Intelligent control method for low water level operation of pipe network based on hot spot identification of external water intrusion

Abstract

The application provides a kind of intelligent regulation and control method for pipe network low water level operation based on external water intrusion hot spot identification, liquid level and water quality linkage sensing equipment is arranged in the external water intrusion hotspot area, the edge computing node uses variational mode decomposition and mutual information entropy fusion algorithm to calculate the external water intrusion intensity index in real time. The cloud-based long short-term memory network and deep deterministic policy gradient reinforcement learning with fusion attention mechanism construct an intelligent decision model, and output the optimal pumping frequency of the pump station. After receiving the instruction, the pump station controller actively reduces the pipe network water level below the design fullness in the dry season, and uses the low water level condition to expose the structural defects of the pipe network. When the conductivity or COD concentration significantly rebounds, the system automatically triggers the frequency locking logic to maintain the low water level strategy. The edge-cloud collaborative microservice grid architecture and LSM-Tree distributed storage engine are adopted, and the automatic switching of flood prevention mode and the safety boundary constraint mechanism are integrated.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of pipeline network operation at low water levels, and in particular to an intelligent control method for pipeline network operation at low water levels based on the identification of external water intrusion hotspots. Background Technology

[0002] In urban sewage pipe networks in the northern plains, the persistently high water level is a common and challenging problem. Due to high groundwater levels or the infiltration of external water (such as river water or groundwater) through structural defects like pipe joint gaps and cracks in the pipe walls, the sewage level is often only 0.5 to 2 meters below the manhole opening. This high water level operation presents multiple drawbacks: First, the high water level masks structural defects in the pipes, making it difficult for maintenance personnel to promptly detect and repair cracks or loose joints using conventional methods, leading to long-term and accelerated escalation of potential hazards. Second, the significant infiltration of external water significantly dilutes the sewage concentration, resulting in lower levels of indicators such as chemical oxygen demand (COD) in the incoming sewage, making it difficult to meet performance evaluation standards. Furthermore, existing sewage pipe networks lack scheduling buffer space, and the coordination and scheduling capabilities between pump stations are weak. Once rainfall or increased external water intrusion occurs, the pipe network water level rises rapidly, forcing high-concentration sewage to overflow directly, causing serious environmental pollution.

[0003] In existing technologies, there have been some research and applications addressing the issues of external water intrusion and pipeline network scheduling. For example, existing technology CN115048759B discloses a method for accurately estimating the inflow rate and inflow point of external water in sewage pipelines based on model prediction. This technology focuses on quantitatively calculating the inflow rate and locating the inflow point by establishing a mathematical model, mainly addressing the issues of "measurement" and "location." It lacks direct linkage with pump station operation and control, and cannot suppress external water in real time through control measures. Existing technology CN118551510A proposes a method for identifying key areas of external water intrusion. It assesses infiltration defects and cross-connection defects by constructing a water quality and quantity balance model to determine key investigation areas. This technology also focuses on guidance for "identification" and "investigation." The existing technology CN104846926A does not address how to reduce infiltration by actively lowering the water level during operation; Existing technology CN104846926A relates to a system and method for measuring external intrusion in a pipe network using a volumetric method, which mainly uses physical devices (water tank, sensor) to measure the volume of external intrusion under negative pressure conditions. It is a detection method and does not have the ability to intelligently control the entire network; Existing technology CN120103881A discloses a method for regulating the water level of an internal ditch in a drainage system. Its focus is on preventing backflow of river water by regulating the sluice gates in the river channel, emphasizing "external prevention". However, it lacks specific control strategies for how the pumping stations inside the sewage pipe network respond to external water intrusion and how to maintain low water level operation.

[0004] In summary, existing sewage pumping station scheduling relies heavily on fixed liquid level thresholds for start-stop control, lacking the ability to perceive and respond to dynamic changes in external water intrusion. This traditional "passive" scheduling cannot flexibly adjust operating strategies, making it difficult to achieve the ideal operating state of "low water level, high concentration." Furthermore, existing control systems often lack the ability to deeply clean data for complex operating conditions (such as high-frequency oscillation noise caused by pumping station start-stop), and their software deployment architecture often employs general virtualization solutions, which are insufficient to meet millisecond-level real-time control requirements when edge resources are limited. Therefore, there is an urgent need for a technical solution that can deeply integrate liquid level-water quality linkage sensing with dynamic pumping station frequency control, and achieve intelligent operation of the pipe network at low water levels through advanced algorithms and architecture. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent control method for low-water-level operation of pipe networks based on the identification of external water intrusion hotspots. This method solves the technical problems of existing sewage pipe networks, which are difficult to expose defects due to long-term high-water-level operation, dilute sewage concentration due to external water infiltration, and lack of dynamic response capability to external water intrusion in traditional pump station scheduling, thus making it impossible to achieve improved quality and efficiency at low water levels.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an intelligent control method for low water level operation of pipe networks based on the identification of external water intrusion hotspots, the method comprising: S1: Install liquid level-water quality linkage sensing devices on the inspection well sections in the hot spots of external water intrusion, and simultaneously collect liquid level time series data and conductivity time series data characterizing water quality concentration on the pipe network section, and encapsulate the collected data into binary protocol data packets and transmit them to the edge computing node through the industrial Internet of Things gateway; S2: Denoise and extract features from liquid level time series data and conductivity time series data at edge computing nodes. Based on the negative correlation between liquid level fluctuation and conductivity dilution effect, identify external water or groundwater intrusion behavior and calculate the real-time external water intrusion intensity index. S3: Input the real-time external water intrusion intensity index into the preset intelligent decision mapping model. This model is generated based on the nonlinear mapping relationship between the external water intrusion intensity and the pumping frequency of the pumping station, and outputs the target pumping frequency command. S4: The pump station PLC controller receives the pumping frequency command of the target pump station, adjusts the speed of the variable frequency pump set, and drives the water level of the pipeline network to drop below the design fullness of the pipeline network during the dry season, using the low water level operation condition to expose the structural defects of the pipeline network. S5: During operation, when an abnormal increase in the conductivity of the pipeline network or an increase in the concentration of chemical oxygen demand is detected, it is determined that the infiltration of groundwater or external water has decreased. The system locks the current or lower low water level operation target and maintains the low water level strategy to continuously increase the concentration of sewage collection.

[0007] In the preferred embodiment, step S2, which involves detecting external water or groundwater intrusion and denoising the data, employs a variational mode decomposition algorithm, including: Set the number of modes in the variational mode decomposition and the quadratic penalty factor; The liquid level time series data and the conductivity time series data are decomposed into several intrinsic mode function components respectively; Calculate the correlation coefficient between each intrinsic mode function component and the original signal, and remove random noise components with correlation coefficients lower than a preset threshold and high-frequency oscillation components caused by the start and stop of the pump station; The remaining effective intrinsic mode function components are subjected to Hilbert transform to reconstruct pure liquid level trend signals and conductivity trend signals, which are used as basic characteristic data for identifying external water intrusion.

[0008] In the preferred scheme, the calculation of the real-time external water intrusion intensity index in step S2 adopts a fusion algorithm based on mutual information entropy: Calculate the mutual information value of the pure liquid level trend signal and the pure conductivity trend signal within the sliding time window to quantify the coupling degree between the increase in liquid level and the decrease in concentration. Construct a feature evaluation matrix that includes normalized values ​​of liquid level amplitude, reciprocal normalized values ​​of conductivity, and mutual information values; The information entropy and objective weight of each indicator in the feature evaluation matrix are calculated using the entropy weight method. Based on objective weights, each indicator is weighted and summed to generate a real-time external water intrusion intensity index between 0 and 1. The higher the index, the greater the amount of external water or groundwater intrusion, and the higher the risk of the current high water level masking defects.

[0009] In the preferred embodiment, the intelligent decision mapping model in step S3 is constructed using a long short-term memory network that incorporates an attention mechanism: The model input layer receives a feature vector composed of the real-time external water intrusion intensity index, pipeline liquid level, and pump station current frequency at the current time and the past N time steps. Long Short-Term Memory (LSTM) network layers extract the temporal dependencies of feature vectors to capture the lag effect of external water intrusion processes; The attention mechanism layer calculates the contribution weights of features at different time steps to the current decision state, with a focus on enhancing the feature weights at moments of sudden changes in conductivity. The fully connected layer outputs the predicted optimal pumping frequency for the next moment, enabling an adaptive response to the dynamic changes in external water intrusion.

[0010] In the preferred scheme, the construction and training process of the intelligent decision mapping model adopts a deep deterministic policy gradient reinforcement learning algorithm: Construct a reinforcement learning environment that includes a state space, an action space, and a reward function; The state space is defined as the current pipe network liquid level, conductivity, and real-time external water intrusion intensity index; The operating space is defined as the frequency adjustment step size of the pump station frequency converter; The design reward function provides a positive reward when the pipeline liquid level is lower than the design fullness and the conductivity or chemical oxygen demand concentration increases, and a negative penalty when overflow occurs or the pump station energy consumption increases sharply. The Actor-Critic network structure continuously interacts and iterates with the environment, updating network parameters until the model's output strategy maximizes the cumulative reward.

