Three-dimensional collaborative management and control method and system for gas turbine power plant

Abstract

The application is suitable for the technical field of power system monitoring, and provides a gas turbine power plant three-dimensional collaborative management and control method and system, wherein the gas turbine power plant three-dimensional collaborative management and control method comprises the following steps: step S1: constructing a three-dimensional digital model fused with historical data archives; step S2: fault prediction based on a generative adversarial network and a mechanism model; and step S3: collaborative maintenance and closed-loop treatment based on augmented reality. The application associates and binds multi-source data with corresponding components in a three-dimensional geometric model by constructing a three-dimensional digital model fused with historical data archives. When maintenance is performed, field personnel can identify a target pipeline through an augmented reality device, and can superimpose and display historical data archives and real-time state information of the pipeline in a display interface.

Description

Technical Field

[0001] This invention belongs to the field of power system monitoring technology, and in particular relates to a three-dimensional collaborative management and control method for gas turbine power plants. Background Technology

[0002] As a key power energy facility, the pipeline system of a gas turbine power plant undertakes the core task of transporting high-temperature and high-pressure media. The safety and reliability of pipeline operation are directly related to the overall production efficiency and service life of the power plant.

[0003] Currently, pipeline management primarily relies on distributed control systems to collect real-time operating parameters, trigger over-limit alarms through preset thresholds, and perform maintenance in conjunction with fixed-cycle periodic inspections. However, this model has certain limitations. For example, there is a lack of effective correlation between operating data and the spatial location of the pipeline itself, making it difficult for maintenance personnel to intuitively locate abnormal areas in three-dimensional space. Although some research has explored the application of 3D visualization technology in power plant management, existing solutions mostly remain at the level of static model display, failing to integrate real-time IoT monitoring data, historical maintenance records, and 3D geometric models, and especially lacking the ability to predict fault evolution based on artificial intelligence.

[0004] Therefore, the technical problem that this invention aims to solve is how to construct a three-dimensional control architecture that can integrate multi-source heterogeneous data, realize dynamic prediction of pipeline health status, and support seamless collaboration between on-site and remote operations, so as to dynamically and in real-time locate abnormal areas. Summary of the Invention

[0005] The purpose of this invention is to provide a three-dimensional collaborative management and control method for gas turbine power plants, aiming to solve the above-mentioned problems.

[0006] The present invention provides a three-dimensional collaborative management and control method for gas turbine power plants, the method comprising: Collect operating parameters, non-destructive testing data and fault records of gas turbine power plant pipelines; associate and bind the collected data with the corresponding pipeline components in the three-dimensional geometric model of the gas turbine power plant, and establish a historical data archive for each pipeline segment, including original installation stress data, weld location data, material creep record data and historical overheating zone coordinate data, so that the historical state change process of the pipeline can be traced back through the time axis; A fault prediction model based on a generative adversarial network (GAN) is constructed and trained. The GAN includes a generator and a discriminator. Real-time operating parameters are input into the trained fault prediction model and combined with a thermodynamic mechanism model for joint simulation to predict the health distribution of the pipeline network in the future time period and mark and display potential fault points. In response to the predicted potential fault points, a work order containing maintenance information is generated; when on-site personnel identify the target pipeline using augmented reality equipment, the historical data archive and real-time status information of the pipeline are overlaid on the display interface; after the maintenance is completed, the measured data is uploaded to the three-dimensional digital model, and the historical data archive is updated.

[0007] As a further aspect of the present invention, the process of establishing a historical data archive for each pipeline segment includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

[0008] As a further aspect of this invention, the training objective in constructing and training a fault prediction model based on generative adversarial networks includes three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

[0009] As a further aspect of the present invention, the process of joint simulation using a thermodynamic mechanism model includes: Real-time operating parameters are input into a trained generator, which simulates the degradation path of equipment under specific operating conditions to generate potential fault feature data. The generated fault feature data is then compared and fused with the calculation results of the thermodynamic mechanism model to predict the health distribution of the pipeline network in the future time period.

[0010] As a further aspect of the present invention, the process of marking and displaying potential fault points includes: Responding to user-inputted extreme working condition assumption analysis commands, an accident chain inference model based on multi-physics coupling simulates the sequence of potential risk points and accident propagation paths under extreme working conditions under the action of multi-physics coupling. The calculation basis of the accident chain inference model is that the failure probability distribution at future time moments is jointly determined by the stress field distribution state at the current time moment, as well as the rate of change of temperature, pressure, and flow velocity, and is calculated by a multi-physics coupling operator based on the fusion of finite element analysis and machine learning. The multi-physics coupling operator can be constructed in the following way: using finite element analysis as the physics solver to generate training samples, learning the coupling relationship between physics fields through a machine learning model, and using the trained surrogate model to replace finite element analysis for rapid prediction.

[0011] As a further aspect of the present invention, the process of marking and displaying potential fault points also includes: Data from acoustic sensors deployed on the pipeline is collected. When a minor leak occurs, the precise three-dimensional coordinates of the leak point are calculated and marked using the time difference of sound waves arriving at different sensors and the spatial topology of a three-dimensional geometric model. The calculation of the leak point coordinates is based on the following: the arrival time difference between any two sensors is calculated based on the time it takes for each sensor to receive the same leaking acoustic event, and then multiplied by the speed of sound propagation in the medium to obtain the distance difference between the two sensors and the leak point. Based on the distance difference data of at least three sets of sensor pairs, combined with the known spatial coordinates of each sensor, the spatial point that minimizes the sum of squared positioning errors is determined by solving for the intersection of hyperboloids or by using a nonlinear least squares optimization method. This point is then used as the precise three-dimensional coordinates of the leak point.

[0012] As a further aspect of the present invention, the method further includes: a remote expert generating a virtual identifier with pointing and annotation meanings through a control terminal, and transmitting the virtual identifier to the augmented reality device of the on-site personnel through WebSocket or a real-time communication protocol based on TCP / IP; The virtual identifier is precisely anchored to a specific component in the augmented reality device's field of view by on-site personnel in the following way: the on-site environment is reconstructed in real time based on a simultaneous localization and mapping algorithm, and the spatial pose of the target component is determined by combining preset marker points or natural feature point recognition technology. The three-dimensional coordinates of the virtual identifier are then mapped to the local coordinate system of the augmented reality device, thereby achieving precise alignment and real-time tracking between the virtual identifier and the physical component.

