Dynamic leakage monitoring and production-marketing difference prediction method and system for DMA (direct memory access) partitioned water supply pipe network based on multi-modal data fusion and intelligent decision

Abstract

The invention discloses a multi-modal data fusion and intelligent decision-making-based leakage monitoring and production and marketing difference prediction method and system for a DMA (direct memory access) partitioned water supply pipe network. The method comprises the steps of collecting sensor, vision and text data; preprocessing data through cleaning, complementing and feature engineering; constructing a multi-modal data fusion model, a leakage positioning and evaluation model, a maintenance strategy generation model and a self-evolution mechanism; and realizing leakage early warning, maintenance scheduling and monitoring optimization according to model output. The system correspondingly comprises a data acquisition module, a preprocessing module, an A model module and a decision application module. The problems that a traditional method is single in data, low in prediction precision and not intelligent in decision making are solved, accurate leakage monitoring and intelligent management are achieved, and the operation and maintenance efficiency and economic benefits of a water supply system are improved.

Description

Technical Field

[0004] This invention belongs to the field of intelligent operation and maintenance technology of water supply networks. Specifically, it relates to a method and system for monitoring leakage and predicting production and sales difference in DMA (District of Meters) water supply networks based on multimodal data fusion and intelligent decision-making. This method is used to achieve efficient detection, location and prediction of leakage in water supply networks and improve the management efficiency of water supply systems. Background Technology

[0005] Currently, leakage in water supply networks seriously affects water resource utilization efficiency and the economic benefits of water supply companies. Traditional leakage monitoring methods mainly include flow balance methods and pressure analysis methods, but they have the following drawbacks:

[0006] 1. Data limitation: Relying solely on pressure and flow data collected by sensors cannot fully reflect the status of the pipeline network and has insufficient ability to detect hidden leaks (such as underground micro-seepage).

[0007] 2. Low prediction accuracy: Prediction methods based on statistical or physical models are difficult to adapt to the complexity and dynamic changes of pipeline network operation, and the prediction errors of leakage rate and production-sales difference are large.

[0008] 3. Insufficient intelligent decision-making: The lack of an automated decision-making mechanism makes it impossible to generate the optimal maintenance strategy in real time based on the leakage situation, resulting in delayed maintenance response and increased water waste.

[0009] The limitations of existing technologies necessitate an urgent need in this field for a leakage monitoring and prediction solution that integrates multi-source data and possesses intelligent decision-making capabilities. Summary of the Invention

[0010] Purpose of the invention

[0011] This invention aims to provide a method and system for monitoring leakage and predicting production and sales differences in DMA zoned water supply networks based on multimodal data fusion and intelligent decision-making. By integrating sensor data, visual data, and text data, and combining the A1 algorithm, it achieves accurate detection, location, prediction, and intelligent decision-making of leakage, thereby reducing water supply losses and improving operation and maintenance efficiency.

[0012] Technical solution

[0013] The method of the present invention includes the following steps:

[0014] 1. Data Acquisition:

[0015] Sensor data acquisition: Electromagnetic flow meters are deployed at the inlet and outlet of the DMA zone to collect the influent flow rate Q. in and water flow rate Q out Pressure sensors are installed along the pipeline to collect pressure P, and temperature sensors are connected to obtain ambient temperature T. The sampling interval is 15 minutes.

[0016] Visual data acquisition: Using drones to inspect and capture images of the pipeline area every month, and then connecting them to InSAR satellite data to obtain information on surface deformation.

[0017] Text data acquisition: The repair work order is structured and user-reported leaks are converted into text data using voice recognition technology.

[0018] 2. Data preprocessing:

[0019] Data cleaning: Applying the 3σ principle to sensor data. in The isolated forest algorithm is used to remove outliers; visual images are converted to grayscale and denoised; and stop words are removed and stemming is performed on text data.

[0020] Data completion: Short-term sensor data were obtained using linear interpolation. Long-term missing data is filled using ARIMA model prediction; missing visual data is estimated using the mean of neighboring regions; and key information in text data is manually filled in.

[0021] Feature engineering:

[0022] Sensor data: Calculate the flow difference ΔQ = Q in -Q out Pressure volatility Pressure-flow correlation coefficient Features such as...

[0023] Visual data: Features such as water accumulation area and crack length were extracted using the YOLOv8 model. Surface deformation was obtained by combining InSAR data.

[0024] Text data: Named entity recognition is performed using BERT, and the reasons for omissions are classified using FastText.