[0011] In the preferred embodiment, this method employs an edge-cloud collaborative microservice mesh architecture for deployment, and the steps include: S61: On the edge computing node, a process isolation environment is built using the Linux kernel namespace technology, and the resource usage quota of each process is limited using control group technology. In this isolation environment, a signal processing computing instance compiled based on the local instruction set is run directly. This instance has an embedded linear algebra operation library, which is used to execute the iterative solution logic of the alternating direction multiplier method of the variational mode decomposition algorithm in step S2, avoiding the performance loss caused by the virtualization layer. S62: Deploy the service mesh control plane on the cloud server, inject the Sidecar proxy container into the Pods of the intelligent decision mapping model service and data storage service, take over all network communication traffic between services, and configure circuit breaker and degradation policies to prevent cascading failures. S63: Edge computing nodes communicate with cloud servers via the gRPC high-performance remote procedure call protocol. It defines Protocol Buffers files to describe interface data structures, serializes the collected feature data into binary streams for transmission, and achieves millisecond-level data transmission and command issuance.

[0012] In the preferred embodiment, the data storage service is deployed using a distributed time-series database based on the LSM-Tree storage engine. The specific implementation steps include: S71: Construct a columnar storage structure with timestamp and sensor ID as the key, and first write the real-time collected liquid level, conductivity and pump station frequency data into the MemTable table in memory. S72: When the amount of data in the MemTable reaches a preset threshold, it is frozen and converted into an immutable SSTable file, which is then sequentially written to the disk storage area to improve the throughput during high-concurrency writes by utilizing the sequential disk write characteristics. S73: Periodically performs hierarchical compression and merging operations on SSTable files on the disk, uses Bloom filters to quickly determine if data exists, cleans up redundant data marked for deletion, and performs downsampling aggregation on historical millisecond-level data exceeding 3 months to convert it into minute-level statistics to optimize storage space.

[0013] In the preferred embodiment, the execution logic of the pump station PLC controller in step S4 includes a safety boundary constraint module: After receiving the pumping frequency command from the target pumping station, the safety boundary constraint module is first invoked to verify the command value; If the command value causes the water level in the pump station's suction pool to fall below the minimum flooding depth, the command value will be forcibly corrected to shutdown or minimum frequency operation to prevent cavitation. If the rate of change of the command value exceeds the maximum allowable acceleration / deceleration slope of the inverter, the command is smoothed and filtered, and the frequency setting value is issued in a step-by-step manner.

[0014] In the preferred embodiment, the method also includes an automatic switching step for flood prevention mode during the rainy season: The system obtains gridded rainfall forecast data for the next two hours from the meteorological department in real time through the API interface; When the rainfall forecast data exceeds the preset flood control threshold, or when a sudden increase in the reading of the on-site rain gauge is detected, the low water level concentration logic based on the intelligent decision mapping model in step S3 is interrupted. Automatically switch to flood control mode, control the pumping station to operate at full load, lower the water level in the pipeline network to near the bottom elevation of the pipeline, free up the storage capacity of the pipeline network, and ensure the safety of overflow.

[0015] In the preferred embodiment, the specific execution logic of the low water level strategy in step S5 includes: S101: Set the threshold for the positive rate of change of conductivity and the threshold for the positive rate of change of chemical oxygen demand; S102: During the process of pumping up and lowering the water level at the pumping station, the first-order difference values ​​of conductivity and chemical oxygen demand are calculated in real time using a sliding window method. S103: If the first-order difference value is positive and exceeds the set threshold, a frequency lock command is generated to freeze the current PID control parameters, prohibit the pump station frequency from being lowered, and force the pump station to maintain the current high-frequency pumping state until the liquid level reaches the physical limit or the conductivity no longer increases, thereby confirming that the external water intrusion has been suppressed to the greatest extent.

[0016] This invention provides an intelligent control method for low-water-level operation of pipe networks based on the identification of external water intrusion hotspots. This application employs a method that deploys level-water quality linkage sensors in areas prone to external water intrusion and combines edge computing with cloud-based decision-making to fundamentally transform the pipe network operation mode. By collecting real-time level and conductivity data and using variational mode decomposition algorithms to remove pump station start-up and shutdown interference and random noise, combined with mutual information entropy algorithms to accurately quantify the intensity of external water intrusion, the system can accurately identify intrusion behaviors masked by high water levels. Based on this, by establishing a nonlinear mapping model between the intensity of external water intrusion and the pumping frequency, the system no longer relies on a single fixed level control but dynamically adjusts the pumping frequency according to the real-time intensity of external water intrusion, proactively lowering the pipe network water level below the design fill level during the dry season. This low-water-level operation strategy directly reduces the pressure difference between the inside and outside of the pipe network, significantly reducing groundwater infiltration and thus increasing the collection concentration of wastewater. Furthermore, the low water level exposes defects such as pipe joints and cracks, providing favorable conditions for maintenance personnel to inspect and repair them.

[0017] This application features in-depth optimization of the control algorithm and system architecture, significantly improving the system's response speed and decision-making level. At the algorithm level, a Long Short-Term Memory (LSTM) network with an attention mechanism is introduced to capture the time-delay effect of external water intrusion, and a Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm is used for model training. This enables the control strategy to adapt to complex nonlinear changes in the pipeline network, avoiding the lag and oscillation of traditional PID control under complex operating conditions. In particular, when an abnormal increase in conductivity or COD concentration is detected, the system can intelligently lock onto a low water level target and continuously suppress external water intrusion through a proactive strategy of "maintaining a low water level," ensuring the achievement of the quality and efficiency improvement goals.

[0018] This application employs advanced computer implementation schemes for software deployment and data storage, ensuring the system's efficient and stable operation. It adopts an edge-cloud collaborative architecture, utilizing Linux namespaces and control group technology to build a lightweight isolated environment at the edge, directly running high-performance computing instances compiled based on local instruction sets. This avoids the performance loss of traditional virtualization layers and ensures the real-time performance of feature extraction at the edge. At the cloud, a microservice mesh governance architecture and a distributed time-series database based on the LSM-Tree storage engine are used. Through MemTable memory table buffering and SSTable disk sequential write mechanisms, the system effectively handles the high-concurrency write requirements of massive sensor data, and optimizes query and storage efficiency using layered compression and Bloom filters. Furthermore, the system integrates safety boundary constraints and automatic switching functions for flood prevention during the rainy season. This prevents pump station cavitation and frequent start-ups and shutdowns, while prioritizing the overcurrent safety of the pipeline network during the rainy season, resolving the conflict between flood prevention and concentration enhancement. This approach has high practical value and promising prospects for widespread application. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the system architecture of this invention; Figure 2 This is the logic control flowchart of the present invention; Figure 3 This is a diagram showing the control interface of the software interface of this invention. Detailed Implementation

[0020] Example 1 like Figure 1-3 As shown, a smart control method for low-water-level operation of a pipe network based on the identification of external water intrusion hotspots is presented. The method includes: S1: Install liquid level-water quality linkage sensing devices on the inspection well sections in the hot spots of external water intrusion, and simultaneously collect liquid level time series data and conductivity time series data characterizing water quality concentration on the pipe network section, and encapsulate the collected data into binary protocol data packets and transmit them to the edge computing node through the industrial Internet of Things gateway; S2: Denoise and extract features from liquid level time series data and conductivity time series data at edge computing nodes. Based on the negative correlation between liquid level fluctuation and conductivity dilution effect, identify external water or groundwater intrusion behavior and calculate the real-time external water intrusion intensity index. S3: Input the real-time external water intrusion intensity index into the preset intelligent decision mapping model. This model is generated based on the nonlinear mapping relationship between the external water intrusion intensity and the pumping frequency of the pumping station, and outputs the target pumping frequency command. S4: The pump station PLC controller receives the pumping frequency command of the target pump station, adjusts the speed of the variable frequency pump set, and drives the water level of the pipeline network to drop below the design fullness of the pipeline network during the dry season, using the low water level operation condition to expose the structural defects of the pipeline network. S5: During operation, when an abnormal increase in the conductivity of the pipeline network or an increase in the concentration of chemical oxygen demand is detected, it is determined that the infiltration of groundwater or external water has decreased. The system locks the current or lower low water level operation target and maintains the low water level strategy to continuously increase the concentration of sewage collection.

[0021] This implementation first constructs a sensing network at the physical deployment level. In hotspot areas with a high risk of external water intrusion, specifically, level and water quality linkage sensors are deployed at the cross-sectional locations of inspection wells. These linkage devices achieve strict time synchronization, simultaneously collecting time-series data on water level at the pipeline cross-section and time-series data on conductivity, which characterizes water quality concentration. To improve the real-time performance and bandwidth utilization of data transmission, the collected raw data is not transmitted in plaintext but is encapsulated into binary protocol data packets through an industrial IoT gateway. This binary format compresses the data volume, and the data is then stably transmitted to edge computing nodes through the industrial IoT gateway, ensuring data integrity and low-latency arrival in the complex underground pipeline network communication environment.

