Water environment monitoring station operation and maintenance simulation system based on digital twinning

Abstract

This invention discloses a digital twin-based water environment monitoring station operation and maintenance simulation system, belonging to the field of environmental monitoring and operation and maintenance management technology. The system includes a multi-source sensing unit for the station, a digital twin modeling engine, an operation and maintenance simulation decision-making platform, and a station execution feedback unit, constructing a three-layer collaborative architecture. A high-fidelity digital twin is constructed through multi-source data fusion and generative 3D reconstruction. A two-stage anomaly detection model accurately identifies equipment and water quality anomalies, and an improved quantile regression model is used to predict future trends, ultimately generating an optimal operation and maintenance strategy and forming a closed loop. This invention solves the problems of passive operation and maintenance and extensive resource allocation in existing systems, upgrading the operation and maintenance mode from passive repair to proactive predictive maintenance, improving operation and maintenance efficiency, and reducing operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and operation and maintenance management technology, and more specifically, to a water environment monitoring station operation and maintenance simulation system based on digital twins. Background Technology

[0002] Water environment monitoring stations are core infrastructure for watershed water quality supervision, pollution early warning, and ecological assessment. Their stable operation directly affects the continuity and reliability of monitoring data. Current monitoring station operation and maintenance management mainly relies on periodic manual inspections and remote status monitoring, which presents the following prominent problems: First, operation and maintenance decisions lack the ability to accurately predict the health status of station equipment, often adopting a post-failure repair model, resulting in long periods of data loss and delayed emergency response; Second, the geographically dispersed nature of monitoring stations and the significant differences in environmental conditions mean that the allocation of operation and maintenance resources lacks a global optimization basis, leading to low efficiency in manpower and spare parts scheduling; Third, existing simulation systems mostly focus on simulating single physical processes, failing to construct digital twins of all station elements and multi-physical field coupling, thus unable to achieve real-time mapping and forward-looking projection of station operating status; Fourth, anomaly detection methods rely on fixed threshold rules, making it difficult to capture latent fault precursors under the coupling of multiple parameters, resulting in high false alarm and false negative rates.

[0003] Therefore, it is necessary to design a water environment monitoring station operation and maintenance simulation system that integrates digital twins, multi-source data intelligent planning, and predictive operation and maintenance decision-making to achieve high-fidelity mapping of station operation status, dynamic assessment of equipment health, and proactive optimization of operation and maintenance strategies. Summary of the Invention

[0004] This invention proposes a water environment monitoring station operation and maintenance simulation system based on digital twins. In response to the problems of passive operation and maintenance response, extensive resource allocation, insufficient simulation fidelity and weak anomaly detection capabilities of existing monitoring stations, a three-layer collaborative architecture of "physical station perception layer - digital twin modeling layer - operation and maintenance simulation decision layer" is constructed to realize full-element mirror mapping of the operation status of monitoring stations, future situation prediction and intelligent generation of operation and maintenance strategies.

