Multi-parameter fusion pipe network overflow prediction system and method

Abstract

The invention discloses a multi-parameter fusion pipe network overflow prediction system and method, relates to the technical field of pipe network overflow prediction, and aims at overflow risks caused by equipment faults, unstable communication and extreme weather in a combined drainage system. A system scheme integrating five steps of predictive maintenance, reliable real-time control, digital twinning check, layered emergency plan and cloud edge cooperation is provided; in the first step, equipment health is monitored and faults are predicted to ensure that key pump stations and gates are available; 2, generating a fault-tolerant scheduling instruction by adopting reliable optimization and multi-agent reinforcement learning; 3, pre-judging and correcting the deviation by utilizing digital twinning check before execution; 4, dealing with a single-point fault and communication paralysis by using a layered distributed emergency mechanism; 5, cross-regional parallel simulation and self-adaptive scheduling are realized through cloud edge collaboration and federated learning; overflow can be reduced, environmental pollution and waterlogging loss under the rainstorm condition can be reduced, and efficient collaborative management of the drainage system is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of pipe network overflow prediction, and in particular to a pipe network overflow prediction system and method based on multi-parameter fusion. Background Art

[0002] Driven by the dual impacts of modern urbanization and frequent extreme weather events, large-scale urban combined sewer networks are frequently impacted by high-intensity, short-duration rainfall events. This causes a rapid mixing of rainwater and sewage within a region and leads to sudden surges in network load. Firstly, after a downpour, the flow distribution in each sub-area can drastically change, making it easy for local sections of the pipeline to overflow and for storage reservoirs to run out of capacity. Secondly, if downstream pumping stations and sluice gates cannot be started, stopped, or switched promptly, water can accumulate upstream and even backflow. These issues not only cause sewage overflows to pollute water bodies and streets, but also lead to urban flooding, impacting public transportation and infrastructure safety. To mitigate these risks, relevant departments typically utilize intelligent sensors and real-time control technologies to schedule pumping station start-up and stop-down times and sluice gate openings. However, with the continuous expansion of urban areas and the intensification of climate change, traditional network control methods based on manual or simple automated rules are no longer able to cope with the uncertainties brought about by sudden increases in rainfall intensity, unstable communications, and failures of critical equipment. A more reliable, systematic solution is urgently needed to ensure the coordinated operation of drainage infrastructure under high loads and effectively prevent overflow incidents.

[0003] A search revealed a Chinese invention patent application with publication number CN119416988A, which provides a real-time optimization control method, device, equipment, and medium for a drainage system, relating to the technical field of drainage systems. The method includes: obtaining historical data on the drainage system, including historical status data of the drainage system, historical rainfall data in the area where the drainage system is located, and historical control data of multiple control devices within the drainage system; determining the overflow volume of the drainage system under various scenarios based on the historical control data, historical status data, historical rainfall data, and multiple future random scenarios; determining candidate control strategies for multiple control facilities in the drainage system based on the overflow volumes of the various scenarios and the corresponding probability of occurrence of each scenario; and optimizing the candidate control strategies to obtain an optimized target control strategy. This method improves the robustness of the real-time control strategy and effectively alleviates the phenomenon of urban waterlogging and a sharp increase in outlet overflow volume when drainage facilities suddenly fail.

[0004] Combining the above practical application scenarios and existing technologies:

[0005] The most prominent technical problem in current practical applications is the insufficient reliability and unexpected failure of the execution equipment (such as pumping stations, gates, etc.), which will cause the overall scheduling of the drainage system to become ineffective at critical moments.

[0006] Specifically, pumping stations may suddenly shut down due to overload, electromechanical failure, or unstable power supply. Gates may become stuck with foreign objects or severely worn, preventing them from operating as instructed, thus affecting the safe discharge of water from the pipe network. Furthermore, if network delays or communication interruptions occur during peak rainstorms, remote control commands, even if issued, may not reach the execution end, rendering emergency dispatch measures ineffective. When these failures are compounded by unstable communications, water levels upstream of the drainage network can rise rapidly, causing road flooding or sewage overflows. Downstream or surrounding water bodies bear a greater burden, leading to deteriorating water quality and increased flood pressure. Failure to promptly activate backup pumping stations or switch to diversion strategies can result in not only environmental pollution and traffic paralysis, but also increased difficulty for water management departments in emergency response. In severe cases, this can cause significant damage to municipal infrastructure and the lives and property of residents. Summary of the Invention

[0007] (1) Technical problems solved

[0008] In response to the shortcomings of the existing technology, the present invention provides a multi-parameter fusion pipeline overflow prediction system and method, by proposing the integration of predictive maintenance, reliable real-time control, digital twin verification, layered emergency plans and cloud-edge collaboration; the first step is to monitor equipment health and predict failures to ensure the availability of key pump stations and gates; the second step is to use reliable optimization and multi-agent reinforcement learning to generate fault-tolerant scheduling instructions; the third step is to use digital twin verification to predict and correct deviations before execution; the fourth step is to use a layered distributed emergency mechanism to deal with single point failures and communication paralysis; the fifth step is to achieve cross-regional parallel simulation and adaptive scheduling through cloud-edge collaboration and federated learning; it can reduce overflow and reduce environmental pollution and waterlogging losses under heavy rain conditions, and realize efficient collaborative management of drainage systems; it solves the technical problems recorded in the background technology.

[0009] (2) Technical solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0011] A multi-parameter fusion pipeline overflow prediction method, including:

[0012] When the operating parameters of the execution equipment are collected and the fault risk threshold is determined to be triggered, the machine learning algorithm is used to calculate the health index and failure probability of the operating parameters and output a predictive maintenance warning;

[0013] After receiving the health index and failure probability and detecting changes in pipeline network monitoring data, it initiates reliable optimization and multi-agent reinforcement learning to iteratively solve the gate opening and pump station start and stop vectors. In the event of a failure, it automatically triggers the fault tolerance mechanism and outputs updated optimal control instructions.

[0014] The digital twin platform receives the optimal control vector and pipe network monitoring data, and combines it with the equipment health index to quickly simulate and evaluate the deviation of the hydraulic process. If the deviation exceeds the corresponding threshold, it will be fed back to the control layer to trigger recalculation or fine-tuning.

[0015] If the digital twin verification confirms that the optimal control vector is feasible and needs to be formally distributed to each execution device, the hierarchical distributed coordination mechanism monitors the communication quality and shares the health index and failure probability. If the main network or key equipment fails, the sub-area controller independently activates the emergency plan and dynamically triggers fault-tolerant recalculation;

[0016] When the scale of urban pipeline networks expands or cross-regional watersheds require collaborative management, the cloud-edge collaborative architecture is enabled to upload the local reinforcement learning parameters and simulation measurements of each sub-area, form a global strategy through federal aggregation, and send it to the edge nodes.

[0017] Preferably, a real-time sensor is deployed on the execution device, and multi-dimensional sensor data of each execution device in a time series is regularly collected through a distributed sensor network, and the original collected data is pre-processed to form a standardized data set;

[0018] Based on the standardized data set, a health index calculation model and a failure probability prediction model are introduced to quantify the health status of the equipment and output the failure probability according to the health index failure probability function.

[0019] Preferably, an uncertainty set is introduced to characterize the potential failure mode of the execution equipment, and a double-layer extreme value problem is solved and optimized under the control instruction vector. The weights or ranges of the equipment health index vector and the failure probability vector are dynamically adjusted in the worst-case analysis to obtain the optimal control vector.