[0025] 3. AI Model Building and Application:

[0026] Multi-model data fusion model: This model concatenates sensor, visual, and text feature vectors, employing a Transformer architecture (multi-head attention mechanism formula: ...). (V, etc.) analysis, outputting the leakage probability.

[0027] Leakage location and assessment model:

[0028] GNN model: Models the pipeline network as a graph structure and utilizes GCN (Geometry Network Concept) for node feature updates. Alternatively, the GAT model can be used to predict the probability of node leakage, and the TDOA algorithm (d = υΔt) can be used to locate the leakage point.

[0029] Maintenance strategy generation model:

[0030] Reinforcement learning model: The state space is sensor status, leakage location, etc., and the action space is maintenance operation. The reward function is R=EC (E is the repair efficiency and C is the cost). The PPO algorithm is used to generate maintenance strategies.

[0031] NLP model: Automatically generates repair work orders that include information on defects.

[0032] Model self-evolution mechanism: The iCaRL algorithm is used to achieve online learning, and uncertainty sampling is used for active learning.

[0033] 4. Decision-making and application:

[0034] Leakage warning: When the leakage probability exceeds the threshold (e.g., 0.7), an early warning is triggered, and relevant personnel are notified according to the level.

[0035] Maintenance scheduling: Allocate resources and plan maintenance routes according to maintenance strategies.

[0036] Monitoring optimization: The reinforcement learning agent dynamically adjusts the sensor sampling frequency and deployment location.

[0037] The present invention also provides a system for implementing the above method, including a data acquisition module, a data preprocessing module, an A1 model module, and a decision application module, wherein the modules work together to achieve leakage monitoring and prediction functions.

[0038] Beneficial effects

[0039] 1. High-precision forecasting: Combining multimodal data fusion with AI algorithms significantly improves the accuracy of leakage rate and production-sales difference forecasting. The RMSE for leakage rate forecasting can be reduced to 0.8%, and the RSE for production-sales difference forecasting can be improved. 2 Increased to 0.88.

[0040] 2. Intelligent decision-making: Automatically generates maintenance strategies, reducing maintenance response time from 4 hours to 1.5 hours and improving operation and maintenance efficiency.

[0041] 3. Resource optimization: By dynamically adjusting monitoring strategies through reinforcement learning, the number of sensors can be reduced by 25%-30%, thus lowering deployment costs.

[0042] 4. High adaptability: The model's self-evolution mechanism enables the system to quickly adapt to changes in the pipeline network and maintain long-term effectiveness. Attached Figure Description