[0005] This invention provides a water environment monitoring station operation and maintenance simulation system based on digital twins. The system includes: a multi-source sensing unit for the station, a digital twin modeling engine, an operation and maintenance simulation decision platform, and a station execution feedback unit. The multi-source sensing unit at the site is deployed in various functional areas of the water environment monitoring station and the surrounding water area. It integrates water quality sensing array, hydrological sensing array, meteorological sensing array, equipment status sensing array, video monitoring equipment that collects visible light video frame images, and panoramic camera that collects panoramic images to obtain multi-source heterogeneous sensing data streams. The digital twin modeling engine communicates with the multi-source sensing units of the site, performs spatiotemporal alignment and multimodal fusion processing on multi-source heterogeneous sensor data streams, constructs a high-fidelity digital twin of the monitoring site, and generates a multi-dimensional situation map of the site's operating status. The engine integrates a multi-source data intelligent fusion module, a digital twin scene generation module, a site health status assessment module, a future situation prediction module, and a twin calibration and optimization module. Multi-source data intelligent fusion module: Based on a unified spatiotemporal benchmark, it performs time synchronization, spatial registration and confidence weighted fusion of water quality sensor data, hydrological sensor data, meteorological sensor data, equipment status sensor data and video monitoring image data to construct a data cube of the entire site's operating status and generate a full-domain operating status map including water quality distribution field, hydrodynamic parameter field and equipment temperature field. The digital twin scene generation module takes panoramic images captured by a panoramic camera as input and employs a digital twin construction pipeline based on generative 3D reconstruction to convert panoramic images into editable 3D Gaussian sputtering and collision mesh representations. It then generates diverse digital twin variant scenes through prompt-driven scene editing, forming a high-fidelity site digital twin. The digital twin scene generation module also supports the stitching of multiple digital twin variant scenes. Through panoramic feature matching and geometric iteration nearest-point fine registration, it merges multiple sub-region digital twins into a unified simulation environment. Site health status assessment module: The visible light video frame images collected by the video surveillance equipment are input into the pre-trained two-stage anomaly detection model. The first stage uses the YOLOv8n model to locate the site equipment and key water areas and extract the region of interest. The second stage uses the RexNet-150 model to perform normal or abnormal binary classification of the extracted region of interest and outputs the site equipment health status label and the water quality abnormal event label. Future Situation Simulation Module: Taking the overall operational status map, site equipment health status indicators, and third-party meteorological and hydrological forecast data as input, it uses an improved quantile regression convolutional low-rank model to generate probabilistic prediction intervals for key operational parameters in the future period. Combined with the digital twin scenario variants output by the digital twin scenario generation module, it performs multi-scenario Monte Carlo simulations and generates a site operational status simulation report. Twin calibration and optimization module: Using the observation data transmitted back in real time by the multi-source sensing unit of the site as the calibration benchmark, the module calculates the residual consistency score between the digital twin output and the physical entity observation, and uses the residual-driven adaptive calibration algorithm to dynamically adjust the model parameters of the digital twin so that the digital twin continuously approximates the actual operating state of the physical site. The operation and maintenance simulation decision-making platform, deployed on a cloud server, interacts bidirectionally with the digital twin modeling engine and the site execution feedback unit via a secure communication link. The platform receives site operation status simulation reports and site equipment health status indicators from the digital twin modeling engine, and based on a pre-set operation and maintenance strategy knowledge base, generates an optimal set of operation and maintenance strategies, including inspection path planning, spare parts pre-allocation plans, and equipment preventative maintenance work orders, which are then distributed to the site execution feedback unit for execution. Simultaneously, the platform provides users with a 3D visualized operation and maintenance simulation interface, early warning information push notifications, and historical operation and maintenance performance analysis reports. The site execution feedback unit includes a drone inspection execution terminal, a site local controller, and a mobile terminal for maintenance personnel. It is used to receive and execute the optimal set of operation and maintenance strategies issued by the operation and maintenance simulation decision platform, and to send the execution results and on-site feedback data back to the operation and maintenance simulation decision platform to form an operation and maintenance closed loop.

[0006] By adopting the above solution, the beneficial effects achieved by the present invention are as follows: This invention achieves high-fidelity real-time mapping between the physical entities and virtual twins of water environment monitoring stations by constructing a digital twin modeling engine. The multi-source data intelligent fusion module performs unified spatiotemporal alignment and confidence-weighted fusion of multimodal heterogeneous data such as water quality, hydrology, meteorology, equipment status, and video images, generating a comprehensive operational status map of the station. This solves the fragmented situational awareness problem caused by data silos and inconsistent spatiotemporal benchmarks in traditional monitoring systems. The digital twin scene generation module adopts a pipeline based on generative 3D reconstruction, supporting the automatic construction of editable high-fidelity 3D simulation scenes from panoramic images. It also generates diverse digital twin variant scenes through prompt-driven editing, providing a rich simulation data foundation for operation and maintenance simulations under different working conditions, significantly improving the coverage and robustness of the simulation system.

[0007] This invention achieves accurate identification of the health status of site equipment and water quality anomalies through a two-stage anomaly detection model. The model first uses YOLOv8n for target-level region of interest extraction, effectively filtering out background noise interference. Then, RexNet-150 is used for refined binary classification of the focused region, significantly improving both the precision and recall of anomaly detection. This is particularly suitable for monitoring site scenarios where anomaly samples are sparse and the background is complex and variable. Furthermore, the future situation projection module employs an improved quantile regression convolutional low-rank model to directly generate multi-step forward probabilistic prediction intervals, avoiding the strong assumptions about error distribution inherent in traditional point prediction methods. A conformal prediction calibration mechanism ensures the theoretical coverage of the prediction intervals, providing reliable quantitative uncertainty information for operation and maintenance decisions.