[0022] Once the data reaches the edge computing node, the system performs denoising and feature extraction operations. The core basis for this step lies in the significant differences in the physicochemical properties of urban sewage and intrusive water bodies. Specifically, urban sewage typically has high conductivity, while external water such as rainwater, river water, or shallow groundwater has relatively low conductivity. Therefore, when external water intrusion occurs, the liquid level in the pipe network tends to rise, while the conductivity of the mixed sewage tends to decrease, exhibiting a significant negative correlation over time. To quantify this characteristic, the edge computing node uses the Pearson correlation coefficient formula to calculate the correlation between liquid level and conductivity, as follows: ; In the above formula, The Pearson correlation coefficient represents the relationship between the liquid level data sequence and the conductivity data sequence, and its value ranges from negative one to positive one. Representing the Liquid level measurements at each time point This represents the arithmetic mean of the liquid level within that time window. Representing the Conductivity measurements at various time points This represents the arithmetic mean of the conductivity within that time window. This represents the total number of data sampling points within the sliding time window. When the correlation coefficient is calculated... When the level approaches negative one and the rate of change of the liquid level is positive, the system determines that there is a strong external water intrusion.

[0023] Based on the identification of intrusion behavior, the system further calculates a real-time external water intrusion intensity index, which serves as the direct basis for subsequent control measures. The calculation model for the real-time external water intrusion intensity index employs a weighted fusion algorithm, normalizing the fluctuation range of the liquid level and the degree of dilution of conductivity before synthesis. Real-time external water intrusion intensity index The calculation formula is as follows: ; In this formula, express Real-time external water intrusion intensity index at any given moment. This represents the measured liquid level at the current moment. and These are the highest and lowest recorded liquid levels for this pipe section during its historical operating cycle, respectively. This item is used to characterize the relative height of the liquid level. The measured conductivity at the current moment. This is the baseline conductivity of domestic sewage at the minimum flow rate during the dry season at night in this region. This parameter is used to characterize the degree to which the water quality is diluted. and These are the weighting coefficients for the liquid level factor and the water quality factor, respectively, and their sum equals one. Using this formula, the system can integrate the physical dimension of liquid level rise and the chemical dimension of concentration decrease into a dimensionless intensity index.

[0024] After obtaining the real-time external water intrusion intensity index, the system inputs it into a pre-set intelligent decision mapping model. This model, during its training phase, has learned the nonlinear relationship between the external water intrusion intensity and the optimal pumping frequency of the pumping station under historical operating conditions. Based on the currently input intensity index, the model infers a target pumping frequency command that can effectively suppress the external water intrusion. This command aims to break away from the traditional fixed-level start-stop mode and instead adopt a dynamic response mode.

[0025] After receiving the target pumping frequency increase command, the PLC controller of the pumping station executes specific control actions by adjusting the speed of the variable frequency pump set. During the dry season, the core objective of the control strategy is to drive the operating water level of the pipeline network down to below the design fullness of the pipeline. The design fullness refers to the ratio of the cross-sectional area of ​​the water flow in the pipeline to the total cross-sectional area of ​​the pipeline. Controlling the water level below this threshold creates a non-full pipe flow state. This low water level operation not only reduces the pressure of sewage leakage from the pipeline, but more importantly, it exposes structural defects such as pipe joints and cracks that were originally submerged to the air or within the visible range, providing physical conditions for subsequent pipeline inspection and repair. The PLC controller uses an incremental PID algorithm to accurately approximate the target frequency, and its control law is as follows: ; In the formula, Indicates the first The incremental output value of the inverter frequency at each sampling moment. This represents the deviation between the target water level and the actual water level at the current moment. and These are the deviation values ​​from the previous moment and the two moments before, respectively. This is a proportionality coefficient that determines the response speed. These are integral coefficients used to eliminate steady-state errors. The differential coefficient is used to suppress overshoot. This algorithm allows the pump station to smoothly adjust its speed, avoiding drastic fluctuations in the pipe network level.

[0026] During continuous system operation, a closed-loop feedback regulation mechanism is formed. When the monitoring system detects an abnormal increase in the network conductivity or an upward trend in the chemical oxygen demand (COD) concentration, this physical phenomenon indicates an increase in the proportion of raw sewage in the mixed wastewater, meaning that the infiltration of external clean water, i.e., groundwater or external water, is decreasing. At this point, the system determines that the current low-water-level operation strategy is effective and triggers a locking logic to lock the current low-water-level operation target or automatically set a lower target water level value. By maintaining this high-intensity low-water-level pumping strategy, the system continuously reduces the pressure within the network to inhibit the infiltration of external groundwater, thereby continuously increasing the collection concentration of wastewater and ensuring that the influent water quality of the wastewater treatment plant meets the standards.

[0027] Example 2 In step S2, the detection of external water or groundwater intrusion and the data denoising adopt variational mode decomposition algorithm, including: setting the number of modes of variational mode decomposition and the quadratic penalty factor; The liquid level time series data and the conductivity time series data are decomposed into several intrinsic mode function components respectively; Calculate the correlation coefficient between each intrinsic mode function component and the original signal, and remove random noise components with correlation coefficients lower than a preset threshold and high-frequency oscillation components caused by the start and stop of the pump station; The remaining effective intrinsic mode function components are subjected to Hilbert transform to reconstruct pure liquid level trend signals and conductivity trend signals, which are used as basic characteristic data for identifying external water intrusion.

[0028] The detection and data denoising stage for external water or groundwater intrusion employs a variational mode decomposition (VMD) algorithm. This stage aims to extract the true trends in liquid level and water quality changes from complex signals affected by pump station start-up and shutdown interference and random noise. The system first initializes the core parameters of VMD, including the number of modes determining the number of decomposition layers and a quadratic penalty factor used to balance data fidelity and the bandwidth of the decomposed signal. The number of modes is typically set based on the spectral complexity of the signal, while the quadratic penalty factor determines the bandwidth of the mode components; a larger penalty factor ensures the accuracy of the reconstructed signal. The mathematical essence of VMD is to construct and solve a constrained variational problem whose objective is to decompose the original signal into a specified number of eigenmode function components such that the sum of the estimated bandwidths of all components is minimized. The constraint is that the sum of all components equals the original signal. The mathematical expression of this constrained variational problem is as follows: ; In the above formula, The decomposition yields the first Each intrinsic mode function component The preset number of modes. Representing the The center frequency of each component. This is the Dirac function, also known as the unit impulse function. The imaginary unit is represented by the asterisk (*). The asterisk (*) indicates a convolution operation. Indicates time Find the partial derivative. (Term) Represents the components of the intrinsic mode function Perform a Hilbert transform and construct an analytic signal. This is the exponential term, used to shift the spectrum of each mode to the baseband. This represents the square of the L2 norm, used to measure the bandwidth of a signal. The physical meaning of the entire formula is to ensure that all sub-signals decomposed are within acceptable limits. Under the premise of being able to reconstruct the original signal, find a set of center frequencies. And the corresponding sub-signals with limited bandwidth make the frequency of each sub-signal pass through the most compact way.

[0029] To solve the constrained variational problem, the system introduces Lagrange multipliers and a quadratic penalty term, transforming it into an unconstrained variational problem, and then uses the alternating direction multiplier method for iterative solution. During the iteration process, the liquid level time series data and conductivity time series data are decomposed separately, yielding a series of intrinsic mode function components with different center frequencies. Changes in water level and conductivity caused by external water intrusion typically exhibit slow trend terms, belonging to low-frequency signals, while drastic fluctuations in liquid level caused by pump station start-up and shutdown, as well as random measurement errors of the sensors themselves, manifest as high-frequency oscillation signals or irregular white noise. To eliminate these invalid components, the system calculates the Pearson correlation coefficient between each intrinsic mode function component and the original input signal. The formula for calculating the correlation coefficient is as follows: ; In this formula, Indicates the first The correlation coefficients between the intrinsic mode function components and the original signal. For the first A sequence of intrinsic mode function components. This is the original time-series data sequence of liquid level or conductivity. and These are the mean values ​​of the component sequence and the original signal sequence, respectively. and These are the standard deviations of the component sequences and the original signal sequences, respectively. This represents the expected value of a mathematical operation. The system will calculate the expected value. Compared with the preset filtering threshold, when If the value is below the threshold, the component is determined to be random noise or high-frequency interference and is removed from the signal reconstruction set.

[0030] After removing invalid components, the system performs Hilbert transform processing on the remaining valid intrinsic mode function components. The Hilbert transform is used to construct the analytical form of the signal, thereby accurately extracting the instantaneous amplitude and frequency of the signal and eliminating phase ambiguity caused by spurious components. For each retained valid component... Its Hilbert transform is defined as: ; In the formula, Represents the Hilbert transform operator. The variable is the integral variable. Through the Hilbert transform, the real signal can be extended to the complex plane, thus more accurately describing the signal's envelope characteristics. Finally, the system linearly superimposes and reconstructs all the effective components after filtering and transformation, obtaining a pure liquid level trend signal and conductivity trend signal that have been freed from high-frequency noise and pump station oscillation interference. These two reconstructed trend signals not only retain the slow change characteristics caused by external water intrusion but also greatly improve the signal-to-noise ratio and accuracy of subsequent calculations of the real-time external water intrusion intensity index.

[0031] In the preferred scheme, the calculation of the real-time external water intrusion intensity index in step S2 adopts a fusion algorithm based on mutual information entropy: Calculate the mutual information value of the pure liquid level trend signal and the pure conductivity trend signal within the sliding time window to quantify the coupling degree between the increase in liquid level and the decrease in concentration. Construct a feature evaluation matrix that includes normalized values ​​of liquid level amplitude, reciprocal normalized values ​​of conductivity, and mutual information values; The information entropy and objective weight of each indicator in the feature evaluation matrix are calculated using the entropy weight method. Based on objective weights, each indicator is weighted and summed to generate a real-time external water intrusion intensity index between 0 and 1. The higher the index, the greater the amount of external water or groundwater intrusion, and the higher the risk of the current high water level masking defects.

