A three-dimensional visualization device intelligent management platform for petrochemical enterprises

Abstract

This invention discloses a 3D visualization-based intelligent equipment management platform for petrochemical enterprises, relating to the field of intelligent equipment management technology. The platform includes a visualization platform with the following modules: a 3D twin mapping module, which, based on digital twins and image classification algorithms, reconstructs the appearance features and operating status of petrochemical enterprise equipment in 3D visualization, constructing a dynamic 3D mapping model of equipment appearance and operating status. This invention utilizes multi-source perception fusion and dynamic 3D reconstruction technology to spatiotemporally align and fuse multimodal data, constructing a dynamic 3D mapping model of equipment appearance and operating status. This model presents the geometric shape of the equipment and intuitively maps physical attributes to a 3D surface, enabling accurate identification and location of hidden anomalies such as minor leaks and early faults. It overcomes the visual blind spots of traditional manual inspections, providing precise status perception capabilities for equipment management.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment management technology, specifically a three-dimensional visualization intelligent equipment management platform for petrochemical enterprises. Background Technology

[0002] With the rapid development of the petrochemical industry, the production process is becoming increasingly complex, and the requirements for the safety, stability and efficiency of equipment operation are constantly increasing. In order to achieve comprehensive management and monitoring of various equipment in petrochemical enterprises, traditional management methods can no longer meet the needs of modernization. By using 3D modeling and virtual reality technology, various data of equipment and production environment can be presented to managers in real time and accurately, enabling intuitive display of information such as production line, equipment operating status, and environmental parameters.

[0003] In existing technologies, large petrochemical enterprises have a wide variety of equipment with complex layouts and visual blind spots in the field environment. Manual inspections are difficult to identify minor leaks or early faults in real time. Even if an anomaly is initially located, there is a lack of dynamic projection capabilities for equipment aging trends, making it difficult to predict the path and time window of fault evolution. At the same time, during emergency response, the dynamic changes in the field environment, such as wind direction, temperature, and explosion risk, make it difficult to optimize the emergency response path in real time, resulting in delayed scheduling decisions and affecting the efficiency and safety of emergency response. To address these issues, a three-dimensional visualization intelligent equipment management platform for petrochemical enterprises is proposed. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention is implemented through the following technical solution: a three-dimensional visualization equipment intelligent management platform for petrochemical enterprises, comprising a visualization platform, wherein the visualization platform is communicatively connected to the following modules: The 3D twin mapping module, based on digital twin and image classification algorithms, reconstructs the appearance features and operating status of petrochemical enterprise equipment in 3D visualization, builds a dynamic 3D mapping model of equipment appearance and operating status, realizes accurate identification and location of minor leaks or early faults in equipment, achieves synchronous visualization mapping of equipment appearance and operating status, and improves the identification accuracy of minor faults. The aging trend prediction module combines reinforcement learning and image synthesis algorithms to simulate the aging process of equipment under multiple operating conditions, generate future state image sequences, dynamically predict and predict the degradation trend of equipment, realize the visualization prediction of the equipment aging process, and provide an intuitive basis for predictive maintenance. The anomaly diagnosis and classification module uses an image classification algorithm based on convolutional neural networks to compare the current equipment image with historical images in real time, identify existing abnormal features, locate and mark the abnormal equipment points, realize real-time intelligent identification and accurate location of equipment anomalies, and improve the efficiency of fault detection. The predictive maintenance decision module is used to integrate aging simulation results and anomaly diagnosis conclusions to generate equipment health scores and maintenance recommendations, automatically trigger early warnings, realize predictive maintenance decisions for equipment, improve fault response speed and maintenance efficiency, realize quantitative assessment of equipment health status and automatic generation of maintenance recommendations, and improve response speed. The emergency response optimization module combines the fruit fly algorithm with digital twin dynamic simulation, accesses real-time environmental data such as wind direction, temperature, and explosion risk, updates hazardous areas and safe passages in the 3D scene, and dynamically generates the optimal emergency response path based on early warning information and real-time environmental changes. This avoids risky areas, comprehensively improves the emergency dispatch efficiency and safety assurance capabilities of the petrochemical plant area, and achieves real-time optimization of emergency paths in dynamic environments.

[0005] Preferably, the three-dimensional twin mapping module includes a multi-source perception fusion unit and a dynamic three-dimensional reconstruction unit; The multi-source sensing fusion unit is used to collect multimodal data of various types of equipment in petrochemical enterprises in real time using laser scanning and photogrammetry technology, including appearance images, temperature distribution and vibration, to form the initial input of equipment operating status and to build an initial equipment twin model to achieve accurate fusion of multi-source data. The dynamic 3D reconstruction unit is used to introduce image classification algorithms to classify and identify the surface features of the equipment in real time. Combined with multimodal data and the initial twin model of the equipment, it performs dynamic mapping between visual features and operating status to form a dynamic 3D mapping model of the equipment appearance and operating status, thereby improving the real-time performance and accuracy of the model representation.

[0006] Preferably, the multi-source sensing fusion unit performs the following steps: By deploying a laser scanner array in the petrochemical plant area, high-precision point cloud data is collected on the appearance of the equipment according to a preset scanning cycle. Simultaneously, an infrared thermal imager is triggered to obtain the temperature distribution spectrum of the equipment surface. Furthermore, a distributed vibration sensor array is used to collect vibration spectrum data during equipment operation, forming a multi-modal dataset. This provides a comprehensive and synchronous multi-source data foundation for equipment condition monitoring. The collected point cloud data, temperature distribution map and vibration spectrum are spatiotemporally aligned. The point cloud registration algorithm based on the iterative nearest point is used to fuse the point cloud data collected in multiple batches into a unified world coordinate system. A preliminary three-dimensional geometric model of the equipment containing geometric shape and physical attributes is constructed to achieve high-precision spatial fusion and unified coordinate expression of multiple batches of point cloud data. Based on the three-dimensional geometric model of the equipment, the temperature distribution spectrum is texture-mapped, and the temperature data is superimposed on the corresponding area of ​​the equipment surface in pseudo-color form. At the same time, the vibration spectrum data is mapped into the dynamic displacement vector of the key parts of the equipment, forming an initial equipment twin model containing multi-dimensional information of geometry, temperature and vibration. A three-dimensional visualized digital twin basic model reflecting the multi-dimensional physical state of the equipment is then constructed.

[0007] Preferably, the dynamic three-dimensional reconstruction unit performs the following steps: Based on the initial twin model of the equipment, local image blocks on the surface of the equipment are extracted and input into a pre-trained lightweight convolutional neural network classifier to classify and identify the feature states of the equipment surface in real time, and generate a label map of the equipment surface state. The feature states include rust, cracks and media leakage, thereby realizing the automated identification and classification of equipment surface features and improving the efficiency and accuracy of inspection. The surface condition label map of the equipment obtained by classification and identification is fused with the temperature distribution map in the multimodal data. By constructing a state correlation matrix, the mapping relationship between the visual features of the equipment surface and temperature and vibration anomalies is established, potential fault feature points are identified, a multi-dimensional data correlation analysis mechanism is established, and early potential fault areas are accurately located. Based on the identified fault feature points, the initial equipment twin model is dynamically and locally refined and reconstructed. In the model, the abnormal area is refined with high-density mesh and texture enhancement. The appearance features of the equipment and the internal operating status are mapped synchronously, and a dynamic three-dimensional mapping model is output in real time to realize the fine three-dimensional reconstruction of the abnormal area and intuitively present the changes in the internal state of the equipment.

[0008] Preferably, the aging trend prediction module includes a multi-condition aging simulation unit and a future state synthesis unit; The multi-condition aging simulation unit is used to construct an aging path model of the equipment under different conditions based on historical operating data and equipment type using reinforcement learning algorithms, to simulate the equipment deterioration process, predict the future deterioration path, realize intelligent prediction of equipment aging path under multiple conditions, and improve the ability to predict deterioration trends. The future state synthesis unit is used to apply image synthesis technology to simulate the changes in the appearance of the equipment at future points in time, generate a sequence of equipment aging state images, identify potential faults in advance, assist in predictive maintenance decisions, realize the visualization synthesis of the future aging state of the equipment, and assist in the early identification of potential faults.