[0013] The present invention also provides a three-dimensional collaborative management and control system for gas turbine power plants. This system is used to implement the aforementioned three-dimensional collaborative management and control method for gas turbine power plants. The system includes: The model building module is used to collect operating parameters, non-destructive testing data and fault records of gas turbine power plant pipelines; it associates and binds the collected data with the corresponding pipeline components in the three-dimensional geometric model of the gas turbine power plant, and establishes a historical data archive for each pipeline segment, including original installation stress data, weld location data, material creep record data and historical overheating area coordinate data, so that the historical state change process of the pipeline can be traced back through the time axis; The prediction and simulation module is used to construct and train a fault prediction model based on a generative adversarial network, which includes a generator and a discriminator. Real-time running parameters are input into the trained fault prediction model and combined with a thermodynamic mechanism model for joint simulation to predict the health distribution of the pipeline network in the future time period and mark and display potential fault points. The collaborative handling module is used to generate a work order containing maintenance information in response to the predicted potential fault points; when on-site personnel identify the target pipeline through augmented reality equipment, the historical data archive and real-time status information of the pipeline are overlaid on the display interface; after the maintenance is completed, the measured data is uploaded to the three-dimensional digital model and the historical data archive is updated.

[0014] As a further aspect of the present invention, the process of establishing a historical data archive for each pipeline segment includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

[0015] As a further aspect of the present invention, the method further includes: in the process of constructing and training a fault prediction model based on a generative adversarial network, the training objective includes three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a 3D digital model that integrates historical data archives, linking multi-source data (operating parameters, non-destructive testing data, fault records, etc.) with corresponding components in the 3D geometric model. During maintenance, on-site personnel use augmented reality equipment to identify target pipelines, and the historical data archives and real-time status information of the pipeline are overlaid on the display interface. This seamless integration of data and physical entities allows on-site personnel to obtain contextual information instantly, greatly improving the intuitiveness and accuracy of fault location and solving the problem of difficulty in intuitively locating abnormal areas in traditional methods. Attached Figure Description

[0017] Figure 1 A flowchart of the three-dimensional collaborative management and control method for gas turbine power plants.

[0018] Figure 2 This is a block diagram of the composition structure of a three-dimensional collaborative control system for gas turbine power plants. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] Figure 1 This is a flowchart of a three-dimensional collaborative management and control method for gas turbine power plants, provided as an embodiment of the present invention. The method includes: Step S1: Construct a three-dimensional digital model that integrates historical data archives; Specifically, in constructing a 3D digital model integrating historical data archives, the first step is to collect operating parameters, non-destructive testing (NDT) data, and fault records for the gas turbine power plant pipelines. Operating parameter collection can be achieved by deploying various sensors on the pipelines, such as temperature sensors, pressure sensors, and flow sensors, which monitor the state of the internal medium in real time. NDT data can be obtained manually and periodically using equipment such as ultrasonic testing instruments and X-ray inspection machines to inspect pipeline wall thickness and weld quality. Fault records can be obtained from the power plant's operation and maintenance logs, accident reports, etc. This data collection forms the basis for subsequent analysis and prediction. For example, a traditional data acquisition system can be used, storing sensor data in a relational database, storing NDT reports in document format, and recording fault records in text format in spreadsheets.

[0022] Subsequently, the collected data is associated and bound to the corresponding pipeline components in the 3D geometric model of the gas turbine power plant. The 3D geometric model, constructed using CAD software, accurately reflects the spatial location, dimensions, and topology of the pipelines. Data association and binding can be achieved by establishing foreign key relationships in the database, linking operating parameters, monitoring data, and fault records to unique identifiers representing specific pipeline components in the 3D model. For example, each pipeline component can be assigned a unique ID, and all related data can be tagged with this ID, allowing queries to match data with components in the 3D model using the ID.

[0023] Furthermore, a historical data archive is established for each pipeline segment, containing initial installation stress data, historical weld location data, material creep record data, and historical overheat zone coordinate data. Initial installation stress data can be obtained through stress measurements or finite element analysis during installation. Historical weld location data can be obtained through manual recording or scanning weld markings. Material creep record data can be estimated through periodic material sampling analysis or empirical models based on operating time and temperature. Historical overheat zone coordinate data can be obtained by periodically scanning and recording the locations of abnormally high-temperature areas using an infrared thermometer. This data can be organized into a structured archive, for example, in the form of a file directory, with each pipeline corresponding to a folder, and each folder containing multiple sub-files named by time, each storing data at a specific point in time.

[0024] Therefore, the timeline allows for tracing back the historical changes in the pipeline's status. This means that users can specify a point in time or a time period, and the system can display all historical data and status information of the pipeline within that time period. For example, a simple query interface can be developed that allows users to input a date range, and the system can then retrieve and display all records within the corresponding time period from the historical data archive, thereby enabling the tracing of the pipeline's historical status.

[0025] Step S2: Fault prediction based on generative adversarial networks and mechanistic models; Specifically, in fault prediction based on generative adversarial networks (GANs) and mechanistic models, a fault prediction model based on a GAN is first constructed and trained. This GAN consists of a generator and a discriminator. The generator is designed to learn the degradation patterns of pipelines under different operating conditions and generate simulated fault feature data. The discriminator is trained to distinguish these simulated data from real pipeline fault data. For example, a standard GAN architecture can be used, with both the generator and discriminator composed of multilayer perceptrons or convolutional neural networks. Through iterative training, the generator can produce simulated data that is statistically similar to real fault data.

[0026] Next, real-time operating parameters are input into the trained fault prediction model. These parameters include pipeline temperature, pressure, flow rate, vibration, etc. These parameters are acquired from the power plant's real-time monitoring system via a data interface and formatted before being input into the fault prediction model. For example, real-time sensor data can be directly used as the model's input feature vector.

[0027] Furthermore, co-simulation is performed using a thermodynamic mechanism model. After obtaining the prediction results from the generative adversarial network model, they can be compared or fused with the calculation results from the thermodynamic mechanism model. The thermodynamic mechanism model can calculate physical quantities such as stress distribution and temperature field inside the pipeline based on the pipeline's material properties, geometric dimensions, and operating conditions. For example, a preliminary failure probability can be predicted by the generative adversarial network model, and then this probability can be used as an input parameter for the thermodynamic mechanism model. The boundary conditions of the mechanism model can be adjusted, and another simulation can be performed to verify the physical rationality of the prediction results.

[0028] This allows for the prediction of the pipeline network's health distribution over a future time period and the marking of potential failure points. The health distribution can be a numerical value representing the probability of a pipeline failure at a future point in time. Potential failure points are highlighted on the 3D digital model using color changes, flashing indicators, or specific icons. For example, the system can calculate the failure probability of each pipeline segment within the next week and map these probabilities to different colors, such as green for healthy, yellow for warning, and red for high risk, then display the corresponding pipeline segments on the 3D model using these colors.