[0020] Preferably, a multi-agent reinforcement learning strategy is adopted to configure an independent agent for each sub-area to be responsible for local scheduling and maintain interaction with the central controller; based on the observation vector, the agent's strategy function outputs the local control fine-tuning amount, which is used to locally correct the initial control instructions given by the central controller, and updates its strategy function through a distributed reinforcement learning algorithm; the observation vector includes local water level, flow and corresponding equipment health index, and fault probability information.

[0021] Preferably, based on the loaded multi-source data, the digital twin model simulates the dynamics of water flow, pressure, and potential overflow points in the pipe network, and embeds the opening and closing inertia, maximum opening and closing rate, and pumping station pumping power curve of the actuator equipment into the digital twin model;

[0022] In the digital twin model, a functional factor determined by the health index and failure probability of the equipment is introduced to adapt to the health attenuation effect of the execution equipment.

[0023] Preferably, after the digital twin completes the simulation, the simulation output state is compared with the desired scheduling target to form deviation information and quantified by a real-time deviation metric;

[0024] When certain components of the real-time deviation measurement exceed the preset threshold, the digital twin platform immediately generates risk warnings and deviation information, triggering the correction or recalculation of the control instruction vector. If it is also detected that a certain device has shown obvious hysteresis in this simulation, the fault tolerance mechanism or other emergency strategies will be activated.

[0025] Preferably, the urban drainage network is divided into several sub-areas, each sub-area is configured with a local sub-area controller, and each sub-area controller synchronously receives the final control instruction vector and the corresponding equipment health index or failure probability;

[0026] Monitor and dynamically record communication quality indicators. If the main communication channel is found to be interrupted or the delay exceeds the threshold, it will automatically switch to the backup channel, or the central controller will determine that the main communication is abnormal and enter the emergency coordination mode.

[0027] Preferably, when the backup failure probability reaches a high-risk range, or the communication quality is lower than a threshold, or the digital twin platform indicates that the overflow risk of the sub-area is too high and the central dispatch has not had time to update the instructions, the local independent emergency plan is triggered;

[0028] When a network interruption occurs or the central command cannot be issued in time, the sub-area controller switches to offline operation mode. If the communication is not completely interrupted, the adjacent sub-area controllers can exchange limited information through a temporary topology.

[0029] Preferably, when network communication is restored, the status data from the sub-areas are comprehensively analyzed, and the central controller performs global fault-tolerant recalculation to obtain an updated global instruction vector, which is broadcast to each sub-area via all available communication paths;

[0030] For local operations or collaborative operations with adjacent sub-areas that have been carried out in emergency situations, the central controller will use them as additional experience samples after communication is restored, incorporate them into the offline retraining or online learning process of multi-agent reinforcement learning, and continuously iterate and update the strategies of each sub-area agent.

[0031] Preferably, several multi-level partitions are formed based on the sub-areas, and the key data of each sub-area controller flows between the cloud and the edge side;

[0032] The cloud and edge nodes are configured to respectively undertake large-scale data aggregation and real-time control instruction execution, and evaluate the connection quality between the cloud and the edge in real time. If the network conditions are poor, the edge nodes automatically enable local scheduling and emergency functions, and synchronize updates with the cloud after communication is restored.

[0033] Preferably, a federated learning framework is used to regularly upload the model parameter sets distributed across edge nodes to the cloud for aggregation. The health index vectors, failure probability vectors, and digital twin simulation errors collected locally by each sub-area are transmitted to the cloud, and a distributed training algorithm is used to iteratively upgrade the global reinforcement learning model.

[0034] After receiving the updated global model parameters, the edge node performs local fine-tuning and combines local live online training to improve local control quality.

[0035] Preferably, a complex large-scale pipeline network is divided into several parallel simulation units. When the edge node completes the local simulation, the local simulation results and related boundary conditions are uploaded to the cloud for combined evaluation to form a global performance verification index.

[0036] Based on the fused global performance verification indicators and the control instructions returned by the sub-areas, the cross-regional global instructions are merged and output for the cross-regional scheduling plan at the large-scale watershed level. These instructions are distributed to the edge nodes of each sub-area and executed after fine-tuning with local reinforcement learning.

[0037] A multi-parameter fusion pipeline overflow prediction system, including:

[0038] The health assessment unit collects the operating parameters of the execution equipment and determines that the fault risk threshold is triggered. It then uses the machine learning algorithm to calculate the health index and failure probability of the operating parameters and outputs a predictive maintenance warning;

[0039] The scheduling and reinforcement learning unit receives the health index and failure probability and detects changes in pipeline network monitoring data. It then initiates reliable optimization and multi-agent reinforcement learning to iteratively solve the gate opening and pump station start and stop vectors. In the event of a failure, it automatically triggers the fault tolerance mechanism and outputs updated optimal control instructions.

[0040] The digital twin simulation verification unit receives the optimal control vector and pipe network monitoring data from the digital twin platform, and uses the equipment health index to quickly simulate and evaluate the deviation of the hydraulic process. If the deviation exceeds the corresponding threshold, it will be fed back to the control layer to trigger recalculation or fine-tuning;

[0041] Hierarchical distributed collaborative emergency unit: If the digital twin verification confirms that the optimal control vector is feasible and needs to be formally distributed to each execution device, the hierarchical distributed collaborative mechanism monitors the communication quality and shares the health index and failure probability. If the main network or key equipment fails, the sub-area controller independently activates the emergency plan and dynamically triggers fault-tolerant recalculation;

[0042] Cloud-edge collaboration and federated learning units. When the scale of urban pipeline networks expands or cross-regional watersheds require collaborative management, the cloud-edge collaboration architecture is enabled to upload the local reinforcement learning parameters and simulation measurements of each sub-area, form a global strategy through federated aggregation, and send it to the edge nodes.

[0043] (3) Beneficial effects

[0044] The present invention provides a multi-parameter fusion pipeline overflow prediction system and method, which has the following beneficial effects:

[0045] By accurately quantifying the health status and failure probability of executing equipment, and combining this with reliable control and multi-agent reinforcement learning, digital twin simulation, layered distributed emergency response, and large-scale cloud-edge collaboration, the following significant benefits are achieved in urban drainage overflow prevention and control:

[0046] Relying on predictive maintenance and detection of equipment health, maintenance or load reduction arrangements can be triggered at the budding stage of equipment failure, avoiding sudden failure of key pumping stations and gates during peak rainstorms; it can not only significantly enhance the reliability of scheduling plans, but also provide reliable optimal instructions. * (t) works together with the multi-agent reinforcement learning strategy in a dynamic and uncertain environment, taking into account both global optimization and local adaptation: when the health index H(t) decreases or the failure probability P(t) increases, reliable control will automatically increase the equipment redundancy coefficient, while the sub-area agents will quickly correct local operations based on experience, achieving a creative fusion of high fault tolerance and fast scheduling.

[0047] The rapid performance check of digital twin simulation provides prior verification and deviation detection, exposing execution layer delays or communication failures that may lead to overflow in the virtual environment in advance, and forming a closed loop with reliable control, so that efficient corrections can be completed before actual instructions are issued.