[0009] Preferably, the multi-condition aging simulation unit performs the following steps: The system extracts the operating parameter sequences of different equipment types under various working conditions and the corresponding equipment status detection records at different time points from the historical database, constructs a sample dataset of equipment aging process, and performs normalization and feature engineering on the sample data to provide a high-quality training data foundation for the reinforcement learning model, ensuring the accuracy and reliability of aging path learning. A reinforcement learning model based on deep Q-network is constructed, taking the current aging state of the equipment as the environmental state, maintenance operations and working condition adjustments as the action space, and the equipment performance degradation rate as the reward function. The optimal aging path strategy under different working conditions is obtained through offline training, realizing adaptive learning of the equipment aging process and intelligent decision-making of the optimal maintenance strategy. The real-time collected equipment operation data is input into the trained reinforcement learning model. Combined with the current operating conditions, the model generates degradation path prediction curves for the equipment at multiple future time points. The output includes equipment aging trend prediction results containing aging rate, key failure modes, and expected remaining lifespan, enabling accurate prediction of the future aging state of the equipment and providing a reliable decision-making basis for predictive maintenance.

[0010] Preferably, the future state synthesis unit performs the following steps: Based on the equipment aging trend prediction results, the surface state labels and key degradation parameters of the equipment at different prediction time points are extracted, and the input condition vector of the conditional generative adversarial network is constructed, which includes quantitative indicators of corrosion area ratio, crack density and propagation direction, so as to realize the quantitative expression of aging feature parameters and provide a precise conditional control basis for image synthesis. The current real surface image of the device and the conditional vector are input into the generator of the trained conditional generative adversarial network. Through multi-layer cascaded convolution and deconvolution operations, the surface state simulation image of the device at future time nodes is synthesized pixel by pixel, generating a high-fidelity simulation image of device aging, which intuitively presents the future degradation pattern. The synthesized simulated images from different time points are arranged in chronological order to construct an image sequence of equipment aging status. This sequence is then spatiotemporally mapped with a dynamic 3D mapping model to dynamically display the deterioration and evolution of the equipment's appearance over time in a 3D scene. This enables 3D visualization and simulation of the equipment aging process, providing intuitive decision support for predictive maintenance.

[0011] Preferably, the abnormal diagnosis and classification module performs the following steps: High-resolution surface images of key parts of the equipment are extracted in real time from the dynamic 3D mapping model and registered pixel-level with historical images of the same parts of the same equipment in the historical image database. This eliminates image shifts caused by differences in shooting angle and lighting conditions, ensures consistency of image comparison benchmarks, and eliminates interference from environmental factors on diagnostic results. The registered current image and historical images are input into a Siamese convolutional neural network based on an attention mechanism. The network extracts image features through a dual-branch structure, calculates the difference feature maps between feature maps, identifies abnormal feature regions such as tiny leaks and new cracks, enhances the ability to perceive tiny anomalies, and significantly improves the detection rate of early faults. The system spatially locates the identified abnormal feature areas, transforms the pixel coordinates of the abnormal areas into the world coordinate system of the dynamic 3D mapping model, marks the abnormal points in the 3D scene with highlighted boxes, and generates an abnormal diagnosis record containing the abnormal type, confidence level, and discovery time. This enables intuitive visualization of the abnormal location and information, facilitating quick location and verification by maintenance personnel.

[0012] Preferably, the predictive maintenance decision module performs the following steps: By weighting and fusing the output estimated remaining lifespan with the output anomaly type and severity, a multi-level evaluation index system for equipment health is constructed. The equipment health score of each device at the current moment is calculated to achieve a quantitative assessment of the equipment health status and provide a unified data foundation for maintenance decisions. Based on the comparison between the equipment health score and the preset multi-level warning thresholds, when the equipment health score is lower than the first-level warning threshold, preventive maintenance suggestions are automatically triggered, and a maintenance work order draft containing the maintenance type, suggested time window and required spare parts is generated, so as to realize the automatic identification of maintenance needs and the standardized generation of work orders, thereby improving response efficiency. The draft maintenance work order is pushed to the review interface of the operation and maintenance management system. At the same time, the warning information and maintenance suggestions are overlaid on the corresponding equipment model in the 3D scene in the form of visual labels. This allows operation and maintenance personnel to view maintenance details and confirm execution in the 3D environment, realizing intuitive presentation of warning information and closed-loop management of maintenance process, and improving the convenience of operation.

[0013] Preferably, the emergency response optimization module performs the following steps: Real-time access to meteorological monitoring stations and gas sensor networks deployed in the plant area allows for the acquisition of current wind direction and speed, ambient temperature, and the distribution of flammable and toxic gas concentrations. Combined with the location of hazardous sources in the dynamic three-dimensional mapping model, the risk rate is dynamically analyzed to achieve real-time perception and dynamic updates of hazardous areas, ensuring the timeliness and accuracy of risk warnings. A dynamic path planning model is constructed based on the fruit fly optimization algorithm. The real-time updated danger zone is used as an obstacle constraint, and the safety exit and emergency assembly point are used as target nodes. The optimal emergency response path that avoids all risk zones is searched and generated in the three-dimensional scene to ensure that the escape path always avoids the current risk zone and improves the safety of emergency evacuation. The generated optimal emergency response path is overlaid onto the 3D scene of the plant as a 3D dynamic navigation line. The path is updated in real time as environmental data changes, and the path guidance information is pushed to the handheld terminals of on-site personnel and the visualization platform. This enables dynamic optimization and scheduling of emergency response, synchronizes the path information between on-site personnel and the monitoring center, and ensures efficient collaboration in emergency command.

[0014] Compared with the prior art, the beneficial effects of the present invention are: (I) This intelligent management platform for three-dimensional visualization equipment for petrochemical enterprises uses multi-source perception fusion and dynamic three-dimensional reconstruction technology to spatiotemporally align and fuse multi-modal data such as laser scanning, infrared thermal imaging and vibration sensing to construct a dynamic three-dimensional mapping model of equipment appearance and operating status. It accurately presents the geometric shape of the equipment and intuitively maps physical properties such as temperature distribution and vibration displacement to the three-dimensional surface, enabling accurate identification and location of hidden anomalies such as minor leaks and early failures. It overcomes the visual blind spots of traditional manual inspection and provides accurate status perception capabilities for equipment management.

[0015] (II) This three-dimensional visualization equipment intelligent management platform for petrochemical enterprises integrates reinforcement learning and image synthesis algorithms. Based on historical operating data, it constructs an aging path model of equipment under multiple operating conditions. Through deep Q-network learning of the optimal aging strategy, it predicts the future deterioration trend and remaining life of the equipment. Furthermore, by combining conditional generative adversarial networks, it transforms abstract numerical predictions into intuitive future state image sequences. It dynamically displays the deterioration process of the equipment appearance over time in a three-dimensional scene, enabling maintenance personnel to predict the fault evolution path and time window in advance, providing a scientific basis for predictive maintenance.

[0016] (III) This intelligent management platform for three-dimensional visualization equipment for petrochemical enterprises adopts a twin convolutional neural network based on the attention mechanism to perform pixel-level registration and difference analysis on current equipment images and historical images. By calculating the difference feature maps between feature maps, it accurately identifies early abnormal features such as minor leaks and new cracks. The identified abnormal areas are accurately located to the world coordinate system of the three-dimensional model through coordinate transformation and are displayed intuitively in the three-dimensional scene in the form of highlighted marked boxes. It also generates diagnostic records containing abnormality type, confidence level and discovery time, which greatly improves the ability to discover hidden faults and ensures that abnormalities can be captured and dealt with in a timely manner in the bud stage.