[0008] The operation and maintenance simulation decision platform of this invention integrates digital twin simulation results with an operation and maintenance strategy knowledge base. Based on the predicted equipment health degradation trend and potential abnormal events, it can automatically generate an optimal set of operation and maintenance strategies, including inspection path planning, spare parts pre-allocation, and preventive maintenance work orders. Through the site execution feedback unit, it forms a closed-loop verification and continuous optimization, realizing the upgrade of the operation and maintenance of monitoring sites from "passive response maintenance" to "proactive predictive maintenance", effectively reducing the average fault repair time and the total life cycle operation and maintenance cost of the site. Attached Figure Description

[0009] Figure 1 This is a system architecture diagram of a water environment monitoring station operation and maintenance simulation system based on digital twin proposed in this invention; Figure 2 This is a schematic diagram of the reasoning process of the two-stage anomaly detection model in this invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0011] Example 1: according to Figure 1 This invention provides a digital twin-based water environment monitoring station operation and maintenance simulation system, applicable to automatic water quality monitoring stations at river sections, buoy monitoring platforms in lakes and reservoirs, and comprehensive monitoring stations at drinking water sources. The system includes a multi-source sensing unit for the station, a digital twin modeling engine, an operation and maintenance simulation decision platform, and a station execution feedback unit.

[0012] The site's multi-source sensing unit specifically includes: a water quality sensor array, containing an online five-parameter water quality analyzer (temperature, pH, dissolved oxygen, conductivity, turbidity, measurement accuracy ±0.1℃ / ±0.05 pH / ±0.1 mg / L / ±1% / ±2%), an online permanganate index analyzer (measurement range 0~20 mg / L, accuracy ±5%), an online ammonia nitrogen analyzer (measurement range 0~10 mg / L, accuracy ±5%), and an online total phosphorus and total nitrogen analyzer (measurement range 0~5 mg / L, accuracy ±5%); a hydrological sensor array, containing a radar water level gauge (measurement range 0~30 m, accuracy ±3 mm) and an acoustic Doppler current profiler (flow velocity measurement range ±5 m / s, accuracy ±0.5%); and a meteorological sensor array, containing an ultrasonic anemometer (wind speed range 0~60 m / s, accuracy ±0.2 m / s) and an atmospheric temperature and humidity sensor (accuracy ±0.2℃ / ±2%). RH), photosynthetically active radiation sensor (accuracy ±5%); equipment status sensor array, including vibration sensors, current transformers, and temperature sensors deployed in the water pump, air compressor, autosampler, and communication cabinet; video monitoring equipment, including three 4-megapixel starlight-level network PTZ cameras to collect visible light video frame images of the equipment area inside the station, the water sampling platform, and the surrounding water area; panoramic camera (Insta360 Titan, 11K resolution) to collect panoramic images.

[0013] The digital twin modeling engine is deployed on the site's edge computing server, configured with an Intel Xeon Silver 4310 processor, 64 GB of memory, an NVIDIA RTX 4090 GPU, and a 2 TB NVMe solid-state drive. It connects to the various sensor arrays, video surveillance equipment, and panoramic cameras of the site's multi-source sensing unit via industrial Ethernet, with a data acquisition frequency adjustable from 1 Hz to 0.1 Hz. The engine integrates a multi-source data intelligent fusion module, a digital twin scene generation module, a site health status assessment module, a future situation prediction module, and a twin calibration and optimization module. The modules interact with each other through a shared memory mechanism, with a processing latency of ≤200 ms.

[0014] The operation and maintenance simulation decision-making platform is deployed on the Huawei Cloud Stack enterprise cloud platform, configured with a 32-core, 128 GB memory virtual server. It transmits data bidirectionally to the digital twin modeling engine via a 5G communication terminal. The platform's backend integrates a PostgreSQL relational database, a ChromaDB vector database, and a MinIO object storage to store historical operation data, digital twin model versions, and an operation and maintenance strategy knowledge base. The frontend provides a WebGL-based 3D visualization operation and maintenance simulation interface, supporting real-time browsing of digital twin scenarios, parameter curve display, and early warning information pop-ups.

[0015] The site execution feedback unit includes: a DJI Mavic 3T drone inspection execution terminal, equipped with an infrared thermal imaging and visible light dual-light gimbal, which performs equipment appearance inspection and checks for abnormal floating objects on the water surface according to the planned path; a Siemens S7-1500 site local controller, which is electrically connected to actuators such as water pumps, valves, and air compressors, and receives and executes equipment control commands; and a maintenance personnel mobile terminal, which runs a maintenance work order management application, receives preventive maintenance tasks, and uploads on-site handling photos and maintenance records.

[0016] Example 2: This embodiment, based on Embodiment 1, elaborates in detail on the specific implementation process of each core module in the digital twin modeling engine.