[0032] In step S2, to more accurately quantify the severity of external water intrusion, the system employs a fusion algorithm based on mutual information entropy to calculate the real-time external water intrusion intensity index. This process is based on the pure liquid level trend signal and pure conductivity trend signal reconstructed from the aforementioned variational mode decomposition. Mutual information entropy, as an information theory indicator, can effectively measure the nonlinear dependency between two random variables. The system first sets a sliding time window of fixed length, within which it extracts the pure liquid level trend signal sequence and the pure conductivity trend signal sequence. To calculate the mutual information value, the system uses histogram statistics or kernel density estimation to estimate the marginal probability distribution of the two signal sequences and their joint probability distribution. The formula for calculating the mutual information value is expressed as follows: ; In the above formula, Represents the trend signal of pure liquid level With pure conductivity trend signal The mutual information values ​​between them. and These represent the value space of the liquid level signal and the conductivity signal within the current time window, respectively. The liquid level signal value is indicated. The marginal probability, The conductivity signal is represented by a value of The marginal probability. Indicates the liquid level value And the conductivity is taken as The joint probability. Usually, the base is 2 or the natural constant is taken. The mutual information value is the logarithm to the base. This mutual information value quantifies the degree of coupling between the two physical processes of liquid level rise and concentration decrease over time. A larger mutual information value indicates a stronger correlation between liquid level fluctuations and conductivity changes, and a higher synchronicity between the water level rise caused by external water intrusion and the concentration dilution effect.

[0033] After obtaining the mutual information value, the system constructs a multi-dimensional feature evaluation matrix to comprehensively assess the intrusion intensity from different dimensions. This matrix contains three core indicators: normalized liquid level amplitude, normalized inverse conductivity, and the calculated mutual information value. Liquid level amplitude directly reflects the occupancy of the pipe network capacity by external water intrusion; the normalized inverse conductivity characterizes the degree of water dilution, as lower conductivity implies a higher proportion of external water, and its inverse shows a positive correlation; the mutual information value represents the confidence level of the correlation between the two. Feature Evaluation Matrix The first in The first indicator in the The values ​​of each sampling point, after standardization, are denoted as follows: To eliminate the influence of dimensions, each indicator is mapped to the range of zero to one.

[0034] Subsequently, the system objectively calculates the weights of these three indicators using the entropy weight method to avoid subjective bias from manually setting the weights. The core idea of ​​the entropy weight method is to use information entropy to measure the dispersion of the indicators; the greater the dispersion of an indicator, the more information it provides, and the greater its corresponding weight should be. First, the weights of the three indicators are calculated. Information entropy of each indicator The calculation formula is as follows: ; In the formula, Indicates the first Information entropy of each indicator This represents the total number of sample points within the sliding window. Indicates the first The first indicator in the The proportion of the value of a sampling point to the sum of the values ​​of all sampling points of the indicator, i.e. .constant This is an adjustment coefficient used to ensure that the information entropy value is between zero and one. If If it is zero, then define It is zero.

[0035] After calculating the information entropy of each indicator, the system further calculates the objective weight of each indicator. The weights are calculated based on the redundancy of information entropy, using the following formula: ; In the formula, For the first The weighting coefficients for each indicator are calculated, with the denominator being the sum of the information entropy redundancy of all three indicators. This formula ensures that the sum of the weights of all indicators equals one. In this way, the system can dynamically allocate weights based on the fluctuation characteristics of the current data. For example, when the liquid level is high but fluctuates little, while the conductivity drops sharply, the weight of conductivity-related indicators will be automatically increased, thus more sensitively detecting intrusion events.

[0036] Finally, based on the calculated objective weights, the various characteristic indicators are weighted and summed to generate the final real-time external water intrusion intensity index. The calculation formula is as follows:

[0037] In the formula, Represents the normalized value of the liquid level amplitude. This represents the normalized value of the inverse of conductivity. This represents the normalized value of the mutual information value. , , These are the corresponding objective weights. (Generated) It is a value between zero and one. When the index approaches one, it indicates that the current intrusion of external water or groundwater is extremely large, and due to the high water level, the risk of pipeline defects being covered is extremely high. The system needs to immediately adopt a high-intensity pumping strategy to lower the water level. When the index approaches zero, it indicates that the pipeline network is operating relatively smoothly, with no obvious signs of external water intrusion.

[0038] In the preferred embodiment, the intelligent decision mapping model in step S3 is constructed using a long short-term memory network that incorporates an attention mechanism: The model input layer receives a feature vector composed of the real-time external water intrusion intensity index, pipeline liquid level, and pump station current frequency at the current time and the past N time steps. Long Short-Term Memory (LSTM) network layers extract the temporal dependencies of feature vectors to capture the lag effect of external water intrusion processes; The attention mechanism layer calculates the contribution weights of features at different time steps to the current decision state, with a focus on enhancing the feature weights at moments of sudden changes in conductivity. The fully connected layer outputs the predicted optimal pumping frequency for the next moment, enabling an adaptive response to the dynamic changes in external water intrusion.

[0039] In the specific implementation of step S3, the intelligent decision mapping model is constructed as a deep neural network architecture, the core of which consists of an input layer, a long short-term memory network layer, an attention mechanism layer, and a fully connected output layer cascaded together. The model input layer is responsible for receiving and preprocessing time-series data; specifically, it receives data including the current time and past data. A sequence of feature data at each time step. A feature vector for each time step. It consists of physical quantities in three key dimensions: the real-time external water intrusion intensity index calculated from the preceding steps. Real-time liquid level at pipeline cross-section and the current operating frequency of the pumping station The mathematical expression of this eigenvector is as follows: ; In the above formula, the subscript Represents a discrete time step index, with values ​​ranging from arrive ,in This refers to the current moment. This represents the vector transpose operation. These continuous... The feature vectors of each time step together form a feature matrix, which serves as the input tensor for subsequent neural network layers. This aims to provide the model with a sufficiently long historical perspective to capture the dynamic evolution of the pipeline network's operating status.

[0040] The Long Short-Term Memory (LSTM) network layer aims to address the vanishing or exploding gradient problems that traditional recurrent neural networks (RNNs) often encounter when processing long sequences of data, thereby effectively extracting the long-term dependencies of feature vectors over time. This layer introduces a unique memory unit structure, using forget gates, input gates, and output gates to finely control the flow of information. To avoid conflict with the letter representing conductivity in the preceding steps... And letters representing liquid level This can cause confusion; therefore, the following approach is adopted. Representing the internal state of a memory cell, using This indicates the output state of the hidden layer. Long Short-Term Memory (LSTM) network units in... The time-based update calculation process is described by the following set of formulas: ; ; ; ; ; ; In the above set of formulas, This represents the Sigmoid activation function, which outputs a gating coefficient between zero and one. The Hadamard product represents the matrix multiplication operation, which is an element-wise multiplication operation. , , These represent the activation vectors of the forget gate, input gate, and output gate, respectively, which determine the proportion of historical information retained and the degree to which new input information is accepted. , , , This is the corresponding weight matrix. , , , This is the corresponding bias vector. Through this recursive process, the model can capture the hysteresis effect of external water intrusion events on the pipe network level and concentration from the sequence data. That is, the current state depends not only on the current input, but also on the cumulative rainfall or infiltration behavior over a period of time.

[0041] The attention mechanism layer, located after the Long Short-Term Memory (LSTM) network layer, plays a crucial role in dynamically calculating the contribution weights of hidden state features at different time steps to the current decision state. Since external water intrusion is often accompanied by abrupt changes in conductivity, these key moment features are vital for decision-making. The attention mechanism automatically identifies and focuses on these critical time points, assigning them greater weight while suppressing noise interference from irrelevant time steps. The attention weights are calculated using a scoring function mechanism, as shown in the following formula: ; ; In the formula, Representing historical moments Hidden state With the current moment Hidden state The correlation score between them. , It is a learnable linear transformation matrix. For parameter vectors. These are attention weight coefficients normalized using the Softmax function; these coefficients directly reflect historical moments. The importance of the features to the generation of current regulatory instructions. Based on the calculated weights, the hierarchical structure further generates contextual semantic vectors. : ; Finally, the fully connected layer receives the context semantic vector generated by the attention mechanism. This is mapped to the final output space, which predicts the optimal pumping frequency for the next moment. This process is achieved through a nonlinear transformation, designed to fit the highly complex hydraulic characteristics of the pipe network. The output calculation formula is: ; In the formula, This is the pumping frequency command for the target pumping station at the next moment, output by the model. and These are the weight matrix and bias term of the fully connected layer, respectively. The activation function for the output layer is typically a ReLU or Sigmoid function, linearly scaled according to the inverter's physical frequency range. This frequency command not only considers the current instantaneous state but also incorporates attention-weighted historical trend information, thereby achieving an adaptive response to dynamic changes in external water intrusion. This ensures that the pumping station maintains low water level operation while avoiding cavitation or energy waste due to overreaction.

