Digital twin water plant cluster comprehensive visual management system

Abstract

The invention relates to the technical field of data management, and discloses a digital twin water plant cluster comprehensive visual management system. The system comprises a processing module, a modeling module, a process flow topology self-adaptive visual management interface and a flow control module, wherein the processing module carries out standardization processing on multi-source parameter data of a water plant to generate real-time data flow; the modeling module constructs a three-dimensional digital twinborn model by utilizing an oblique photography technology; the extraction module extracts spatio-temporal characteristics through multi-channel data fusion to obtain early warning information, the calculation module diagnoses abnormity based on a ternary association network and a heterogeneous graph attention algorithm, and the coding module generates a digital watermark health state code to form a predictive maintenance scheme. The panoramic display of the running state of the water plant, the accurate identification of the abnormal mode and the optimal configuration of maintenance resources are realized. According to the invention, panoramic display of the running state of the water plant, accurate identification of the abnormal mode and optimal configuration of the maintenance resources are realized, so that the overall running efficiency and management level of the water plant cluster are improved.

Description

Technical Field

[0001] The present application relates to the field of data management technology, and in particular to a digital twin water plant cluster integrated visualization management system. Background Art

[0002] With the continuous advancement of informatization in the water industry, traditional water plant monitoring systems have gradually evolved from early single SCADA systems to comprehensive monitoring platforms integrating multiple functions. Existing water plant monitoring systems primarily collect equipment operating data through distributed control systems (DCS) or programmable logic controllers (PLCs), using two-dimensional graphical interfaces to display process flows and key parameters, enabling real-time monitoring and remote control of water plant production processes. At the same time, some advanced water plants have begun to incorporate three-dimensional modeling technology, providing a three-dimensional display of plant layouts and key equipment, enhancing visualization. In terms of data analysis, existing technologies already enable basic historical data query, statistical analysis, and trend display, providing a certain level of decision support for operations management. Furthermore, some water plants have established simple rule-based alarm systems that issue alerts when parameters exceed preset thresholds, prompting operations and maintenance personnel to take timely action.

[0003] However, existing water plant monitoring systems still suffer from numerous shortcomings. First, subsystems are often independent of each other, resulting in severe data silos, making effective information integration and sharing difficult. Second, existing 3D visualization technologies are mostly limited to static display, lacking dynamic correlation with real-time data and failing to truly reflect the operating status of equipment. Furthermore, alarm mechanisms generally rely on simple threshold comparison methods and lack comprehensive analysis capabilities for multi-dimensional data, resulting in high false alarm rates and low diagnostic accuracy. Furthermore, existing systems rely primarily on manual judgment for fault diagnosis and lack intelligent abnormal pattern recognition and root cause analysis capabilities, making predictive maintenance difficult. Finally, water groups with multiple water plants lack a unified cluster management platform, hindering collaborative optimization and resource sharing across plants and resulting in low management efficiency. These issues severely hinder improvements in water plant operational efficiency and management capabilities. Summary of the Invention

[0004] This application provides a digital twin water plant cluster comprehensive visualization management system, which is used to achieve a panoramic display of the water plant's operating status, accurate identification of abnormal patterns, and optimal allocation of maintenance resources, thereby improving the overall operating efficiency and management level of the water plant cluster.

[0005] This application provides a digital twin water plant cluster integrated visualization management system, which includes:

[0006] In the technical solution provided by the present application, the present invention obtains real-time data streams and standardized data sets by standardizing the multi-source parameter data collected by sensors of water plant equipment, thereby solving the problems of inconsistent data formats and uneven quality in traditional water plant monitoring systems, and providing a high-quality data foundation for subsequent analysis; based on the standardized data set, three-dimensional modeling is performed using oblique photography technology to obtain a three-dimensional digital twin model of the water plant, which overcomes the defects of limited expression ability of traditional two-dimensional visualization interfaces and realizes high-precision digital reconstruction of the physical space of the water plant; a process flow topology adaptive visualization framework is constructed according to the digital twin model to obtain a cluster water plant visualization management interface, breaking through the limitations of traditional monitoring system interfaces that are fixed and difficult to adapt to different process flows, and flexibly responding to various water plant layouts through an adaptive framework; real-time data streams are input into the visualization management interface, and spatiotemporal features are extracted through a multi-channel data fusion mechanism to obtain early warning information, thus getting rid of the The traditional system relies only on a simple model of single threshold judgment to achieve collaborative analysis and anomaly detection of multi-dimensional data; it constructs a process-equipment-environment ternary association network for early warning information, and uses heterogeneous graph attention network to calculate anomaly patterns to obtain diagnostic results. Among them, heterogeneous graph attention network, as an advanced artificial intelligence algorithm, can automatically learn complex interaction patterns between different types of nodes and capture implicit associations that are difficult to discover by traditional methods. It has significant advantages under the complex working conditions of water plants. By simultaneously processing three types of heterogeneous data: process parameters, equipment status and environmental conditions, the algorithm achieves accurate positioning and root cause analysis of faults, greatly improving the accuracy and interpretability of diagnosis; based on the diagnostic results, it generates a visual code of the equipment health status based on digital watermarks to obtain a predictive maintenance plan, which changes the traditional passive maintenance model and realizes the transition from "fault repair" to "predictive maintenance", significantly reducing the unplanned downtime and maintenance costs of equipment. Overall, the present invention has constructed a complete technical solution from data acquisition, three-dimensional modeling, visualization to intelligent diagnosis and predictive maintenance, forming a real-time interactive closed loop between the physical world and the digital space. Through artificial intelligence algorithms, it conducts in-depth mining and analysis of massive multi-source heterogeneous data of water plants, realizes intuitive display and accurate perception of complex process states, and significantly improves the safety and stability of water plant operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 This is a schematic diagram of an embodiment of the digital twin water plant cluster comprehensive visualization management system in the embodiment of this application. DETAILED DESCRIPTION

[0009] An embodiment of the present application provides a digital twin water plant cluster integrated visualization management system. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or apparatus.

[0010] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the digital twin water plant cluster comprehensive visualization management system includes:

[0011] The processing module is used to standardize the multi-source parameter data collected by the water plant equipment sensors to obtain real-time data streams and standardized data sets;

[0012] The modeling module is used to perform 3D modeling based on standardized data sets using oblique photography technology to obtain a 3D digital twin model of the water plant;

[0013] A construction module is used to build an adaptive visualization framework for process flow topology based on the digital twin model, and obtain a visual management interface for the cluster water plant;

[0014] The extraction module is used to input the real-time data stream into the visual management interface, extract the spatiotemporal features through the multi-channel data fusion mechanism, and obtain early warning information;

[0015] The calculation module is used to construct a process-equipment-environment ternary association network for early warning information, calculate abnormal patterns using a heterogeneous graph attention network, and obtain diagnostic results;

[0016] The encoding module is used to generate a visual encoding of the equipment health status based on a digital watermark based on the diagnosis results to obtain a predictive maintenance plan.

[0017] It is understandable that the execution subject of this application can be a digital twin water plant cluster integrated visualization management system, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0018] Specifically, the processing module first receives multi-source parameter data collected by various sensors throughout the water plant. These include water quality parameters (turbidity, pH, residual chlorine, ammonia nitrogen, and temperature), flow parameters (inlet and outlet flow, instantaneous flow), pressure parameters (intake pressure, storage pressure, and supply pressure), equipment operating parameters (pump operating status, fault status, and valve switch position), and energy consumption data (total power and segmented power). This raw data is transmitted to the processing module via a multi-layered data acquisition network. Data cleaning is then performed to remove outliers, duplicates, and missing values. The cleaned data is then converted to a unified format and standardized. This data is then organized and categorized according to temporal, spatial, and process relationships to form a structured database. Data integrity and accuracy are verified through a data reception and verification mechanism. Finally, time series analysis and feature extraction are performed on the verified, reliable data. Statistical features are calculated using a sliding window algorithm to generate a real-time data stream and standardized dataset. The modeling module uses oblique photography technology to comprehensively map the water plant. In specific operations, drones and multi-angle photography equipment were used to acquire high-resolution 1:500 image data, capturing spatial information such as the water plant's geographic location, topography, and building layout. The acquired raw image data was then registered and fused with joystick aerial survey image data. Spatial correspondences were established using a point-matching algorithm, generating a fused image in a unified coordinate system. Based on the fused image, a 3D reconstruction algorithm was used to generate a point cloud model of the water plant. Multi-view stereo matching was used to calculate spatial point coordinates, yielding 3D point cloud data with centimeter-level accuracy. The point cloud data was then geometrically processed, and a 3D geometric model of the water plant scene was constructed using a gridding algorithm and surface reconstruction methods. This resulted in a 3D model of the process equipment. The 3D model of the process equipment was spatially mapped with process flow data from a standardized dataset, establishing a data-model correspondence and forming a dynamically linked digital model. Finally, material mapping and texture mapping were performed on the dynamically linked digital model, and rendering optimization techniques were used to manipulate lighting and shadow effects, resulting in a 3D digital twin of the water plant.