[0017] The specific workflow of the multi-source data intelligent fusion module is as follows: Based on the IEEE 1588 precision time synchronization protocol, nanosecond-level timestamps are added to each frame of data from the water quality sensor array, hydrological sensor array, meteorological sensor array, equipment status sensor array, and video monitoring equipment; for sensor sampling points with irregular spatial distribution, the Bayesian Skrigin interpolation algorithm is used to reconstruct the spatial grid of the 100 m × 100 m water area around the monitoring station, with the grid resolution set to 2 m × 2 m; based on the factory calibration certificates and historical operational stability records of each sensor, a sensor confidence vector is constructed, where the confidence value is in the range of [0,1]; for the same type of parameter values ​​from multiple sensors within the same grid, a weighted average fusion is performed according to the confidence value, and finally a full-domain operational status map containing water quality parameter fields (dissolved oxygen concentration distribution, turbidity distribution), hydrodynamic parameter fields (flow velocity vector distribution, water level elevation), and equipment thermal fields (cabinet temperature distribution, pump vibration intensity distribution) is generated, with an update cycle of 10 seconds.

[0018] The digital twin scene generation module adopts a digital twin construction pipeline based on generative 3D reconstruction, and its specific implementation steps are as follows: Step G1: The panoramic images captured by the panoramic camera are transmitted in real time to the digital twin modeling engine via the site's multi-source sensing unit; the panoramic image sequence captured by two panoramic cameras deployed on the top of the monitoring station and the water sampling platform is used as input, and the scene geometry and texture are reconstructed from coarse to fine using the Marble multimodal world model, outputting a 3D Gaussian sputtering representation file (PLY format) and a collision mesh representation file (OBJ format); the 3D Gaussian sputtering representation is used to achieve lighting and material rendering that is highly consistent with the real scene, and the collision mesh representation is used to support collision detection and interactive operations in subsequent physical simulations; Step G2: Use a 3D format conversion tool to convert the 3D Gaussian sputtering PLY data and collision mesh OBJ data output in Step G1 into the universal scene description USD format, and import them into the NVIDIA Isaac Sim physics simulation engine to ensure the compatibility of the twin with downstream simulation modules. Step G3: Edit the reconstructed scene based on natural language prompts to generate digital twin variant scenes; the specific prompt template is: "Monitoring station scene, [season] lighting conditions, water surface exhibits [water quality status] characteristics, equipment appearance is [newness / oldness]"; by replacing the semantic slot values ​​in square brackets (season: spring / summer / autumn / winter; water quality status: clear / slightly eutrophic / high turbidity; newness / oldness: factory condition / one year of use with rust / covered with stains), 12 different digital twin variant scenes are generated. Each variant scene retains the topology and equipment layout of the original scene, with differences only in visual texture, ambient lighting, and some semantic elements; Step G4: For multi-area site environments including the station building interior, water intake platform, and shore buffer zone, since a single panoramic image cannot cover the entire area, multiple digital twin variant scene stitching is performed. First, local feature extraction and matching are performed on the overlapping images of adjacent areas (panoramic view of the station building interior and panoramic view of the water intake platform): The SuperPoint model is used to extract the 256-dimensional feature descriptor and corresponding key point positions of each image, and the LightGlue model is used for feature matching to obtain a set of matching point pairs. Based on the matching point pairs, the essential matrix is ​​decomposed using the eight-point method to obtain the relative rotation matrix. With unit translation vector Then, using the known camera installation height... Meters (vertical distance from the ground plane to the camera's optical center) and the three-dimensional ground plane point set reconstructed in step G1 The metric scaling factor is calculated using the following formula. : ; in, This is a metric for scaling factors; is the known installation height of the panoramic camera relative to the ground plane; n is the ground plane normal vector; Indicates transpose; This is the set of ground points obtained by triangulation in a unit-scale coordinate system; For ground points; Indicates to Each ground point calculate The set of results obtained; This indicates the median calculation, after removing outliers; the calculated value... The value is approximately 3.21, representing the unit translation vector. Restored to metric-level translation vector , forming a rough registration pose ; Step G5: In coarse registration pose Based on this, fine-grained point-to-plane iterative nearest-point registration is performed on the dense point cloud of the overlapping region between the two areas; the fine pose to be optimized is defined. ,by Using the initial value, minimize the geometric distance error function. After convergence, obtain the fine registration pose. Then, apply the 3D Gaussian sputtering kernel of each sub-region through... The coordinates are uniformly transformed to a global coordinate system and merged into a globally unified digital twin simulation environment, which is then imported into the Isaac Sim platform for subsequent modules to use. The Marble multimodal world model is a spatial intelligent model whose core function is to automatically convert various input signals, such as text, single images, panoramic images, multi-view images, or videos, into a complete three-dimensional world. This model uses three-dimensional Gaussian sputtering as the primary scene representation format, capable of outputting Gaussian sputtering files in PLY format and collision mesh files in OBJ format, and supports semantic and geometric level scene editing based on natural language prompts. In this embodiment, the Marble multimodal world model receives panoramic image sequences from a water environment monitoring station acquired by a panoramic camera, and automatically generates a digital twin and multiple variant scenes through steps G1 to G3.