[0043] Appendix Figure 1 Flowchart of a DMA-based method for monitoring leakage and predicting production-sales gap in regional water supply networks based on multimodal data fusion 1. Overall Architecture: This flowchart adopts a modular and layered design, clearly showing the complete technical process from data collection to decision application. The layers are connected by data flow arrows, reflecting the logical coherence of the method and the closed-loop characteristics of the data flow. 2. Data Acquisition Layer: Sensor data acquisition: The influent flow rate is collected using an electromagnetic flowmeter. and water flow rate The pressure sensor acquires the pipeline pressure P, and the temperature sensor records the ambient temperature T. The sampling frequency is 15 minutes, providing basic physical parameters for monitoring the pipeline operation status. Visual data acquisition: UAV inspections acquire high-resolution images of the pipeline area, while InSAR satellite data provides information on surface deformation. Combining the two can capture abnormal surface changes caused by potential leaks. Text data collection: Structured processing of historical maintenance work orders, extraction of key information such as the location of the leak and the repair time, and real-time processing of user leak reports via voice recognition technology, converting them into text data to supplement leak clues provided by human feedback. 3. Data preprocessing layer: Data cleaning: For sensor data, first apply the 3σ principle. Gross errors are identified by using the isolated forest algorithm to detect abrupt outliers; visual images are converted to grayscale and denoised to improve feature extraction accuracy; stop words are removed from text data and stemming is performed to standardize semantic expression. Data completion: Linear interpolation was used to address short-term (≤2 hours) sensor data loss. For long-term missing data, ARIMA model is used to predict and fill in missing values; for visual data, missing values ​​are estimated by the mean of features in neighboring regions; for text data, when key information is missing, it is marked as pending and manually supplemented. Feature engineering: Sensor characteristics: Calculate the flow difference Pressure volatility Basic characteristics such as minimum nighttime flow rate; derived pressure-flow correlation coefficient. , Complex features, etc. Visual features: The YOLOv8 model was used to extract quantitative features such as water accumulation area and crack length from the images. Surface deformation was calculated using InSAR data to construct a three-dimensional spatial feature vector. Text data: Named entity recognition is performed using the BERT model to extract key entities such as leakage location, pipe diameter, and material; the FastText model is used to classify and encode the causes of leakage, forming a text semantic feature vector. 4. AI Model Layer Multi-model data fusion model: The preprocessed sensor, vision, and text feature vectors are concatenated into a high-dimensional feature vector. Utilizing the Transformer architecture (multi-head attention mechanism formula: It automatically learns the weights of different modal data, outputs the leakage probability distribution, and determines whether there is leakage in the DMA partition. Leakage location and assessment model: GNN model: The pipeline network is abstracted as a graph structure, where nodes contain features such as pressure, flow rate, and pipe material, and edges represent pipeline connections. The GCN algorithm is used (node ​​feature update formula: Predict the probability of node leakage, combined with the TDOA algorithm. Locate the coordinates of the leak. CV model: Based on YOLOv8, target detection is performed on UAV images to identify water accumulation areas and crack locations; combined with InSAR surface deformation data, the scope of leakage impact is comprehensively assessed and the area of ​​the affected region is calculated. Maintenance strategy generation model: Reinforcement learning model: The state space is defined as sensor state, leakage location, maintenance resource distribution, etc., and the action space is to adjust the sensor sampling frequency and schedule maintenance personnel. The reward function is R=EC (E is the repair efficiency and C is the cost). The PPO algorithm is used to train the agent to generate the optimal maintenance path. NLP Model: Based on the leakage location and assessment results, automatically generate a maintenance work order containing leakage location, cause analysis, and suggested maintenance methods, and use the BERT model to generate natural language descriptions. Model self-evolution mechanism: The iCaRL algorithm is used to enable online learning of the model, continuously receiving new data to update the model parameters; through uncertainty sampling, unlabeled abnormal samples are actively identified, triggering the manual labeling process, expanding the training dataset, and improving the model's ability to identify new leakage patterns. 5. Decision Application Layer Leakage warning: When the leakage probability output by the multimodal data fusion model exceeds a set threshold (e.g., 0.7), the system automatically triggers a warning mechanism. Based on the magnitude of the leakage probability, it is divided into three warning levels: high, medium, and low, and relevant management personnel are notified via SMS, APP push, etc. Maintenance scheduling: Based on the maintenance strategy, the optimal maintenance plan is generated from the model output, and maintenance personnel, equipment and materials are scheduled. The best maintenance route is planned using a geographic information system, and the maintenance progress is tracked in real time to ensure that maintenance work is carried out efficiently. Monitoring optimization: The reinforcement learning agent dynamically adjusts the sensor sampling frequency and deployment location based on the real-time operation status of the pipeline network; while ensuring monitoring accuracy, it reduces sensor energy consumption and equipment maintenance costs, and achieves a reasonable allocation of monitoring resources.