[0042] In the preferred scheme, the construction and training process of the intelligent decision mapping model adopts a deep deterministic policy gradient reinforcement learning algorithm: Construct a reinforcement learning environment that includes a state space, an action space, and a reward function; The state space is defined as the current pipe network liquid level, conductivity, and real-time external water intrusion intensity index; The operating space is defined as the frequency adjustment step size of the pump station frequency converter; The design reward function provides a positive reward when the pipeline liquid level is lower than the design fullness and the conductivity or chemical oxygen demand concentration increases, and a negative penalty when overflow occurs or the pump station energy consumption increases sharply. The Actor-Critic network structure continuously interacts and iterates with the environment, updating network parameters until the model's output strategy maximizes the cumulative reward.

[0043] In the construction and training phase of the intelligent decision mapping model, the system employs the Deep Deterministic Policy Gradient (DDPG) algorithm. This algorithm is a deep reinforcement learning method based on the Actor-Critic architecture, capable of effectively handling decision problems in high-dimensional continuous action spaces. In its implementation, the first step is to construct a reinforcement learning interactive environment that includes a state space, an action space, and a reward function.

[0044] State space is the foundation for an intelligent agent's perception of its environment. In this implementation, the state space consists of three physical quantities: the current pipeline liquid level, conductivity, and the real-time external water intrusion intensity index. The state vector at time t is Its mathematical expression is as follows: ; In the formula, This indicates the real-time liquid level height of the pipeline cross-section at the current moment. This represents the conductivity value monitored at the current moment. This is the real-time external water intrusion intensity index calculated in the previous steps. This represents the transpose operation of a vector. These three components together describe the current hydraulic and water quality status of the pipe network, as well as the intensity of external disturbances.

[0045] The action space defines the control behaviors that an agent can exert on its environment. Given that the pump station frequency converter requires fine and continuous adjustment, the action space is set as the frequency adjustment step size of the pump station frequency converter, denoted as... This action determines whether the pump station's operating frequency will increase or decrease relative to the current frequency at the next moment. The value range is physically constrained within the maximum allowable acceleration / deceleration rate range of the frequency converter, i.e. The pump station frequency at the next moment. From the current frequency With action The result of superposition is: ; The reward function is the core signal guiding the agent to learn the optimal strategy. Its design aims to balance multiple objectives, including low water level operation, concentration enhancement, overflow prevention safety, and energy conservation. It consists of a weighted sum of positive reward terms and negative penalty terms, and its piecewise function form is designed as follows: ; In the above formula, This is a positive reward coefficient for concentration increase, used when the pipeline liquid level is below the design fullness height. And the conductivity shows an increasing trend. At that time, the system is encouraged to continue to maintain or optimize the current strategy. The reward coefficient is used to maintain a low water level; the lower the water level, the greater the reward value. This is the overflow penalty coefficient, applied when the liquid level reaches the warning level. At that time, a huge negative penalty is imposed to forcefully avoid overflow accidents. This is the energy consumption penalty coefficient. This represents the current instantaneous power of the pumping station, used to suppress unnecessary high-frequency operation under non-specific operating conditions in order to save energy.

[0046] The core of the deep deterministic policy gradient algorithm lies in the collaborative iteration of the Actor-Critic network structure. The Actor network, or policy network, is responsible for iterating based on the current state. Output deterministic actions Its parameters are denoted as The Critic network, or value network, is responsible for evaluating actions. In state The degree of superiority or inferiority is determined by the output Q value, whose parameter is denoted as... During training, the Critic network minimizes the Bellman error loss function. To update the parameters: ; ; In the formula, The batch size for sampling from the experience playback buffer. For the target Q value, For instant rewards, This is the discount factor. Q' and These are the Critic target network and the Actor target network, respectively. The Actor network updates its parameters by maximizing the Q-value of the Critic network's output, and its objective function is... The gradient is calculated as follows: ; Through continuous interaction and iteration, the Actor network's strategy gradually approaches the global optimum, enabling the pump station control strategy output by the model to maximize long-term cumulative rewards and ultimately achieve the ideal state of the pipeline network operating at low water level and high concentration.

[0047] In the preferred embodiment, this method employs an edge-cloud collaborative microservice mesh architecture for deployment, and the steps include: S61: On the edge computing node, a process isolation environment is built using the Linux kernel namespace technology, and the resource usage quota of each process is limited using control group technology. In this isolation environment, a signal processing computing instance compiled based on the local instruction set is run directly. This instance has an embedded linear algebra operation library, which is used to execute the iterative solution logic of the alternating direction multiplier method of the variational mode decomposition algorithm in step S2, avoiding the performance loss caused by the virtualization layer. S62: Deploy the service mesh control plane on the cloud server, inject the Sidecar proxy container into the Pods of the intelligent decision mapping model service and data storage service, take over all network communication traffic between services, and configure circuit breaker and degradation policies to prevent cascading failures. S63: Edge computing nodes communicate with cloud servers via the gRPC high-performance remote procedure call protocol. It defines Protocol Buffers files to describe interface data structures, serializes the collected feature data into binary streams for transmission, and achieves millisecond-level data transmission and command issuance.

[0048] In specific implementation step S61, to simultaneously ensure high-performance computing and system stability on resource-constrained edge computing nodes, the system employs a lightweight isolation scheme based on native Linux kernel features. This scheme does not rely on traditional hypervisors but directly calls the Linux kernel's namespace mechanism to construct a process isolation environment. The namespace mechanism logically isolates system resources such as process IDs, network protocol stacks, file system mount points, inter-process communication, and hostnames, giving the signal processing computing instances running within it an independent system view, thus achieving an effect similar to an independent operating system. Simultaneously, the system utilizes control group technology for fine-grained resource management of processes within this isolated environment. Control group technology limits CPU time slices, memory usage limits, and disk I / O bandwidth at the kernel level, preventing the entire edge node system from crashing due to sudden high loads from a single computing task. The signal processing computing instances running in this isolated environment are binary executables compiled directly from the native instruction set of the edge node's CPU architecture. These instances statically link to a hardware-optimized linear algebra library, fully utilizing the CPU's single-instruction multiple-data extended instruction set to accelerate matrix operations. The core task of this computational example is to execute the iterative solution logic of the alternating direction multiplier method in the variational mode decomposition algorithm. In each iteration, the frequency domain update formula for the eigenmode function components is as follows: ; In the above formula, Indicates the first After the nth iteration The spectrum of each modal component The spectrum of the original signal. For the spectrum of other modal components, For the spectrum of Lagrange multipliers, As a secondary penalty factor, For the first The center frequency of each mode. Since this update process involves a large number of complex division and accumulation operations, and requires parallel computation of all frequency points in the frequency domain, high-performance computing instances compiled with native instruction sets can avoid the instruction translation overhead of interpreted languages ​​or virtualization layers, significantly improving the iteration convergence speed.

[0049] In implementation step S62, the cloud server employs a microservice mesh architecture to manage complex distributed services. The system deploys an independent service mesh control plane responsible for managing and distributing traffic control policies. In the data plane, the system uses a sidecar proxy pattern, where a lightweight network proxy container is injected into each Pod unit of the intelligent decision mapping model service and data storage service. This sidecar proxy handles all network communication traffic entering and leaving the Pod, allowing business containers to focus solely on core logic without handling complex network interactions. To prevent a cascading avalanche effect caused by a slow or failed microservice, the sidecar proxy is configured with a circuit breaker degradation strategy. This strategy is triggered based on error rate statistics within a sliding time window, and the circuit breaker trigger condition can be described logically as follows: ; In the formula, This represents the request error rate within the current time window. Count the failed requests. For the total number of requests, This is a preset circuit breaker threshold. When the detected error rate exceeds this threshold, the Sidecar agent will immediately cut off the call chain to the faulty service and directly return a preset degraded response or cached data, thereby protecting the upstream service from being blocked.

[0050] In specific implementation step S63, the communication link between the edge computing node and the cloud server adopts the gRPC high-performance remote procedure call protocol based on the HTTP / 2 standard. Unlike traditional text transmission protocols, gRPC utilizes the multiplexing feature of HTTP / 2, allowing multiple requests and responses to be processed concurrently on a single TCP connection, greatly reducing the overhead of network connection establishment. At the data serialization level, the system defines Protocol Buffers files to strictly describe the data structure of the interface. Protocol Buffers uses variable-length encoding technology to serialize the collected liquid level, conductivity, and characteristic data into a compact binary stream. Compared to text formats such as JSON, this binary stream removes redundant information such as field names, significantly reducing the data payload size transmitted over the network. After the serialized binary data stream is transmitted to the cloud over the network, the cloud service can quickly restore the original data object using pre-generated deserialization code. The encoding and decoding efficiency of the entire process is extremely high, thereby achieving millisecond-level data transmission latency and ensuring that the pump station control commands generated in the cloud can be issued and executed in real time.

[0051] In the preferred embodiment, the data storage service is deployed using a distributed time-series database based on the LSM-Tree storage engine. The specific implementation steps include: S71: Construct a columnar storage structure with timestamp and sensor ID as the key, and first write the real-time collected liquid level, conductivity and pump station frequency data into the MemTable table in memory. S72: When the amount of data in the MemTable reaches a preset threshold, it is frozen and converted into an immutable SSTable file, which is then sequentially written to the disk storage area to improve the throughput during high-concurrency writes by utilizing the sequential disk write characteristics. S73: Periodically performs hierarchical compression and merging operations on SSTable files on the disk, uses Bloom filters to quickly determine if data exists, cleans up redundant data marked for deletion, and performs downsampling aggregation on historical millisecond-level data exceeding 3 months to convert it into minute-level statistics to optimize storage space.