[0019] according to Figure 2 The site health status assessment module adopts a two-stage anomaly detection model, and its specific implementation steps are as follows: Step H1: The video surveillance equipment acquires visible light video frame images in real time, processes them through the multi-source data intelligent fusion module, and stores them in the global operation status map; reads the visible light video frame images (acquired by cameras inside the station building and water intake platform of the video surveillance equipment, resolution 1920×1080, sampling frame rate 1 fps) from the global operation status map, and inputs them into the YOLOv8n target detection model; the YOLOv8n model is pre-trained on the COCO dataset, the detection confidence threshold is set to 0.25, and outputs the bounding box coordinates (top left corner x, y, width, height) and COCO category labels of all detected targets in each frame; Step H2: Iterate through all detected targets output by the YOLOv8n target detection model, and only retain bounding boxes whose category labels belong to the predefined set of interest categories C={"monitoring instruments","water sampling pipelines","water surface","floating objects"}; for example, irrelevant targets with category labels such as "person" and "chair" are directly discarded. Step H3: Based on the preserved bounding box coordinates, use OpenCV library functions to crop the corresponding regions from the original image to obtain the region of interest (ROI) images; scale each ROI image to 224×224 pixels and normalize it according to the mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] of the ImageNet dataset; Step H4: The preprocessed region-of-interest images are sequentially input into the RexNet-150 classification model. This model is initialized with weights pre-trained on ImageNet and fine-tuned on a self-built site maintenance image dataset. The self-built dataset contains 273,897 samples collected from 10 different monitoring sites. The abnormal samples cover categories such as equipment damage, pipe leakage, abnormal floating objects on the water surface, and precursors to algal blooms. The samples are divided into training, validation, and test sets in an 8:1:1 ratio. The Adam optimizer is used for fine-tuning training with an initial learning rate of 3×10⁻⁶. -4 The batch size is 16, and the training is conducted for 10 rounds. The model with the highest F1 score on the validation set is selected as the final model. The RexNet-150 model outputs a binary classification confidence score for each region of interest, representing the probability that the region belongs to the "abnormal" category. Step H5: Mark regions of interest with confidence scores exceeding the judgment threshold of 0.7 as anomalous events; for regions marked as anomalous, draw a red highlighted border on the original video frame image based on its bounding box coordinates, and overlay an anomalous type label; if the anomalous type is water quality anomalous (algal bloom, oil film), further retrieve the time series data of dissolved oxygen, turbidity, and chlorophyll a parameters of the corresponding water area from the water quality sensor array for the past 1 hour; if the sensor data also deviates from the historical baseline, then raise the confidence level of the anomalous event to "confirmed anomalous", and output the site equipment health status label (normal / attention / abnormal level three) and the water quality anomalous event label.