[0019] The construction module develops a cluster water plant visualization management platform based on a 3D digital twin model of the water plant. The digital twin model analyzes the process flow topology and uses graph theory algorithms to extract the connections between process units, resulting in a process flow topology diagram. A three-tier visualization framework is designed based on the process flow topology diagram, assigning functional nodes to the front-end WebGL rendering layer, the middleware data processing layer, and the back-end service layer, forming the basic visualization architecture. Subsequently, a 3D scene visualization module is developed, implementing interactive features such as zooming, panning, rotating, and first-person roaming, creating a 3D interactive scene. The 3D interactive scene is integrated with a dynamic process flow graphical representation system, displaying process parameter status through color changes at process nodes and dynamic particle flow effects, creating a process flow visualization interface. Multidimensional data charting functionality is then added, presenting key indicators of water quality, flow rate, and energy consumption through visualization components such as bar charts, line charts, and heat maps, forming a comprehensive analysis dashboard. Finally, a unified management system for the water plant cluster is constructed. Using a regional partitioning algorithm and a process type grouping mechanism, unified visualization and comparative analysis of multiple water plants are achieved, resulting in a cluster water plant visualization management interface. The extraction module inputs real-time data streams into the visualization management interface to extract spatiotemporal features. First, a real-time monitoring mechanism for process parameters is established. By tracking and updating the visual interface through parameter changes, real-time monitoring data is generated. Then, a multi-level threshold warning system is set up, and key parameters such as water quality, flow, and pressure are divided into normal ranges, warning thresholds, and alarm thresholds, and a threshold rule library is established. Real-time comparison is performed between the real-time monitoring data and the threshold rule library to identify abnormal parameters that exceed the normal range and generate preliminary warning signals. Subsequently, a time series analysis is performed on the preliminary warning signals. The persistence of the abnormality is determined by the data change trends at multiple consecutive time points to form time domain feature data. Multi-channel data fusion is performed based on the time domain feature data and the spatial layout information of the process units, and the operational correlation between the process units is extracted to obtain spatiotemporal feature data. Finally, the spatiotemporal feature data is subjected to abnormal classification processing, and is graded according to the degree of abnormality, scope of impact, and urgency to generate warning information.

[0020] The computation module constructs a process-equipment-environment ternary association network based on the warning information to calculate anomaly patterns. First, feature extraction is performed, organizing process parameter nodes, equipment status nodes, and environmental condition nodes into a hierarchical structure. Three types of connection relationships are defined between the nodes to form an initial ternary association structure. A heterogeneous graph network is then constructed, where process nodes include inlet water quality, treatment unit status, and outlet water parameter attributes; equipment nodes include operating status, vibration values, and energy consumption attributes; and environmental nodes include temperature, humidity, and external condition attributes. This creates a multi-dimensional feature heterogeneous graph. A three-layer heterogeneous graph attention network is then applied. The first layer is a node type feature encoding layer. Three different weight matrices are used to process the features of the three node types, respectively, to obtain type-aware node representations. This type-aware node representation is then fed into the second multi-head attention layer. By computing eight attention heads in parallel, the interaction patterns between nodes of different types are learned, resulting in node features with enhanced relationships. The third layer, a global graph pooling layer, aggregates the entire graph information using a weighted summation approach. Residual connections are used to preserve the original features, resulting in an anomaly pattern representation vector. Finally, a similarity calculation is performed with a predefined fault pattern library to match the most similar fault type and cause, resulting in a diagnosis result. The encoding module generates a visual code of the equipment health status based on the diagnosis result and formulates a predictive maintenance plan. The diagnostic results are then used to quantitatively assess the equipment health index, calculating it based on three dimensions: fault type, fault severity, and development trend. A digital watermark encoding rule is then developed to map the health index to a specific visual encoding pattern. This visual code of the equipment health status is created using a combination of three visual elements: color, shape, and density. The visual code of the equipment health status is embedded into the corresponding device in the digital twin model and displayed on the 3D model surface through a transparency overlay, forming a visual health status indicator. Based on the visual health status indicator and historical maintenance records, an equipment life prediction model is established. Remaining service life estimates are used to determine maintenance priorities and generate a prioritized maintenance list. A personalized maintenance plan is generated for the equipment on the prioritized maintenance list. Maintenance time and required resources are scheduled through task decomposition and resource matching, resulting in a detailed maintenance task list. Finally, the detailed maintenance task list is linked to the maintenance knowledge base, extracting corresponding maintenance steps and safety precautions based on the equipment type and fault characteristics, and formulating a predictive maintenance plan.

[0021] For example, when a water plant's flocculation sedimentation tank experienced an abnormality, the processing module first collected sensor data and discovered that the effluent turbidity gradually increased from a normal value of 0.5 NTU to 3.2 NTU, exceeding the preset warning value of 2.0 NTU. The modeling module precisely located the equipment within the 3D digital twin. The construction module visually displayed a red alert indicating an abnormal condition in the flocculation sedimentation tank area through a visual management interface. The extraction module analyzed and found that turbidity had been increasing continuously for nearly four hours, with a temporal correlation with a decrease in dosing rate. The calculation module, through ternary correlation network analysis, discovered abnormal vibration values ​​and flow rate fluctuations in the flocculant dosing pump, forming a clear correlation pattern with the increased turbidity, diagnosing the dosing pump as a fault. Finally, the encoding module assessed the equipment's health index as "dangerous," marked it on the 3D model with a digital watermark encoding a flashing red outline, and generated a targeted maintenance plan, including spare part replacement, pipeline inspection, and calibration testing, to prevent further deterioration of water quality.

[0022] In the embodiment of the present application, the present invention obtains real-time data streams and standardized data sets by standardizing the multi-source parameter data collected by sensors of water plant equipment, thereby solving the problems of inconsistent data formats and uneven quality in traditional water plant monitoring systems, and providing a high-quality data foundation for subsequent analysis; based on the standardized data set, three-dimensional modeling is performed using oblique photography technology to obtain a three-dimensional digital twin model of the water plant, which overcomes the defects of limited expression ability of traditional two-dimensional visualization interfaces and realizes high-precision digital reconstruction of the physical space of the water plant; a process flow topology adaptive visualization framework is constructed according to the digital twin model to obtain a cluster water plant visualization management interface, breaking through the limitations of traditional monitoring system interfaces that are fixed and difficult to adapt to different process flows, and flexibly responding to various water plant layouts through an adaptive framework; real-time data streams are input into the visualization management interface, and spatiotemporal features are extracted through a multi-channel data fusion mechanism to obtain early warning information, breaking away from traditional The system relies only on a simple model of single threshold judgment to achieve collaborative analysis and anomaly detection of multi-dimensional data; it constructs a process-equipment-environment ternary association network for early warning information, and uses heterogeneous graph attention network to calculate anomaly patterns to obtain diagnostic results. Among them, heterogeneous graph attention network, as an advanced artificial intelligence algorithm, can automatically learn complex interaction patterns between different types of nodes and capture implicit associations that are difficult to discover by traditional methods. It has significant advantages under complex working conditions of water plants. By simultaneously processing three types of heterogeneous data: process parameters, equipment status and environmental conditions, the algorithm achieves precise positioning and root cause analysis of faults, greatly improving the accuracy and interpretability of diagnosis; based on the diagnostic results, it generates a visual code of the equipment health status based on digital watermarks to obtain a predictive maintenance plan, which changes the traditional passive maintenance model and realizes the transition from "fault repair" to "predictive maintenance", significantly reducing the unplanned downtime and maintenance costs of equipment. Overall, the present invention has constructed a complete technical solution from data acquisition, three-dimensional modeling, visualization to intelligent diagnosis and predictive maintenance, forming a real-time interactive closed loop between the physical world and the digital space. Through artificial intelligence algorithms, it conducts in-depth mining and analysis of massive multi-source heterogeneous data of water plants, realizes intuitive display and accurate perception of complex process states, and significantly improves the safety and stability of water plant operations.

[0023] In a specific embodiment, the processing module is configured to:

[0024] Water quality parameters, flow parameters, pressure parameters, equipment operation parameters and energy consumption data are collected through a multi-level data acquisition network to obtain the original data set;

[0025] Apply data cleaning algorithms to the original data set to process outliers, duplicate values, and missing values ​​to obtain preliminarily processed data;

[0026] Perform unified format conversion and data standardization on the data after preliminary processing to obtain standardized data with unified format;

[0027] Classify and organize standardized data in a unified format according to temporal, spatial and process relationships to obtain a structured database;

[0028] Establish a data reception and verification mechanism based on a structured database to verify data integrity and accuracy, and obtain verified and trusted data;

[0029] Perform time series analysis and feature extraction on verified trusted data, calculate statistical features through sliding window algorithm, and obtain real-time data stream and standardized data set.

[0030] Specifically, the processing module first collects key water plant data through a multi-layered data acquisition network. This multi-layered data acquisition network consists of a field layer, a control layer, and a management layer, forming a clearly defined data transmission system. The field layer uses technologies such as industrial Ethernet, wireless transmission, and fiber optics to acquire raw data from sensors and transmit it to a programmable logic controller (PLC). The control layer ensures that the PLC collects equipment operating status and production process data in real time. The management layer establishes a water plant data center to store and process the collected data. Water quality parameters collected by the sensor network include turbidity, pH, residual chlorine, ammonia nitrogen, and temperature, which directly reflect the effectiveness of water treatment. Flow parameters include inlet and outlet flow rates and instantaneous flow rate, which are used to monitor water balance. Pressure parameters include intake pressure, storage pressure, and supply pressure, which are related to system operation status. Equipment operating parameters include pump operation status, fault status, and valve switch position, reflecting equipment health. Energy consumption data includes total and segmented power consumption, which is used for energy efficiency analysis. This data is transmitted in real time through the field sensor network at a preset collection frequency (typically every 5 seconds), and aggregated into a raw data set.