[0048] Layered distributed collaboration and emergency plan switching ensure that even if central communications fail or a single point of equipment fails completely, each sub-district can rely on trained reinforcement learning models and local emergency plans to maintain minimum viable drainage capacity. If the digital twin detects that emergency measures are insufficient, reliable recalculation is triggered and new instructions are dynamically broadcast, reflecting the unique synergy of multi-factor linkage and real-time response. For larger-scale or cross-regional pipeline network scenarios, a cloud-edge collaborative architecture and federated learning are introduced to distribute control and simulation tasks of the ocean sub-districts in parallel between the cloud and edge nodes. This ensures globality and high computing power while retaining rapid response at the sub-district level.

[0049] Through this hierarchical aggregation approach, the reinforcement learning models of each sub-region can share and absorb richer cross-regional experience, and continuously iterate and upgrade under the action of the federal aggregation function to achieve true cross-basin integrated overflow prevention management.

[0050] Overall, by integrating machine learning, hierarchical distributed collaboration, and digital twins, it embodies a creative closed loop from predictive maintenance to real-time simulation to multi-level fault-tolerant scheduling. It can not only identify fault risks in advance on the equipment side, but also respond and correct errors quickly at the control decision-making level. It can also rely on distributed emergency mechanisms and cloud-edge collaboration to ensure the stable operation of the drainage system under extreme conditions, greatly improving the overflow prevention efficiency and fault-tolerant resilience of the urban drainage network. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the process of the pipeline network overflow prediction method based on multi-parameter fusion of the present invention;

[0052] Figure 2 This is a schematic diagram of the system of the multi-parameter fusion pipeline overflow prediction method of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] See also Figure 1 The present invention provides a multi-parameter fusion pipeline overflow prediction method, including:

[0055] Step 1: When the sensor network regularly collects the core operating parameters of the device If it is determined that the fault risk threshold has been triggered, the machine learning module is called to analyze the operating parameters. Dynamically calculate the health index H(t) and failure probability P(t), generate maintenance warnings and output equipment health reports when the corresponding thresholds are exceeded, and then pass the health index H(t) and failure probability P(t) to the downstream scheduling link;

[0056] The step 1 includes the following:

[0057] Step 101: Sensor data acquisition and preliminary processing

[0058] Deploy multiple types of real-time sensors on the execution equipment (including pump stations, gates, etc.), such as current sensors, voltage sensors, opening and closing torque sensors, temperature sensors, etc.; regularly collect multi-dimensional sensor data of each execution equipment in time series through the distributed sensor network. i (t), denoted as:

[0059] X i (t) = [x i1 (t),xi2 (t),…,x iM (t)]

[0060] Where i represents the execution device number, t represents the time, M represents the number of sensor channels configured for device i, and x im (t) represents the data collected by the mth sensor at time t;

[0061] The original collected data is preliminarily cleaned and denoised, and invalid or abnormal jump data are eliminated to form a standardized data set for further analysis by machine learning models The standardization process includes normalization, extreme value clipping and other operations to ensure that sensor data of different dimensions are comparable;

[0062] During use, by establishing a preliminary cleaning and standardization mechanism for multi-source sensor data, the accuracy of subsequent health status assessments can be effectively improved, and the interference of outliers on the assessment results can be reduced. In the initial processing of sensor data, a set of adaptive standardization rules is maintained for each device, so that in the scenario of expanding the scale of multiple devices, it can still automatically adapt to the data distribution characteristics of new devices.

[0063] Step 102: Calculate health index and predict failure probability

[0064] After obtaining and cleaning the standardized dataset Based on this, a health index calculation model and a failure probability prediction model are introduced to quantify the health status of the equipment at time t. The health index is recorded as H i (t), the calculation formula is as follows:

[0065]

[0066] H i (t) represents the health index of the i-th executive equipment (such as a pump station or gate) at time t; is the standardized sensor data vector corresponding to the i-th device at time t, and the vector dimension is M, that is:

[0067]

[0068] in is the mth sensor channel data after denoising and normalization;

[0069] W i is the sensor weight matrix of the i-th device, with a size of M × M, which can be set as a symmetric positive definite or semi-positive definite matrix; the elements can be estimated by optimization methods based on historical fault data or prior experiments, or determined by engineering experience when there are few devices;

[0070] express The Euclidean norm (L2 norm) of φ i (τ) represents the stress factor of the i-th device at time τ, which can be formed by weighted average of information such as the start-stop frequency, load amplitude, and temperature limit of the device;

[0071] Denotes the mapping function for the sensing vector of the slave device at time τ Extract additional correction factors, for example:

[0072]

[0073] Where γ1 and γ2 are adjustable positive real coefficients, or are adaptively configured according to the device type and operating environment; i (τ) is combined to comprehensively reflect the superposition of the stress load and the abnormality of the sensor data at the current moment; the specific form of the function Ω(·) can be customized by the user according to the actual equipment characteristics;

[0074] α1, α2, α3, and α4 are all positive real coefficients. Different values ​​can be used for different devices i, and corresponding fine-tuning can also be performed under different working conditions of the same device;

[0075] In order to further quantify the failure probability, the following failure probability function is established to output the failure probability P i (t):

[0076] P i (t)=1-Ψ[H i (t)]

[0077] Where: P i (t) The probability of failure of device i at time t; Ψ[·] is a function that maps the health index to the device health index, which can be a monotonically decreasing form, such as Ψ[z] = κ1 / (1+exp(-κ2z); κ1 and κ2 are positive real coefficients, which means that when the health index is high, the failure probability is low, and vice versa;

[0078] When used, by weighting the multi-source sensor data and combining it with the exponential function and the integral term, the dynamic contribution of each sensor channel can be fully considered in the fault trend prediction. Through nonlinear coupling, it can provide a more sensitive recognition capability for long-term fatigue and sudden overload of the execution equipment. i (t) function, the failure probability can be directly connected to the reliability control algorithm in the next step, so that a more reasonable redundancy coefficient can be set according to the equipment availability information in the second step.

[0079] Step 2: When the health index H(t) and failure probability P(t) are received and the rainfall or flow forecast data is updated, the reliable control module is started to integrate the worst-case analysis and multi-agent reinforcement learning to obtain the optimal control vector u for the gate opening and the start and stop of the pump station. * (t) Perform online solution and trigger the fault-tolerance mechanism in the event of a failure, forming a dynamically iterative scheduling instruction set;

[0080] The second step includes the following:

[0081] Step 201: Real-time reliability scheduling and worst-case optimization

[0082] The central controller combines the health index H i (t) and failure probability P i (t), and real-time pipe network monitoring data (such as water level, flow, and rainfall forecasts), to build a reliable scheduling optimization model, where:

[0083] To cope with possible performance degradation or failure of the equipment, the model adopts worst-case analysis and introduces the uncertainty set ε to characterize the potential failure mode of the execution equipment.

[0084] Let u(t)=[u1(t),u2(t),…,u N (t)] T is the control instruction vector for N execution devices at time t (such as gate opening, pump station start and stop quantity); H(t) = [H1(t), H2(t), ..., H N (t)] T and P(t)=[P1(t),P2(t),…,P N (t)] T are the equipment health index vector and failure probability vector obtained from step 1 respectively;

[0085] During the optimization process, a disturbance variable ε is introduced for each device. i (t)∈ε, which is used to characterize the possible performance degradation or delay. The uncertainty set ε can be defined based on the failure probability vector P(t) and engineering experience. For example, the device with a higher failure probability takes a larger value within the uncertainty range.