[0029] Considering that severe pipeline failures in gas turbine power plants are low-probability events and the sample size of real historical failure data is limited, at least one of the following strategies is further adopted to ensure the training effect of generative adversarial networks under limited sample conditions: (1) Based on the transfer learning method, the generator is pre-trained using public datasets or simulation datasets with similar working conditions or similar material properties, and then fine-tuned using real historical fault data to reduce the dependence on the sample size of the target domain. (2) Introduce a data augmentation architecture to apply physically feasible perturbations to a limited number of real fault samples, including but not limited to adding noise, stretching the time axis, and scaling the amplitude, to generate an augmented sample set for training; (3) A few-shot learning framework is adopted, which combines generative adversarial networks with meta-learning. By learning prior distributions on multiple tasks, the model can quickly adapt to new failure modes with only a few samples. (4) Use thermodynamic mechanism models to generate a large amount of simulated fault data that conforms to physical laws, and use it as auxiliary training samples for generative adversarial networks to expand the scale of the training dataset.

[0030] To ensure the real-time performance of fault prediction and extreme condition simulation, the system employs at least one of the following computational acceleration strategies: (1) Layered computing architecture: The prediction task is divided into a lightweight real-time prediction layer and a heavyweight offline inference layer. After the real-time running parameters are input, the lightweight version generator (such as a network compressed by model pruning, quantization or knowledge distillation) is used to make a fast prediction in seconds and output the preliminary health distribution; when high-precision prediction is required or a specific threshold is triggered, the complete generative adversarial network and thermodynamic mechanism model are called for fine simulation. (2) Proxy model replacement: The computationally intensive thermodynamic mechanism model or multiphysics coupling model is trained by neural network to build a proxy model, and the model inference replaces the traditional numerical solution, reducing the computation time from minutes to seconds; (3) Parallel computing and hardware acceleration: The reasoning process of generative adversarial networks is accelerated in parallel by using graphics processors or tensor processors, while a distributed computing architecture is adopted for numerical calculation tasks such as finite element analysis to achieve parallel processing of multiple tasks. (4) Pre-calculation and result caching: For typical working scenarios, the calculation results of thermodynamic mechanism models and multi-physics coupling operators are pre-calculated and cached. During real-time prediction, the similarity between the current working parameters and the cached results is used for fast matching and interpolation to avoid repeated calculations. (5) Incremental update architecture: The health prediction of the pipeline network is performed incrementally. For areas where there is no significant change in operating conditions, the prediction results of the previous time step are used. Only areas where the parameters change drastically are recalculated, which effectively reduces the computational load.

[0031] Step S3: Collaborative inspection and closed-loop handling based on augmented reality; Specifically, in terms of augmented reality-based collaborative maintenance and closed-loop handling, the system first generates a work order containing maintenance information based on predicted potential fault points. Once the fault prediction model identifies a potential fault point, the system can automatically trigger the work order generation process. The work order includes the specific location of the fault point, the predicted fault type, suggested maintenance measures, a list of required tools, and safety precautions. For example, the system can automatically fill in a preset work order template based on the prediction results and then send it to maintenance management personnel for review and dispatch.

[0032] Subsequently, when on-site personnel identify the target pipeline using augmented reality (AR) equipment, the historical data archives and real-time status information of the pipeline are overlaid on the display interface. Wearing AR glasses or using an AR tablet, the AR equipment can identify the pipeline when its camera is pointed at it. It then retrieves the associated historical data archives (such as installation stress and weld records) and real-time operating parameters (such as current temperature and pressure) from the 3D digital model. This information is then overlaid in real-time on the on-site personnel's field of vision as virtual labels, charts, or text boxes, aligned with the physical pipeline. For example, when on-site personnel point the AR equipment at a section of pipeline, a virtual label appears on the screen, labeled "Pipeline Segment A-101," along with the date and result of its most recent non-destructive testing, and the current real-time temperature reading.

[0033] After maintenance is completed, the measured data is uploaded to the 3D digital model, and the historical data archive is updated. Following the completion of the maintenance work, on-site personnel will upload the measured data obtained during the maintenance process, such as new wall thickness measurements, photos of repaired welds, and serial numbers of replaced components, to the system via augmented reality equipment or a compatible mobile terminal. This data will be automatically associated with the corresponding pipeline components, updating their information in the 3D digital model and the historical data archive. For example, maintenance personnel can use augmented reality equipment to take photos of the repaired pipeline and input new wall thickness measurements. After receiving this data, the system will automatically update the historical data archive of that pipeline segment and simultaneously update the attribute information of that pipeline segment in the 3D model.

[0034] The overall technical concept of this invention demonstrates significant technological contributions. Traditional gas turbine power plant pipeline management methods have shortcomings in terms of the correlation between data and spatial location, making it difficult for maintenance personnel to intuitively locate abnormal areas. This invention achieves seamless integration of data and physical entities by associating and binding multi-source data with a three-dimensional geometric model, and using augmented reality equipment to overlay and display historical data archives and real-time status information on-site. This allows on-site personnel to obtain contextual information in real time, greatly improving the intuitiveness and accuracy of fault location.

[0035] In terms of fault prediction, existing technologies mostly rely on alarms triggered by exceeding preset thresholds, which is a reactive approach. This invention, however, constructs and trains a fault prediction model based on generative adversarial networks and mechanistic models. This model can predict the health distribution of pipeline networks over future time periods and mark potential fault points. This proactive predictive capability enables power plants to shift from passive response to proactive prevention, identifying risks before faults occur and thus avoiding unplanned outages and significant economic losses.

[0036] Furthermore, traditional maintenance decision-making lacks data support, leading to over-maintenance or under-maintenance. This invention, responding to predicted potential fault points, automatically generates work orders containing maintenance information and integrates historical data archives and real-time status information for collaborative on-site handling. This data-driven maintenance decision-making ensures targeted and efficient maintenance work, avoiding resource waste.

[0037] Finally, addressing the issues of information silos and low fault handling efficiency between on-site inspections and remote experts, this invention utilizes augmented reality devices to achieve real-time overlay display and data uploading of on-site information, forming a closed-loop management system encompassing data collection, prediction, maintenance, and feedback. This closed-loop management architecture not only ensures real-time data updates and continuous model optimization but also provides a solid foundation for future intelligent operation and maintenance. Compared to existing technologies, the method of this invention represents a technological advancement in data integration, intelligent prediction, decision support, and collaborative efficiency.

[0038] In a preferred embodiment of the present invention, step S1, which involves establishing a historical data archive for each pipeline segment, further includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

[0039] In this embodiment, the pipeline design stress threshold data refers to the maximum allowable stress value or safe stress range determined during the pipeline design phase based on factors such as material properties, working medium, temperature, and pressure. It serves as a benchmark for evaluating the pipeline's structural integrity and safety margin. This data can originate from pipeline design drawings, material specifications, engineering calculation reports, or industry standards. Alternatively, it can be obtained through simulation tools such as finite element analysis, which calculate the theoretical stress distribution of the pipeline under simulated operating conditions and extract the stress thresholds for key points.