[0031] Applying a data cleaning algorithm to the raw dataset is a critical step in ensuring data quality. The data cleaning algorithm first detects outliers—data points that significantly deviate from the normal range. This outlier detection uses the 3-Sigma principle and boxplots to identify data outside the normal range. For example, if a pH sensor reading suddenly jumps from a normal value of 7.2 to 14.0, the cleaning algorithm marks it as an outlier and replaces it with the previous valid value or the average of the nearest valid values. The cleaning algorithm also identifies and resolves duplicate values ​​by comparing data values ​​at adjacent time points. When the data for multiple consecutive time points are identical and the parameter is expected to fluctuate, it identifies duplicate values ​​caused by sensor lag and generates reasonable data points using linear interpolation. For missing values, the algorithm imputes them using methods such as the mean or median interpolation within the time window or forward filling to ensure data continuity. After these cleaning steps, pre-processed data is generated. This pre-processed data undergoes unified format conversion and data standardization. Unified format conversion is a key step in resolving data format inconsistencies across systems and devices. The processing module has designed a unified data transmission protocol and format to convert data in various non-standard formats (such as proprietary formats of different equipment manufacturers) into standardized JSON or XML formats. For example, proprietary binary format data output by a certain brand of PLC and ASCII text data output by another brand of flow meter are both converted into a unified JSON format, containing fields such as measurement value, timestamp, device ID and unit. Data standardization converts parameters of different dimensions and ranges to the same scale, usually using minimum-maximum normalization or Z-score normalization methods. For example, turbidity data (usually ranging from 0-100 NTU) and pH data (ranging from 0-14) are converted to a unified range of 0-1 through minimum-maximum normalization, so that different parameters can be compared and analyzed on the same scale. After these processes, standardized data with a unified format is obtained.

[0032] Categorizing and organizing standardized data in a uniform format according to temporal, spatial, and process relationships is a core step in building a structured database. Temporal classification organizes data along a timeline, establishing a data structure with multi-level time indexes based on year, month, day, hour, minute, and second, facilitating rapid retrieval and analysis of data by time period. Spatial classification, based on the physical layout of the water plant, organizes data according to spatial dimensions such as plant area, building, process unit, and equipment location, forming a spatially correlated data structure. Process relationship classification focuses more on the logical associations of data within the water treatment process, categorizing data according to the entire process flow from raw water intake to factory water supply, covering all process units such as distribution wells, pre-ozonation contact tanks, flocculation sedimentation tanks, V-type filters, and lift pump rooms. This multi-dimensional classification and organization creates a structured database with a complete structure and clear relationships, supporting multi-angle data retrieval and analysis.

[0033] Establishing a data reception and verification mechanism based on a structured database is an important step in ensuring data reliability. The data reception mechanism uses real-time data stream processing technology to receive continuous data streams from various data sources through a message queue system (such as Kafka), ensuring the stability of data reception under high concurrency. The data verification mechanism verifies the integrity and accuracy of the data through multiple verification methods. Integrity verification checks the integrity of the data packet to ensure that the data is not lost or damaged during transmission; range verification ensures that the data value is within a reasonable range that is physically possible; consistency verification checks whether the logical relationship between related parameters conforms to physical laws, such as verifying whether the difference between the inlet flow rate and the outlet flow rate is within a reasonable range; and time series continuity verification checks whether the data timestamps are continuous and increment in sequence. Once the data passes these verifications, it is marked as trusted data and stored in a high-reliability database, providing a high-quality data foundation for subsequent analysis.

[0034] Time series analysis and feature extraction of verified, trusted data are the final steps in generating real-time data streams and standardized datasets. Time series analysis first arranges data in chronological order and then explores temporal patterns in the data through methods such as trend analysis, seasonality analysis, and periodicity analysis. Sliding window algorithms play a key role in this process. They slide a fixed-length time window (such as 10 minutes, 1 hour, or 24 hours) across the time series and calculate statistical features of the data within the window, including the mean, median, maximum, minimum, standard deviation, and coefficient of variation. These statistical features effectively capture the data's central tendency, dispersion, and variability. As the window continues to slide, the newly generated statistical features form a new time series, forming the real-time data stream. Simultaneously, the original verified data and the extracted features are combined to form a structured, standardized dataset, supporting subsequent modeling and analysis.

[0035] In a specific embodiment, the modeling module is used to:

[0036] Multi-angle oblique photography was carried out on the water plant, and high-resolution 1:500 image data was acquired through drones to obtain the original image data of the water plant;

[0037] The original image data is registered and fused with the joystick aerial survey image data, and the spatial correspondence is established through the same-name point matching algorithm to obtain a fused image in a unified coordinate system;

[0038] Based on the fused images, a 3D reconstruction algorithm is used to generate a point cloud model of the water plant. The spatial point coordinates are calculated through multi-view stereo matching to obtain high-precision 3D point cloud data.

[0039] Perform geometric processing on the 3D point cloud data, construct a 3D geometric model of the water plant scene through meshing algorithms and surface reconstruction methods, and obtain a 3D model of the process equipment;

[0040] Spatially map the 3D model of process equipment with the process flow data in the standardized data set, establish a data-model correspondence, and obtain a dynamically associated digital model;

[0041] Material mapping and texture mapping are performed on the dynamically associated digital model, and the lighting and shadow effects are processed through rendering optimization technology to obtain a three-dimensional digital twin model of the water plant.

[0042] Specifically, the modeling module first comprehensively maps the water plant using multi-angle oblique photography. Oblique photography is an aerial photogrammetry method that simultaneously captures vertical and oblique imagery. It typically uses a five-lens camera (one vertical lens and four oblique lenses, facing forward, backward, left, and right), capable of capturing comprehensive three-dimensional information of ground targets. In practice, a drone equipped with a high-precision camera flies at an altitude of 50-100 meters above the water plant according to a pre-set route, capturing high-resolution image data at a scale of 1:500. The scale of 1:500 here means that 1 centimeter on the map represents 5 meters on the ground. This accuracy allows for clear capture of details within the water plant, buildings, structures, and major equipment. During the photography process, the drone performs multiple overlapping flights to ensure that each ground point appears in at least three images from different angles, thereby acquiring sufficient perspective information. This raw image data contains rich texture information and preliminary spatial position information, providing the raw material for subsequent 3D reconstruction. Registering and fusing this raw image data with the joystick aerial survey image data is a key step in obtaining accurate spatial data in a unified coordinate system. Joystick aerial surveying uses specialized aerial survey equipment controlled by a joystick to acquire high-precision aerial imagery. These images typically have precise geographic coordinate information and high spatial resolution. The registration and fusion process first identifies keypoints—corresponding physical points in both image sets. Keypoint matching algorithms use feature detection methods such as SIFT (Scale-Invariant Feature Transform) or SURF (Speeded Up Robust Features) to extract keypoints (such as corners and edge intersections) from the images. Feature descriptors are then calculated and matched based on the similarity of these descriptors. In water plant scenarios, building corners, road intersections, and prominent equipment are often good candidates for keypoints. After matching, spatial transformation algorithms (such as affine or perspective transformations) are used to transform the original image data into the coordinate system of the joystick aerial survey imagery to achieve spatial alignment. During the fusion process, weights are calculated based on image quality, clarity, and shooting angle to ensure uniformity and integrity of the fused image. The resulting fused image has a unified geographic coordinate system, laying the foundation for subsequent 3D reconstruction.

[0043] Generating a point cloud model of a water plant using a 3D reconstruction algorithm based on fused images is a critical step in the conversion from 2D images to 3D data. This 3D reconstruction algorithm is primarily based on the principles of multi-view stereo vision, employing computer vision methods to recover 3D structure from multi-angle 2D images. The specific processing steps include camera calibration, feature matching, triangulation, and point cloud optimization. Camera calibration determines the shooting position, angle, and internal parameters of each image; feature matching identifies corresponding physical points across different images; and triangulation calculates the coordinates of points in 3D space based on known camera parameters and matching points. For example, once a flocculation sedimentation tank at a water plant is identified in multiple images, the 3D coordinates of the tank can be accurately calculated by calculating the difference in imaging position (parallax) from different viewpoints and combining the camera parameters. Multi-view stereo matching uses semi-global matching (SGM) or multi-view stereo matching (MVS) algorithms to calculate the optimal depth value for each pixel, creating a dense depth map that is then converted into a 3D point cloud. The resulting point cloud data contains a large number of discrete 3D points, each with precise XYZ coordinates and RGB color information, forming a preliminary 3D representation of the water plant. Geometric processing of three-dimensional point cloud data is a necessary step in constructing a continuous surface model. Although the original point cloud data contains rich spatial information, it is merely a discrete set of points, lacking a continuous surface or topological structure. Geometric processing begins with point cloud filtering to remove noise and outliers. Common methods include statistical outlier filtering and radius outlier filtering. The point cloud is then downsampled to reduce the data volume while maintaining geometric features. Voxel grid downsampling is typically used to divide the space into small cubes (voxels), replacing the original point set with the center point of each voxel. The point cloud is converted into a triangular mesh model using a gridding algorithm. Commonly used algorithms include Poisson surface reconstruction and greedy projection triangulation. During the surface reconstruction process, issues such as holes and non-manifold regions must be addressed to ensure the integrity and correctness of the model. Process equipment in water plants, such as pumps, valves, and flow meters, often require individual and detailed modeling, including steps such as equipment contour extraction, feature recognition, and parametric modeling. The three-dimensional models of these process equipment need to meet centimeter-level accuracy requirements to accurately reflect the geometric shape and size of the actual equipment, providing a basis for subsequent data binding and simulation analysis.