[0017] (iv) This intelligent management platform for three-dimensional visualization equipment for petrochemical enterprises accesses meteorological monitoring and gas sensor network data in real time, combines the location of hazardous sources in the three-dimensional model, dynamically calculates the diffusion direction and range of hazardous gases, generates high-risk areas that are updated in real time, constructs a dynamic path planning model based on the fruit fly optimization algorithm, uses high-risk areas as dynamic obstacle constraints, quickly searches for the optimal emergency response path in the three-dimensional grid map, and completes path replanning within 0.5 seconds after environmental data is updated. The generated three-dimensional dynamic navigation line is pushed to the handheld terminals of on-site personnel and the large screen of the control center in real time, supports automatic detection and replanning of deviation from the path, realizes dynamic optimization and scheduling of the entire process from environmental perception to path navigation, and comprehensively improves emergency response efficiency and safety assurance capabilities. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the workflow of a three-dimensional visualization equipment intelligent management platform for petrochemical enterprises according to the present invention; Figure 2 This is a schematic diagram of the module structure of a three-dimensional visualization equipment intelligent management platform for petrochemical enterprises according to the present invention. Detailed Implementation

[0019] 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 some embodiments of the present invention, but not all embodiments.

[0020] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a three-dimensional visualization equipment intelligent management platform for petrochemical enterprises, including a visualization platform, which has the following communication connection modules: The 3D twin mapping module, based on digital twin and image classification algorithms, reconstructs the appearance features and operating status of petrochemical enterprise equipment in 3D visualization, builds a dynamic 3D mapping model of equipment appearance and operating status, realizes accurate identification and location of minor leaks or early faults in equipment, achieves synchronous visualization mapping of equipment appearance and operating status, and improves the identification accuracy of minor faults. The 3D twin mapping module includes a multi-source perception fusion unit and a dynamic 3D reconstruction unit. The multi-source sensing fusion unit utilizes laser scanning and photogrammetry technologies to collect multimodal data from various types of equipment in petrochemical enterprises in real time, including appearance images, temperature distribution, and vibration. This forms the initial input for equipment operating status and constructs an initial equipment twin model, achieving accurate fusion of multi-source data. Through a laser scanner array deployed in the petrochemical plant area, high-precision point cloud data is collected from the equipment appearance according to a preset scanning cycle. Simultaneously, an infrared thermal imager is triggered to acquire the temperature distribution spectrum of the equipment surface. A distributed vibration sensor array is used to collect vibration spectrum data during equipment operation, forming a multimodal dataset. This provides a comprehensive and synchronous multi-source data foundation for equipment condition monitoring. The collected point cloud data... Data, temperature distribution maps, and vibration spectra are spatiotemporally aligned. A point cloud registration algorithm based on iterative nearest point is used to fuse point cloud data collected in multiple batches into a unified world coordinate system. A preliminary three-dimensional geometric model of the equipment, including geometric shape and physical attributes, is constructed to achieve high-precision spatial fusion and unified coordinate expression of multiple batches of point cloud data. Based on the three-dimensional geometric model of the equipment, texture mapping is performed on the temperature distribution map, and the temperature data is superimposed on the corresponding area of ​​the equipment surface in pseudo-color form. At the same time, the vibration spectrum data is mapped into the dynamic displacement vector of key parts of the equipment, forming an initial equipment twin model containing multi-dimensional information of geometry, temperature, and vibration. A three-dimensional visualized digital twin basic model reflecting the multi-dimensional physical state of the equipment is constructed. It should be noted that a laser scanner array is deployed around key equipment in the petrochemical plant area. Specifically, a phase-type laser scanner, model FARO Focus S, is used with a 1 / 4 resolution (6.136mm per meter of scanning point spacing). The scan quality is set to 3× to improve the signal-to-noise ratio. The preset scanning cycle is a panoramic scan every 4 hours, simultaneously triggering an infrared thermal imager. A FLIR A700 fixed-mount thermal imager is selected, with a thermal sensitivity of less than 50mK and a temperature range of -20℃ to 150℃, acquiring temperature distribution maps of the equipment surface with a spatial resolution of 640×480 pixels. Simultaneously, a distributed vibration sensor array is deployed at the connection between the equipment bearing housing and flange, using PCB Piezotronics 352C33 accelerometers with a sensitivity of 100mV / g, continuously acquiring vibration spectrum data during equipment operation. All three data acquisition devices are connected to the same timing server via an industrial Ethernet switch, based on IEEE... The 1588 precision time protocol achieves microsecond-level synchronization, ensuring accurate alignment of multimodal data along the time axis and forming an initial multimodal dataset containing geometry, temperature, and vibration data. The acquired multimodal data undergoes unified spatiotemporal alignment and fusion processing. The raw point cloud data acquired by the laser scanner array is imported into a 3D data processing workstation, employing a point cloud registration algorithm based on the iterative nearest point algorithm. Specifically, the nearest point search radius is set to 0.05m, the maximum number of iterations is set to 50, and the convergence threshold is set to 1×10⁻⁶. -6The process involves gradually fusing point cloud data collected in multiple batches into a pre-defined world coordinate system. The origin of the coordinate system is set as the factory's baseline control point. Using the Gauss-Kruger projection coordinate system, a preliminary 3D geometric model of the equipment is constructed after registration. The model's accuracy is controlled within ±2mm. The surface mesh uses a triangular facet structure with a maximum side length not exceeding 5cm. Subsequently, the registered temperature distribution map is texture-mapped. Based on the correspondence between pixels and spatial points, the temperature data is superimposed onto the model surface in pseudo-color. The temperature display range is set from -20℃ to 150℃, using rainbow color mapping, with blue for low-temperature areas and red for high-temperature areas. Simultaneously, the vibration spectrum data is processed to extract the dominant frequency and amplitude within the equipment's operating frequency range, converting them into key vibration parameters for the equipment. The dynamic displacement vector of the part is set with an amplitude range of 0.01mm to 0.5mm, and the direction is determined according to the sensor installation orientation. This results in an initial device twin model containing geometric shape, temperature field distribution, and vibration displacement field. The model file is exported using the open-source Collada format, preserving complete material and animation information. After the initial device twin model is built, it enters a continuous operation and dynamic update phase. In each subsequent scan cycle, newly acquired multimodal data is processed according to the above process and fused with historical models for updates. During point cloud registration, the nearest point search radius of the iterative nearest point algorithm is adjusted to 0.02m to improve local registration accuracy, the maximum number of iterations is reduced to 30 to balance computational efficiency, and the convergence threshold is maintained at 1×10⁻⁶. -6 The m remains unchanged, the mapping method between the temperature distribution map and the vibration data remains unchanged, but the temperature display range is dynamically adjusted according to seasonal changes, and the amplitude range of the vibration displacement vector is automatically updated according to the real-time operating status of the equipment. When the vibration exceeds the standard, the upper limit of the amplitude range can be temporarily increased to 1.0 mm to fully record the abnormal data. The model update frequency is consistent with the scanning cycle, that is, a complete equipment twin model version is generated every 4 hours. Historical versions are archived and stored in the industrial time series database according to the timestamp for subsequent aging trend analysis and abnormal diagnosis. The dynamic 3D reconstruction unit incorporates image classification algorithms to perform real-time classification and recognition of equipment surface features. Combining multimodal data and an initial equipment twin model, it dynamically maps visual features to operational states, forming a dynamic 3D mapping model