[0052] In the specific implementation steps of the data storage service, particularly for step S71, the system first constructs an efficient columnar storage architecture. The core of this architecture lies in its use of a combination key strategy combining timestamps and unique sensor identifiers. In this columnar storage structure, data in the same column is stored contiguously on the physical medium. This is particularly advantageous for time-series data analysis because queries often involve continuously reading a specific indicator, such as liquid level or conductivity, over a period of time, rather than reading an entire row of records. To achieve high-throughput writes, the system maintains a data structure called MemTable in memory. MemTable is typically organized using ordered data structures such as skip lists or red-black trees. All real-time collected data on liquid level, conductivity, and pump station frequency are first written to this memory buffer. Since the random read / write speed of memory is much higher than that of disk, this mechanism greatly reduces write latency, ensuring that the system can handle the peak pressure of massive concurrent data uploads from sensors.

[0053] Following step S72, the system performs a conversion from memory to persistent storage. When the amount of data accumulated in the MemTable reaches a preset storage threshold, such as tens to hundreds of megabytes, the system triggers a freeze operation. At this time, the current MemTable is marked as read-only, and a new MemTable is immediately created to receive subsequent write requests, thus ensuring the continuity of the write service. The frozen read-only MemTable is then serialized and flushed to disk, forming an immutable SSTable file, i.e., a sorted string table file. In the SSTable file, data is stored strictly in key-value order. The key advantage of this flushing process is that it transforms random memory write operations into sequential disk write operations. In traditional mechanical hard drives or solid-state drives, sequential write performance is far superior to random write performance. This design maximizes the disk's input / output bandwidth, thereby significantly improving the overall throughput of the system in high-concurrency write scenarios.

[0054] Subsequently, in step S73, the system periodically performs hierarchical compression and merging operations in the background to maintain the health of the storage system and optimize query performance. Over time, a large number of SSTable files are generated on the disk, requiring the scanning of multiple files when reading data, increasing the risk of read amplification. The hierarchical compression and merging mechanism merges multiple low-level and overlapping SSTable files into a new, high-level and ordered SSTable file. To quickly determine whether a specific key exists in an SSTable file during merging or querying, thereby avoiding unnecessary disk reads, the system introduces a Bloom filter. A Bloom filter is a highly space-efficient probabilistic data structure. Assume the bit array length of the Bloom filter is... The number of hash functions is The number of elements inserted is The false positive rate for determining whether an element exists is... It can be described by the following formula: ; In the above formula, This is the base of the natural logarithm. This formula shows that by appropriately setting the length of the bit array... and the number of hash functions This allows the false positive rate to be kept extremely low. Using a Bloom filter, the system can eliminate non-existent keys within milliseconds, significantly improving query efficiency. Furthermore, during the merging process, the system identifies and thoroughly cleans up redundant data marked for deletion, freeing up storage space. To address the long-term storage costs of historical data, the system implements a downsampling aggregation strategy. For high-frequency, millisecond-level raw data exceeding three months, the system converts it into minute-level statistical data. Assuming the original data sequence is... The aggregation time window is Aggregated minute-level data values The mean aggregation algorithm is used to calculate: ; In the formula, Indicates in The number of raw data points included within the time window. For the first in this window Each original sample value is used. Through this calculation, the data density is significantly reduced, preserving long-term trend characteristics while significantly reducing the storage footprint of historical data, thus achieving a balance between storage costs and query demands.

[0055] In the preferred embodiment, the execution logic of the pump station PLC controller in step S4 includes a safety boundary constraint module: After receiving the pumping frequency command from the target pumping station, the safety boundary constraint module is first invoked to verify the command value; If the command value causes the water level in the pump station's suction pool to fall below the minimum flooding depth, the command value will be forcibly corrected to shutdown or minimum frequency operation to prevent cavitation. If the rate of change of the command value exceeds the maximum allowable acceleration / deceleration slope of the inverter, the command is smoothed and filtered, and the frequency setting value is issued in a step-by-step manner.

[0056] In the specific implementation of step S4, the programmable logic controller (PLC) of the pumping station does not blindly execute control commands from the cloud or edge, but instead has a built-in high-priority safety boundary constraint module. This module is located between the output of the intelligent decision mapping model and the input of the frequency converter, acting as the last line of defense for the physical system. When the PLC receives the command to increase the pumping frequency of the target pumping station, it first calls this module to perform logical verification of the safety and feasibility of the command.

[0057] The first layer of verification involves the minimum submersion depth constraint to prevent cavitation. Cavitation is a major safety hazard in pump station operation. When the water level in the suction tank is too low, the pressure at the pump impeller inlet may drop below the saturated vapor pressure of water, causing water vaporization and the formation of bubbles. When these bubbles burst, they generate a violent impact that damages the impeller. To avoid this, the safety module reads the water level sensor data from the suction tank in real time and compares it with preset physical limit parameters. The corrected safety frequency command... The calculation logic is as follows: ; In the above formula, This represents the measured liquid level height in the suction tank at the current moment. The minimum submersion depth specified by the pump manufacturer, which is the critical liquid level to prevent air from being drawn into the impeller. As a safety margin, it is usually set to 10 to 20 centimeters to offset the risk of misjudgment caused by liquid level fluctuations. This is the minimum permitted operating frequency for the pumping station. This is the original target frequency output by the intelligent model. The formula indicates that once the liquid level falls below the safety threshold, the system will forcibly override the intelligent command, ordering the pump station to shut down or maintain a minimum frequency standby, thereby physically preventing cavitation conditions.

[0058] The second layer of verification involves smoothing and filtering the maximum acceleration / deceleration slope of the frequency converter. Because sudden changes in motor speed generate significant mechanical stress and cause harmonic impacts on the power grid, the frequency converter has strict hardware limitations on the rate of frequency change. If the intelligent model's output command requires a drastic frequency jump within a very short time, the safety module will activate a step-wise smoothing algorithm, decomposing a large frequency jump into several small frequency increments, which are then sent out sequentially according to time steps. Actual output frequency per execution cycle The calculation is as follows: ; In the formula, This is the actual output frequency of the previous control cycle. The target frequency is after liquid level verification. This is the maximum allowable frequency change of the frequency converter within one control cycle, and this value is determined by the acceleration and deceleration time parameters of the frequency converter. This is a cutoff function used to limit the frequency deviation within an allowable slope range. If the difference between the target frequency and the current frequency exceeds... The system will only follow The step size is adjusted. By issuing the set value in this stepwise manner, the pump station's speed change curve will present a smooth slope rather than a steep step, thus effectively protecting the electromechanical equipment and extending its service life.

[0059] In the preferred embodiment, the method also includes an automatic switching step for flood prevention mode during the rainy season: The system obtains gridded rainfall forecast data for the next two hours from the meteorological department in real time through the API interface; When the rainfall forecast data exceeds the preset flood control threshold, or when a sudden increase in the reading of the on-site rain gauge is detected, the low water level concentration logic based on the intelligent decision mapping model in step S3 is interrupted. Automatically switch to flood control mode, control the pumping station to operate at full load, lower the water level in the pipeline network to near the bottom elevation of the pipeline, free up the storage capacity of the pipeline network, and ensure the safety of overflow.

[0060] In the implementation of the automatic switching steps for flood prevention modes during the rainy season, the system constructs a forward-looking meteorological early warning and rapid response mechanism to address the flow safety issues of sewage pipe networks under extreme weather conditions. The system establishes an encrypted connection with a professional database from the meteorological department through a standardized application programming interface (API) to retrieve gridded rainfall forecast data for the next two hours at pre-defined time intervals (e.g., every minute). This gridded data has high spatial resolution and can accurately cover the catchment area served by the pumping stations. After parsing, the acquired data stream is mapped into the hydraulic model of the pipe network to assess the impact of upcoming rainfall on the network load.

[0061] To ensure the timeliness and accuracy of flood control mode switching, the system employs a dual-trigger logic that combines predicted and measured data. On one hand, the system calculates the predicted cumulative rainfall within the future time window; on the other hand, it connects to a high-precision tipping bucket rain gauge deployed on-site to monitor the current instantaneous rainfall intensity in real time. The flood control mode switching signal... Determined by the following logical formula: ; In the above formula, for The mode switching signal at any time triggers flood control mode when its value is 1, and maintains normal operation when its value is 0. Indicates from the current moment To the future The total cumulative rainfall forecast for the period. This is a preset cumulative rainfall flood control threshold, which is set according to the upper limit of the design drainage capacity of the pipe network. This is the real-time reading from the on-site rain gauge. The rate of change of rainfall intensity is called instantaneous rainfall intensity. This is the instantaneous rainfall intensity threshold, used to capture short-term heavy rainfall events. This "OR" logic design ensures that the system's defense mechanism can be activated immediately, whether it's a long-term trend warning from a weather forecast or a sudden localized rainstorm.