[0086] Let Γ[u(t),H(t),ε i (t)] means that under a given control instruction u(t), considering the health state H(t) of device i and the disturbance ε i (t), a comprehensive measurement function for overflow risk or scheduling cost (which may include penalty items, drainage benefits, overflow penalties, etc.);

[0087] Based on the above definition, reliability scheduling optimization can be expressed as the following two-level extreme value problem, and then the optimal control vector u is obtained * (t):

[0088]

[0089] in: Find the optimal control instructions for the corresponding central controller; = represents a worst-case evaluation of all possible uncertain disturbances; Γ[·] is a comprehensive performance loss function or cost function, where a larger value indicates a higher overflow risk or operating cost; the comprehensive performance loss function Γ[·] consists of three parts: an overflow risk penalty, which applies a proportionally larger penalty to the excess flow when the predicted flow exceeds the available capacity of the partition, forcing the optimization result to avoid overflow as much as possible; an energy consumption / regulation cost, which introduces a cost penalty for starting and stopping pump stations or large changes in gate opening, encouraging solutions to be as energy-efficient and stable as possible while ensuring overflow prevention; and a worst-case cost, which combines the device response hysteresis or degradation ε with the current health index H(t). When a device is in a weakened state, the penalty for relying on it is increased, thereby automatically avoiding or reducing reliance on unreliable devices during optimization.

[0090] The health index vector H(t) and the failure probability vector P(t) guide the dynamic adjustment of the weight or range of each device in the uncertainty set ε. For example, if the failure probability P i (t) is higher, the disturbance ε i The value space of (t) can be relaxed, which means greater failure uncertainty.

[0091] During use, through a double-layer extreme value design, step 201 can still ensure the safety margin of pipeline network scheduling under adverse conditions, avoiding overflow uncontrolled due to single-point equipment failure or performance degradation. The health index vector H(t) and the failure probability vector P(t) are used to dynamically adjust the uncertainty boundary to form an adaptive scheduling mechanism. The failure probability P(t) is used to guide the online adjustment of the uncertainty set ε, giving the scheduling model stronger risk prevention capabilities when the equipment health decline trend is obvious. The comprehensive performance loss function Γ[·] through the comprehensive measurement function can flexibly integrate multi-dimensional objectives (such as overflow control, energy consumption constraints, emission compliance, etc.), improving the system's adaptability to complex working conditions.

[0092] Step 202: Multi-agent reinforcement learning strategy

[0093] After completing the scheduling optimization, the central controller generates the initial control instruction u *(t) It may be necessary to make secondary corrections at a finer granularity (e.g., for each sub-area or distributed control unit). To this end, a multi-agent reinforcement learning (MARL) strategy is adopted to configure an independent agent for each sub-area (or key pipeline section) and maintain interaction with the central controller:

[0094] Defining an Agent Set Where K is the number of sub-areas, and each sub-area has a corresponding reinforcement learning agent responsible for local scheduling;

[0095] Agent a k The observation vector o at time t k (t) Including local water level, flow and corresponding equipment health index H i (t), failure probability P i (t) other information;

[0096] Agent a k The policy function π k (o k (t)) Output local control fine-tuning amount Δu k (t), used to give the initial control instruction u to the central controller * (t) Perform local correction, where: the policy function π k o k (t) is essentially the mapping relationship used by the k-th sub-region agent to determine the fine-tuning control command when given the observation state - it uses the local observation vector, usually using a deep neural network To approximate; each sub-area agent can be imagined as a driver, o k (t) is the dashboard information he sees (water level, flow, equipment status), Δu k (t) is his fine-tuning operation on the steering wheel and accelerator, and the policy function π k It is his brain that learns the most correct fine-tuning plan through continuous training to ensure safe and economical driving in any road conditions (i.e. uncertain rain conditions or equipment health). k It is no longer a manually written rule, but an optimal mapping learned through a large amount of simulation training. It can adapt to complex and changeable pipeline network conditions and continuously improve scheduling effects.

[0097] To coordinate multi-agent goals, set a joint reward function Typical objectives might include:

[0098] Low overflow rate reward; Minimum energy consumption or pump station operating cost reward; Equipment health protection reward (based on the equipment health index H i (t) or failure probability P i (t) dynamic weighting);

[0099] Through distributed reinforcement learning algorithms (which can be based on policy gradients or value iteration), each agent continuously updates its policy function π in a large number of offline simulations and online feedback. k (·), while ensuring the global reliability goal, improve local scheduling flexibility and the ability to respond to emergencies.

[0100] When in use, with the help of multi-agent reinforcement learning, each sub-area can quickly adjust local control in the face of actual environmental disturbances and equipment failures, reducing communication dependence on the central controller; and the optimal control vector u output in step 201 * (t) forms a global-local dual-layer control model. Centralized scheduling ensures macroscopic safety margins, while sub-regional reinforcement learning enables the system to adapt in real time. By incorporating the failure probability P(t) and health index H(t) into the multi-agent reward design, agents are enabled to focus not only on drainage scheduling objectives but also on the maintenance and failure risks of the executing equipment, truly achieving a deep integration of reinforcement learning and equipment health management. Local incremental fine-tuning avoids overlap or conflict between agents, improving the stability of multi-agent parallel decision-making.

[0101] Step 3: Generate the optimal control vector u * After (t), the digital twin platform receives the corresponding instructions and synchronously reads the health index H(t) and failure probability P(t) and pipe network sensor information, calls fast hydraulic simulation and real-time deviation measurement E(t) to perform a pre-verification of the actual executability. If the real-time deviation measurement E(t) exceeds the threshold, it will be fed back to the reliable control module and trigger the strategy recalculation;

[0102] The step three includes the following:

[0103] Step 301: Digital twin model initialization and real-time data integration

[0104] The digital twin platform includes a global hydraulic model of the pipe network system, a dynamic characteristic model of the execution equipment, and a calibrated fault simulation module. Its core is to use the same parameters and data objects as steps one and two, such as the execution equipment health index vector H(t) and the control instruction vector u * (t), etc., to achieve high-fidelity reproduction of the current pipeline network status.

[0105] To ensure the consistency between the digital twin model and the real environment at time t, the following multi-source data needs to be loaded:

[0106] Equipment health information; real-time measurement data: such as water level, flow, and rainfall observations at each monitoring point; control instructions from the previous cycle; network communication delay or failure events: used to synchronize communication status and evaluate whether instructions can be issued in a timely manner.

[0107] The digital twin model uses a high-order hydraulics description that combines partial differential equations with nonlinear flux functions to simulate the dynamics of water flow, pressure, and potential overflow points in the pipe network. Simultaneously, characteristics such as the opening and closing inertia of the actuator equipment, the maximum opening and closing rate, and the pumping station power curve are embedded in the model to more realistically depict potential execution delays.

[0108] In order to adapt to the health decay effect of the execution equipment, the health index H for device i in the model is i (t) and failure probability P i (t) Jointly determined introduction function factor Λ i (t):

[0109] Λ i (t) = 1-Θ(H i (t),P i (t))

[0110] Where Θ[·] is a monotonically increasing function, which is used when the health index H i (t)Decrease or failure probability P i When (t) becomes high, the function factor Λ i (t) is reduced accordingly, reflecting the impact of reduced equipment efficiency or potential jamming in the simulation;

[0111] When used, the digital twin model can accurately simulate the health degradation and potential impact of failure of the execution equipment, avoiding simulation distortion caused by the separation of the model and actual engineering data, and introducing the functional factor Λ i (t) By closely linking the health status with the equipment’s operating efficiency, digital mapping and real-time tracking of the physical characteristics of the execution layer can be achieved.