of equipment appearance and operational status. This improves the real-time performance and accuracy of the model. Based on the initial equipment twin model, it extracts local image blocks from the equipment surface and inputs them into a pre-trained lightweight convolutional neural network classifier. This classifies and recognizes the feature states of the equipment surface in real time, generating a surface status label map. Feature states include rust, cracks, and media leakage, achieving automated identification and classification of equipment surface features, improving inspection efficiency and accuracy. The surface status label map of the equipment obtained by class recognition is fused with the temperature distribution map in the multimodal data through feature layer fusion. By constructing a state correlation matrix, the mapping relationship between the visual features of the equipment surface and temperature and vibration anomalies is established, potential fault feature points are identified, a multi-dimensional data correlation analysis mechanism is established, and early potential fault areas are accurately located. Based on the identified fault feature points, the initial equipment twin model is dynamically and locally refined and reconstructed. In the model, the abnormal area is refined with high-density mesh and texture enhancement display, and the equipment appearance features and internal operating status are mapped synchronously. The real-time updated dynamic three-dimensional mapping model is output to realize the refined three-dimensional reconstruction of the abnormal area and intuitively present the changes in the internal state of the equipment. It should be noted that after the initial twin model of the equipment was constructed, surface local image blocks of key parts of the equipment were extracted from the model according to a 4-hour scanning cycle. During the extraction process, typical parts prone to degradation, such as flange connections, weld areas, and pipe bends, were located based on the model's world coordinate system. A standard image block of 256×256 pixels was cropped centered on each key part to ensure complete coverage of the equipment surface feature area. The extracted image blocks were then input into a pre-trained lightweight convolutional neural network classifier. This classifier adopted the MobileNetV3 architecture and performed transfer learning on a training set containing more than 50,000 on-site images of petrochemical equipment. The confidence threshold for image classification was set to 0.85. When the classification confidence score exceeds the threshold, the system automatically identifies three characteristic states: rust, cracks, and media leakage. It then generates a surface condition label map containing category labels and confidence scores at the corresponding location in the model, maintaining the same spatial resolution as the original image. The system performs feature layer fusion processing between the identified surface condition label map and the concurrently collected temperature distribution map. Based on a unified world coordinate system, the two types of data are spatially registered at the pixel level to ensure that each pixel in the label map corresponds to the same physical location on the equipment surface. An 11×11 pixel sliding window is constructed and iterates through the registered fused data. Within each window, the pixel percentage of the three labels (rust, cracks, and leakage) is calculated, and the calculation is performed simultaneously. The temperature gradient value and the difference between the temperature and the normal operating condition baseline value in the corresponding area are used to determine the presence of potential fault characteristics in the area when both the label pixel ratio and the temperature deviation value exceed 8℃ within the window. This is done through a preset state association matrix. The weight coefficients of the association matrix are trained based on more than 30,000 sets of historical fault case data. The identified potential fault feature points are recorded in three-dimensional coordinates with a coordinate accuracy of ±2mm. Based on the coordinates of the identified potential fault feature points, a dynamic local refinement and reconstruction process is triggered on the initial equipment twin model. Within a spherical area with a radius of 0.3m around the fault feature point, the original triangular mesh with a maximum side length of 5cm is refined to a maximum side length of 1cm, increasing the mesh density in the refined area. The density is increased to 5 times that of the original to ensure that the geometry of micro-cracks and local corrosion can be accurately represented. In terms of texture mapping, a semi-transparent fault feature highlighting layer is superimposed on the encrypted area. The corrosion area is marked with brown highlight and the transparency is set to 30%. The crack area is marked with red highlight and the transparency is set to 20%. The leakage area is marked with yellow highlight and flashing display with a flashing frequency of 1Hz. After completing the local refinement, the updated model is synchronously mapped with the temperature distribution spectrum and vibration displacement field data. The temperature display range is dynamically adjusted according to the current season, and the vibration displacement vector amplitude is automatically updated according to real-time monitoring data. The final output is a dynamic three-dimensional mapping model that is updated every 4 hours. Historical versions are archived and stored according to timestamps. The aging trend prediction module is used to combine reinforcement learning and image synthesis algorithms to simulate the aging process of equipment under multiple working conditions, generate future state image sequences, dynamically predict and predict the degradation trend of equipment, realize the visualization prediction of the equipment aging process, and provide intuitive basis for predictive maintenance. The aging trend prediction module includes a multi-working-condition aging simulation unit and a future state synthesis unit. The multi-condition aging simulation unit utilizes reinforcement learning algorithms to construct aging path models for equipment under different operating conditions based on historical operating data and equipment type. This simulates the equipment degradation process, predicts future degradation paths, and enables intelligent prediction of equipment aging paths under multiple operating conditions, improving the ability to predict degradation trends. It extracts operating parameter sequences for different equipment types under various operating conditions and corresponding time-point equipment status detection records from historical databases to construct a sample dataset of the equipment aging process. The sample data is then normalized and feature-engineered to provide a high-quality training data foundation for the reinforcement learning model, ensuring the accuracy and reliability of aging path learning. This is achieved by constructing a deep Q-network-based model. The network's reinforcement learning model takes the current aging state of the equipment as the environmental state, maintenance operations and operating condition adjustments as the action space, and the equipment performance degradation rate as the reward function. Through offline training, it obtains the optimal aging path strategy under different operating conditions, realizing adaptive learning of the equipment aging process and intelligent decision-making of the optimal maintenance strategy. Real-time collected equipment operation data is input into the trained reinforcement learning model, and combined with the current operating conditions, it generates degradation path prediction curves for multiple future time nodes of the equipment. The output includes the aging rate, key failure modes, and expected remaining lifespan of the equipment aging trend prediction results, realizing accurate prediction of the future aging state of the equipment and providing a reliable decision-making basis for predictive maintenance. It should be noted that the operating parameter sequences of different equipment types over the past 36 months were extracted from the industrial time-series database. These included continuous monitoring data such as pressure, temperature, flow rate, and vibration amplitude, with a sampling frequency of once per hour. Simultaneously, equipment condition monitoring records at corresponding time points were extracted, containing quantitative indicators such as wall thickness measurements, crack length, and corrosion depth. The operating parameters and condition monitoring records were aligned according to timestamps to construct a sample dataset of equipment aging processes. The dataset covers eight major equipment categories, including centrifugal pumps, compressors, heat exchangers, and reactors. Each category contains no fewer than 200 complete aging process samples. The sample data were normalized, mapping each operating parameter to the [0, 1] interval to eliminate the influence of dimensions. Simultaneously, feature engineering is performed to extract 12-dimensional derived features, forming a standardized model input data format. These 12-dimensional derived features include pressure fluctuation rate, temperature change rate, vibration trend slope, pressure peak factor, maximum temperature gradient, vibration spectrum centroid shift, cumulative runtime, start-stop count statistics, load fluctuation amplitude, medium flow rate change rate, wall thickness reduction rate, and crack propagation rate. A reinforcement learning model based on a deep