the water level threshold correction value is H'=H×(1-0.2×I(R≥50)); The jacking coefficient threshold correction value is K'=K×(1-0.1×I(R≥50)); Wherein I is an indicator function, taking 1 if the condition is met, otherwise taking 0.

10. The method of claim 1, wherein, Also includes: The alarm confidence is calculated by the following formula: C=0.6×sensor data confidence+0.4×image / pipeline correlation; Wherein the sensor data confidence is based on the historical stability record of the sensor in the digital twin system; The image / pipeline correlation is the overlap degree of the image water accumulation area and the digital twin system pipe network coverage area; When C≥0.7, it is confirmed that the alarm is effective.

Images