[0062] once When the value changes to 1, the system immediately executes an interrupt operation, forcibly terminating the low-water-level concentration logic based on the intelligent decision mapping model in step S3. At this time, the decision weight of the agent is reset to zero, and the target setpoint of the PID controller is bypassed. The system control strategy instantly switches to flood control and drainage mode, issuing a full-load operation command to the pumping station PLC. In this mode, the output frequency of the pumping station frequency converter... It is directly set to the maximum allowed power frequency: ; In the formula, This refers to the rated maximum frequency of the water pump motor. This is an indicator function. This refers to the lowest dead water level in the pipe bottom elevation or suction tank. The goal of this control logic is to utilize the valuable time window before rainfall convergence to pump out the existing sewage in the pipe network as quickly as possible, forcibly lowering the network water level to near the pipe bottom elevation. This operation frees up a significant amount of storage capacity in the pipe network, essentially creating an online storage tank, thus reserving ample buffer space for the upcoming peak flow during the rainy season, minimizing the risk of pipe network overflow, and ensuring the overflow safety of the drainage system.

[0063] In the preferred embodiment, the specific execution logic of the low water level strategy in step S5 includes: S101: Set the threshold for the positive rate of change of conductivity and the threshold for the positive rate of change of chemical oxygen demand; S102: During the process of pumping up and lowering the water level at the pumping station, the first-order difference values ​​of conductivity and chemical oxygen demand are calculated in real time using a sliding window method. S103: If the first-order difference value is positive and exceeds the set threshold, a frequency lock command is generated to freeze the current PID control parameters, prohibit the pump station frequency from being lowered, and force the pump station to maintain the current high-frequency pumping state until the liquid level reaches the physical limit or the conductivity no longer increases, thereby confirming that the external water intrusion has been suppressed to the greatest extent.

[0064] In specific implementation step S5, the core logic of the system's strategy to maintain a low water level lies in verifying and reinforcing the current control effect through real-time monitoring of the dynamic feedback of water quality indicators. In sub-step S101, the system first initializes the operating parameters, setting the thresholds for the positive change rate of conductivity and the positive change rate of chemical oxygen demand (COD). These two thresholds are key critical points for determining whether the pipeline network operation has shifted from "mixing and dilution-dominated" to "raw water concentration-dominated." Conductivity, as a physical quantity reflecting the ion concentration in wastewater, usually indicates a decrease in the proportion of low-conductivity external water or groundwater mixed in when its value increases; COD, as a chemical indicator reflecting the degree of organic pollution in water bodies, also indicates an increase in the raw wastewater collection rate when its concentration increases. The purpose of setting these two thresholds is to filter out conventional water quality fluctuation noise and ensure that the subsequent frequency locking mechanism is triggered only when there is a significant and continuous improvement trend in water quality.

[0065] In sub-step S102, when the pumping station performs high-frequency pumping to lower the water level in the pipe network according to the preceding instructions, the system activates the sliding window calculation engine to process the water quality data in real time. The system maintains a fixed-length time window, within which time-series data of conductivity and chemical oxygen demand are continuously collected. To eliminate the interference of instantaneous measurement noise on trend judgment, the system first calculates the weighted average of the data within the window, and then calculates the first-order difference value based on the smoothed data. Taking conductivity as an example, the formula for calculating its first-order difference value is as follows: ; In the above formula, Indicates time The first-order difference of conductivity reflects the rate of change of conductivity over time. For the current moment The mean conductivity within the sliding window, The previous sampling time The average conductivity within the sliding window. This represents the sampling time interval. Similarly, the first-order difference value of chemical oxygen demand... The same mathematical formula is used for the calculation. This calculation process quantifies in real time the recovery rate of water quality concentration within the pipe network as the water level drops.

[0066] In sub-step S103, the system logically compares the calculated first-order difference value with a preset threshold and executes a control decision accordingly. The specific discrimination and execution logic is as follows: ; In the formula, The threshold for the positive rate of change of conductivity. This is the threshold for the positive rate of change of chemical oxygen demand. When monitored... Greater than or Greater than When the current water level reduction operation significantly reduces the infiltration of external water, the sewage concentration in the pipe network rises rapidly, indicating a significant external water intrusion suppression effect. At this point, the system immediately generates a frequency lock command, forcibly freezing the parameter output of the current PID controller and prohibiting the pump station's frequency converter from performing routine frequency reduction operations based on the liquid level deviation. The pump station is forced to maintain its current high-frequency pumping state, without slowing down even when the liquid level is close to the target value, aiming to continuously maintain the maximum pressure difference between the inside and outside of the pipe network using high-intensity pumping. This lock state will continue until the termination condition is met: either the pipe network liquid level reaches the physical limit minimum water level, or the first-order difference value of the monitored water quality indicator is no longer positive, indicating that the concentration increase has reached the limit equilibrium point under this operating condition. This closed-loop logic ensures that external water intrusion is suppressed to the greatest extent, thereby achieving intelligent deep regulation based on water quality feedback.

[0067] Example 3 To further illustrate this point, in conjunction with Example 1, this example combines... Figure 1 System architecture principles, Figure 2 Logic control flow and Figure 3 The software interface control demonstration details the construction and operation process of an intelligent control system for low water level operation of a pipe network based on the identification of external water intrusion hotspots.

[0068] The physical sensing layer of this system first deploys level-water quality linked sensing devices at the manhole sections of inspection areas identified as hotspots for external water intrusion in the urban sewage pipe network. These devices work in conjunction with ground-based tipping bucket rain gauges to simultaneously collect time-series data on pipe network levels, time-series data on conductivity (characterizing water quality concentration), and real-time rainfall data. To adapt to the complex underground communication environment and improve transmission efficiency, the collected raw data is encapsulated into compact binary protocol data packets on-site and sent to edge computing nodes via an industrial IoT gateway. At the edge computing layer, the system does not rely on traditional heavy virtualization technologies. Instead, it utilizes Linux kernel namespace technology to build a lightweight process isolation environment and combines it with cgroups technology to strictly limit resource quotas. Within this isolated environment, a high-performance C++ signal processing computation instance compiled based on the local CPU instruction set runs directly. This instance embeds a linear algebra arithmetic library specifically designed to execute variational mode decomposition (VMD) algorithms. After the data is received, this calculation instance uses the iterative logic of the alternating direction multiplier method to decompose the liquid level and conductivity data into several intrinsic mode function components, calculates the correlation coefficient to remove pump station start-up and shutdown oscillations and random noise, and reconstructs a pure trend signal through Hilbert transform. Subsequently, the system calculates the real-time external water intrusion intensity index based on the mutual information entropy fusion algorithm. The higher the index, the greater the amount of external water intrusion and the higher the risk of it being masked by high water levels.

[0069] Preprocessed feature data at the edge (including real-time external water intrusion intensity index, pipeline liquid level, pump station frequency, etc.) is transmitted to the cloud decision layer in milliseconds via a high-performance remote procedure call protocol based on HTTP / 2, gRPC, in the form of a binary stream serialized using Protocol Buffers. The cloud server cluster adopts a microservice mesh architecture, with a Sidecar proxy container taking over inter-service traffic and configuring circuit breaker and degradation strategies to ensure high availability. The data is first written to a distributed time-series database based on the LSM-Tree storage engine, using MemTable buffering and SSTable disk sequential write mechanisms to handle high-concurrency writes, and periodically using Bloom filters for layered compression and merging. At the same time, the core model in the intelligent decision service Pod begins to work. This model uses a Long Short-Term Memory (LSTM) network with an attention mechanism to capture the time-lag effect of external water intrusion, and combines it with a Deep Deterministic Policy Gradient (DDPG) reinforcement learning agent for policy optimization, outputting the target pump station pumping frequency command for the next time step.

[0070] After receiving the target frequency command from the cloud, the pump station PLC controller first verifies it using its built-in safety boundary constraint module. The system continuously checks whether the water level in the suction tank is below the minimum submersion depth and whether the required frequency change rate exceeds the maximum allowable acceleration / deceleration rate of the inverter. If the safety boundary is triggered, the system will forcibly correct the command to shutdown, minimum frequency operation, or smoothing filtering to prevent cavitation and equipment damage. After successful verification, the command drives the variable frequency pump set to actively lower the network water level to below the design fullness in dry season mode, using the low water level condition to expose structural defects in the network. During operation, the system executes a closed-loop feedback maintenance strategy, calculating the first-order difference values ​​of conductivity and chemical oxygen demand (COD) in real time using a sliding window method. Once the first-order difference values ​​of these water quality indicators are detected to be positive and exceed preset thresholds, indicating that the low water level strategy effectively suppresses external water, the system immediately triggers frequency locking logic, freezes PID parameters, and forcibly maintains the current high-frequency pumping state until the water level reaches its physical limit or the water quality stops rising, thus achieving continuous concentration. In addition, the system also integrates a flood control mode switching module, which obtains gridded rainfall forecast data for the next two hours from the meteorological department in real time via API. When the predicted rainfall exceeds the threshold or the reading of the on-site rain gauge suddenly increases, the system automatically interrupts the intelligent model logic and switches to flood control mode, controlling the pumping station to operate at full load to free up the reservoir capacity of the pipeline network.

[0071] All the above-mentioned operational statuses and data flows are visualized on the dashboard software interface of the dispatch center. This interface uses a dark-colored GIS map as its core, overlaid with the pipeline topology, and generates a dynamic heat map based on the intrusion intensity index calculated at the edge. Red areas visually indicate high-intensity intrusion hotspots. Operators can click on specific inspection well icons on the GIS map to view real-time data on liquid level and conductivity, as well as trend curves after VMD denoising; clicking on the pump station icon displays the current frequency command interaction status and inverter operating parameters on the right panel. The status bar at the top of the interface displays in real-time whether the system is in "dry season low water level mode," "low water level maintenance strategy," or "flood control and drainage mode," and provides traffic light-style status indications based on meteorological warning information. The alarm information bar at the bottom scrolls to display key events such as "Region C triggered low water level maintenance strategy, frequency locked" or "Meteorological warning detected, about to switch to flood control mode," achieving closed-loop management across the entire chain from bottom-level perception, intelligent decision-making, execution control to human-machine interaction.