Q-network is constructed. The model adopts a three-layer fully connected neural network structure, with 128 nodes in the input layer corresponding to the feature vector dimension of the current aging state of the equipment. The hidden layer nodes are 256 and 128 respectively, the activation function is ReLU, and the output layer has 64 nodes. The action space for operation and condition adjustment should be maintained. The equipment performance degradation rate is used as the reward function. The equipment performance degradation rate is calculated based on the wall thickness reduction rate and crack propagation rate. The reward value ranges from [-1, 0]. The faster the degradation, the lower the reward value. The optimal aging path strategy under different operating conditions is obtained through offline training. An experience replay mechanism is used during training. The replay buffer capacity is set to 10,000 experience data points, the batch size for each sampling is 64, the learning rate is set to 0.001, the discount factor γ is set to 0.95, and the number of training iterations is 5000 rounds until the loss function converges to below 0.01. The real-time collected equipment operation data is processed by the same normalization and feature engineering and then input into the system. The trained reinforcement learning model, combined with current operating conditions, generates a degradation path prediction curve for the equipment over the next 24 months at a monthly granularity. The prediction output includes aging rate indicators, specifically the monthly average wall thickness reduction rate of 0.02mm to 0.15mm and the monthly average crack propagation rate of 0.5mm to 3.0mm. Key failure modes are identified, including corrosion perforation, fatigue fracture, and sealing failure. The estimated remaining life is output in monthly units, with calculation accuracy controlled within ±2 months. The prediction results are stored in a real-time database in a structured data format for use by the predictive maintenance decision-making module. They are also displayed in a visualization interface as trend curves, allowing maintenance personnel to view the predicted equipment aging status at different time points. The future state synthesis unit is used to apply image synthesis technology to simulate the changes in the appearance of equipment at future points in time, generating a sequence of images showing the equipment's aging state. This allows for the early identification of potential faults, assists in predictive maintenance decisions, and enables the visualization and synthesis of the equipment's future aging state. Based on the predicted aging trend, it extracts surface state labels and key degradation parameters of the equipment at different predicted time points, constructs input condition vectors for a conditional generative adversarial network, and includes quantitative indicators such as the percentage of rust area, crack density, and propagation direction. This achieves a quantitative expression of aging characteristic parameters, providing precise conditional control for image synthesis and enabling the current state of the equipment to be visualized. The real surface image and conditional vector at each time point are input into the generator of the trained conditional generative adversarial network. Through multi-layer cascaded convolution and deconvolution operations, the surface state simulation image of the device at future time points is synthesized pixel by pixel, generating a high-fidelity device aging simulation image, which intuitively presents the future degradation pattern. The synthesized simulation images at different time points are arranged in chronological order to construct a sequence of device aging state images, and spatiotemporally correlated with a dynamic 3D mapping model. The device appearance degradation evolution process over time is dynamically displayed in a 3D scene, realizing the 3D visualization and deduction of the device aging process, and providing intuitive decision support for predictive maintenance. It should be noted that in practical applications, the surface condition labels and key degradation parameters of the equipment at four predicted time points (6th, 12th, 18th, and 24th months) are extracted from the structured data output by the multi-condition aging simulation unit. Specifically, the extracted quantitative indicators include the percentage of rusted area, crack density, and crack propagation direction. The percentage of rusted area is expressed as a percentage and calculated based on the pixel ratio of rusted areas per unit area on the equipment surface. Crack density is defined as the total length of cracks per square meter of surface area, in mm / m². The crack propagation direction is recorded in degrees, ranging from 0° to 180°, based on the predicted main direction angle of historical crack paths. These three indicators are concatenated in a preset order to construct a 12-dimensional conditional vector. Each time point corresponds to a set of 3-dimensional vectors, for a total of four sets of vectors. These vectors are input into the conditional generative adversarial network. To ensure image synthesis quality, all input vectors are normalized before being fed into the generator, mapping the values ​​of each indicator to the [-1, 1] interval. The spatial resolution between the conditional vectors and the equipment surface image is determined. The resolution is kept consistent at 256×256 pixels. The real surface image of the device at the current moment and the constructed conditional vector are input into the generator of the trained conditional generative adversarial network. The generator adopts the U-Net architecture, which contains 8 convolutional layers and 8 deconvolutional layers. It retains image detail information through skip connections. In the generation process, the conditional vector is first expanded into an embedding vector that matches the size of the image feature map through a fully connected layer. Then, it is spliced ​​and fused with the image features extracted by the convolutional layer in the channel dimension. The generator synthesizes a simulated image of the device's surface state at a future time node pixel by pixel. The image resolution is kept at 256×256 pixels and the color channels are RGB three channels. In the synthesis process, the rusted area is presented as a brown pixel cluster, the cracked area is presented as a line of dark pixels, and the leaked area is presented as a bright spot. The quality of the synthesized image is verified by a discriminator. The discriminator adopts the PatchGAN structure, which divides the image into 16×16 patches for realism and fakeism discrimination to ensure the authenticity and rationality of the synthesized image in local details.The simulated images obtained from the synthesis at four time points (6th, 12th, 18th, and 24th months) are arranged in chronological order to construct an image sequence of the equipment aging state. Subsequently, based on a unified world coordinate system, the image sequence is spatiotemporally correlated with a dynamic 3D mapping model. Specifically, the pixel coordinates of each simulated image are transformed to the surface texture coordinates of the dynamic 3D mapping model through a camera projection matrix, achieving precise alignment between the 2D image and the 3D model. In the 3D visualization scene, the degradation evolution of the equipment appearance over time is dynamically displayed in the form of a timeline slider. When the slider slides to the corresponding time point, the surface texture of the dynamic 3D mapping model is automatically replaced with the synthesized image of that time point, while the key degradation parameter values ​​of that point are overlaid. The 3D scene supports rotation, scaling, and translation operations, making it convenient for maintenance personnel to observe the equipment aging trend from any angle. The entire dynamic display process runs smoothly at 30 frames per second, and the switching latency between historical and predicted nodes is less than 200 milliseconds, providing an intuitive visualization basis for predictive maintenance decisions. The anomaly diagnosis and classification module uses an image classification algorithm based on convolutional neural networks to compare the current equipment image with historical images in real time, identify existing abnormal features, locate and mark the abnormal equipment points, realize real-time intelligent identification and accurate location of equipment anomalies, and improve the efficiency of fault detection. The predictive maintenance decision module is used to integrate aging simulation results and anomaly diagnosis conclusions to generate equipment health scores and maintenance recommendations, automatically trigger early warnings, realize predictive maintenance decisions for equipment, improve fault response speed and maintenance efficiency, realize quantitative assessment of equipment health status and automatic generation of maintenance recommendations, and improve response speed. The emergency response optimization module combines the fruit fly algorithm with digital twin dynamic simulation, accesses real-time environmental data such as wind direction, temperature, and explosion risk, updates hazardous areas and safe passages in the 3D scene, and dynamically generates the optimal emergency response path based on early warning information and real-time environmental changes. This avoids risky areas, comprehensively improves the emergency dispatch efficiency and safety assurance capabilities of the petrochemical plant area, and achieves real-time optimization of emergency paths in dynamic environments.