[0044] Spatial mapping of 3D models of process equipment with process flow data from standardized datasets is a core step in realizing digital twin functionality. The spatial mapping process first establishes a unique correspondence between the physical equipment and the virtual model. Each 3D model is assigned a globally unique identifier (GUID) and associated with the equipment data collected by the processing module. Data binding rules are then defined to clarify which data items are mapped to which model attributes or components. For example, a 3D model of a water pump may require binding to various data such as speed, flow, power, and temperature, while a valve model requires binding to data such as opening and pressure differential. The mapping process also includes data conversion, converting real-time data into a form suitable for visualization. For example, converting valve opening percentage to the valve opening angle in the 3D model. After spatial mapping is completed, the 3D model is no longer a static geometric representation, but a dynamic entity that reflects real-time operating status. When real-time data changes, the corresponding attributes of the 3D model are also updated synchronously. For example, the color of the water pump model changes with temperature, and the opening of the valve model adjusts according to the control signal, thus establishing a true data-model correspondence. Material mapping and texture mapping of dynamically associated digital models are crucial for enhancing visual effects and information presentation. Material mapping applies the collected texture images of the actual equipment surface to the surface of the three-dimensional model to enhance realism and recognition. This process includes steps such as UV unfolding, texture projection and stitching. UV unfolding maps the surface of the three-dimensional model to a two-dimensional plane to create a texture coordinate system; texture projection projects the actual photo onto the model surface; stitching handles the continuity problem of the texture boundary. For standard equipment, you can also use the preset material library to apply the physical properties of different materials such as metal, plastic, and glass. In terms of rendering optimization, the visual quality and information expression ability of the model are improved by configuring the lighting model (such as the Phong lighting model or physically based rendering), shadow calculation (such as shadow mapping or ambient occlusion) and visual effect processing. These processing enables the three-dimensional model to not only accurately reflect the geometric shape, but also show the material properties and light and shadow effects, enhance the user's intuitive understanding of the water plant environment, and ultimately form a three-dimensional digital twin model of the water plant with high visual fidelity and data correlation.

[0045] Taking the ozone contact tank at a water plant as an example, the modeling process was implemented as follows: First, a drone was flown over the ozone contact tank from multiple angles, capturing high-resolution images from five different angles (vertical, 45° forward, 45° backward, 45° left, and 45° right). Ten images were taken from each angle, ensuring 80% overlap. These original images were registered with previously acquired joystick aerial survey images. Feature points were extracted using the SIFT algorithm, successfully matching 37 points of the same name, including the four corners of the contact tank and the inlet and outlet, establishing accurate spatial correspondence. Based on the registered fused images, the MVS algorithm was used for 3D reconstruction, generating an initial point cloud containing approximately 2 million points at a density of 5 cm / point. The point cloud data was filtered and downsampled, retaining approximately 500,000 key points. A triangulated mesh model was then generated using the Poisson surface reconstruction algorithm, accurately recreating the contact tank's geometry, including the tank structure, internal baffles, and agitation equipment. Next, a mapping relationship was established between the 3D model and real-time data, including inlet flow data, tank water level, dissolved oxygen concentration, pH value, and ozone dosage. Finally, appropriate material mapping and textures were applied to set the water surface reflection effect and the concrete material of the tank body. Dynamic water flow particle effects were added to intuitively display the direction and speed of the water flow. The completed digital twin model can reflect the operating status of the ozone contact tank in real time. When the dissolved oxygen concentration is too low or the pH value is abnormal, the corresponding location on the model will be highlighted with warning colors, helping operators quickly identify the problem and take appropriate measures.

[0046] In a specific embodiment, a building block is provided for:

[0047] Analyze the process topology structure of the digital twin model, extract the connection relationship between process units through graph theory algorithms, and obtain the process topology diagram;

[0048] Design a three-layer visualization framework based on the process topology diagram, allocate functional nodes of the front-end WebGL rendering layer, the middleware data processing layer, and the back-end service layer to obtain the basic visualization architecture;

[0049] Develop a 3D scene visualization module for the basic visualization architecture, and obtain a 3D interactive scene by implementing zooming, translation, rotation, and first-person roaming interactive functions;

[0050] Integrate the 3D interactive scene with the dynamic graphic representation system of the process flow, and display the process parameter status through the color change of process nodes and dynamic particle flow effects to obtain a process flow visualization interface;

[0051] Add multi-dimensional data chart display function to the process flow visualization interface, present key indicators of water quality, flow rate, and energy consumption through bar charts, line charts, and heat map visualization components to obtain a comprehensive analysis panel;

[0052] A unified management system for water plant clusters is constructed based on a comprehensive analysis panel. The visualization and comparative analysis of multiple water plants are achieved through the regional division algorithm and process type grouping mechanism, resulting in a visual management interface for clustered water plants.

[0053] Specifically, the construction module performs a process topology analysis on the digital twin model. Process topology analysis refers to the process of mathematically describing and analyzing the connection relationships between various process units in the entire water treatment process. In specific operations, the construction module extracts process unit information from the digital twin model, including key treatment units such as the water distribution well, pre-ozone contact tank, flocculation sedimentation tank, V-type filter, lift pump room, post-ozone contact tank, activated carbon filter, backwash pump room, ultrafiltration membrane tank, clear water tank, suction well, and water supply pump room. Then, a process flow diagram is constructed using a graph theory algorithm, in which each process unit is a node in the graph and the water flow direction and pipeline connection are the edges in the graph. The graph theory algorithm mainly uses a directed graph structure, traversing the entire process through depth-first search or breadth-first search to identify water flow paths, parallel units, and series units. The algorithm also calculates critical paths, bottleneck units, and loop paths to provide a basis for subsequent optimization. The output of the process flow topology analysis is a process flow topology diagram, which clearly shows the entire process from raw water intake to factory water supply, including the logical relationship between the main line process and auxiliary systems (such as dosing system, sludge discharge system, etc.).

[0054] Designing a three-tier visualization framework based on the process flow topology is key to building an efficient visualization system. This three-tier architecture is a classic software design pattern. In the digital twin water plant system, it is specifically divided into a front-end WebGL rendering layer, a middleware data processing layer, and a back-end service layer. The front-end WebGL rendering layer is responsible for 3D scene rendering and interaction. WebGL is a JavaScript API based on OpenGL ES that can directly leverage GPU acceleration to render complex 3D scenes. The middleware data processing layer handles data conversion, caching, and streaming, and includes components such as data format converters, real-time data caching, historical data indexing, and data aggregation calculations. The back-end service layer provides data storage and computing services, including relational databases, time-series databases, file storage systems, and a computing engine. During the functional node allocation process, the corresponding components and processing logic of each node and edge in the process flow topology are determined based on their characteristics. For example, parameters with high frequency changes (such as flow rate and pressure) are assigned to the real-time data stream processing component, data requiring historical trend analysis (such as water quality indicators) is assigned to the time-series database, and complex computational requirements (such as efficiency evaluation) are assigned to the computing engine. This layered architecture ensures the scalability of the system and supports the access and management of multiple water plants.

[0055] The 3D scene visualization module, implemented based on WebGL technology, uses a series of graphics processing and rendering techniques to present the 3D model of the water plant in a browser. The module implements basic interactive functions such as zooming, panning, rotating, and first-person roaming. Zooming is achieved by adjusting the parameters of the camera's view frustum, allowing users to view the entire water plant layout from a macro perspective and observe equipment details from a micro perspective. Panning is achieved by changing the camera's position, maintaining the same viewing direction and allowing the scene to move within the viewport. Rotation is achieved by modifying the camera's orientation, rotating the scene around a specified point or axis. First-person roaming simulates the user's perspective as they walk through the water plant, allowing keyboard controls for forward, backward, and turn directions to create an immersive experience. Implementing these interactive functions requires a carefully designed interaction controller that handles mouse and keyboard events and converts them into corresponding camera parameter changes. Furthermore, to improve rendering performance, the module employs level of detail (LOD) technology, dynamically adjusting model accuracy based on viewing distance, displaying a high-precision model close up and simplifying it to a low-precision model at a distance, achieving a balance between visual quality and operational efficiency. The dynamic graphical representation system for process flows is a visualization method that transforms abstract data into intuitive graphics. In the digital twin water plant system, process parameter status is primarily expressed through color changes at process nodes and dynamic particle flow effects. This color change utilizes color mapping technology, mapping parameter values ​​within different ranges to a predefined color spectrum. For example, pH values ​​are mapped to a gradient from red (acidic) to green (neutral) to blue (alkaline), enabling operators to easily assess water quality. The dynamic particle flow effect is implemented using particle system technology, generating particles representing water flow within pipes and equipment. Particle attributes such as number, velocity, and color are linked to actual flow rate, flow velocity, and water quality, dynamically simulating water flow conditions. During the integration process, the mapping between geometric objects in the 3D scene and functional units in the process flow must be established to ensure accurate data mapping to visualization elements. Furthermore, data changes at different timescales must be handled, such as millisecond-level changes in equipment status and hourly-level changes in water quality trends. Multi-timescale rendering technology is employed to ensure the consistency and accuracy of the visual presentation.