[0072] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A smart control method for low-water-level operation of pipe networks based on the identification of external water intrusion hotspots, characterized by: The method includes: S1: Install liquid level-water quality linkage sensing devices on the inspection well sections in the hot spots of external water intrusion, and simultaneously collect liquid level time series data and conductivity time series data characterizing water quality concentration on the pipe network section, and encapsulate the collected data into binary protocol data packets and transmit them to the edge computing node through the industrial Internet of Things gateway; S2: Denoise and extract features from liquid level time series data and conductivity time series data at edge computing nodes. Based on the negative correlation between liquid level fluctuation and conductivity dilution effect, identify external water or groundwater intrusion behavior and calculate the real-time external water intrusion intensity index. S3: Input the real-time external water intrusion intensity index into the preset intelligent decision mapping model. This model is generated based on the nonlinear mapping relationship between the external water intrusion intensity and the pumping frequency of the pumping station, and outputs the target pumping frequency command. S4: The pump station PLC controller receives the pumping frequency command of the target pump station, adjusts the speed of the variable frequency pump set, and drives the water level of the pipeline network to drop below the design fullness of the pipeline network during the dry season, using the low water level operation condition to expose the structural defects of the pipeline network. S5: During operation, when an abnormal increase in the conductivity of the pipeline network or an increase in the concentration of chemical oxygen demand is detected, it is determined that the infiltration of groundwater or external water has decreased. The system locks the current or lower low water level operation target and maintains the low water level strategy to continuously increase the concentration of sewage collection.

2. The intelligent control method for low water level operation of pipe networks based on external water intrusion hotspot identification as described in claim 1, characterized in that: Step S2 The detection and data denoising of external water or groundwater intrusion in China employs variational mode decomposition algorithms, including: Set the number of modes in the variational mode decomposition and the quadratic penalty factor; The liquid level time series data and the conductivity time series data are decomposed into several intrinsic mode function components respectively; Calculate the correlation coefficient between each intrinsic mode function component and the original signal, and remove random noise components with correlation coefficients lower than a preset threshold and high-frequency oscillation components caused by the start and stop of the pump station; The remaining effective intrinsic mode function components are subjected to Hilbert transform to reconstruct pure liquid level trend signals and conductivity trend signals, which are used as basic characteristic data for identifying external water intrusion.

3. The intelligent control method for low water level operation of pipe networks based on external water intrusion hotspot identification as described in claim 2, characterized in that: In step S2, the real-time external water intrusion intensity index is calculated using a fusion algorithm based on mutual information entropy. Calculate the mutual information value of the pure liquid level trend signal and the pure conductivity trend signal within the sliding time window to quantify the degree of coupling between the increase in liquid level and the decrease in concentration. Construct a feature evaluation matrix that includes normalized values ​​of liquid level amplitude, inverse normalized values ​​of conductivity, and mutual information values; The information entropy and objective weight of each indicator in the feature evaluation matrix are calculated using the entropy weight method. Based on objective weights, each indicator is weighted and summed to generate a real-time external water intrusion intensity index between 0 and 1. The higher the index, the greater the amount of external water or groundwater intrusion, and the higher the risk of the current high water level masking defects.

4. The intelligent control method for low water level operation of pipe networks based on external water intrusion hotspot identification as described in claim 1, characterized in that: The intelligent decision mapping model in step S3 is constructed using a long short-term memory network that incorporates an attention mechanism: The model input layer receives a feature vector composed of the real-time external water intrusion intensity index, pipeline liquid level, and pump station current frequency at the current time and the past N time steps. Long Short-Term Memory (LSTM) network layers extract the temporal dependencies of feature vectors to capture the lag effect of external water intrusion processes; The attention mechanism layer calculates the contribution weights of features at different time steps to the current decision state, with a focus on enhancing the feature weights at moments of sudden changes in conductivity. The fully connected layer outputs the predicted optimal pumping frequency for the next moment, enabling an adaptive response to the dynamic changes in external water intrusion.

5. The intelligent control method for low-water-level operation of a pipe network based on the identification of external water intrusion hotspots as described in claim 4, characterized in that: intelligent... The decision mapping model is constructed and trained using a deep deterministic policy gradient reinforcement learning algorithm. Construct a reinforcement learning environment that includes a state space, an action space, and a reward function; The state space is defined as the current pipe network liquid level, conductivity, and real-time external water intrusion intensity index; The operating space is defined as the frequency adjustment step size of the pump station frequency converter; The design reward function provides a positive reward when the pipeline liquid level is lower than the design fullness and the conductivity or chemical oxygen demand concentration increases, and a negative penalty when overflow occurs or the pump station energy consumption increases sharply. The Actor-Critic network structure continuously interacts and iterates with the environment, updating network parameters until the model's output strategy maximizes the cumulative reward.

6. The intelligent control method for low-water-level operation of a pipe network based on the identification of external water intrusion hotspots as described in claim 1, characterized in that: This method employs an edge-cloud collaborative microservice mesh architecture for deployment, and the steps include: S61: On the edge computing node, a process isolation environment is built using the Linux kernel namespace technology, and the resource usage quota of each process is limited using control group technology. In this isolation environment, a signal processing computing instance compiled based on the local instruction set is run directly. This instance has an embedded linear algebra operation library, which is used to execute the iterative solution logic of the alternating direction multiplier method of the variational mode decomposition algorithm in step S2, avoiding the performance loss caused by the virtualization layer. S62: Deploy the service mesh control plane on the cloud server, inject the Sidecar proxy container into the Pods of the intelligent decision mapping model service and data storage service, take over all network communication traffic between services, and configure circuit breaker and degradation policies to prevent cascading failures. S63: Edge computing nodes communicate with cloud servers via the gRPC high-performance remote procedure call protocol. It defines Protocol Buffers files to describe interface data structures, serializes the collected feature data into binary streams for transmission, and achieves millisecond-level data transmission and command issuance.

7. The intelligent control method for low water level operation of a pipe network based on the identification of external water intrusion hotspots as described in claim 6, characterized in that: The data storage service is deployed using a distributed time-series database based on the LSM-Tree storage engine. The specific implementation steps include: S71: Construct a columnar storage structure with timestamp and sensor ID as the key, and first write the real-time collected liquid level, conductivity and pump station frequency data into the MemTable table in memory. S72: When the amount of data in the MemTable reaches a preset threshold, it is frozen and converted into an immutable SSTable file, which is then sequentially written to the disk storage area to improve the throughput during high-concurrency writes by utilizing the sequential disk write characteristics. S73: Periodically performs hierarchical compression and merging operations on SSTable files on the disk, uses Bloom filters to quickly determine if data exists, cleans up redundant data marked for deletion, and performs downsampling aggregation on historical millisecond-level data exceeding 3 months to convert it into minute-level statistics to optimize storage space.

8. The intelligent control method for low water level operation of pipe networks based on external water intrusion hotspot identification as described in claim 1, characterized in that: The execution logic of the pump station PLC controller in step S4 includes a safety boundary constraint module: After receiving the pumping frequency command from the target pumping station, the safety boundary constraint module is first invoked to verify the command value; If the command value causes the water level in the pump station's suction pool to fall below the minimum flooding depth, the command value will be forcibly corrected to shutdown or minimum frequency operation to prevent cavitation. If the rate of change of the command value exceeds the maximum allowable acceleration / deceleration slope of the inverter, the command is smoothed and filtered, and the frequency setting value is issued in a step-by-step manner.

9. The intelligent control method for low water level operation of pipe networks based on external water intrusion hotspot identification as described in claim 1, characterized in that: The method also includes an automatic switching step for flood prevention mode during the rainy season: The system obtains gridded rainfall forecast data for the next two hours from the meteorological department in real time through the API interface; When the rainfall forecast data exceeds the preset flood control threshold, or when a sudden increase in the reading of the on-site rain gauge is detected, the low water level concentration logic based on the intelligent decision mapping model in step S3 is interrupted. Automatically switch to flood control mode, control the pumping station to operate at full load, lower the water level in the pipeline network to near the bottom elevation of the pipeline, free up the storage capacity of the pipeline network, and ensure the safety of overflow.

10. The intelligent control method for low-water-level operation of a pipe network based on the identification of external water intrusion hotspots according to claim 1, characterized in that: The specific execution logic of the low water level maintenance strategy in step S5 includes: S101: Set the threshold for the positive rate of change of conductivity and the threshold for the positive rate of change of chemical oxygen demand; S102: During the process of pumping up and lowering the water level at the pumping station, the first-order difference values ​​of conductivity and chemical oxygen demand are calculated in real time using a sliding window method. S103: If the first-order difference value is positive and exceeds the set threshold, a frequency lock command is generated to freeze the current PID control parameters, prohibit the pump station frequency from being lowered, and force the pump station to maintain the current high-frequency pumping state until the liquid level reaches the physical limit or the conductivity no longer increases, thereby confirming that the external water intrusion has been suppressed to the greatest extent.

Images