[0044] Figure 2: System structure diagram of DMA-based regional water supply network leakage monitoring and prediction based on multimodal data fusion and intelligent decision-making. 1. Overall Architecture: This system adopts a layered architecture design, consisting of a hardware layer and a software layer. The hardware layer provides computing and storage resources, while the software layer implements core business functions. The software layer is further divided into four modules: data acquisition, preprocessing, AI models, and decision applications. These modules collaborate with each other through data flow and control flow. 2. Hardware layer: Edge computing devices: Deployed in the DMA partition site, they adopt a low-power, high-performance embedded computing platform to run lightweight AI models processed by model compression technology (such as quantized YOLOv8 models), and are responsible for the initial processing and feature extraction of data acquisition, reducing the amount of data transmission and reducing network bandwidth requirements. Data center server cluster: Composed of high-performance servers, it adopts a distributed computing architecture and is responsible for running complex AI model training and inference tasks, storing massive amounts of historical data and model parameters, and supporting large-scale parallel data processing and deep analysis. 3. Software Layer - Data Acquisition Module: Sensor data acquisition unit: Connects to devices such as electromagnetic flowmeters, pressure sensors, and temperature sensors via industrial-grade communication protocols (such as Modbus and MQTT), and collects flow, pressure, and temperature data at a set sampling frequency (15 minutes), performing preliminary format conversion and timestamp synchronization. Visual data acquisition unit: controls the UAV inspection route and shooting parameters, periodically acquires images of the pipeline network area; accesses InSAR satellite data services through a dedicated interface to obtain surface deformation information; and performs format standardization and geographic coordinate calibration on the acquired visual data. Text data acquisition unit: performs structured processing on maintenance work orders and extracts key information fields; processes user reports of leaks in real time through a speech recognition engine, converts speech into text, and performs preliminary semantic analysis to extract key information such as the location and symptoms of the leak. 4. Software Layer - Data Preprocessing Module: Data cleaning unit: Implements parallel processing of the 3σ principle and the isolated forest algorithm, performs outlier detection and removal on multi-source heterogeneous data; performs noise reduction processing such as grayscale conversion and Gaussian filtering on visual images; and performs preprocessing operations such as word segmentation and stop word removal on text data. Data completion unit: Based on the type and duration of missing data, it automatically selects linear interpolation or ARIMA model for data filling; for missing areas in visual data, it completes the data through spatial interpolation of features from neighboring areas; for missing key information in text data, it generates a task list to be processed and pushes it to the manual processing queue. Feature Engineering Unit: For sensor data, calculates various statistical features and derived features; for visual data, calls the YOLOv8 model for object detection and feature extraction; for text data, uses BERT and FastText models for semantic analysis and feature encoding; standardizes and merges the three types of features to construct a unified feature vector. 5. Software Layer - AI Model Module: Multimodal data fusion model unit: Based on the PyTorch framework, the Transformer architecture is implemented. It adopts a multi-head attention mechanism to fuse and analyze multi-source features and outputs the omission probability distribution. It supports two modes: online model update and batch training, which can be dynamically adjusted according to the amount of data and computing resources. Leakage Location and Assessment Model Unit: Integrates GNN and CV models. The GNN model is based on the DGL framework and analyzes the pipeline network structure and predicts leakage points. The CV model is based on the YOLOv8 framework and processes UAV images in real time. The outputs of the two models are fused to comprehensively assess the leakage location and impact range. Maintenance strategy generation model unit: The reinforcement learning algorithm is implemented using the Stable Baselines3 framework to train the agent to generate the optimal maintenance strategy; the NLP model is implemented based on the Hugging Face Transformers library to automatically generate maintenance work orders; and the simulation verification and optimization adjustment of maintenance strategies are supported. Model self-evolution mechanism unit: Implements the incremental learning framework of the iCaRL algorithm, supporting continuous updates of model parameters; develops an uncertainty sampling engine to actively filter samples that require manual annotation; maintains dynamic updates and version management of the training dataset to ensure continuous improvement of model performance. 6. Software Layer - Decision Application Module: Leakage warning unit: Implements threshold triggering mechanism and hierarchical warning rule configuration; supports multi-channel warning information push (SMS, APP, email, etc.); maintains warning logs and processing records, and provides warning statistics and analysis functions. Maintenance scheduling unit: integrates a geographic information system (GIS) to realize maintenance route planning and resource scheduling optimization; develops a maintenance task management system to track maintenance progress and result feedback; and establishes a maintenance knowledge base to accumulate maintenance experience and best practices. Monitoring and optimization unit: Based on the decision-making results of the reinforcement learning agent, dynamically adjust the sensor deployment location and sampling frequency; establish a sensor health status assessment model to predict sensor failure risks; optimize data transmission strategy to reduce communication energy consumption and cost.

[0045] Figure 3 Flowchart of data acquisition, processing, model calculation, and decision output Data Flow Diagram Key Nodes Data collection: Sensor data: Real-time acquisition of pipeline network operating parameters through electromagnetic flow meters, pressure sensors, etc. Visual data: Drones acquire surface images during inspections, while InSAR satellites provide sub-millimeter deformation monitoring. Text data: Structured processing of maintenance records, voice recognition of user leak reports. Data processing: Cleaning: The 3σ principle is used to remove sensor outliers and isolated forests are used to detect mutation data. Completion: Linear interpolation fills in short-term missing values, and the ARIMA model predicts long-term missing values. Feature engineering: Calculate sensor features such as flow difference and pressure fluctuation rate, extract visual target features using YOLOv8, and extract text semantic features using BERT. Model calculation: Multimodal fusion: The Transformer architecture integrates three types of features and outputs the probability of leakage. Leakage location: The GNN model analyzes the pipeline network structure and combines it with the TDOA algorithm (d=vΔt) to locate the leakage point. Impact assessment: CV model analysis of images to determine the extent of water accumulation, and InSAR data to assess ground subsidence. Strategy generation: Reinforcement learning (PPO algorithm) optimizes maintenance paths, and NLP automatically generates maintenance work orders. Decision output: Warning level: A red warning is triggered when the probability of leakage is >0.7, and a notification is sent via SMS / APP. Intelligent scheduling: GIS plans the shortest maintenance route and dynamically allocates personnel and equipment. Monitoring optimization: Reinforcement learning dynamically adjusts the sensor sampling frequency, reducing equipment energy consumption by 30%.