[0056] Adding multidimensional data charting capabilities to the process flow visualization interface is a key measure to enhance data analysis capabilities. Multidimensional data charts are visualizations that simultaneously display data relationships across multiple dimensions. In the digital twin water plant system, these primarily include components such as bar charts, line charts, and heat maps. Bar charts are commonly used to compare quantitative relationships across categories, such as comparing energy consumption across different process units; line charts are used to display changing trends in time series data, such as continuous monitoring of water quality indicators; and heat maps use color intensity to represent data density or intensity, such as the distribution of equipment failure frequencies. These chart components are developed using responsive design principles, automatically adjusting the display based on data characteristics and supporting interactive exploration, such as data filtering, time range selection, and combined indicator analysis. A key design feature is the integration of charts with the 3D scene. When a user selects a process unit or piece of equipment in the 3D scene, the associated data chart automatically updates to display detailed data and analysis results for that object. Similarly, when a user identifies an anomaly in a data chart, the 3D scene automatically locates the corresponding location, enabling two-way interaction. This close integration of multi-dimensional data charts and three-dimensional scenes forms a comprehensive analysis panel, providing operation and maintenance personnel with multi-angle and multi-level data analysis tools.

[0057] The ultimate goal of achieving collaborative management of multiple water plants is to establish a unified management system for water plant clusters based on a comprehensive analysis dashboard. This unified management system integrates data and models from multiple water plants onto a unified platform through a regionalization algorithm and a process type grouping mechanism. The regionalization algorithm divides water plants into distinct regional groups based on geographic location and administrative divisions, facilitating regional-level management and scheduling. The implementation utilizes spatial clustering methods, such as DBSCAN (density-based clustering) or the K-means algorithm, to automatically form regional groups based on the geographic coordinates of the water plants. The process type grouping mechanism categorizes water plants based on their process characteristics (e.g., conventional treatment, advanced treatment, membrane treatment, etc.), facilitating technology comparison and experience sharing among similar water plants. The system provides a multi-level view switching function, allowing users to view an overview of all water plants in a global view or drill down to detailed views of specific regions or individual plants. Cross-plant comparative analysis is a key feature, allowing users to select similar indicators from multiple water plants for side-by-side comparison to identify best practices and optimization opportunities. The system also implements centralized alarm management, uniformly sorting and displaying alarm information from all water plants according to importance and urgency, helping managers respond to the most critical issues promptly. Through these features, the unified water plant cluster management system enables coordinated management and optimized operations across multiple water plants.

[0058] In a specific embodiment, the extraction module is used to:

[0059] Establish a real-time monitoring mechanism for process parameters for real-time data streams, and update the visual interface by tracking parameter changes to obtain real-time monitoring data;

[0060] A multi-level threshold warning system is established based on real-time monitoring data, which divides key parameters of water quality, flow rate and pressure into normal range, warning threshold and alarm threshold, and obtains a threshold rule base;

[0061] Compare real-time monitoring data with the threshold rule library in real time to identify abnormal parameters that exceed the normal range and obtain preliminary warning signals;

[0062] Conduct time series analysis on the preliminary warning signals, determine the persistence of abnormalities through the data change trends at multiple consecutive time points, and obtain time domain feature data;

[0063] Based on the time domain feature data and the spatial layout information of the process units, multi-channel data fusion is performed to extract the operation correlation between the process units and obtain the spatiotemporal feature data;

[0064] The spatiotemporal feature data is classified and processed for abnormality, and graded according to the degree of abnormality, scope of impact and urgency to obtain early warning information.

[0065] Specifically, the real-time process parameter monitoring mechanism is a technical system that continuously collects, processes, and displays key process parameters, aiming to promptly reflect the water plant's operating status. This mechanism obtains real-time data streams from the processing module via a data acquisition interface, with a data update frequency of at least 5 seconds to ensure timely monitoring. Upon receipt, data is first validated to confirm the continuity of timestamps and the validity of the value range. It is then categorized and organized according to process unit and parameter type. Processing logic and display rules are configured for different types of parameters, including inlet and outlet water quality (including turbidity, residual chlorine, pH, ammonia nitrogen, and temperature), flow parameters (inlet and outlet flow rates, instantaneous flow rates), pressure data (intake pressure, storage pressure, and supply pressure), equipment status (pump operating status, valve switch position), power parameters, and energy consumption data. Parameter change tracking utilizes an incremental update strategy, processing and transmitting only changed parameter values ​​to reduce data processing overhead. The visual interface is updated using a push mechanism. When a monitored parameter changes, the new value is pushed to the front-end interface via real-time communication technologies such as WebSocket, triggering a redraw and status update of the corresponding area. This series of processes generates real-time monitoring data containing complete timestamps, parameter IDs, parameter values, and status tags. Establishing a multi-level threshold warning system based on real-time monitoring data is a key step in implementing the early warning function. This multi-level threshold warning system sets multiple levels of judgment criteria for different process parameters to distinguish different operating states and risk levels. In the digital twin water plant system, key parameters such as water quality, flow rate, and pressure are divided into three ranges: normal range, warning threshold, and alarm threshold. The normal range represents the ideal operating range for the parameter; the warning threshold represents the range where the parameter deviates from normal but not yet to a dangerous level; and the alarm threshold indicates that the parameter has deviated to a dangerous level requiring immediate intervention. Threshold setting is based on three data sources: national standards (such as the water quality parameter specifications in the "National Drinking Water Quality Standard" GB5749-2006); historical water plant operating experience (normal fluctuation ranges determined through statistical analysis of historical data); and technical specifications provided by equipment manufacturers (such as the rated flow and pressure ranges of pumps). Threshold setting also takes into account seasonal variations. For example, in areas where raw water quality fluctuates significantly, turbidity thresholds may differ between summer and winter. The system supports dynamic adjustment of thresholds. Operations and maintenance personnel can modify the threshold values ​​of specific parameters based on actual conditions. Modification records are stored and used for subsequent optimization. All set threshold rules are organized into a structured threshold rule library, which contains fields such as parameter ID, parameter name, upper and lower limits of normal range, upper and lower limits of warning thresholds, upper and lower limits of alarm thresholds, applicable period, and last update time.

[0066] Comparing real-time monitoring data against the threshold rule library in real time is the process of generating preliminary warning signals. This real-time comparison utilizes streaming processing. Whenever new monitoring data arrives, it is immediately compared against the corresponding rules in the threshold rule library. The comparison algorithm first identifies the applicable threshold rule based on the parameter ID and timestamp, then determines whether the parameter value is within the normal range. If the parameter value exceeds the normal range but does not reach the alarm threshold, a warning-level signal is generated; if the parameter value exceeds the alarm threshold, an alarm-level signal is generated. To avoid false alarms caused by transient fluctuations, the algorithm also employs debounce processing. This means that a warning signal is only generated if the parameter exceeds the threshold multiple times (typically 3-5 times) in a row. The comparison process also considers the parameter's rate of change. Even if some parameters are within the threshold, rapid changes within a short period of time will trigger an alert. For example, a 20% drop in pump outlet pressure over 10 minutes is considered an anomaly even if it remains within the normal range. The comparison results form a preliminary warning signal, containing information such as the parameter ID, parameter name, current value, normal range, type of violation (above the upper limit or below the lower limit), warning level (warning or alarm), time of occurrence, and duration of the violation.

[0067] Time series analysis of preliminary warning signals is a key tool for assessing the persistence and development trend of anomalies. Time series analysis is a statistical method that studies a sequence of chronologically ordered data points. In water plant early warning systems, it is primarily used to determine whether anomalies are temporary fluctuations or ongoing trends. The analysis constructs a time window for the parameter triggering the warning and collects historical data for a recent period (e.g., 30 minutes, 2 hours, or 24 hours, depending on the parameter type). This data is then smoothed using a moving average to reduce the impact of random fluctuations. The parameter's trend is then calculated, and the slope is derived through linear regression to determine whether the parameter is increasing, decreasing, or stable. The parameter's fluctuation period is also calculated, and regular oscillations are identified using autocorrelation functions or Fourier transforms. For anomalies that have persisted for a period of time, the system also assesses their development speed—the rate at which the parameter deviates from the normal range. Based on this analysis, the system determines the persistence of the anomaly and categorizes it as transient (a brief fluctuation followed by recovery), persistent (continuously exceeding the threshold with no recovery trend), or periodic (recurring with a regular pattern). These temporal characteristics, along with the severity of the anomaly (the extent to which it deviates from the normal range), constitute the time domain feature data. Multi-channel data fusion based on time-domain feature data and spatial layout information of process units is a key technology for achieving global perception. Multi-channel data fusion is the process of comprehensively analyzing data from different sources and types to form a more comprehensive and accurate understanding. In the digital twin water plant system, the fused data primarily comes from two channels: time-domain feature data, which includes parameter anomalies, duration, and trends; and spatial layout information of process units, which describes their positional relationships, connectivity, and water flow directions. The fusion process first establishes a topological diagram of the process units, clarifying upstream and downstream relationships and parameter transmission paths. Then, based on the physical laws of water treatment, a correlation model is established between the parameters, such as the relationship between influent quality and reagent dosage, and the relationship between flocculation effect and sedimentation efficiency. The system then analyzes the location of the abnormal parameters in the process flow, tracing possible upstream causes and predicting potential downstream impacts. The system also detects whether multiple related parameters are abnormal simultaneously to determine whether it is a localized fault or a systemic issue. If multiple related parameters are abnormal simultaneously, the correlation coefficient is calculated to identify the propagation path and core source of the anomaly. The results of the fusion analysis form spatiotemporal feature data, including the process location of the abnormal parameters, the status of the associated parameters, the spatiotemporal propagation characteristics of the anomaly, and the potential root causes.