[0112] After functionally coupling the equipment health information with the failure probability, and then writing the results into the equipment performance parameters in the high-order hydraulic simulation equation, the adaptive ability of the digital twin can be enhanced.

[0113] Step 302: Rapid performance verification and deviation information feedback

[0114] After the digital twin completes a simulation from time t to t+Δt, it is necessary to compare the simulation output state with the expected scheduling target to form deviation information. The real-time deviation metric E(t) is defined to quantify the gap between the actual executable and the theoretical instructions, where:

[0115] Assume that there are several key observation points in the pipe network, and the state vector of each observation point in the digital twin simulation is S sim (t), and the state vector under the theoretical model expectation in step 2 is S exp (t):

[0116]

[0117] Where: E(t) represents the real-time deviation metric calculated based on the digital twin simulation results at the current time t. Its components can correspond to the comprehensive degree of difference in different key areas of the pipeline network or different observation indicators (such as upstream water level, downstream flow, pumping station drainage effect, etc.);

[0118] S sin(τ) S is the simulation output vector of the pipe network system status at time τ (t≤τ≤t+Δt) by the digital twin platform, which contains multi-dimensional information such as water level, flow, and pressure at each observation point; exp (τ) is the expected state vector or theoretical model output at the same time τ under the second step prediction;

[0119] W E is a weighting or mapping matrix used to weight or transform differences in different dimensions; represents the element-wise multiplication (Hadamard product) between the components of the vector; Φ(τ,H(τ),P(τ)) is the correction function vector obtained by introducing the health index H(τ) and the failure probability P(τ) of the execution device at time τ;

[0120] Its function is to make the state difference in the event of equipment attenuation or high failure risk (such as P i (τ) increases, H i When (τ) decreases), additional amplification or reduction is obtained, and the example form can be defined as:

[0121] Φ(τ,H(τ),P(τ))=[Φ1(τ),Φ2(τ),…,Φ M (τ)] T

[0122] where each Φ m (τ) may be some monotonic function that depends on the health index or failure probability associated with device m, such as:

[0123] Φ m (τ)=1+v m ψ(H m (τ),P m (τ))

[0124] v m is an adjustable coefficient, ψ(·) represents a specific function (monotonically increasing or monotonically decreasing) that maps the health or fault information of the equipment, thereby amplifying the corresponding observation difference to provide a warning when the health degradation is obvious;

[0125] The mapping function ψ(·) is designed to convert the health status and failure risk of the executing device into a difference amplification factor, which is used to highlight those devices with deteriorating health or high failure probability in subsequent data processing and verification. Specifically, ψ(·) contains two signals:

[0126] Health degradation amplification: When the equipment health index H(t) is lower, it means that the component wear, fatigue or performance degradation is more serious. At this time, ψ(·) will grow rapidly in an exponential manner, thereby automatically increasing the weight of the monitoring indicators related to the equipment, making the system more sensitive to the errors and fluctuations caused by it. Failure risk amplification: When the equipment failure probability P(t) is higher, it means that the possibility of the equipment failing is increasing. ψ will also exponentially amplify this risk signal, making the system more inclined to reduce its reliance on high-risk equipment or make compensation in advance during control optimization and simulation verification.

[0127] Overall, ψ combines and weights the risks of health degradation and failure probability, controlling the amplification strength through adjustable sensitivity and weight parameters. This allows the identification of vulnerable execution devices that require the most attention and priority during digital twin simulation or scheduling optimization, significantly improving the reliability and overflow prevention of the entire system in the event of sudden failures or adverse operating conditions.

[0128] ||·|| p For L p norm (p>1), the sensitivity to large error components can be adjusted by different p values, and Δt is the short time step used for fast simulation or verification;

[0129] When some components of the real-time deviation metric E(t) exceed the preset threshold, the digital twin platform immediately generates risk warnings and deviation information, which are returned to the reliability real-time control module in step 2, triggering the control instruction vector u * (t) correction or recalculation. If it is detected that a device has shown obvious hysteresis in this simulation (determined by the function factor Λ i (t) drops to an extremely low value), the fault tolerance mechanism or other emergency strategies can be activated;

[0130]

[0131] Where: D(t) represents the deviation information and risk warning vector sent by the digital twin platform to step 2;

[0132] It is a comprehensive function that can include health status, failure probability and verification measure together, and output the direction or magnitude of the correction so that the reliability control module can evaluate the necessity of secondary optimization.

[0133] Comprehensive Function The purpose is to superimpose four types of information, namely digital twin calibration error, equipment health attenuation, failure risk and functional degradation, according to different weights and sensitivities, to generate risk warnings or "corrective instruction suggestions" for each device or each sub-area.

[0134] Deviation amplification: When the simulation error E(t) is large, this value is amplified exponentially to remind the control layer to focus on potential scheduling failures caused by the gap between simulation and expectations.

[0135] Health decay: When the health index H(t) decreases, a penalty that increases proportionally with 1-H(t) is generated, informing the control layer that the reliability of the device has significantly decreased.

[0136] Failure risk: When the failure probability P(t) increases, its impact is also amplified exponentially, so as to prioritize reducing the scheduling dependence on high-risk equipment.

[0137] Functional degradation compensation: Combined with the performance factor Λ(t) simulated in the digital twin, the reduced availability due to equipment aging or performance degradation is directly compensated proportionally.

[0138] Finally, the weighted sum of these four parts forms the comprehensive prompt value D of each device i (t), the control decision module adjusts or recalculates the scheduling strategy accordingly, thereby giving priority to reducing the load of high-risk and high-deviation devices in the next round of execution instructions and compensating for their performance deficiencies.

[0139] Thus, the comprehensive function It integrates multi-source health, simulation, and risk information into a single improvement recommendation, allowing the system to quickly make the most targeted strategy corrections when there is uncertainty or equipment unreliability.

[0140] When used, by combining with the digital twin model established in step 301, step 302 can evaluate the pipe network efficiency that may be generated after the execution of the instruction at this moment within a shorter simulation cycle (such as seconds to minutes), and quantify the difference between it and the theoretical expectation. If the difference is found to be too large, the deviation information is immediately fed back to the control layer in the second step, allowing the system to recalculate or fine-tune the strategy before the instruction is officially issued, greatly reducing the overflow risk caused by the lag of the execution layer. Differentiated considerations are made according to the specific key areas of the urban pipe network (such as flood-prone areas, sensitive water bodies, etc.), and the real feasibility of the evaluation system is refined.

[0141] Step 4: When the digital twin verifies the optimal control vector u *(t) When feasible and formally issued to each execution device, the hierarchical distributed coordination mechanism monitors communication quality and shares key data; if the main network is interrupted or key equipment fails, the sub-area controller independently executes the emergency plan according to the trained strategy and triggers reliable recalculation or multi-sub-area linkage again when local measures fail;

[0142] The step 4 includes the following contents:

[0143] Step 401: Initialization and communication management of the hierarchical distributed collaboration mechanism

[0144] The urban drainage network is divided into several sub-areas, each of which is equipped with a local sub-area controller. A central controller is set up at the macro level to coordinate the global drainage strategy. The sub-area controllers maintain real-time data exchange with the central controller through the main communication network, and a local communication network is established within the sub-area to ensure a stable connection between the equipment (such as pump stations and gates) and the local controllers. Each sub-area controller synchronously receives and stores the final control instruction vector u after successful verification in the third step. * (t), and the corresponding equipment health index H i (t) or failure probability P i (t);

[0145] Sub-intervals share necessary information such as pipeline capacity, reservoir water level, and downstream discharge permits to facilitate global scheduling;

[0146] Improve the stability of information transmission in the event of heavy rain or equipment failure through redundant communication links (such as wired + wireless combination).