[0046] Figure 4 System module architecture diagram System Module Architecture Description Hardware layer: Edge computing devices: Deployed on-site in the pipeline network, running lightweight models (such as compressed YOLOv8) to complete data preprocessing. Data center servers: support the training of complex models (such as GNN, Transformer) and store all historical data. Software layer: Data acquisition module: Multi-source device access interface, supports protocols such as Modbus and MQTT, and realizes data synchronization and timestamp calibration. Data processing module: pipelined processing flow, supporting distributed computing to accelerate feature engineering. AI model module: Modular design (fusion, localization, evaluation, strategy, evolution), supporting hot-swappable algorithm upgrades. Decision application modules: visual early warning interface, maintenance work order management system, and sensor status monitoring panel. Closed-loop feedback: The maintenance results and the new data-driven model self-evolution (iCaRL online learning + active learning) continuously optimize prediction accuracy. Detailed Implementation

[0047] Taking a city's DMA partition as an example, the specific implementation steps are as follows:

[0048] 1. Data Acquisition Phase: Electromagnetic flow meters are installed at the inlet and outlet of this zone, pressure sensors are deployed along the pipeline, drone inspections are conducted monthly and InSAR data is collected, and maintenance work orders and leak reporting telephone texts are also collected.

[0049] 2. Data preprocessing stage: The collected data is processed according to the cleaning, completion, and feature engineering methods described above to form an effective feature dataset.

[0050] 3. AI Model Application Stage: The processed data is input into a multimodal fusion model to determine whether there is any leakage. If there is leakage, GNN and CV models are used to locate and evaluate the leakage. Maintenance strategies are generated through reinforcement learning and NLP models.

[0051] 4. Decision-making and execution phase: Based on the model output, perform leakage early warning, maintenance scheduling, and monitoring and optimization operations.

[0052] Actual operation has verified that this invention effectively improves the leakage monitoring and management level of this zone.

Claims

1. A method for monitoring leakage and predicting production-sales difference in DMA-based zoned water supply networks based on multimodal data fusion and intelligent decision-making, characterized in that, Includes the following steps Data acquisition steps: Collect sensor data, visual data, and text data from the DMA partition. The sensor data includes inlet flow rate, outlet flow rate, pressure, and temperature data. The visual data includes UAV inspection images and InSAR satellite data. The text data includes maintenance work orders and user leak reporting phone text. Data preprocessing steps: The collected data is cleaned, completed, and feature-engineered. The cleaning adopts the 3σ principle and the isolated forest algorithm. The completion uses linear interpolation and the ARIMA model for sensor data. The feature engineering calculates sensor data features, extracts visual data features, and analyzes text data features. AI model construction and application steps: Construct a multimodal data fusion model, a leakage location and evaluation model, a maintenance strategy generation model, and a model self-evolution mechanism. The multimodal data fusion model adopts the Transformer architecture, the leakage location and evaluation model includes a GNN model and a CV model, and the maintenance strategy generation model includes a reinforcement learning model and an NLP model. Decision-making and application steps: Based on the model output, perform leakage early warning, maintenance scheduling, and monitoring optimization.

2. The method according to claim 1, characterized in that, The sensor data acquisition interval is 15 minutes, the UAV inspection cycle for the visual data is once a month, and the InSAR satellite data update cycle is 1 month.

3. The method according to claim 1, characterized in that the 3σ principle is achieved by calculating the mean and standard deviation of the data. To determine outliers, the linear interpolation formula is as follows:

4. The method according to claim 1, characterized in that, The GNN model models the pipeline network as a graph structure and uses the GCN or GAT algorithm. The CV model identifies image features based on YOLOv8 and combines InSAR to evaluate the leakage range.

5. The method according to claim 1, characterized in that, The reinforcement learning model uses sensor status and leakage location as the state space, maintenance operation as the action space, leakage repair efficiency minus maintenance cost as the reward function, and the PPO algorithm to generate maintenance strategies.

6. A DMA-based regional water supply network leakage monitoring and production-sales difference prediction system based on multimodal data fusion and intelligent decision-making, implementing the method of any one of claims 1-5, characterized in that it comprises: Data acquisition module: used to acquire sensor data, visual data, and text data. Data preprocessing module: Used to clean, complete, and perform feature engineering on the collected data. AI Model Module: Used to build and run multimodal data fusion models, leakage location and assessment models, maintenance strategy generation models, and model self-evolution mechanisms; Decision Application Module: Used for leakage early warning, maintenance scheduling, and monitoring optimization based on model output.

Images