[0068] Anomaly classification of spatiotemporal feature data is a key step in generating final warning information. Anomaly classification involves categorizing anomalies into different types and levels based on their characteristics, allowing for appropriate response measures. The classification process begins by assessing the severity of the anomaly. By calculating the percentage or standard deviation of the parameter's deviation from the normal range, the severity of the anomaly is quantified as mild, moderate, or severe. The impact of the anomaly is then assessed, determining whether it is a single piece of equipment anomaly, a single process unit anomaly, or a chain reaction of multiple units. This impact assessment builds on the spatiotemporal fusion analysis conducted in the previous step by tracing the anomaly's propagation path through the process flow. The urgency of the anomaly is then assessed, taking into account its development speed, duration, and potential impact on effluent quality. The urgency is categorized into three levels: general attention, requiring action, and emergency response. Based on a comprehensive assessment of the anomaly's severity, scope, and urgency, the system categorizes warnings into three levels: general notification, major warning, and emergency alarm. Different levels of warning utilize different notification methods and response processes. General notifications are displayed only on the interface, major warnings are notified via text message, and emergency alarms trigger visual and audible alarms and automatically notify management personnel by phone. The final generated warning information includes abnormality description, classification results, recommended treatment measures and related auxiliary information, providing comprehensive decision support for operation and maintenance personnel.

[0069] In a specific embodiment, the computing module is configured to:

[0070] Extract features from warning information, organize process parameter nodes, equipment status nodes, and environmental condition nodes into a hierarchical structure, and define three types of connection relationships between nodes: process flow edges, equipment dependency edges, and environmental impact edges, to obtain an initial ternary association structure.

[0071] Based on the initial ternary association structure, a heterogeneous graph network structure is constructed, in which the process node contains the attributes of influent water quality, treatment unit status, and effluent parameters; the equipment node contains the attributes of operating status, vibration value, and energy consumption; and the environment node contains the attributes of temperature, humidity, and external conditions, resulting in a multi-dimensional feature heterogeneous graph.

[0072] A three-layer heterogeneous graph attention network is applied to the multi-dimensional feature heterogeneous graph. The first layer is the node type feature encoding layer. The features of process nodes, equipment nodes, and environment nodes are processed through three different sets of weight matrices to obtain type-aware node representations.

[0073] The type-aware node representation is input into the second-layer multi-head attention layer. By parallel computing 8 attention heads, each attention head independently learns the interaction pattern between nodes of different types. The results are concatenated and transformed nonlinearly to obtain the node features with enhanced relationships.

[0074] The node features with enhanced relationships are processed through the third global graph pooling layer, which uses weighted summation to aggregate the entire graph information and combines residual connections to retain the original features to obtain the abnormal pattern representation vector;

[0075] The similarity between the abnormal pattern representation vector and the predefined fault pattern library is calculated, and the most similar fault type and cause are matched using the cosine similarity measurement to obtain the diagnosis result.

[0076] Specifically, the calculation module performs feature extraction on the warning information generated by the extraction module and constructs a ternary association structure. Feature extraction is the process of extracting useful information from raw data. In the digital twin water plant system, feature extraction converts warning information into a structured graph data model. In specific operations, the calculation module divides the key elements of the water plant into three types of nodes: process parameter nodes, equipment status nodes, and environmental condition nodes. Process parameter nodes represent key indicators in the water treatment process, such as turbidity, pH value, residual chlorine content, etc.; equipment status nodes represent the operating status of water plant equipment, such as the switch status and operating parameters of equipment such as water pumps, valves, and flow meters; and environmental condition nodes include external factors such as temperature, humidity, and changes in water quality of the water source. These nodes are organized according to a hierarchical structure to form a clear hierarchical relationship. At the same time, three types of edges are defined to represent the associations between nodes: process flow edges describe the connection between water flow direction and process units, such as the process flow from the coagulation tank to the sedimentation tank; equipment subordinate edges represent the ownership relationship between equipment and process units, such as the flocculant dosing pump belongs to the coagulation process unit; environmental impact edges represent the impact of environmental factors on the process or equipment, such as the impact of ambient temperature on biological treatment efficiency. In this way, the operating status of the water plant is represented as a network structure composed of nodes and edges, that is, the initial ternary association structure. Based on the initial ternary association structure, the calculation module further constructs a heterogeneous graph network structure. A heterogeneous graph refers to a graph structure containing multiple types of nodes and edges. The main difference from an ordinary graph is that the nodes and edges have different types and attributes. During the construction process, each node type is assigned specific attributes: process nodes include attributes for influent quality (such as raw water turbidity and pH), treatment unit status (such as flocculation effect and filter resistance), and effluent parameters (such as effluent turbidity and residual chlorine content); equipment nodes include attributes for operating status (such as on / off status and operating time), vibration values ​​(such as motor vibration amplitude), and energy consumption (such as power and accumulated power); and environmental nodes include attributes for temperature, humidity, and external conditions (such as rainfall and changes in raw water source). These attribute values ​​are derived from real-time and historical data collected by the processing modules. Edges between nodes are also assigned specific attributes, such as flow rate and velocity for process flow edges, control relationship and response time for equipment dependency edges, and impact degree and lag time for environmental impact edges. In this way, the initial structured information is transformed into a multidimensional feature heterogeneous graph containing rich semantic information, providing a foundation for subsequent in-depth analysis.

[0077] Applying a three-layer heterogeneous graph attention network to multi-dimensional feature heterogeneous graphs is the core technology for achieving intelligent diagnosis. The heterogeneous graph attention network is a deep learning model specifically designed for heterogeneous graph data, capable of capturing the complex relationships between nodes of different types. The primary function of the first node type feature encoding layer is to transform the heterogeneous features of different node types into a common feature space. Specifically, three sets of weight matrices are designed for each of the three types of nodes (process, equipment, and environment), each specifically processing the features of the corresponding node type. For example, the raw feature vector of a process node undergoes a linear transformation using a process-specific weight matrix and a nonlinear activation function (such as ReLU) to generate a preliminary representation of the process node. Device and environment nodes undergo similar processing, but using their own dedicated weight matrices. This type-aware processing preserves the characteristics of each node type while mapping them into a unified feature space, facilitating subsequent interaction analysis. Through this layer of processing, each node obtains a type-aware node representation that preserves its own unique feature information while enabling interaction with nodes of other types.

[0078] The type-aware node representations are fed into the second-layer multi-head attention layer. The multi-head attention mechanism is a technique that simultaneously learns feature relationships from multiple perspectives. In the digital twin water plant system, eight parallel attention heads are used for computation. Each attention head independently learns the interaction patterns between different types of nodes. For example, the first attention head might focus on the relationship between process parameters and equipment status, the second attention head might focus on the impact of environmental conditions on process performance, and other attention heads might capture relationships in other dimensions. The computation of each attention head involves three steps: first, the query vector, key vector, and value vector are calculated. The similarity between the query vector and the key vector is then calculated as the attention weight. Finally, the attention weight is weighted and summed with the value vector to produce the attention output. The outputs of the eight attention heads are concatenated into a wider feature vector, which is then processed through a nonlinear transformation (such as a fully connected layer with ReLU activation) to produce relationship-enhanced node features. These features not only contain information about the node itself but also incorporate information from related nodes, forming a global understanding of the water plant's operating status. Processing these relationship-enhanced node features through the third-layer global graph pooling layer is the final step in generating the abnormal pattern representation vector. Global graph pooling is the process of aggregating information from the entire graph into a fixed-length vector. This is implemented in the digital twin water plant system using a weighted summation approach. Unlike simple averaging or summing, weighting nodes is assigned different weights based on their importance, with higher weights assigned to important nodes (such as critical equipment that triggers early warnings). Weight calculations are based on node centrality metrics (such as degree centrality and eigenvector centrality) and anomaly scores, ensuring that key information is retained during the aggregation process. A residual connection mechanism is also introduced during the aggregation process, adding the original node features with those processed by the first two layers. This design effectively prevents information loss during multi-layer transmission and preserves the important information in the original features. Global graph pooling aggregates the scattered node features in a heterogeneous graph into a fixed-dimensional vector. This vector encodes global information about the water plant's current abnormal state, known as the anomaly pattern representation vector. The predefined fault pattern library is a knowledge base containing common fault pattern feature vectors, their corresponding descriptions, causes, and solutions. It is constructed from historical water plant failure cases, equipment maintenance records, and expert experience. Similarity calculation uses the cosine similarity metric, which measures the similarity between two vectors by calculating the cosine of the angle between them. Cosine similarity ranges from -1 to 1, with values ​​closer to 1 indicating greater similarity. The calculation process first calculates the cosine similarity between the current anomaly pattern representation vector and each fault pattern vector in the fault pattern library. The top N (usually N = 3) fault patterns with the highest similarity are then selected as candidate patterns.If the highest similarity exceeds a preset threshold (typically 0.8), the diagnostic result for that fault mode is directly adopted. If the highest similarity is below the threshold but multiple similar fault modes exist, the common characteristics of these candidate modes are comprehensively analyzed to generate a combined diagnostic result. If all similarities are low, the fault type is marked as unknown, the characteristics are recorded, and expert intervention is prompted. The final diagnostic result includes complete information such as fault type, possible cause, scope of impact, severity, and recommended measures, providing a scientific basis for subsequent maintenance decisions.