[0147] Use a multi-channel network to monitor and dynamically record communication quality indicators (such as delay, packet loss rate, etc.). If the main communication channel is interrupted or the delay exceeds the threshold, it will automatically switch to the backup channel to ensure that the minimum necessary data can still be transmitted between the sub-area controller and the central controller. A communication quality function can be defined:

[0148] Ψ comm (t) = exp(-γη(t))

[0149] Where η(t) is a dynamic indicator representing the comprehensive packet loss rate and delay, and γ is a positive real constant.

[0150] If comm (t) drops below a certain threshold, the central controller determines that the main communication is abnormal and switches to emergency coordination mode.

[0151] When used, the hierarchical distributed structure is combined with the fault-tolerant communication protocol, so that the system can still maintain high reliability in a large-area multi-subarea topology, avoiding a single communication failure that causes global control paralysis, and the communication quality function Ψ comm(t) Incorporating parameters such as failure probability into the hierarchical scheduling evaluation can achieve a unified design of coordinated fault tolerance at the network layer and the device layer, rather than just local optimization at the pipeline network or device level.

[0152] Step 402: Local independent emergency plan execution

[0153] When the central controller detects the failure probability P of a key device i (t) has reached the high risk range, or the communication quality function Ψ comm When (t) is lower than the threshold, or the digital twin platform indicates that the overflow risk of the sub-area is too high and the central dispatch has not had time to update the instructions, the local independent emergency plan is triggered; the local emergency plan is stored inside the sub-area controller and includes operational strategies such as activating backup pumps, closing non-critical branch gates, and prioritizing the pumping of critical trunk lines.

[0154] In the second step, each sub-region agent {a k The local scheduling strategy has been trained using multi-agent reinforcement learning. When a network outage occurs or central instructions cannot be issued in a timely manner, the sub-zone controller can switch to offline operation mode and make operational decisions based on the trained strategy:

[0155]

[0156] Among them k (t) represents the water level, flow rate, and execution equipment health index H(t) and failure probability P that can be observed by the sub-area controller. i (t) and other information; π k (·) is the strategy function of k agents in the sub-area;

[0157] If communication is not completely interrupted, adjacent sub-zone controllers can exchange limited information through a temporary topology, such as exchanging boundary water levels or warning signals, without relying on the global scheduling of the central controller. In this case, multiple sub-zones can coordinate operations based on local emergency plans, such as temporarily shutting down the upstream sub-zone and accelerating pumping and drainage in the downstream sub-zone, to reduce the impact of the overflow.

[0158] During use, even if the support of the central controller is lost, the sub-area can still rely on the local reinforcement learning strategy trained in the second step to make emergency decisions independently, ensuring the most basic overflow prevention and control capabilities. Through local collaboration, multiple adjacent sub-areas can maintain the most core data interoperability in the case of communication degradation, alleviating the risk of pipeline network out of control caused by the lack of global instructions; the trained multi-agent reinforcement learning strategy is made into an offline emergency plan at the sub-area level, combined with the real-time fault probability P(t) to judge the degree of failure, thereby improving the adaptability and accuracy of emergency operations.

[0159] Step 403: Global Fault Tolerance Coordination and Feedback

[0160] When network communication is restored or the key indicators of the sub-area can still be obtained under limited bandwidth, the central controller will conduct a comprehensive analysis of the status data from the sub-area, including: local emergency execution status Current equipment health index vector H(t), failure probability vector P(t); latest rainfall monitoring or other external forecast information;

[0161] Subsequently, the central controller calls the second step to perform global fault-tolerant recalculation and obtains the updated global instruction vector u * (t+δ) and broadcast it to each sub-area through all available communication paths;

[0162] For local operations or collaborative operations with adjacent sub-areas that have been carried out in emergency situations, the central controller will use them as additional experience samples after communication is restored, incorporate them into the offline retraining or online learning process of multi-agent reinforcement learning, and continuously iterate and update the strategies of each sub-area agent; by comparing the operation records during the emergency with the simulation output of the digital twin platform, the local emergency plan can be further improved or revised.

[0163] When in use, after the emergency operation is completed, the system can quickly return to the globally optimal scheduling track, rather than being stuck in a fragmented local mode for a long time. Emergency decision-making data from sudden extreme failure scenarios can in turn become important supplementary experience for reinforcement learning, making multi-agent strategies more adaptable in similar extreme situations. Through fault-tolerant recalculation and emergency operation feedback loops, the hierarchical distributed mechanism and central scheduling form a mutually reinforcing cycle, not only ensuring the timeliness of emergency measures, but also enabling the continuous improvement of the overall scheduling strategy.

[0164] Step 5: When the urban pipe network is expanded or connected across regions and efficient overflow prevention and control is required, the cloud-edge collaborative architecture is enabled to strengthen the local learning parameters Θ of each sub-area. k and simulation measure E k (t) Upload to the cloud and use the federated aggregation function Forming a global strategy and sending it to edge nodes enables the merging of multi-level partitioned parallel simulation and cross-regional scheduling, ultimately ensuring efficient and robust operation of large-scale pipeline networks.

[0165] The step five includes the following:

[0166] Step 501: Initialize the cloud-edge layered architecture and divide it into multiple levels

[0167] Based on the sub-areas established in the fourth step, further subdivision or merging can be used to form several multi-level partitions, such as the regional level, sub-area level, and local node level. Each level maintains communication with the previous or next level through the cloud or edge nodes, allowing key data from each sub-area controller (equipment health index H(t), failure probability P(t), digital twin verification results, etc.) to flow efficiently between the cloud and the edge. The cloud is responsible for global data storage and batch model training, enabling unified processing of historical observation data, health information, and failure probabilities from each partition. Edge nodes, in response to the more real-time local control needs, combine multi-agent reinforcement learning from the second step to perform small-scale online training and fast inference on the edge side to avoid over-reliance on cloud bandwidth.

[0168] By comparing with the existing communication quality function Ψ in the fourth step comm (t) (monitoring packet loss rate and latency) is integrated to evaluate the connection quality between cloud and edge. If the network condition is poor, the edge node automatically enables local scheduling and emergency functions, and then synchronizes updates with the cloud after communication is restored.

[0169] When in use, multi-level partitioning not only retains the hierarchical distribution advantages of the fourth step, but also provides a scalable architecture for larger-scale pipelines, avoiding global failures caused by overload of nodes at a certain layer or network bottlenecks. The cloud and edge nodes have clear division of labor, which can not only efficiently process large-scale data, but also maintain timely response to local emergencies.