[0021] Example 2, as Figure 1 , Figure 2As shown, based on Embodiment 1, the present invention provides a technical solution: the anomaly diagnosis classification module performs the following steps: extracting high-resolution surface images of key parts of the equipment in real time from the dynamic three-dimensional mapping model, performing pixel-level registration with historical images of the same parts of the same equipment in the historical image database, eliminating image offset caused by differences in shooting angle and lighting conditions, ensuring the consistency of image comparison benchmark, eliminating the interference of environmental factors on the diagnostic results, inputting the registered current image and historical image into a twin convolutional neural network based on the attention mechanism, extracting image features through the network's dual-branch structure, calculating the difference feature maps between feature maps, identifying abnormal feature areas such as small leaks and new cracks, enhancing the perception of small anomalies, significantly improving the detection rate of early faults, spatially locating the identified abnormal feature areas, converting the pixel coordinates of the abnormal areas to the world coordinate system of the dynamic three-dimensional mapping model, marking the abnormal points in the three-dimensional scene with highlighted boxes, and generating an anomaly diagnosis record containing anomaly type, confidence level and discovery time, realizing intuitive visualization of anomaly location and information, facilitating quick location and verification by maintenance personnel; It should be noted that after the dynamic 3D mapping model is updated every 4 hours, an anomaly diagnosis process is automatically triggered. First, high-resolution surface images of key areas such as equipment flange connections and weld areas are extracted from the current model. The image resolution is 256×256 pixels. Simultaneously, historical images of the same equipment and the same area acquired in the previous scanning cycle are retrieved from the historical image database. A feature point detection algorithm based on scale-invariant feature transformation is used for pixel-level registration, with a feature point matching threshold set to 0.75. Homography matrix transformation is used to eliminate image shifts caused by differences in shooting angle and lighting conditions, ensuring precise pixel-level alignment between the two images, with registration accuracy controlled within 1 pixel. The registered current image and historical image are then input into a Siamese convolutional neural network based on an attention mechanism for anomaly identification. This network uses a dual-branch structure with shared weight parameters, and the backbone network is a ResNet-50 architecture. It is pre-trained on a dedicated dataset containing 80,000 petrochemical equipment images. The network enhances its sensitivity to minute changes through an attention module, extracting images from both images separately. After obtaining the 512-dimensional feature map, the difference feature map between the two is calculated. The difference threshold is set to 0.35. When the response value of a continuous 5×5 pixel area in the difference feature map exceeds this threshold, the system determines that there is an anomaly in the area. The smallest identifiable anomaly size is 0.5mm×0.5mm, which can accurately capture early fault features such as tiny leaks and new cracks. After the anomaly feature area is identified, spatial positioning and 3D annotation are immediately performed. The pixel coordinates of the anomaly area are transformed into the world coordinate system of the dynamic 3D mapping model through the camera projection matrix. The coordinate transformation accuracy reaches ±2mm. In the 3D scene, the anomaly point is marked with a red semi-transparent highlight mark box. The size of the mark box is automatically adjusted according to the actual size of the anomaly area. The minimum display size is 5cm×5cm. At the same time, an anomaly diagnosis record containing anomaly type, confidence score and discovery time is generated. The confidence threshold is set to 0.85. All records are stored in the industrial time series database by timestamp and pushed to the operation and maintenance personnel interface through the visualization platform. Detailed diagnostic information can be viewed by clicking in the 3D scene. The predictive maintenance decision-making module performs the following steps: It weights and fuses the output estimated remaining lifespan with the output anomaly type and severity to construct a multi-level evaluation index system for equipment health. It calculates the equipment health score for each device at the current moment, achieving a quantitative assessment of equipment health status and providing a unified data foundation for maintenance decisions. Based on the equipment health score, it compares it with preset multi-level warning thresholds. When the equipment health score is lower than the first-level warning threshold, it automatically triggers preventative maintenance recommendations, generating a draft maintenance work order containing maintenance type, recommended time window, and required spare parts. This achieves automatic identification of maintenance needs and standardized generation of work orders, improving response efficiency. The draft maintenance work order is pushed to the operation and maintenance management system's review interface. Simultaneously, the warning information and maintenance recommendations are overlaid as visual labels on the corresponding equipment model in the 3D scene, allowing operation and maintenance personnel to view maintenance details and confirm execution in the 3D environment. This achieves intuitive presentation of warning information and closed-loop management of the maintenance process, improving operational convenience. It should be noted that the predictive maintenance decision module receives the estimated remaining life data output from the aging trend projection module and the anomaly type and severity information output from the anomaly diagnosis module in real time. The estimated remaining life is output on a monthly basis, with the calculation accuracy controlled within ±2 months. The anomaly types include three categories: corrosion perforation, fatigue fracture, and sealing failure. The severity is divided into three levels: mild, moderate, and severe, based on the area of ​​the anomaly region and the characteristic response value. Then, the normalized value of the remaining life, the weight score of the anomaly type, and the weight score of the severity are calculated. A weighted fusion algorithm is used to construct a multi-level evaluation index system for equipment health. The weight coefficients are trained based on more than 30,000 sets of historical failure case data, with the remaining life weight accounting for 40%, the anomaly type weight accounting for 30%, and the severity weight accounting for 30%. After fusion calculation, an equipment health score for each device at the current moment is generated. The score ranges from 0 to 100, with the lower the score, the worse the equipment health status. The formula for calculating the equipment health score is as follows: ; ; ; ; In the formula: Rate the health of the equipment; The remaining lifetime weighting coefficient has a value of 0.40. This is the normalized value of the remaining lifetime, ranging from 0 to 100. To estimate the remaining lifespan; This is the baseline value for the maximum lifespan of the equipment, preset according to the equipment type and industry standards, used to map the remaining lifespan to a percentage-based scoring range; This is the weighting coefficient for the anomaly type, with a value of 0.30. The weight score for the anomaly type ranges from 0 to 100; These are abnormal types, including three categories: corrosion perforation, fatigue fracture, and sealing failure. The abnormal type mapping function determines the benchmark score corresponding to different types based on the statistics of historical failure cases. The lower the score, the greater the impact of the type on the health of the equipment. The abnormal type of corrosion perforation has a weight of 60 points, the abnormal type of fatigue fracture has a weight of 50 points, and the abnormal type of sealing failure has a weight of 70 points. This is the severity weighting coefficient, with a value of 0.30; The severity score is a weighted score, with a value ranging from 0 to 100; The severity level of the anomaly is categorized into three levels: mild, moderate, and severe, determined by a combination of the area of ​​the anomaly and the characteristic response value. The severity mapping function is based on the statistical analysis of historical failure cases to determine the baseline score corresponding to different levels. The lower the score, the higher the severity. The severity weight is 80 points for mild, 60 points for moderate, and 40 points for severe. Based on a comparison of the equipment health score with preset multi-level early warning thresholds, the early warning thresholds are set to three levels: Level 1 warning threshold is 75 points, Level 2 warning threshold is 60 points, and Level 3 warning threshold is 45 points. When the equipment health score is lower than the Level 1 warning threshold of 75 points, the system automatically triggers a preventive maintenance recommendation process. The generated maintenance work order includes three core components: maintenance type, recommended time window, and required spare parts. The maintenance type is determined based on the anomaly diagnosis results as corrosion treatment, crack repair, or seal replacement; the recommended time window is set to within the next 1 to 3 months based on the remaining life prediction results; the required spare parts are matched with the corresponding model and specifications from the equipment ledger database. The maintenance work order draft is stored in a structured data format, including fields such as equipment number, anomaly location coordinates, and maintenance priority, ensuring seamless data integration with the operation and maintenance management system. The draft maintenance work order is pushed to the review interface of the operation and maintenance management system. At the same time, the warning information and maintenance suggestions are superimposed on the corresponding equipment model in the 3D scene in the form of visual labels. The visual labels use a semi-transparent background design. The first-level warning is displayed as a yellow label, the second-level warning as an orange label, and the third-level warning as a red label. The label size is 10cm×10cm and it floats 0.3m above the equipment model. The operation and maintenance personnel can click on the label in the 3D environment to view the maintenance details, including the equipment health score, anomaly type, suggested maintenance time window, required spare parts list, etc. After confirmation of execution, the system automatically assigns the maintenance task to the corresponding maintenance team and marks the equipment as under maintenance in the model. After the maintenance is completed, the operation and maintenance personnel upload the maintenance record and close the work order, realizing the closed-loop management of predictive maintenance. The emergency response optimization module performs the following steps: It accesses real-time data from meteorological monitoring stations and gas sensor networks deployed in the plant area to obtain current wind direction and speed, ambient temperature, and the concentration distribution of combustible and toxic gases. Combined with the location of hazardous sources in the dynamic 3D mapping model, it dynamically analyzes the hazard rate to achieve real-time perception and dynamic updates of hazardous areas, ensuring the timeliness and accuracy of risk warnings. Based on the fruit fly optimization algorithm, it constructs a dynamic path planning model, using the real-time updated hazardous areas as obstacle constraints and safety exits and emergency assembly points as target nodes. It searches and generates the optimal emergency response path in the 3D scene, avoiding all risk areas, ensuring that escape routes always avoid the current risk area and improving the safety of emergency evacuation. The generated optimal emergency response path is overlaid onto the plant's 3D scene as a 3D dynamic navigation line. The path is updated in real-time as environmental data changes, and path guidance information is simultaneously pushed to on-site personnel's handheld terminals and visualization platforms, achieving dynamic optimization and scheduling of emergency response. This ensures synchronization of path information between on-site personnel and the monitoring center, guaranteeing efficient collaboration in emergency command. It should be noted that meteorological monitoring stations and a gas sensor network are deployed at key locations within the factory area to collect real-time data on current wind direction, wind speed, ambient temperature, and the concentration distribution of combustible and toxic gases. The meteorological monitoring stations utilize ultrasonic anemometers with a wind speed measurement range of 0-60 m / s and an accuracy of ±0.1 m / s, and a wind direction measurement accuracy of ±2°. The gas sensor network employs electrochemical and catalytic combustion sensors, with a combustible gas detection range of 0-100% LEL and a toxic gas detection accuracy of ±1% FS. Data is acquired once per second. All monitoring data is transmitted in real-time via industrial Ethernet to the emergency response optimization module, where it is spatially fused with the hazard location in the dynamic 3D mapping model. The hazard location is precisely determined based on the coordinates of the equipment's 3D model, with a coordinate accuracy of ±2 mm. The diffusion direction and range of hazardous gases are calculated based on real-time wind direction and speed, and combined with the gas concentration distribution to generate dynamic hazard zones. The hazard rate calculation formula is as follows: , For risk rate, For gas concentration, A threshold concentration is used; when the hazard rate exceeds 0.8, it is automatically marked as a high-risk area, displayed in real-time as a red semi-transparent block in the 3D scene. Based on the real-time updated hazard area data, a dynamic path planning model is constructed using the fruit fly optimization algorithm. The fruit fly optimization algorithm parameters are set as follows: population size 50, maximum number of iterations 100, initial search step size 0.5m, and dynamic adjustment factor 0.95. The algorithm uses all high-risk areas in the current 3D scene as dynamic obstacle constraints, with obstacle boundaries extended outward by a 1.5m safety distance to ensure personnel detour. The factory's preset safety exits and emergency assembly points are used as target nodes. The target node coordinates are precisely located based on the world coordinate system. The algorithm searches and generates the optimal emergency response path from any location to the target node in the 3D mesh map. The path search range covers the entire factory area, and the mesh resolution is set to 0.5m × 0.5m × 0.5m to ensure that the path planning accuracy meets the actual escape requirements. After each environmental data update, path replanning is completed within 0.5 seconds, generating a path that avoids all obstacles. The optimal path for dynamic risk areas is calculated with a total path length accuracy controlled within ±0.5m. The generated optimal emergency response path is overlaid onto the 3D scene of the plant area as a 3D dynamic navigation line. The navigation line is highlighted in green with a line width of 5cm. Directional arrows are placed every 2m along the path to indicate the escape direction. The path is updated in real time as environmental data changes, with the update frequency consistent with the sensor data acquisition frequency to ensure the real-time and accuracy of the path information. Simultaneously, the path guidance information is pushed to on-site personnel's handheld terminals and a visualization platform via a wireless network. The handheld terminals are explosion-proof tablets with an 8-inch screen, supporting 3D map display and path guidance. The path update delay is less than 1 second. The current wind vane, danger zone boundaries, and safety exit locations are simultaneously overlaid in the 3D scene. This allows maintenance personnel to monitor personnel positions and path execution status in real time on the control center's large screen. When on-site personnel deviate from the planned path, an automatic voice prompt is issued and the path is replanned, achieving dynamic optimization and scheduling of the entire emergency response process.