[0079] Taking the flocculation and sedimentation process unit at a water plant as an example, the computational module's workflow is as follows: First, it receives an alert from the extraction module indicating that the effluent turbidity in the sedimentation tank has consistently exceeded the specified standard (normally <1 NTU, currently 3.5 NTU). The computational module performs feature extraction and organizes the relevant information into a ternary association structure: process parameter nodes include influent turbidity (35 NTU), dosing rate (18 mg / L), mixing intensity (60 rpm), sedimentation time (2 hours), and effluent turbidity (3.5 NTU); equipment status nodes include the flocculation agitator (normal), dosing pump (operating but with elevated vibration), and sedimentation zone scraper (normal); and environmental nodes include water temperature (22°C) and raw water pH (7.2). These nodes are connected by process flow edges (raw water → flocculation → sedimentation), equipment dependency edges (the dosing pump belongs to the flocculation unit), and environmental influence edges (water temperature affects flocculation efficiency). A multidimensional feature heterogeneous graph was then constructed, adding detailed attributes to each node, such as the dosing pump's operating status (on), vibration value (8.2 mm / s, exceeding the normal range by 5 mm / s), and energy consumption (2.1 kW). A three-layer heterogeneous graph attention network was then applied. The first layer transformed the heterogeneous features of different nodes into a unified space. The second layer's eight attention heads focused on different types of associations. For example, the third head specifically learned the relationship between equipment status and process performance, capturing the strong correlation between dosing pump vibration and effluent turbidity. The third layer performed pooling to obtain a vector representing the current abnormal state. Finally, a cosine similarity calculation showed that this vector had a similarity of 0.92 with the "flocculant dosing pump failure causing unstable dosing" pattern in the fault database, significantly higher than other fault patterns. The diagnosis indicated that the fault was a dosing system failure caused by excessive vibration due to dosing pump bearing wear, which affected dosing accuracy. The recommendation was to inspect and replace the dosing pump bearings, while also temporarily adjusting the dosing rate to compensate for the instability.

[0080] In a specific embodiment, the encoding module is used to:

[0081] The diagnostic results are used to quantitatively evaluate the equipment health index, which is calculated based on three dimensions: fault type, fault severity, and development trend.

[0082] Based on the device health index, a digital watermark encoding rule is constructed, which maps the health index into a specific visual encoding pattern. Through the combination of three visual elements: color, shape, and density, a visual encoding of the device health status is obtained.

[0083] The visual code of the equipment health status is embedded into the corresponding equipment of the digital twin model, and displayed on the surface of the three-dimensional model through transparency overlay to obtain a visual health status mark;

[0084] Establish an equipment life prediction model based on visual health status indicators and historical equipment maintenance records, determine maintenance priorities through remaining service life estimation, and obtain a priority maintenance list;

[0085] Generate personalized maintenance plans for equipment in the priority maintenance list, schedule maintenance time and required resources through task decomposition and resource matching algorithms, and obtain a detailed maintenance task list;

[0086] The detailed maintenance task list is associated with the maintenance knowledge base, and the corresponding maintenance steps and safety precautions are extracted according to the equipment type and fault characteristics to obtain a predictive maintenance plan.

[0087] Specifically, the encoding module performs a quantitative assessment of the equipment health index (EHI) based on the diagnostic results generated by the calculation module. This serves as the foundation for predictive maintenance. The EHI is a comprehensive indicator that describes the current state and reliability of the equipment. It typically ranges from 0 to 100, with 100 indicating a completely healthy device and 0 indicating complete failure. This quantitative assessment is conducted along three dimensions: The fault type dimension assigns a baseline score based on the nature and scope of the fault. For example, a minor fault (e.g., a slight parameter deviation) receives a baseline score of 80-90, a moderate fault (e.g., a worn but still functional component) receives a baseline score of 50-80, and a severe fault (e.g., a damaged critical component) receives a baseline score of 0-50. The fault severity dimension corrects for the degree to which a parameter deviates from its normal value. Factors such as the ratio of the deviation to the threshold, duration, and frequency are converted into correction coefficients using linear or nonlinear mapping. The development trend dimension analyzes the historical rate and acceleration of change in the fault parameters to determine whether the fault is stable, slowly deteriorating, or rapidly deteriorating, and further adjusts the EHI accordingly. The comprehensive calculation uses a weighted average method. The scores of the three dimensions are multiplied by their corresponding weights and then summed to obtain the final EHI. This index intuitively reflects the health status of the device. Establishing a digital watermark encoding rule based on the device health index is a key step in achieving intuitive status display. In this system, a digital watermark refers to a visual mark superimposed on the 3D device model to visually indicate the device's health status. The encoding rule is designed based on a combination of three visual elements: color, shape, and density to ensure clear communication. Color coding uses a gradient from green (healthy) to yellow (attention), and then to red (dangerous), mapping the device health index to specific colors: health indexes of 80-100 are mapped to green, 60-80 to yellow-green, 40-60 to yellow, 20-40 to orange, and 0-20 to red. Shape coding uses different graphical patterns to indicate fault types, such as waves for vibration anomalies, arrows for pressure anomalies, stars for temperature anomalies, and checkered patterns for electrical faults. Density coding uses the density of the pattern to indicate the fault's development trend, with higher density indicating faster development: stable faults use low-density patterns, slowly deteriorating faults use medium-density patterns, and rapidly deteriorating faults use high-density patterns. These three visual elements combine to form a complete visual code of the equipment health status. For example, a code with a red, star-shaped, high-density pattern indicates a severe, rapidly deteriorating temperature anomaly.

[0088] The embedding process first accurately locates the 3D model of the target device within the digital twin model using its device ID or spatial coordinates. A corresponding texture material, including color, shape, and density information, is then generated based on the visual encoding of the device's health status. These textures are then applied to the surface of the device's 3D model using a transparency overlay. An appropriate transparency setting (typically 30%-70%) ensures that details of the original device model remain visible and that the health status indicator is clearly legible. To enhance the visual effect, a flashing or pulsating effect is added to devices with severe faults. This effect is achieved by periodically adjusting the transparency or brightness, with the flashing frequency increasing with severity. For large devices (such as sedimentation tanks and filtration tanks), the health status indicator is applied to key areas or the overall outline of the device. For smaller devices (such as pumps and valves), it may cover the entire device model. The display of the indicator also takes into account the user's viewing angle and focal distance. When the user zooms out to view the entire device, only the critical fault indicator is displayed. When the user zooms in to focus on a specific area, the health status indicators for all levels within that area are displayed. This allows operations and maintenance personnel to intuitively identify problematic devices and their health status.

[0089] A device life prediction model is established based on visual health status indicators and historical maintenance records. This mathematical model predicts the future state of a device based on historical data. In digital twin water plant systems, it is primarily used to estimate the remaining useful life of a device. The model construction begins by collecting three types of historical data: health status data (including historical health index changes, fault types, and frequencies), maintenance records (including maintenance time, type, replacement parts, and effects), and operating environment data (including load, temperature, and humidity). Time series-based prediction methods, such as moving averages, exponential smoothing, or the more complex ARIMA (Autoregressive Integrated Moving Average) model, are then used to analyze historical health index changes and predict future trends. Maintenance activities are treated as intervention events in the time series to analyze the extent and duration of their impact on the device health curve. Based on this, a degradation model is applied to calculate the device degradation rate. The remaining useful life of the device is estimated by combining the manufacturer's design life and industry experience. Finally, maintenance priority is determined based on the remaining useful life, current health index, and fault impact. A weighted scoring approach is typically used, assigning different weights to multiple factors according to their importance. The total score is then calculated and ranked to form a prioritized maintenance list.

[0090] Generating personalized maintenance plans for equipment on the priority maintenance list is a crucial step in achieving precise maintenance. A personalized maintenance plan is a detailed maintenance plan tailored to specific equipment and specific fault types, taking into account multiple factors, including equipment characteristics, fault nature, and resource constraints. The plan generation process begins with task decomposition, breaking down maintenance work into basic operational units such as inspection, testing, cleaning, adjustment, and replacement. The prerequisites and consequences of each unit are clearly defined. Resource matching is then performed, assigning appropriate technicians (based on skill level and expertise), tools and equipment (such as specialized tools and testing equipment), and spare parts and materials (such as bearings and seals) to each unit's requirements. Time planning then proceeds, estimating the standard execution time for each unit and, taking into account inter-operational dependencies, using the critical path method or Gantt chart technique to schedule the operations and overall timeline. Resource allocation also requires consideration of the water plant's operational schedule, maximizing the need to schedule maintenance activities during periods of low equipment load or outages to minimize disruption to normal production. For complex maintenance tasks requiring coordination across multiple departments or external suppliers, coordination mechanisms and contingency plans must be designed. The resulting maintenance plan contains a detailed list of maintenance tasks, clearly indicating who performs what operations using what resources and when, providing clear guidance for maintenance execution.

[0091] Linking the detailed maintenance task list with the maintenance knowledge base is the final step in developing a complete predictive maintenance plan. The maintenance knowledge base is a structured collection of knowledge encompassing equipment maintenance standards, operating procedures, technical requirements, and safety precautions. It comprises technical documentation provided by equipment manufacturers, industry standards, and company-wide maintenance experience. The linkage process begins by searching the knowledge base for relevant maintenance procedures and technical key points based on equipment type (e.g., pumps, valves, filters) and fault characteristics (e.g., vibration, leakage, and reduced efficiency). The search results then extract the appropriate maintenance steps for the situation, including specific operating methods, technical parameters, quality standards, and acceptance requirements. Safety precautions related to maintenance operations are also extracted, including personal protective equipment requirements, operational risks, safe operating procedures, and emergency response measures. For simultaneous maintenance of multiple pieces of equipment, the locational relationships and functional dependencies between the pieces of equipment are analyzed to optimize maintenance routes and operational sequences. All extracted information is integrated with the task list generated in the previous step to form a complete predictive maintenance plan, encompassing operational guidance, technical requirements, and safety assurance. This plan is distributed to maintenance personnel via mobile devices, providing real-time guidance and verification during execution to ensure maintenance quality and safety.