[0040] The cumulative fatigue data resulting from each unit start-up and shutdown refers to the accumulated fatigue damage caused by the rapid changes in parameters such as temperature and pressure during the start-up and shutdown of a gas turbine power plant unit. This causes the pipeline materials to endure cyclical stress, gradually reducing the fatigue life of the material. It can be calculated and estimated by recording the number of cycles for each start-up and shutdown, the temperature / pressure change curves during the start-up and shutdown process, and combining this data with the material's SN curve or a fatigue damage accumulation model.

[0041] The coordinate data of overheated areas recorded in each infrared temperature measurement refers to the areas where the local temperature exceeds the design allowable range or normal operating temperature, detected by infrared thermal imagers and other equipment during pipeline operation. The precise location information of these areas in three-dimensional space is recorded. Through regular or irregular infrared thermal imager inspections, combined with image processing and 3D model matching technology, overheated areas can be automatically identified and their three-dimensional coordinates extracted.

[0042] The purpose of storing the above data in a time-series manner is to ensure that all pipeline health-related data can be organized and queried in chronological order, thereby enabling the tracing of historical pipeline status changes, damage accumulation processes, and maintenance events. A time-series database can be used for storage, as it natively supports timestamp indexing and efficient time-series data querying.

[0043] Ultimately, a complete lifecycle health record for the pipeline is created, integrating key data from the pipeline's design, installation, operation, and maintenance processes to form a comprehensive, dynamic, and traceable health file. This provides complete data support for pipeline status assessment, fault prediction, and maintenance decisions. A unified data management platform can be built to integrate these various data types and provide a visual interface to display the pipeline's health status evolution trend.

[0044] In a preferred embodiment of the present invention, the fault prediction model based on generative adversarial networks is constructed and trained in step S2, and its training objective includes three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

[0045] In this embodiment, a fault prediction model based on a generative adversarial network (GAN) is constructed and trained. The GAN is a deep learning model consisting of a generator and a discriminator, which improve performance through adversarial training. The generator is typically a deep neural network, such as a convolutional neural network or a recurrent neural network, used to learn the distribution of data and generate new samples; the discriminator is also a deep neural network used to distinguish real data from fake data generated by the generator. Alternatively, the generator can employ a variational autoencoder structure, combining the advantages of adversarial training; the discriminator can employ a multilayer perceptron or a more complex network structure to adapt to different types of data discrimination tasks. The training objective is an optimization function or loss function that guides the model's learning process. It defines the metric that the model needs to minimize or maximize during training to achieve the expected performance. The training objective can be constructed by combining different loss functions, such as cross-entropy loss and mean squared error loss, and weighted according to the specific task and data characteristics of the model.

[0046] Specifically, the first part maximizes the probability that the discriminator classifies real historical fault data as true, aiming to ensure that the discriminator can accurately identify real historical fault data and has the ability to distinguish between real and false fault data. When training the discriminator, real historical fault data can be used as positive sample input, and a binary cross-entropy loss function is used, with the goal of making the discriminator output a probability value close to 1.

[0047] The second part minimizes the probability that the discriminator classifies the simulated fault data generated by the generator as true. This aims to encourage the generator to produce more realistic simulated fault data, enabling it to "deceive" the discriminator and thus improve the quality of the generated data. When training the generator, the simulated fault data generated by the generator can be input into the discriminator, and a binary cross-entropy loss function can be used, with the goal of making the discriminator output a probability value close to 0.

[0048] The third part adds a physical consistency constraint term based on thermodynamic mechanisms. This aims to ensure that the simulated fault data generated by the generator is physically reasonable and conforms to the laws of thermodynamics, avoiding the generation of data that violates physical laws. The physical consistency constraint term can be a regularization term that calculates the difference between the generated data and the prediction results of the thermodynamic mechanism model and adds it to the generator's loss function to penalize generated data that does not conform to physical laws.

[0049] The constraint generator generates simulated fault data that conforms to the laws of thermodynamics, aiming to achieve physical consistency and ensure that the simulated data is correct in accordance with fundamental thermodynamic principles such as energy conservation and entropy increase. This can be achieved by embedding thermodynamic equations into the generator's loss function or by correcting the generated data as a post-processing step to ensure that it satisfies these equations.

[0050] Through the above technical solution, this invention significantly improves the accuracy and reliability of fault prediction models in the three-dimensional collaborative management and control method for gas turbine power plants. By refining the training objective of the generative adversarial network into three parts, and in particular by introducing a physical consistency constraint term based on thermodynamic mechanisms, the problem of potential physical inconsistencies in traditional generative models when simulating fault data is effectively solved. This enables the generator to produce simulated fault data that is both highly realistic and strictly conforms to the laws of thermodynamics. When these high-quality simulated data are co-simulated with thermodynamic mechanism models, they can more accurately reflect the actual operating status and potential degradation trends of the pipeline network, thereby achieving accurate prediction of the pipeline health distribution over future time periods and more reliably marking and displaying potential fault points. This avoids prediction bias caused by physical inconsistencies in the simulated data, providing stronger technical support for preventive maintenance and safe operation of gas turbine power plants.

[0051] In a preferred embodiment of the present invention, step S2, which combines a thermodynamic mechanism model with joint simulation, further includes: Real-time operating parameters are input into a trained generator, which simulates the degradation path of equipment under specific operating conditions to generate potential fault feature data. This generated fault feature data is then compared and fused with the calculation results of a thermodynamic mechanism model to predict the health distribution of the pipeline network over a future time period. The specific implementation of this comparison and fusion is as follows: The potential fault feature data generated by the generative adversarial network is compared and fused with the physical field data calculated by the thermodynamic mechanism model based on the same operating parameters. The fusion process adopts at least one of the following methods: (1) Weighted average fusion: The output results of the generator and the calculation results of the thermodynamic mechanism model are assigned weight coefficients respectively, and the two are weighted averaged to obtain the fused prediction results. Among them, the weight coefficients are dynamically adjusted according to the confidence of the generator under the current operating conditions. The confidence is determined by the output probability of the discriminator and the historical prediction accuracy. The higher the confidence, the greater the weight of the generator result; (2) Kalman filter fusion: The calculation results of the thermodynamic mechanism model are used as the state prediction value based on physical laws, and the output results of the generator are used as the observation value. The two are iteratively updated by the Kalman filter algorithm to obtain the optimal fusion estimation result; (3) Bayesian fusion: Construct conditional probability distribution models for the generator output and the thermodynamic mechanism model calculation results respectively, calculate the posterior probability distribution based on Bayes' theorem, and take the state with the largest posterior probability as the final fusion prediction result; (4) Machine learning fusion: The output of the generator and the calculation results of the thermodynamic mechanism model are used together as input features, and the fused prediction results are output through a pre-trained fusion network. The fusion network adopts a multilayer perceptron or a neural network structure based on an attention architecture, and its training objective is to minimize the error between the fusion result and the actual fault data; (5) Consistency verification and selection: The output results of the generator are compared with the calculation results of the thermodynamic mechanism model point by point. For regions where the deviation between the two is less than the preset threshold, the average of the two is taken as the fusion result. For regions where the deviation between the two is greater than the preset threshold, the calculation results of the thermodynamic mechanism model are used first, and the region is marked for subsequent manual verification. At the same time, the difference is used as a feedback signal for further optimization training of the generator.