[0170] Step 502: Federated Learning-Driven Distributed Reinforcement Learning Model Training

[0171] In the second step, each sub-region reinforcement learning agent has a preliminary local strategy π k (·), if the city scale is further expanded, the federated learning framework can be used to set the model parameter set Θ scattered at each edge node k Regularly upload to the cloud for aggregation:

[0172]

[0173] in It is a federated aggregation function used to merge models from different sub-regions. The aggregated global parameters are then distributed to each sub-region to achieve cross-regional experience sharing. The most commonly used form of the federated aggregation function is weighted averaging, which linearly combines the model parameters of each sub-region according to their local sample size or importance weight.

[0174] The high-dimensional data collected locally by each sub-area, such as the health index vector H(t), failure probability vector P(t), and digital twin simulation error, are transmitted to the cloud through encrypted or secure channels;

[0175] The cloud uses efficient distributed training algorithms to iteratively upgrade the global reinforcement learning model in a very short time;

[0176] After receiving the updated global model parameters, the edge nodes perform local fine-tuning and combine it with local live online training to further improve the local control quality.

[0177] When used, each sub-region can share the learning results of other regions, promoting the rapid convergence and generalization of the overall strategy in ultra-large-scale pipe network environments and avoiding overfitting of local single points. Compared with simply centralizing all data in the cloud, this step enables distributed training on edge nodes, reducing communication and computing pressure.

[0178] Step 503: Merge cloud-edge parallel simulation evaluation and cross-region scheduling

[0179] Based on the model of digital twin simulation and rapid performance verification in the third step, the complex large-scale pipeline network is divided into several parallel simulation units. Each edge node is responsible for the real-time simulation of the local partition, and the cloud is responsible for the simulation of the global coupling effect across partitions. When the edge node completes the local simulation, the local simulation results are compared with the relevant boundary conditions S sim (t) Upload to the cloud for combined evaluation to form global performance verification indicators:

[0180]

[0181] E global (t) represents the local performance verification index E of each sub-area in the time period [t, t+Δt] k (τ) is the global performance verification vector formed after synthesis; Δt is the global parallel simulation or verification time window;

[0182] E k (τ) is the local performance check result of sub-area k at time τ (refer to E in the third step k (t) definition), which may include multi-dimensional measurements of the sub-area's overflow risk, hydraulic balance deviation, energy consumption deviation, etc. represents the set of performance check vectors of all sub-areas in the same period τ;

[0183] Γ[{E k (τ)},A] is the "multi-subregion collaborative aggregation function", which is used to integrate the local verification results of each subregion and take into account the adjacency relationship between subregions:

[0184]

[0185] Where: A is the adjacency or coupling relationship matrix (size K × K), reflecting the connection strength between different sub-regions in terms of hydraulic transmission, geographical proximity or discharge coupling;

[0186] Ak,· represents the k-th row vector of matrix A; is the element-wise (Hadamard) product operation, used for E k (τ) differentially amplifies or suppresses in different dimensions, thereby reflecting the proportion or weight of sub-region k in the global deviation; W E is a weighting or mapping matrix used to re-weight or transform different deviation dimensions when calculating global indicators; ||·|| p , is L p Norm (p>1), which determines the sensitivity of the final merge result to large deviation components;

[0187] According to the fused global performance calibration index E global (t) and the control instructions returned by the sub-area The cloud scheduling engine can call the second step again to perform unified scheduling, merging and optimization of ultra-large-scale regions, and output global instructions across regions. It is distributed to the edge nodes of each sub-area, and then fine-tuned and executed in combination with local reinforcement learning. In this case, each sub-area controller first runs a lightweight digital twin in parallel on the edge node, performs a short-term simulation of the local control instructions and health status, and outputs the deviation vector E k (t), and then the deviation vector E k (t) and the necessary boundary conditions are uploaded to the cloud; the cloud calculates the sub-interval coupling matrix A and the weighted mapping W E For all {E k (t)} are fused and the global deviation E is calculated. global (t), and combined with the health index H i (t), failure probability P i (t) and the uncertainty disturbance set ε call the robust optimization module to recalculate the cross-region optimal control instructions Finally, the cloud sends the global command to each sub-area, and then the local reinforcement learning strategy π k Generate fine-tuning increment Δu k (t), forming the final execution command

[0188] During use, cloud-edge parallel simulation can process simulation tasks for sub-regions and execution devices in a short period of time, effectively reducing the computing pressure on a single node. Global scheduling and merging can solve cross-regional watershed hydraulic correlation issues, providing a more accurate global optimal solution to prevent downstream overflows or upstream waterlogging. Extending the rapid performance verification method of the third step digital twin to cloud-edge parallel collaborative simulation scenarios allows each sub-region in a large-scale pipeline network to independently and efficiently simulate local hydraulic processes while performing global coupling assessments in the cloud. The use of a local result + global fusion model accelerates the simulation iteration cycle while maintaining overall control over the worst-case overflow risk.

[0189] See also Figure 2 The present invention provides a multi-parameter fusion pipeline overflow prediction system, including:

[0190] The health assessment unit collects the operating parameters of the execution equipment and determines that the fault risk threshold is triggered. It then uses the machine learning algorithm to calculate the health index and failure probability of the operating parameters and outputs a predictive maintenance warning;

[0191] The scheduling and reinforcement learning unit receives the health index and failure probability and detects changes in pipeline network monitoring data. It then initiates reliable optimization and multi-agent reinforcement learning to iteratively solve the gate opening and pump station start and stop vectors. In the event of a failure, it automatically triggers the fault tolerance mechanism and outputs updated optimal control instructions.

[0192] The digital twin simulation verification unit receives the optimal control vector and pipe network monitoring data from the digital twin platform, and uses the equipment health index to quickly simulate and evaluate the deviation of the hydraulic process. If the deviation exceeds the corresponding threshold, it will be fed back to the control layer to trigger recalculation or fine-tuning;

[0193] Hierarchical distributed collaborative emergency unit: If the digital twin verification confirms that the optimal control vector is feasible and needs to be formally distributed to each execution device, the hierarchical distributed collaborative mechanism monitors the communication quality and shares the health index and failure probability. If the main network or key equipment fails, the sub-area controller independently activates the emergency plan and dynamically triggers fault-tolerant recalculation;

[0194] Cloud-edge collaboration and federated learning units. When the scale of urban pipeline networks expands or cross-regional watersheds require collaborative management, the cloud-edge collaboration architecture is enabled to upload the local reinforcement learning parameters and simulation measurements of each sub-area, form a global strategy through federated aggregation, and send it to the edge nodes.

[0195] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0196] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0197] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0198] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0199] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A multi-parameter fusion pipeline overflow prediction method, characterized by: include, When the operating parameters of the execution equipment are collected and the fault risk threshold is determined to be triggered, the machine learning algorithm is used to calculate the health index and failure probability of the operating parameters and output a predictive maintenance warning; After receiving the health index and failure probability and detecting changes in pipeline network monitoring data, it initiates reliable optimization and multi-agent reinforcement learning to iteratively solve the gate opening and pump station start and stop vectors. In the event of a failure, it automatically triggers the fault tolerance mechanism and outputs updated optimal control instructions. The digital twin platform receives the optimal control vector and pipe network monitoring data, and combines it with the equipment health index to quickly simulate and evaluate the deviation of the hydraulic process. If the deviation exceeds the corresponding threshold, it will be fed back to the control layer to trigger recalculation or fine-tuning. If the digital twin verification confirms that the optimal control vector is feasible and needs to be formally distributed to each execution device, the hierarchical distributed coordination mechanism monitors the communication quality and shares the health index and failure probability. If the main network or key equipment fails, the sub-area controller independently activates the emergency plan and dynamically triggers fault-tolerant recalculation; When the scale of urban pipeline networks expands or cross-regional watersheds require collaborative management, the cloud-edge collaborative architecture is enabled to upload the local reinforcement learning parameters and simulation measurements of each sub-area, form a global strategy through federal aggregation, and send it to the edge nodes.