[0022] The following is a detailed description of the workflow of this 3D visualization equipment intelligent management platform for petrochemical enterprises.

[0023] The platform operates in a 4-hour cycle. First, it uses laser scanner arrays, infrared thermal imagers, and distributed vibration sensor arrays deployed around key equipment to achieve microsecond-level synchronization based on the IEEE 1588 precision time protocol. This allows for the acquisition of high-precision point cloud data, temperature distribution maps, and vibration spectrum data, forming a multimodal dataset. The acquired multimodal data is then imported into a 3D data processing workstation. Using an iterative nearest-point point cloud registration algorithm, the point cloud data is fused into a unified world coordinate system, constructing a preliminary 3D geometric model of the equipment. Subsequently, the temperature distribution map is texture-mapped onto the model surface in pseudo-color, and the vibration spectrum data is converted into dynamic displacement vectors for key parts of the equipment, forming an initial equipment twin model that includes geometric shape, temperature field distribution, and vibration displacement field. Based on this initial equipment twin model, 256×256 pixel surface image blocks are extracted from key areas such as flange connections and weld areas, and input into a lightweight convolutional neural network using the MobileNetV3 architecture. In the network classifier, three characteristic states—rust, cracks, and media leakage—are identified with a confidence threshold of 0.85, generating a label map of the equipment surface condition. The label map is then fused with the temperature distribution map collected at the same time using a feature layer. The percentage of label pixels and the temperature deviation are statistically analyzed using an 11×11 pixel sliding window. When the percentage of label pixels in the window exceeds 15% and the temperature deviation exceeds 8℃, potential fault feature points are determined by a state association matrix trained based on more than 30,000 historical fault cases, and the three-dimensional coordinates are recorded with an accuracy of ±2mm. Based on the coordinates of the fault feature points, the triangular mesh is densified from a maximum side length of 5cm to 1cm within a radius of 0.3m, increasing the mesh density by 5 times. A semi-transparent high-brightness display layer is then superimposed, and a dynamic three-dimensional mapping model updated every 4 hours is finally output. Historical versions are archived and stored according to timestamps. While the dynamic 3D mapping model is being updated, the platform initiates a multi-condition aging simulation unit. It extracts aging process sample data from the industrial time-series database, covering 8 major equipment categories and containing at least 200 sets of data for each category, spanning the past 36 months. After normalization and feature engineering, this data is input into a deep Q-network reinforcement learning model. The model has 128 input layer nodes, 256 hidden layer nodes, and 128 output layer nodes. Offline training is performed using the equipment performance degradation rate as the reward function to generate a degradation path prediction curve for the equipment over the next 24 months, outputting the monthly wall thickness. The thinning rate, monthly average crack propagation rate, key failure modes, and expected remaining life with ±2-month accuracy are calculated. The future state synthesis unit extracts the corrosion area ratio, crack density, and crack propagation direction for the 6th, 12th, 18th, and 24th months to construct a 12-dimensional conditional vector, which is input into a conditional generative adversarial network using a U-Net architecture to generate a 256×256 pixel surface state simulation image for the corresponding time node. The image is then spatiotemporally mapped to a dynamic 3D mapping model using a camera projection matrix, and the aging evolution process of the equipment is dynamically displayed in the 3D scene in the form of a timeline slider. The anomaly diagnosis module is automatically triggered after the current model update. It extracts key area images from the model and performs pixel-level registration with historical images from the previous cycle based on scale-invariant feature transformation. The registered image pairs are then input into a Siamese convolutional neural network based on an attention mechanism. The network uses a ResNet-50 architecture, extracts 512-dimensional feature maps, and calculates difference feature maps. Anomaly regions are identified with a threshold of 0.35. It can identify micro-leakages and new cracks as small as 0.5mm × 0.5mm and marks the anomaly regions in the 3D scene with red semi-transparent marker boxes. The size of the marker boxes is automatically adjusted, with a minimum display size of 5cm × 5cm. The predictive maintenance decision module receives remaining life data and anomaly type and severity information in real time. Based on the weight coefficients trained from more than 30,000 historical failure cases, with remaining life weighting at 40%, anomaly type weighting at 30%, and severity weighting at 30%, it calculates and generates an equipment health score from 0 to 100. When the score is below 75, 60, and 45, it triggers yellow, orange, and red warnings respectively. It automatically generates a draft maintenance work order containing maintenance type, suggested time window, and required spare parts, which is displayed as a 10cm×10cm semi-transparent label floating 0.3m above the equipment model. Users can click to view details and confirm execution. The emergency response optimization module accesses real-time data from meteorological monitoring stations and gas sensor networks. Based on real-time wind direction, wind speed, and gas concentration, it calculates the hazard rate. When the hazard rate exceeds 0.8, high-risk areas are marked with red semi-transparent blocks. A dynamic path planning model is constructed using the fruit fly optimization algorithm, with a population size of 50, a maximum number of iterations of 100, an initial search step size of 0.5m, and a dynamic adjustment factor of 0.95. The model completes path replanning within 0.5 seconds with a grid resolution of 0.5m×0.5m×0.5m, generating the optimal emergency response path that avoids high-risk areas. The path is overlaid on the 3D scene as a green highlighted navigation line with a line width of 5cm and directional arrows every 2m. The path is simultaneously pushed to an 8-inch explosion-proof handheld terminal and the control center's large screen with an update delay of less than 1 second. When personnel deviate from the path, an automatic voice prompt is issued and the path is replanned, achieving dynamic optimization and scheduling of the entire emergency response process.