[0052] In this embodiment, real-time operating parameters refer to physical quantity data collected in real time by various sensors and monitoring equipment during the actual operation of the gas turbine power plant pipeline system. These parameters may include, but are not limited to, the temperature, pressure, flow rate, vibration frequency, surface temperature, ambient temperature, and ambient humidity of the medium inside the pipeline. Their function is to provide real operating condition information of the current pipeline system as the input basis for the fault prediction model, ensuring the real-time performance and accuracy of the prediction. These parameters can be collected through distributed control systems (DCS), sensor networks, or Internet of Things (IoT) devices and transmitted to the data processing platform in real time. The trained generator is a core component of a generative adversarial network (GAN), whose main function is to learn and simulate the distribution characteristics of real data, thereby generating new data with similar statistical characteristics to real data. In this invention, the generator has been trained with a large amount of historical operating data, fault data, and simulated data, and has the ability to simulate the degradation process of gas turbine power plant pipeline equipment under specific operating conditions. The training process usually involves adversarial learning with the discriminator. By continuously optimizing the generator's parameters, the generated simulated data can deceive the discriminator as much as possible, making it unable to distinguish between real data and generated data.

[0053] Simulating the degradation path of equipment under specific operating conditions refers to the process by which a generator, based on input real-time operating parameters and its learned degradation patterns, predicts and generates the performance degradation, damage accumulation, or deterioration that the equipment may experience over a future period. This simulation considers the impact of current operating conditions (i.e., specific operating conditions) on the equipment's lifespan and health, such as how factors like high temperature, high pressure, corrosion, and fatigue accelerate equipment aging. The simulation of the degradation path can be represented as a series of time-varying characteristic data, reflecting the gradual evolution of the equipment from a healthy state to a failure state. Generating potential failure characteristic data refers to the various data indicators output by the generator during the simulation of the equipment degradation path, which characterize the potential failures of the equipment. These data can be physical quantities directly related to failure (such as crack propagation rate, material strength reduction, and local stress concentration), or characteristic quantities indirectly reflecting failure trends (such as abnormal vibration spectra, temperature gradient changes, and acoustic emission signal characteristics). These data are a quantitative description of possible future failure scenarios by the generator, providing a basis for subsequent failure prediction and health assessment.

[0054] The calculation results of the thermodynamic mechanism model are mathematical models built based on physical laws and engineering principles, used to describe the thermodynamic behavior of gas turbine power plant pipeline systems under different operating conditions. The calculation results include, but are not limited to, the temperature field distribution, pressure field distribution, fluid velocity distribution, thermal stress distribution, and material creep rate within the pipeline. These results are obtained through numerical methods (such as finite element analysis and computational fluid dynamics) based on fundamental physical laws such as mass conservation, energy conservation, and momentum conservation, combined with material properties, geometric structure, and boundary conditions. The calculation results of the thermodynamic mechanism model provide a physically consistent and reliable benchmark for verifying and correcting the simulation data generated by the generator. Comparison and fusion refers to the process of comparing and integrating the potential fault characteristic data generated by the generator with the calculation results of the thermodynamic mechanism model. The purpose of the comparison is to check whether the generated data conforms to physical laws; for example, whether the temperature rise in a certain region predicted by the generator matches the heat transfer and energy balance calculated by the thermodynamic model for that region. Fusion, based on comparison, modifies or weights the generated data to reflect both the complex patterns learned by the generator from historical data and the physical consistency constraints represented by the thermodynamic mechanism model. This fusion can be achieved through various algorithms such as weighted averaging, Kalman filtering, and data assimilation to obtain more accurate and reliable prediction results. Predicting the health distribution of a pipeline network over a future period refers to assessing the overall health status of the pipeline network at a future point in time or over a period of time based on the compared and fused data. Health distribution can be presented in numerical, color-coded, or risk-level formats, indicating the health level of different areas or components of the pipeline. For example, health can be a score from 0 to 100, or divided into levels such as "healthy," "sub-healthy," "warning," and "faulty." This prediction not only provides a single health value but, more importantly, presents the differences in health status among various parts of the entire pipeline network, thereby identifying potential weaknesses and high-risk areas.

[0055] In a preferred embodiment of the present invention, the step S2 of marking and displaying potential fault points further includes: In response to user-inputted extreme working condition assumption analysis commands, an accident chain inference model based on multi-physics coupling is used to simulate the sequence of potential risk points and accident propagation paths under extreme working conditions under the action of multi-physics coupling. The calculation basis of the accident chain inference model is that the failure probability distribution at future time is determined by the stress field distribution state at the current time, as well as the rate of change of temperature, pressure and flow velocity, and is calculated by a multi-physics coupling operator based on the fusion of finite element analysis and machine learning.

[0056] In this embodiment, the "extreme operating condition hypothesis analysis command in response to user input" refers to the system's ability to receive and process specific input from the user, which is intended to initiate analysis and simulation under specific extreme operating conditions. This allows maintenance personnel to proactively explore the behavior of pipeline systems under abnormal or high-risk scenarios, rather than simply relying on passive predictions based on routine operating data. This command can provide an input module through a user interface (e.g., a graphical user interface), allowing the user to select preset extreme operating condition scenarios (e.g., overload operation, emergency shutdown, cooling system failure, etc.) or customize key input parameters (e.g., sudden increase in ambient temperature, abnormal fluctuations in internal pressure, etc.).

[0057] The "accident chain simulation model based on multi-physics coupling" refers to a mathematical model that comprehensively considers the interactions of multiple physical phenomena (e.g., thermodynamics, fluid dynamics, solid mechanics, chemical reactions, etc.) to predict and analyze how a fault in a device or system propagates from one point to another, forming a chain reaction. This model can more realistically reflect the behavior of complex industrial systems under extreme conditions. The model achieves data exchange and mutual influence by iteratively solving multiple independent physical field models (e.g., a heat conduction model, a fluid dynamics model, and a structural stress model) by sharing boundary conditions or variables.