2. The multi-parameter fusion pipeline overflow prediction method according to claim 1, characterized in that: Deploy real-time sensors on the execution devices, regularly collect multi-dimensional sensor data of each execution device in time series through a distributed sensor network, and pre-process the original collected data to form a standardized data set; Based on the standardized data set, a health index calculation model and a failure probability prediction model are introduced to quantify the health status of the equipment and output the failure probability according to the health index failure probability function.

3. The multi-parameter fusion pipeline overflow prediction method according to claim 2, characterized in that: An uncertainty set is introduced to characterize the potential failure modes of the execution equipment, and a double-layer extreme value problem is solved and optimized under the control instruction vector. The weights or ranges of the equipment health index vector and the failure probability vector are dynamically adjusted in the worst-case analysis to obtain the optimal control vector.

4. The multi-parameter fusion pipeline overflow prediction method according to claim 3 is characterized by: A multi-agent reinforcement learning strategy is used to assign an independent agent to each sub-area to be responsible for local scheduling and maintain interaction with the central controller. Based on the observation vector, the agent's policy function outputs a local control fine-tuning value, which is used to locally correct the initial control instructions given by the central controller. The policy function is then updated through a distributed reinforcement learning algorithm. The observation vector includes local water level, flow rate, and corresponding equipment health index and failure probability information.

5. The multi-parameter fusion pipeline overflow prediction method according to claim 4 is characterized by: Based on loaded multi-source data, the digital twin model simulates the dynamics of water flow, pressure, and potential overflow points in the pipe network. It also embeds the opening and closing inertia of the actuator equipment, the maximum opening and closing rate, and the pumping station power curve into the digital twin model. In the digital twin model, a functional factor determined by the health index and failure probability of the equipment is introduced to adapt to the health attenuation effect of the execution equipment.

6. The multi-parameter fusion pipeline overflow prediction method according to claim 5, characterized in that: After the digital twin completes the simulation, the simulation output state is compared with the expected scheduling target to form deviation information and quantify it by real-time deviation measurement; When certain components of the real-time deviation measurement exceed the preset threshold, the digital twin platform immediately generates risk warnings and deviation information, triggering the correction or recalculation of the control instruction vector. If it is also detected that a certain device has shown obvious hysteresis in this simulation, the fault tolerance mechanism or other emergency strategies will be activated.

7. The multi-parameter fusion pipeline overflow prediction method according to claim 6, characterized in that: The urban drainage network is divided into several sub-areas, each of which is equipped with a local sub-area controller. Each sub-area controller synchronously receives the final control instruction vector and the corresponding equipment health index or failure probability. Monitor and dynamically record communication quality indicators. If the main communication channel is found to be interrupted or the delay exceeds the threshold, it will automatically switch to the backup channel, or the central controller will determine that the main communication is abnormal and enter the emergency coordination mode.

8. The multi-parameter fusion pipeline overflow prediction method according to claim 7, characterized in that: When the backup failure probability reaches a high-risk range, or the communication quality falls below a threshold, or the digital twin platform indicates that the overflow risk of the sub-area is too high and the central dispatcher has not yet updated the instructions, the local independent emergency plan is triggered; When a network interruption occurs or the central command cannot be issued in time, the sub-area controller switches to offline operation mode. If the communication is not completely interrupted, the adjacent sub-area controllers can exchange limited information through a temporary topology.

9. The multi-parameter fusion pipeline overflow prediction method according to claim 8, characterized in that: When network communication is restored, the status data from the sub-zones is comprehensively analyzed, and the central controller performs global fault-tolerant recalculation to obtain an updated global instruction vector, which is then broadcast to each sub-zone through all available communication paths. For local operations or collaborative operations with adjacent sub-areas that have been carried out in emergency situations, the central controller will use them as additional experience samples after communication is restored, incorporate them into the offline retraining or online learning process of multi-agent reinforcement learning, and continuously iterate and update the strategies of each sub-area agent.

10. The multi-parameter fusion pipeline overflow prediction method according to claim 9, characterized in that: Several multi-level partitions are formed based on the sub-zones, and the key data of each sub-zone controller flows between the cloud and the edge side; The cloud and edge nodes are configured to respectively undertake large-scale data aggregation and real-time control instruction execution, and evaluate the connection quality between the cloud and the edge in real time. If the network conditions are poor, the edge nodes automatically enable local scheduling and emergency functions, and synchronize updates with the cloud after communication is restored.

11. The multi-parameter fusion pipeline overflow prediction method according to claim 10, characterized in that: Use the federated learning framework to regularly upload model parameter sets distributed across edge nodes to the cloud for aggregation; The health index vector, fault probability vector, and digital twin simulation error collected locally by each sub-area are transmitted to the cloud, and the global reinforcement learning model is iteratively upgraded using a distributed training algorithm. After receiving the updated global model parameters, the edge node performs local fine-tuning and combines local live online training to improve local control quality.

12. The multi-parameter fusion pipeline overflow prediction method according to claim 11, characterized in that: The complex large-scale pipeline network is divided into several parallel simulation units. When the edge node completes the local simulation, the local simulation results and related boundary conditions are uploaded to the cloud for combined evaluation to form a global performance verification index. Based on the fused global performance verification indicators and the control instructions returned by the sub-areas, the cross-regional global instructions are merged and output for the cross-regional scheduling plan at the large-scale watershed level. These instructions are distributed to the edge nodes of each sub-area and executed after fine-tuning with local reinforcement learning.

13. A multi-parameter fusion pipeline overflow prediction system, characterized by: include, The health assessment unit collects the operating parameters of the execution equipment and determines that the fault risk threshold is triggered. It then uses the machine learning algorithm to calculate the health index and failure probability of the operating parameters and outputs a predictive maintenance warning; The scheduling and reinforcement learning unit receives the health index and failure probability and detects changes in pipeline network monitoring data. It then initiates reliable optimization and multi-agent reinforcement learning to iteratively solve the gate opening and pump station start and stop vectors. In the event of a failure, it automatically triggers the fault tolerance mechanism and outputs updated optimal control instructions. The digital twin simulation verification unit receives the optimal control vector and pipe network monitoring data from the digital twin platform, and uses the equipment health index to quickly simulate and evaluate the deviation of the hydraulic process. If the deviation exceeds the corresponding threshold, it will be fed back to the control layer to trigger recalculation or fine-tuning; Hierarchical distributed collaborative emergency unit: If the digital twin verification confirms that the optimal control vector is feasible and needs to be formally distributed to each execution device, the hierarchical distributed collaborative mechanism monitors the communication quality and shares the health index and failure probability. If the main network or key equipment fails, the sub-area controller independently activates the emergency plan and dynamically triggers fault-tolerant recalculation; Cloud-edge collaboration and federated learning units. When the scale of urban pipeline networks expands or cross-regional watersheds require collaborative management, the cloud-edge collaboration architecture is enabled to upload the local reinforcement learning parameters and simulation measurements of each sub-area, form a global strategy through federated aggregation, and send it to the edge nodes.

Images