[0020] The future situation projection module uses an improved quantile regression convolutional low-rank model, and its specific implementation steps are as follows: Step F1: Use the dissolved oxygen concentration time-series data (sampling interval 10 minutes, a total of 144 historical data points corresponding to the past 24 hours) in the global operational status map as the observation sequence, and set the length of the observation sequence. Target predicting the future Dissolved oxygen concentration values ​​at each time step (corresponding to the next 12 hours); set model size parameters. ,satisfy The construction length is Recovery target vector , among which the former The last 56 observations of the observation sequence are filled in at each position as the known part, and then... Each position represents a predicted value to be recovered; Step F2: From the historical execution database (which stores all time-series data from the past 90 days), select the window size by sliding window step 1 and window length. Extracting the training sample set using method 128 A total of Training samples; the transformation matrix is ​​learned using an optimization algorithm based on projective gradient descent. ,satisfy The learning objective is to minimize Where the kernel size After training converges, save the transformation matrix. For use in subsequent online predictions; Step F3: Set the significance level (Corresponding to the 90% prediction interval), for the estimation of the upper bound of the prediction interval, set the quantile level. ; Set regularization hyperparameters Construct the optimization objective function: ; in, This is the set of indices for the first 56 positions; for Quantile loss at time; Step F4: Iteratively solve using the alternating direction multiplier method, initializing... For zero vectors, the penalty parameter Maximum number of iterations: 500; during updates In the iteration, for Given the location, the updated value is calculated using the following improved quantile regression formula: ; in, This represents the number of quantiles obtained after updating using the improved quantile regression. The value of the nth variable corresponds to the nth variable in the time series. The predicted or recovered value at each moment; This is the gradient update reference value calculated by the nuclear norm proximal operator in the current iteration, i.e., the temporary value given by the proximal gradient step size; For the first The observation values ​​at each time point; The regularization hyperparameter controls the balance between the quantile loss term and the nuclear norm term; in this embodiment, it is set to 1000. As a quantile level, the upper bound of the prediction interval is estimated at 0.95, and the lower bound is estimated at 0.05. The penalty parameter in the alternating direction multiplier method is set to 1.0 in this embodiment; In this embodiment, the kernel size for the convolution operation in the transformation matrix is ​​set to 64. for For unknown locations, directly use Use this as the update value; alternate updates until the change in the objective function value is less than [the change in the objective function value]. Obtain the upper bound sequence of the prediction interval. ; Step F5: Change the quantile level to Repeat steps F3 to F4 to obtain the lower bound sequence of the prediction interval. Together with the upper bound sequence, they form a preliminary probabilistic prediction interval; Step F6: Divide the first 120 data points of the observed sequence into a training set and the last 24 data points into a calibration set; apply steps F3 to F5 to the calibration set to generate prediction intervals for 24 time steps, and calculate the residual for each time step. This forms a set of residual consistency scores. ;Pick of Quantiles as calibration bias Calculated The final forecast interval will be adjusted to , ; Step F7: Call the 12 sets of digital twin variant scenarios generated by the digital twin scenario generation module, and repeat steps F3 to F6 50 times each in each set of scenarios (Monte Carlo sampling) to obtain a total of 600 dissolved oxygen concentration projection trajectories for the next 12 hours; count the percentage of time in each trajectory where the dissolved oxygen concentration is below 5.0 mg / L (preset hypoxia threshold), calculate the probability of hypoxia event occurring in the next 12 hours as 0.23, and predict the maximum impact range to cover 200 m downstream of the monitoring section; integrate the above analysis results into a site operation status projection report and output it to the operation and maintenance simulation decision platform in JSON format.

[0021] The digital twin calibration and optimization module performs calibration on an hourly basis: it reads all sensor observation data returned by the multi-source sensing unit of the site in the past hour and performs parameter-by-parameter residual calculation with the simulation output of the digital twin under the same operating conditions; if the root mean square error of a certain parameter exceeds a preset threshold for three consecutive cycles, the transformation matrix is ​​triggered. Incremental updates, using new observation data to update the matrix By performing a single gradient descent iteration, the operational state of the digital twin is continuously simulated to closely approximate the actual evolution trajectory of the physical site.

[0022] Example 3: This embodiment, based on Embodiment 1 and Embodiment 2, describes the collaborative working mechanism of the operation and maintenance simulation decision-making platform and the site execution feedback unit.

[0023] After receiving the site operation status simulation report output by the future situation simulation module, the operation and maintenance simulation decision platform analyzes the abnormal event prediction information contained therein; when the probability of any abnormal event in the prediction report exceeds 0.3, the platform automatically triggers the operation and maintenance strategy generation process.

[0024] The operation and maintenance strategy knowledge base contains over 200 pre-defined operation and maintenance rules, each including triggering conditions, action templates, and priorities. The platform matches the anomaly types, predicted occurrence times, and impact ranges in the simulation report with the knowledge base rules, activating the corresponding operation and maintenance strategy generation engine. For the predicted "abnormal temperature rise of water pump bearing" event (probability of occurrence 0.42, expected occurrence time 4 hours later), the platform calls a genetic algorithm to solve the inspection path planning problem, with the dual objectives of minimizing the total inspection path duration and covering the most risk points, generating an optimized flight path for the UAV inspection execution terminal. Simultaneously, the platform checks the spare parts inventory database and finds that only 1 spare part of the same model remains, which is below the safety stock threshold of 3 parts. It automatically generates a spare parts pre-allocation request and pushes it to the warehouse management system.

[0025] The generated inspection routes, preventive maintenance work orders (including bearing replacement operation guidelines and safety precautions), and spare parts allocation plans constitute the optimal operation and maintenance strategy set, which is sent to the site execution feedback unit via the 5G link.