[0058] "Simulating the sequence of potential risk points and accident propagation paths under extreme operating conditions and multiphysics coupling" refers to using the aforementioned multiphysics coupling model to dynamically predict and display potential fault points in a pipeline system and their sequence of occurrence under user-defined extreme operating conditions, as well as the path of how a fault spreads from its initial point along the pipeline network, affecting other components or areas. This helps identify weak points in the system and potential cascading failure risks. During the simulation, key parameters such as stress, temperature, and pressure in various parts of the pipeline are monitored in real time. Once a parameter exceeds a preset safety threshold or its degradation rate is abnormal, it is marked as a potential risk point, and its impact on adjacent areas is deduced based on physical laws.

[0059] The calculation basis of the accident chain extrapolation model is: the failure probability distribution at future moments is jointly determined by the current stress field distribution, as well as the rates of change of temperature, pressure, and flow velocity. This clarifies the core physical and mathematical foundation for the accident chain extrapolation model's prediction. It points out that the probability of pipeline components failing at a future point in time does not solely depend on their current static state, but is dynamically influenced by the current stress distribution and the rates of change of key operating parameters (temperature, pressure, and flow velocity). This dynamic consideration makes the prediction more timely and accurate. This calculation basis can be established by building mathematical models based on physical degradation frameworks, such as fatigue life models and creep damage models. These models take parameters such as stress, temperature, and pressure as inputs and combine them with their rates of change to calculate the cumulative damage to the material, thereby deriving the failure probability.

[0060] The phrase "calculated using a multiphysics coupling operator based on the fusion of finite element analysis (FEA) and machine learning" describes the specific technical means used to achieve the above calculations. It combines the high-precision physical field simulation capabilities of finite element analysis (FEA) in handling complex geometries and boundary conditions with the advantages of machine learning (ML) in handling complex nonlinear relationships, pattern recognition, and optimization. This fusion improves computational efficiency, accuracy, and adaptability to unknown conditions. The operator uses finite element analysis as a solver for the physical fields, generating a large amount of simulation data. Then, a machine learning model is used to train this data, learning the complex coupling relationships and failure modes between physical fields. This allows for rapid results in subsequent real-time predictions through the machine learning model, reducing redundant FEA calculations.

[0061] As a specific implementation method, the following training objective can be adopted when constructing and training a fault prediction model based on generative adversarial networks: In the first part, the loss function of the discriminator can be designed as follows: ; in This indicates that the discriminator analyzes real historical fault data. The probability of judging it as true; This is based on all the real data. of Find the expected value, because The function is monotonically increasing, so when When it approaches 1, When the value is close to 0, the value of this term is large (there are few negative ones). The discriminator wants to maximize this term, that is, it wants to give high scores to all real data. This represents random noise, typically a random vector that follows a simple distribution (such as a Gaussian or uniform distribution). Generator According to random noise The simulated fault data created; Indicates the discriminator versus the generator Generated simulated fault data The probability of judging it as true; This applies to all random noise. The generated simulation data, for its Calculate the expected value; It is the total loss of the discriminator, which is minimized. To maximize the probability that the discriminator identifies real historical fault data as true.

[0062] In the second part, the adversarial loss function of the generator can be designed as follows: ; in , , Same meaning as in Part 1, The generator's goal is to fool the discriminator; therefore, it wants the data it generates to be... The discriminator considers it true, that is, it is hoped that... Make it as large as possible, close to 1. Close to 0. It is the adversarial loss of the generator. The negative sign before the formula indicates the generator's objective (maximizing) Convert to Minimize Therefore, minimize This is equivalent to having the generator produce data that is as realistic as possible, making it difficult for the discriminator to distinguish, thus giving these simulated data a high score. (Approximately 1).

[0063] The third part, the physical consistency constraint, can be designed as follows: ; in It is still simulated fault data generated by the generator; It is a feature extraction function. It extracts features from... Specific, key physical quantities are extracted, such as temperature, pressure, and stress. This is a model based on thermodynamic mechanisms. It is an independent, predefined mathematical model that encapsulates the laws of thermodynamics. Given the same simulation data... This model will "calculate" based on the laws of physics what physical quantities such as temperature and pressure should theoretically be observable under these conditions; This refers to the loss of physical consistency. It quantifies the degree to which the generated data violates the laws of physics.

[0064] Finally, the generator's total loss function can be: ; in It is a weighting coefficient used to balance the importance of adversarial loss and physical consistency constraints. In this way, the generator not only generates realistic data during training, but also ensures that this data conforms to the laws of thermodynamics. It is the generator's final, overall loss function. It guides the direction in which the generator learns.

[0065] In a preferred embodiment of the present invention, the step S2 of marking and displaying potential fault points further includes: Data from acoustic sensors deployed on the pipeline is collected. When a minor leak occurs, the precise three-dimensional coordinates of the leak point are calculated and marked by utilizing the time difference of sound waves arriving at different sensors and the spatial topology of a three-dimensional geometric model. The calculation of the leak point coordinates is based on multiplying the difference between the time each sensor receives the sound wave and the time the leak occurs by the speed of sound wave propagation in the medium to obtain the distance from each sensor to the leak point. The precise three-dimensional coordinates of the leak point are determined by finding the spatial point that minimizes the sum of squared errors between this distance and the actual sensor coordinates.

[0066] In this embodiment, acoustic wave sensors are used to monitor the pipeline's operating status in real time and capture acoustic wave signals generated by minute leaks. These sensors can be implemented using various technologies. For example, piezoelectric sensors can be used, which convert pipeline vibrations or medium acoustic waves into electrical signals through the piezoelectric effect; or fiber optic acoustic wave sensors can be used, which utilize the phase or intensity changes of light waves in optical fibers to detect acoustic waves, and have the advantage of being resistant to electromagnetic interference.

[0067] Sound waves propagate at a relatively constant speed in the pipeline medium. Therefore, the difference in distance between the leak point and different sound wave sensors will cause the sound wave signal to arrive at each sensor at different times. This time difference is key information for determining the location of the leak. In practice, this can be achieved by synchronizing each sensor with a high-precision clock, accurately recording the arrival timestamps of the sound wave signals, and then calculating the differences between these timestamps.

[0068] A three-dimensional geometric model provides precise spatial location information and connection relationships for gas turbine power plant pipelines, a necessary prerequisite for three-dimensional leak location calculations. This model can be imported from existing CAD (Computer-Aided Design) or BIM (Building Information Modeling) data, containing detailed geometric data such as pipeline routing, bends, and branches; alternatively, it can be used to create a high-precision model of the actual pipeline using 3D reconstruction techniques such as laser scanning and photogrammetry to obtain its true spatial topology. Based on the principles of sound wave propagation and sensor data, the specific location of the leak point in three-dimensional space is determined and presented intuitively on the user interface. This can be achieved through various location algorithms; for example, trilateration can be used to solve for the intersection points in three-dimensional space based on the distances from at least three sensors to the leak point.