[0092] Taking the water supply pump system of a water plant as an example, the encoding module's workflow is as follows: First, the calculation module receives diagnostic results, indicating that water supply pump No. 2 has bearing wear and a continuously increasing vibration value, currently reaching 9.5 mm / s (normal range < 5 mm / s). A quantitative assessment of the equipment health index is performed, with the fault type being mechanical wear (a moderate fault, with a base score of 65) and the fault severity being a parameter exceeding the specified value by 90% (with a correction factor of -15). Trend analysis shows a vibration growth rate of 0.5 mm / s / day over the past seven days (rapid deterioration, with a correction factor of -10). The resulting equipment health index is 40. A digital watermark encoding rule is constructed based on the health index, mapping an index of 40 to an orange-red color (approaching a critical state). A wavy shape is selected to represent vibration faults, and a high-density pattern is used to indicate a rapidly deteriorating trend. The code is embedded in the digital twin model, and an orange-red wavy, high-density pattern is superimposed on the 3D model of water supply pump No. 2 with 50% transparency. This pattern also flashes twice per second to visually convey the critical state of the equipment. Combining the equipment's historical maintenance records (the last bearing replacement was eight months ago) and the current degradation rate, the remaining service life was predicted to be approximately 10 days. Considering the water pump's criticality (affecting the entire water supply system), it was given the highest maintenance priority. A personalized maintenance plan was generated, broken down into the following steps: pump shutdown, disassembly, inspection, bearing replacement, reassembly, testing, and startup. Two maintenance engineers were assigned specialized bearing disassembly and assembly tools, along with new bearings. The plan was scheduled to be performed during a period of low system load three days from now, with an estimated maintenance time of four hours. Finally, the maintenance knowledge base was linked to extract the corresponding bearing specifications (6308-2RS), installation torque requirements (65 N·m), assembly clearance standards (0.02-0.05 mm), and vibration test methods (three-point method) for the pump model. A complete maintenance plan, including technical details and safety precautions, was generated and distributed to the maintenance team through the system. The planned maintenance time and expected results were marked on the digital twin model.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A digital twin water plant cluster integrated visualization management system, characterized by: The system comprises: The processing module is used to standardize the multi-source parameter data collected by the water plant equipment sensors to obtain real-time data streams and standardized data sets; A modeling module is used to perform three-dimensional modeling based on the standardized data set using oblique photography technology to obtain a three-dimensional digital twin model of the water plant; A construction module is used to construct a process flow topology adaptive visualization framework based on the digital twin model to obtain a cluster water plant visualization management interface; An extraction module, configured to input the real-time data stream into the visual management interface, extract spatiotemporal features through a multi-channel data fusion mechanism, and obtain early warning information; A calculation module is used to construct a process-equipment-environment ternary association network for the warning information, calculate abnormal patterns using a heterogeneous graph attention network, and obtain a diagnosis result; The encoding module is used to generate a visual encoding of the equipment health status based on a digital watermark based on the diagnosis result to obtain a predictive maintenance plan.

2. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The processing module is used to: Water quality parameters, flow parameters, pressure parameters, equipment operation parameters and energy consumption data are collected through a multi-level data acquisition network to obtain the original data set; Applying a data cleaning algorithm to the original data set to process outliers, duplicate values, and missing values ​​to obtain preliminarily processed data; Performing unified format conversion and data standardization on the preliminarily processed data to obtain standardized data in a unified format; Classifying and organizing the standardized data in a unified format according to temporal, spatial and process relationships to obtain a structured database; Establishing a data receiving and verification mechanism based on the structured database to verify the integrity and accuracy of the data and obtain verified and credible data; Time series analysis and feature extraction are performed on the verified trusted data, and statistical features are calculated using a sliding window algorithm to obtain real-time data streams and standardized data sets.

3. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The modeling module is used to: Multi-angle oblique photography was carried out on the water plant, and high-resolution 1:500 image data was acquired through drones to obtain the original image data of the water plant; The original image data is registered and fused with the joystick aerial survey image data, and a spatial correspondence is established through a homonymous point matching algorithm to obtain a fused image in a unified coordinate system; Based on the fused image, a 3D reconstruction algorithm is used to generate a point cloud model of the water plant, and spatial point coordinates are calculated through multi-view stereo matching to obtain high-precision 3D point cloud data; Performing geometric processing on the three-dimensional point cloud data, constructing a three-dimensional geometric model of the water plant scene through a gridding algorithm and a surface reconstruction method, and obtaining a three-dimensional model of the process equipment; Performing spatial mapping between the three-dimensional model of the process equipment and the process flow data in the standardized data set, establishing a data-model correspondence, and obtaining a dynamically associated digital model; Material mapping and texture mapping are performed on the dynamically associated digital model, and the lighting and shadow effects are processed through rendering optimization technology to obtain a three-dimensional digital twin model of the water plant.

4. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The building blocks are used to: Performing a process flow topology structure analysis on the digital twin model, extracting the connection relationship between process units through a graph theory algorithm, and obtaining a process flow topology diagram; Design a three-layer architecture visualization framework based on the process flow topology diagram, allocate functional nodes of the front-end WebGL rendering layer, the middleware data processing layer, and the back-end service layer to obtain a basic visualization architecture; Developing a three-dimensional scene visualization module for the basic visualization architecture, and obtaining a three-dimensional interactive scene by implementing zooming, translation, rotation, and first-person roaming interactive functions; Integrate the three-dimensional interactive scene with a dynamic graphic representation system for the process flow, and display the process parameter status through the color change of process nodes and dynamic particle flow effects to obtain a process flow visualization interface; Add a multi-dimensional data chart display function to the process flow visualization interface, present key indicators of water quality, flow rate, and energy consumption through bar charts, line charts, and heat map visualization components to obtain a comprehensive analysis panel; Based on the comprehensive analysis panel, a unified management system for water plant clusters is constructed, and visualization and comparative analysis of multiple water plants are achieved through a regional division algorithm and a process type grouping mechanism, thereby obtaining a visual management interface for cluster water plants.

5. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The extraction module is used to: Establishing a real-time monitoring mechanism for process parameters for the real-time data stream, and updating a visualization interface by tracking parameter changes to obtain real-time monitoring data; A multi-level threshold warning system is established based on the real-time monitoring data, and key parameters of water quality, flow rate and pressure are divided into normal range, warning threshold and alarm threshold to obtain a threshold rule base; Performing real-time comparison between the real-time monitoring data and the threshold rule library to identify abnormal parameters that exceed the normal range and obtain preliminary warning signals; Performing time series analysis on the preliminary warning signal, judging the persistence of abnormality through the data change trend of multiple consecutive time points, and obtaining time domain feature data; Perform multi-channel data fusion based on the time domain feature data and the spatial layout information of the process units, extract the operation correlation between the process units, and obtain spatiotemporal feature data; The spatiotemporal feature data is subjected to abnormal classification processing, and is graded according to the degree of abnormality, scope of impact, and degree of urgency to obtain early warning information.

6. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The computing module is configured to: Extract features from the warning information, organize process parameter nodes, equipment status nodes, and environmental condition nodes in a hierarchical structure, and define three types of connection relationships between nodes: process flow edges, equipment dependency edges, and environmental impact edges, to obtain an initial ternary association structure; Based on the initial ternary association structure, a heterogeneous graph network structure is constructed, wherein the process node includes inlet water quality, treatment unit status, and outlet water parameter attributes, the equipment node includes operating status, vibration value, and energy consumption attributes, and the environment node includes temperature, humidity, and external condition attributes, thereby obtaining a multidimensional feature heterogeneous graph; Applying a three-layer heterogeneous graph attention network to the multi-dimensional feature heterogeneous graph, wherein the first layer is a node type feature encoding layer, which processes the features of process nodes, equipment nodes, and environment nodes through three different sets of weight matrices to obtain type-aware node representations; The type-aware node representation is input into the second multi-head attention layer. By parallel computing 8 attention heads, each attention head independently learns the interaction pattern between nodes of different types. The results are concatenated and transformed nonlinearly to obtain the node features with enhanced relationships. The node features of the relationship enhancement are processed through the third global graph pooling layer, which uses a weighted summation method to aggregate the entire graph information and combines residual connections to retain the original features to obtain an abnormal pattern representation vector; The abnormal pattern representation vector is similarly calculated with a predefined fault pattern library, and the most similar fault type and cause are matched using cosine similarity measurement to obtain a diagnosis result.

7. The digital twin water plant cluster comprehensive visualization management system according to claim 1 is characterized in that: The encoding module is used to: Perform a quantitative evaluation of the equipment health index on the diagnostic results, and obtain the equipment health index by comprehensively calculating the three dimensions of fault type, fault severity, and development trend; Based on the device health index, a digital watermark encoding rule is constructed, and the health index is mapped into a specific visual encoding pattern. Through the combination of three visual elements: color, shape, and density, a visual encoding of the device health status is obtained; The device health status visual code is embedded into the corresponding device of the digital twin model, and displayed on the surface of the three-dimensional model by transparency superposition to obtain a visual health status mark; Establishing an equipment life prediction model based on the visualized health status indicator and the equipment's historical maintenance records, determining maintenance priorities through remaining service life estimation, and obtaining a priority maintenance list; Generate a personalized maintenance plan for the equipment in the priority maintenance list, arrange maintenance time and required resources through task decomposition and resource matching algorithm, and obtain a detailed maintenance task list; The detailed maintenance task list is associated with a maintenance knowledge base, and corresponding maintenance steps and safety precautions are extracted according to equipment type and fault characteristics to obtain a predictive maintenance plan.

Images