[0026] After receiving the flight path document, the drone inspection terminal takes off autonomously and flies along the route. It uses an infrared thermal imaging gimbal to scan the temperature of the water pump motor and bearings, and transmits the captured visible light and infrared images back to the operation and maintenance simulation decision-making platform in real time for AI-assisted diagnosis. After receiving the preventive maintenance work order, the site's local controller automatically switches the water pump to standby operation and closes the inlet and outlet valves of the faulty pump before the maintenance personnel arrive on site. The maintenance personnel receive the work order information on their mobile terminals. After completing the bearing replacement according to the work order instructions, they take a photo of the repair and upload it through the terminal application, marking the work order status as completed.

[0027] After the work order completion status and the on-site image data collected by the drone inspection are sent back to the operation and maintenance simulation decision-making platform, the platform updates the equipment health record and archives the entire process data of prediction-response-handling of this event into the performance analysis database for subsequent iterative optimization of knowledge base rules, forming a complete predictive operation and maintenance closed loop.

[0028] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A water environment monitoring station operation and maintenance simulation system based on digital twins, characterized in that, The system includes: a site multi-source sensing unit, a digital twin modeling engine, an operation and maintenance simulation decision platform, and a site execution feedback unit; The multi-source sensing unit at the site is deployed in various functional areas of the water environment monitoring station and the surrounding water area. It integrates water quality, hydrology, meteorology and equipment status sensor arrays, video monitoring equipment that collects visible light video frame images and panoramic cameras that collect panoramic images to obtain multi-source heterogeneous sensor data streams. The digital twin modeling engine communicates with the site's multi-source sensing unit to output a site operation status simulation report and site equipment health status indicators. The operation and maintenance simulation decision platform is deployed on a cloud server and interacts bidirectionally with the digital twin modeling engine and the site execution feedback unit. The platform receives the site operation status simulation report and site equipment health status indicators output by the digital twin modeling engine, and generates the optimal operation and maintenance strategy set based on the preset operation and maintenance strategy knowledge base, and sends it to the site execution feedback unit for execution. The site execution feedback unit, including the drone inspection execution terminal, the site local controller, and the maintenance personnel's mobile terminal, receives and executes the optimal operation and maintenance strategy set, and sends the execution results and on-site feedback data back to the operation and maintenance simulation decision platform to form an operation and maintenance closed loop. The digital twin modeling engine also includes the following modules: Multi-source data intelligent fusion module: processes multi-source heterogeneous sensor data streams, constructs a data cube of the entire site's operational status, and generates a map of the entire operational status. Digital twin scene generation module: Taking panoramic images as input, it adopts a digital twin construction pipeline based on generative 3D reconstruction to convert panoramic images into 3D Gaussian sputtering representation and collision mesh representation, generating high-fidelity digital twins and digital twin variant scenes; Site health status assessment module: Input visible light video frame images into a pre-trained two-stage anomaly detection model, and output site equipment health status indicators and water quality anomaly event indicators; Future Situation Simulation Module: Pre-stores third-party meteorological and hydrological forecast data, takes the overall operational status map, station equipment health status indicators and third-party meteorological and hydrological forecast data as input, uses an improved quantile regression convolutional low-rank model to generate probabilistic prediction intervals, and combines digital twin variant scenarios to conduct multi-scenario Monte Carlo simulations to generate a station operational status simulation report. Twin calibration optimization module: Using the observation data transmitted back in real time by the multi-source sensing units at the site as the calibration benchmark, the module dynamically adjusts the model parameters of the digital twin using a residual-driven adaptive calibration algorithm.

2. The system according to claim 1, characterized in that, The digital twin scene generation module constructs a digital twin and performs scene splicing of digital twin variants using the following steps: Step G1: Using the panoramic image as input, the scene geometry and texture are reconstructed using the Marble multimodal world model, and the output is a 3D Gaussian sputtering representation and a collision mesh representation. Step G2: Convert the 3D Gaussian sputtering representation and collision mesh representation into a general scene description format; Step G3: Based on natural language prompts, perform semantic and geometric editing on the reconstructed scene to generate multiple digital twin variant scenes; Step G4: For multiple digital twin variant scenarios, the relative rotation matrix and unit translation vector are obtained by local feature extraction and matching and decomposition of the essential matrix. The metric scale factor is solved by using the camera installation height and ground plane point set to recover the metric translation vector and form a coarse registration pose. The formula for calculating the scaling factor is: ； in, This is a metric for scaling factors; is the known installation height of the panoramic camera relative to the ground plane; n is the ground plane normal vector; Indicates transpose; This is the set of ground points obtained by triangulation in a unit-scale coordinate system; For ground points; Indicates to Each ground point calculate The set of results obtained; This indicates the median calculation; after solving for the metric factor, the formula is used... Translate the unit vector Restored to a metric-level translation vector , forming a coarse registration pose; Step G5: Based on the coarse registration pose, perform fine registration of the dense point cloud in the overlapping area of ​​adjacent sub-regions from point to plane iterative nearest point, unify the 3D Gaussian sputtering kernel of each sub-region to the global coordinate system, and form a stitched global simulation environment.