[0069] The calculation of leak point coordinates is based on the fundamental physical principle of acoustic localization, converting the time difference of sound wave signal arrival into spatial distance. In practical applications, the propagation speed of sound waves in pipeline media (such as steam, water, and natural gas) can be accurately determined through prior experimental measurements, consulting standard data, or online calibration (e.g., conducting acoustic wave transmission and reception tests between two sensors at a known distance). The precise three-dimensional coordinates of the leak point are determined by finding the spatial point that minimizes the sum of squared errors between this distance and the actual sensor coordinates. This is a mathematical optimization method used to find the leak point location that best fits the physical model from multiple sensor data, even with measurement errors. Commonly used solution methods include nonlinear least squares optimization methods such as the Gauss-Newton method or the Levenberg-Marquardt algorithm. These methods iteratively adjust the leak point coordinates until the sum of squared errors between the calculated and actual distances from the sensor to the leak point is minimized.

[0070] In a preferred embodiment of the present invention, the remote collaborative guidance in step S3 further includes: Remote experts generate virtual identifiers with directional and annotation meanings through a control terminal, which are then projected in real time and precisely anchored to specific components in the augmented reality field of view of on-site personnel via a communication protocol.

[0071] In this embodiment, a remote expert generates virtual icons with directional and annotation meanings via a control terminal. The control terminal can be a device for remote operation and information input, such as a computer workstation with a graphical user interface or a portable tablet. Virtual icons are virtual information elements overlaid in an augmented reality environment, providing intuitive visual guidance to on-site personnel. These virtual icons can take various forms, such as an arrow pointing in a specific direction or an outline highlighting a specific area. Through their directional and annotation meanings, these virtual icons clearly convey the expert's intent, such as indicating the location to be inspected, the order of operations, or precautions.

[0072] The generated virtual identifiers are projected in real time via a communication protocol. The communication protocol is a set of rules for data transmission and exchange, ensuring that information can be accurately and efficiently transmitted from the control terminal to the augmented reality device used by on-site personnel. For example, a WebSocket protocol based on the TCP / IP protocol stack can be used to achieve low-latency bidirectional communication.

[0073] The projected virtual identifiers are precisely anchored to specific components within the augmented reality (AR) view of on-site personnel. Precise anchoring refers to accurately positioning and fixing the virtual identifier in three-dimensional space to its corresponding physical component, ensuring consistency with objects in the physical world. This can be achieved through various technologies. For example, the built-in SLAM technology of AR devices, combined with visual recognition algorithms, can be used to perform real-time 3D reconstruction and target recognition of the on-site environment, thereby achieving precise alignment between the virtual identifier and the physical component. Alternatively, tracking technology based on preset marker points or feature points can be employed to determine the spatial location of the physical component by identifying these marker points, thus precisely overlaying the virtual identifier. Specific components refer to actual physical components in the gas turbine power plant's pipeline system, such as a section of pipeline, a valve, a sensor, or a weld area.

[0074] Figure 2 The present invention provides a three-dimensional collaborative management and control system for gas turbine power plants, comprising: (The block diagram of the system is shown below.) Model building module 11 is used to collect operating parameters, non-destructive testing data and fault records of gas turbine power plant pipelines; associate and bind the collected data with the corresponding pipeline components in the three-dimensional geometric model of the gas turbine power plant, and establish a historical data archive for each pipeline segment, including original installation stress data, weld location data, material creep record data and historical overheating area coordinate data, so that the historical state change process of the pipeline can be traced back through the time axis; The prediction and simulation module 12 is used to construct and train a fault prediction model based on a generative adversarial network, which includes a generator and a discriminator. Real-time running parameters are input into the trained fault prediction model and combined with a thermodynamic mechanism model for joint simulation to predict the health distribution of the pipeline network in the future time period and mark and display potential fault points. The collaborative handling module 13 is used to generate a work order containing maintenance information in response to the predicted potential fault points; when on-site personnel identify the target pipeline through augmented reality equipment, the historical data archive and real-time status information of the pipeline are overlaid on the display interface; after the maintenance is completed, the measured data is uploaded to the three-dimensional digital model and the historical data archive is updated.

[0075] Furthermore, the process of establishing historical data archives for each section of pipeline includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

[0076] Specifically, the method further includes: in the process of constructing and training a fault prediction model based on generative adversarial networks, the training objective includes three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

[0077] In this embodiment, the model building module 11 serves as the system's foundational data layer, responsible for establishing and maintaining historical data archives for the pipeline. This module can be a standalone server-side application that interacts with various data acquisition systems and 3D modeling software within the power plant via API interfaces to achieve automated data acquisition and correlation.

[0078] The prediction and inference module 12 is responsible for predicting the health status of the pipeline network and identifying potential failure points. This module can be a cloud-based AI service platform that utilizes high-performance computing resources for training and inference of generative adversarial network models to process large-scale data and provide rapid predictions.

[0079] The collaborative handling module 13 is the core of the system's interaction with on-site maintenance personnel. Its function is to transform the predicted results into executable maintenance work orders and provide on-site information support and remote collaboration capabilities. This module can be a web-based collaborative work platform that supports multi-user access and interfaces with augmented reality (AR) devices to achieve real-time information sharing.

[0080] The connections between the aforementioned modules—namely, the connection between the model building module 11 and the prediction and inference module 12, and the connection between the collaborative processing module 13 and the former two—can be achieved in various ways. For example, standardized communication protocols can be used for data exchange and instruction transmission to ensure smooth data flow and effective function calls.