[0024] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A three-dimensional visualization-based intelligent equipment management platform for petrochemical enterprises, comprising a visualization platform, characterized in that, The visualization platform communication connection includes the following modules: The 3D twin mapping module, based on digital twin and image classification algorithms, reconstructs the appearance features and operating status of petrochemical enterprise equipment in 3D visualization, and builds a dynamic 3D mapping model of equipment appearance and operating status. The aging trend prediction module combines reinforcement learning and image synthesis algorithms to simulate the aging process of equipment under multiple operating conditions, generate future state image sequences, and dynamically predict and maintain the equipment degradation trend. The anomaly diagnosis and classification module uses an image classification algorithm based on convolutional neural networks to compare the current device image with historical images in real time, identify existing abnormal features, and locate and mark the abnormal device locations. The predictive maintenance decision module is used to integrate aging simulation results and anomaly diagnosis conclusions to generate equipment health scores and maintenance recommendations, automatically trigger early warnings, and realize predictive maintenance decisions for equipment. The emergency response optimization module combines the fruit fly algorithm with digital twin dynamic simulation, accesses real-time environmental data, updates dangerous areas and safe passages in the 3D scene, and dynamically generates the optimal emergency response path by combining early warning information and real-time environmental changes to avoid risky areas.

2. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 1, characterized in that: The three-dimensional twin mapping module includes a multi-source perception fusion unit and a dynamic three-dimensional reconstruction unit; The multi-source sensing fusion unit is used to collect multimodal data of various types of equipment in petrochemical enterprises in real time using laser scanning and photogrammetry technology, including appearance images, temperature distribution and vibration, to form the initial input of equipment operating status and to build an initial equipment twin model. The dynamic 3D reconstruction unit is used to introduce an image classification algorithm to classify and identify the surface features of the equipment in real time. It combines multimodal data and the initial twin model of the equipment to perform dynamic mapping between visual features and operating status, forming a dynamic 3D mapping model of the equipment appearance and operating status.

3. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 2, characterized in that: The multi-source sensing fusion unit performs the following steps: By deploying a laser scanner array in the petrochemical plant area, high-precision point cloud data is collected on the appearance of the equipment according to a preset scanning cycle. Simultaneously, an infrared thermal imager is triggered to obtain the temperature distribution spectrum of the equipment surface. A distributed vibration sensor array is used to collect vibration spectrum data during equipment operation, forming a multimodal dataset. The collected point cloud data, temperature distribution map and vibration spectrum are spatiotemporally aligned. The point cloud registration algorithm based on iterative nearest point is used to fuse the point cloud data collected in multiple batches into a unified world coordinate system, and a three-dimensional geometric model of the device containing geometric shape and physical attributes is initially constructed. Based on the three-dimensional geometric model of the equipment, the temperature distribution spectrum is texture-mapped, and the temperature data is superimposed on the corresponding area of ​​the equipment surface in pseudo-color form. At the same time, the vibration spectrum data is mapped into the dynamic displacement vector of the key parts of the equipment, forming an initial twin model of the equipment containing multi-dimensional information of geometry, temperature and vibration.

4. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 2, characterized in that: The dynamic 3D reconstruction unit performs the following steps: Based on the initial twin model of the equipment, local image blocks on the surface of the equipment are extracted and input into a pre-trained lightweight convolutional neural network classifier to classify and identify the feature states of the equipment surface in real time, and generate a label map of the equipment surface state. The feature states include corrosion, cracks and media leakage. The surface condition label map of the equipment obtained by classification and identification is fused with the temperature distribution map in the multimodal data. By constructing a state correlation matrix, the mapping relationship between the visual features of the equipment surface and temperature anomalies and vibration anomalies is established to identify potential fault feature points. Based on the identified fault feature points, the initial equipment twin model is dynamically and locally refined and reconstructed. In the model, abnormal areas are refined with high-density meshes and texture enhancement for display. The appearance features of the equipment and its internal operating status are mapped synchronously, and a dynamic three-dimensional mapping model is output that is updated in real time.

5. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 2, characterized in that: The aging trend prediction module includes a multi-condition aging simulation unit and a future state synthesis unit. The multi-condition aging simulation unit is used to construct an aging path model of the equipment under different conditions based on historical operating data and equipment type using reinforcement learning algorithms, to simulate the equipment deterioration process and predict future deterioration paths. The future state synthesis unit is used to apply image synthesis technology to simulate the changes in the appearance of the equipment at future points in time, generate a sequence of equipment aging state images, identify potential faults in advance, and assist in predictive maintenance decisions.

6. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 5, characterized in that: The multi-condition aging simulation unit performs the following steps: Extract the operating parameter sequences of different equipment types under various working conditions and the corresponding equipment status detection records at time points from the historical database, construct a sample dataset of equipment aging process, and perform normalization and feature engineering processing on the sample data. A reinforcement learning model based on a deep Q-network is constructed, which takes the current aging state of the equipment as the environmental state, maintenance operations and working condition adjustments as the action space, and the equipment performance degradation rate as the reward function. The optimal aging path strategy under different working conditions is obtained through offline training. The real-time collected equipment operation data is input into the trained reinforcement learning model. Combined with the current operating conditions, the model generates a degradation path prediction curve for the equipment at multiple future time points and outputs a prediction result of the equipment aging trend, including aging rate, key failure modes, and expected remaining life.

7. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 5, characterized in that: The future state synthesis unit performs the following steps: Based on the equipment aging trend prediction results, the surface state labels and key deterioration parameters of the equipment at different prediction time points are extracted, and the input condition vector of the conditional generative adversarial network is constructed, which includes quantitative indicators such as the proportion of rust area, crack density and propagation direction. The current real surface image of the device and the conditional vector are input into the generator of the trained conditional generative adversarial network. Through multi-layer cascaded convolution and deconvolution operations, a simulated image of the device's surface state at future time nodes is synthesized pixel by pixel. The synthesized simulated images at different time points are arranged in chronological order to construct an image sequence of equipment aging status. This sequence is then spatiotemporally mapped with a dynamic 3D mapping model to dynamically display the deterioration and evolution of the equipment's appearance over time in a 3D scene.

8. The intelligent management platform for three-dimensional visualization equipment in petrochemical enterprises according to claim 5, characterized in that: The abnormality diagnosis and classification module performs the following steps: High-resolution surface images of key parts of the equipment are extracted in real time from the dynamic 3D mapping model and then registered pixel-level with historical images of the same parts of the same equipment in the historical image database. The registered current image and historical images are input into a Siamese convolutional neural network based on an attention mechanism. The network's dual-branch structure extracts image features, calculates the difference feature maps between feature maps, and identifies abnormal feature regions. The system spatially locates the identified abnormal feature regions, transforms the pixel coordinates of the abnormal regions into the world coordinate system of the dynamic 3D mapping model, marks the abnormal points in the 3D scene with highlighted bounding boxes, and generates an abnormal diagnosis record containing the abnormal type, confidence level, and discovery time.

9. A three-dimensional visualization equipment intelligent management platform for petrochemical enterprises according to claim 8, characterized in that: The predictive maintenance decision module performs the following steps: The projected remaining lifespan is weighted and fused with the anomaly type and severity to construct a multi-level evaluation index system for equipment health, and the equipment health score of each device at the current moment is calculated. Based on the comparison between the equipment health score and the preset multi-level warning thresholds, when the equipment health score is lower than the first-level warning threshold, preventive maintenance suggestions are automatically triggered, and a draft maintenance work order containing the maintenance type, suggested time window and required spare parts is generated. The draft maintenance work order is pushed to the review interface of the operation and maintenance management system, and the warning information and maintenance suggestions are simultaneously overlaid on the corresponding equipment model in the 3D scene in the form of visual labels.

10. A three-dimensional visualization equipment intelligent management platform for petrochemical enterprises according to claim 9, characterized in that: The emergency response optimization module performs the following steps: Real-time access to meteorological monitoring stations and gas sensor networks deployed in the plant area is used to obtain current wind direction and speed, ambient temperature and distribution of flammable and toxic gas concentrations. Combined with the location of hazardous sources in the dynamic three-dimensional mapping model, the hazard rate is dynamically analyzed. A dynamic path planning model is constructed based on the fruit fly optimization algorithm. The real-time updated danger zones are used as obstacle constraints, and safety exits and emergency assembly points are used as target nodes. The optimal emergency response path that avoids all risk zones is searched and generated in the three-dimensional scene. The generated optimal emergency response path is overlaid onto the 3D scene of the plant as a 3D dynamic navigation line. The path is updated in real time as environmental data changes, and the path guidance information is pushed to the handheld terminals and visualization platform of on-site personnel.

Images