3. The system according to claim 1, characterized in that, The site health status assessment module employs a two-stage anomaly detection model, specifically including: Step H1: Input the visible light video frame image into the YOLOv8n target detection model and output the bounding box and category label of the detected target; Step H2: Filter the detected targets by category, retaining only the bounding boxes of the predefined categories of interest; Step H3: Crop the region of interest image based on the retained bounding box coordinates, and scale and normalize it to obtain the preprocessed region of interest image; Step H4: Input the preprocessed region of interest image into the RexNet-150 classification model and output the confidence score of each region of interest belonging to the abnormal category; the RexNet-150 classification model is initialized with transfer learning using weights pre-trained on the ImageNet dataset and fine-tuned on a site operation and maintenance image dataset labeled with normal or abnormal binary classification labels; Step H5: Mark regions of interest whose confidence scores exceed the preset judgment threshold as abnormal events, and perform cross-validation with the water quality sensor values ​​of the corresponding regions to output the health status of the site equipment and the water quality abnormal event identifier.

4. The system according to claim 1, characterized in that, The future situation projection module uses an improved quantile regression convolutional low-rank model to generate probabilistic prediction intervals, specifically including: Step F1: Using the time-series operational parameter sequence in the global operational status map as the observation sequence, construct the recovery target vector; Step F2: Extract the training sample set from the observation sequence and learn the transformation matrix that satisfies the column orthogonality constraint so that the transformed vector has the low-rank convolution property; Step F3: Construct an improved quantile regression convolutional low-rank recovery model, whose objective function is: ； in, This is the complete vector of runtime parameters to be restored; express 3D real vector space; The goal is to be in the real vector space Find the value that minimizes the objective function. ; The transformation matrix; Transformation matrix with vector The product of these terms yields the transformed vector; To generate a linear mapping operator for the convolution matrix, The kernel size; The matrix nuclear norm; It is an index variable; A set of indices for known observations; For vectors The The component, i.e., the model in the th , The estimated recovery value at each time point; For the first The actual observed value at each moment; For regularization hyperparameters; The quantile loss function; This refers to the quantile level parameter; It is a positive part function; Step F4: Iteratively solve using the alternating direction multiplier method. For the index position corresponding to the existing observation in the observation sequence, calculate the recovery value of the current iteration step according to the proximal gradient update rule of the improved quantile regression. This update rule is determined jointly by the current gradient term, the observation value, and the regularization term. For positions without observations, directly use the current gradient term as the recovery value. Step F5: Set the upper quantile level and the lower quantile level respectively, and execute steps F3 to F4 to obtain the upper and lower bound sequences of the prediction interval; Step F6: Using a conformal prediction-based interval calibration method, calculate the calibration bias and adjust the final prediction interval to the calibrated interval; Step F7: Couple the final prediction interval with the digital twin variant scenario, and generate a site operation status simulation report by statistically analyzing the proportion and intensity distribution of operating parameters exceeding the safety threshold through Monte Carlo sampling.

5. The system according to claim 1, characterized in that, The twin calibration and optimization module uses the real-time transmitted observation data as a benchmark to calculate the residual consistency score between the digital twin output and the physical entity observation. The residual consistency score refers to the root mean square error of the residual sequence between the digital twin output value and the physical entity observation value. When the root mean square error exceeds a preset threshold for multiple consecutive periods, the incremental update of the transformation matrix is ​​triggered, so that the digital twin continuously approximates the actual operating state of the physical site.

6. The system according to claim 1, characterized in that, The operation and maintenance simulation decision-making platform automatically matches the rules in the operation and maintenance strategy knowledge base with the abnormal event prediction information in the site operation status simulation report, generates the optimal operation and maintenance strategy set including inspection path planning, spare parts pre-allocation plan and equipment preventive maintenance work order, and sends it to the site execution feedback unit through the 5G link; after the site execution feedback unit executes it, it sends the result back, forming an operation and maintenance closed loop.

Images