[0081] The system of this invention effectively solves the problem of lack of integration and connection between functional links in traditional methods. The model building module ensures the comprehensiveness and traceability of historical data, providing a solid foundation for subsequent analysis. The connection between the prediction and deduction module 12 and the model building module 11 enables fault prediction to make full use of historical data and real-time parameters, significantly improving the accuracy and timeliness of prediction. The close connection between the collaborative handling module 13 and the first two modules realizes the seamless conversion of prediction results into maintenance work orders, and presents key information intuitively to on-site personnel through augmented reality technology, greatly improving maintenance efficiency and accuracy. At the same time, the remote collaborative guidance function breaks down information silos, enabling experts to provide real-time and accurate guidance for on-site operations. Finally, the closed-loop update architecture of maintenance data ensures the continuous optimization of historical data archives and the self-improvement of the system. Through this highly integrated modular system, the pipeline management of gas turbine power plants has achieved a transformation from passive response to proactive prediction, and from decentralized management to collaborative operation, significantly improving the safety, reliability, and economy of power plant operation.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional collaborative control method for gas turbine power plants, characterized in that, The method includes: Collect operating parameters, non-destructive testing data and fault records of gas turbine power plant pipelines; associate and bind the collected data with the corresponding pipeline components in the three-dimensional geometric model of the gas turbine power plant, and establish a historical data archive for each pipeline segment, including original installation stress data, weld location data, material creep record data and historical overheating zone coordinate data, so that the historical state change process of the pipeline can be traced back through the time axis; A fault prediction model based on a generative adversarial network (GAN) is constructed and trained. The GAN includes a generator and a discriminator. Real-time operating parameters are input into the trained fault prediction model and combined with a thermodynamic mechanism model for joint simulation to predict the health distribution of the pipeline network in the future time period and mark and display potential fault points. In response to the predicted potential fault points, a work order containing maintenance information is generated; when on-site personnel identify the target pipeline using augmented reality equipment, the historical data archive and real-time status information of the pipeline are overlaid on the display interface; after the maintenance is completed, the measured data is uploaded to the three-dimensional digital model, and the historical data archive is updated.

2. The three-dimensional collaborative control method for gas turbine power plants according to claim 1, characterized in that, The process of creating a historical data archive for each section of pipeline includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

3. The three-dimensional collaborative control method for gas turbine power plants according to claim 1, characterized in that, In the process of constructing and training a fault prediction model based on generative adversarial networks, the training objectives include three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

4. The three-dimensional collaborative control method for gas turbine power plants according to claim 1, characterized in that, The process of co-simulation using thermodynamic mechanism models includes: Real-time operating parameters are input into a trained generator, which simulates the degradation path of equipment under specific operating conditions to generate potential fault feature data. The generated fault feature data is then compared and fused with the calculation results of the thermodynamic mechanism model to predict the health distribution of the pipeline network in the future time period.

5. The three-dimensional collaborative control method for gas turbine power plants according to claim 1, characterized in that, The process of marking and identifying potential fault points includes: Responding to user-inputted extreme working condition assumption analysis commands, an accident chain inference model based on multi-physics coupling simulates the sequence of potential risk points and accident propagation paths under extreme working conditions under the action of multi-physics coupling. The calculation basis of the accident chain inference model is that the failure probability distribution at future time moments is jointly determined by the stress field distribution state at the current time moment, as well as the rate of change of temperature, pressure, and flow velocity, and is calculated by a multi-physics coupling operator based on the fusion of finite element analysis and machine learning. The multi-physics coupling operator can be constructed in the following way: using finite element analysis as the physics solver to generate training samples, learning the coupling relationship between physics fields through a machine learning model, and using the trained surrogate model to replace finite element analysis for rapid prediction.

6. The three-dimensional collaborative control method for gas turbine power plants according to claim 5, characterized in that, The process of marking potential fault points also includes: Data from acoustic sensors deployed on the pipeline is collected. When a minor leak occurs, the precise three-dimensional coordinates of the leak point are calculated and marked using the time difference of sound waves arriving at different sensors and the spatial topology of a three-dimensional geometric model. The calculation of the leak point coordinates is based on the following: the arrival time difference between any two sensors is calculated based on the time it takes for each sensor to receive the same leaking acoustic event, and then multiplied by the speed of sound propagation in the medium to obtain the distance difference between the two sensors and the leak point. Based on the distance difference data of at least three sets of sensor pairs, combined with the known spatial coordinates of each sensor, the spatial point that minimizes the sum of squared positioning errors is determined by solving for the intersection of hyperboloids or by using a nonlinear least squares optimization method. This point is then used as the precise three-dimensional coordinates of the leak point.

7. The three-dimensional collaborative control method for gas turbine power plants according to claim 1, characterized in that, The method further includes: a remote expert generating a virtual identifier with pointing and annotation meanings through a control terminal, and transmitting the virtual identifier to the augmented reality device of the on-site personnel through WebSocket or a real-time communication protocol based on TCP / IP; The virtual identifier is precisely anchored to a specific component in the augmented reality device's field of view by on-site personnel in the following way: the on-site environment is reconstructed in real time based on a simultaneous localization and mapping algorithm, and the spatial pose of the target component is determined by combining preset marker points or natural feature point recognition technology. The three-dimensional coordinates of the virtual identifier are then mapped to the local coordinate system of the augmented reality device, thereby achieving precise alignment and real-time tracking between the virtual identifier and the physical component.

8. A three-dimensional collaborative control system for gas turbine power plants, the system being used to implement the three-dimensional collaborative control method for gas turbine power plants as described in any one of claims 1 to 7, characterized in that, The system includes: The model building module is used to collect operating parameters, non-destructive testing data and fault records of gas turbine power plant pipelines; it associates and binds the collected data with the corresponding pipeline components in the three-dimensional geometric model of the gas turbine power plant, and establishes a historical data archive for each pipeline segment, including original installation stress data, weld location data, material creep record data and historical overheating area coordinate data, so that the historical state change process of the pipeline can be traced back through the time axis; The prediction and simulation module is used to construct and train a fault prediction model based on a generative adversarial network, which includes a generator and a discriminator. Real-time running parameters are input into the trained fault prediction model and combined with a thermodynamic mechanism model for joint simulation to predict the health distribution of the pipeline network in the future time period and mark and display potential fault points. The collaborative handling module is used to generate a work order containing maintenance information in response to the predicted potential fault points; when on-site personnel identify the target pipeline through augmented reality equipment, the historical data archive and real-time status information of the pipeline are overlaid on the display interface; after the maintenance is completed, the measured data is uploaded to the three-dimensional digital model and the historical data archive is updated.

9. The three-dimensional collaborative control system for gas turbine power plants according to claim 8, characterized in that, The process of creating a historical data archive for each section of pipeline includes: The design stress threshold data of the pipeline, the cumulative fatigue data caused by each unit start-up and shutdown, and the coordinate data of the over-temperature area recorded by each infrared temperature measurement are linked to the historical data archive as supplementary data and stored according to the time dimension to form a full life cycle health record covering design benchmarks, operational damage and abnormal events.

10. The three-dimensional collaborative control system for gas turbine power plants according to claim 8, characterized in that, The method further includes: in the process of constructing and training a fault prediction model based on a generative adversarial network, the training objective includes three parts: The first part is to maximize the probability that the discriminator identifies real historical fault data as true. The second part minimizes the probability that the discriminator identifies the simulated fault data generated by the generator as true. The third part adds a physical consistency constraint term based on thermodynamic mechanisms to constrain the simulated fault data generated by the generator to conform to the laws of thermodynamics.

Images