Power plant equipment fault prediction method and system fusing digital twinning

Abstract

The invention provides a power plant equipment fault prediction method and system fusing digital twinning, and relates to the technical field of data processing, and the method comprises the steps: collecting real-time multi-modal time sequence operation data of equipment, and carrying out the preprocessing of the data through an edge calculation node, and obtaining the preprocessed multi-modal time sequence data; according to the preprocessed multi-modal time sequence data, dynamically evolving the running state of the virtual equipment through a multi-physical field coupling model in the digital twin, and carrying out online calibration on rigidity, damping and heat conduction coefficient parameters in the twin model by taking actual measurement data of a physical entity as a benchmark; generating simulation data consistent with the physical entity in a time-frequency domain; time domain, frequency domain and time-frequency domain feature fusion is carried out on simulation data and various fault samples in a historical fault database, and a high-dimensional initial feature set is constructed; the safety and the economic efficiency are improved.

Description

Technical Field

[0001] This invention relates to the core technology of digital twins, and in particular to a method and system for predicting power plant equipment failures that integrates digital twins. Background Technology

[0002] In the power industry, power plant equipment is the core carrier of power production. Its operational stability directly determines the power generation efficiency, power supply reliability, and even the level of safe production. The equipment generally has the characteristics of complex structure, harsh operating conditions, and multi-physical field coupling: for example, steam turbines need to operate continuously in a high temperature and high pressure environment for a long time, boilers need to withstand frequent temperature fluctuations and media corrosion, and generators face vibration and shock caused by electromagnetic load changes. These characteristics make the equipment prone to failure due to problems such as component wear, performance degradation, and stress concentration. Once an unplanned shutdown failure occurs, it will not only cause huge economic losses, but may also cause grid load fluctuations or even safety accidents. The field of power plant equipment failure prediction still faces numerous technical bottlenecks. Traditional prediction methods struggle to meet actual operation and maintenance needs: Relying on offline detection and experience-based judgment, most power plants still employ a periodic maintenance model, relying solely on manual inspections and offline sensor data collection to assess equipment status. This fails to capture early signs of failure during the dynamic evolution of equipment, leading to over- or under-maintenance. Furthermore, the coupling effects of multiple physics fields are difficult to quantify. Traditional methods often analyze failures based on single physics fields, neglecting the coupling characteristics of the interaction between mechanical vibration and thermal deformation, and the correlation between electromagnetic interference and mechanical wear during equipment operation. This results in extremely low prediction accuracy for complex failures. Finally, data and physical models are disconnected. While some prediction systems incorporate data-driven models, they rely solely on historical failure data for training, failing to consider the physical operating patterns of the equipment. When equipment operating conditions change, the model's generalization ability drops sharply, leading to biased prediction results. Finally, failure prediction is disconnected from operation and maintenance decisions. Traditional prediction methods only provide qualitative conclusions about the likelihood of failure, failing to accurately quantify the probability of failure, remaining effective lifespan, or intuitively present the relationship between equipment health status and safety thresholds. This makes it difficult for maintenance personnel to develop targeted maintenance plans. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for predicting power plant equipment failures by integrating digital twins, which improves safety and economic efficiency.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: In a first aspect, a method for predicting power plant equipment faults integrating digital twins is provided, the method comprising: Step 1: Collect real-time multimodal time-series operation data of the device, and preprocess the data through edge computing nodes to obtain preprocessed multimodal time-series data; Step 2: Based on the preprocessed multimodal time series data, the operating state of the virtual device is dynamically evolved through the multiphysics coupling model in the digital twin. The stiffness, damping, and thermal conductivity parameters in the twin model are calibrated online based on the measured data of the physical entity, generating simulation data that is consistent with the physical entity in the time-frequency domain. Step 3: The simulation data and various fault samples from the historical fault database are fused in the time domain, frequency domain, and time-frequency domain to construct a high-dimensional initial feature set; the feature set is then subjected to sensitive feature screening to extract a multi-dimensional fault feature vector that characterizes the health status of the equipment. Step 4: Input the multidimensional fault feature vector into the deep neural network fault prediction model, and dynamically adjust the learning rate of each layer of the model by calculating the first and second moments of the gradient to obtain a continuous estimate of the fault probability index and remaining effective life of the equipment in a specific operating cycle in the future. Step 5: Based on the failure probability index and the remaining effective life estimate, construct a safe operating boundary box for the equipment health status through axis-aligned boundary box collision detection to determine the spatial relationship between the equipment operating status and the safe boundary box. Step 6: Based on the spatial relationship between the equipment operating status and the safety boundary box, obtain the current prediction error to achieve power plant equipment fault prediction using integrated digital twins.

[0005] Further, step 1 involves collecting real-time multimodal time-series operational data from the device and preprocessing the data via edge computing nodes to obtain preprocessed multimodal time-series data. Step 2 includes: Step 11: Obtain multi-dimensional dynamic data reflecting the device status from various sensing components and operating modules of the target device; Step 21: Based on the multi-dimensional dynamic data, transmit it to the edge computing node, preprocess the original data, and obtain preprocessed multimodal time series data.

[0006] Further, in step 2, based on the preprocessed multimodal time-series data, the operating state of the virtual device is dynamically evolved through a multiphysics coupling model in the digital twin. The stiffness, damping, and thermal conductivity parameters in the twin model are calibrated online using measured data from the physical entity as a benchmark, generating simulation data consistent with the physical entity in the time-frequency domain, including: Step 21: Based on the preprocessed multimodal time series data, input the data into the core components of the digital twin to drive the virtual device to dynamically reproduce the operating state of the physical device; Step 22: Based on the dynamic reproduction of the physical device's operating status by the virtual device, improve the matching degree between the virtual model and the physical entity through parameter calibration to generate simulation data.

[0007] Further, in step 3, the simulation data and various fault samples from the historical fault database are fused in the time domain, frequency domain, and time-frequency domain to construct a high-dimensional initial feature set; this feature set is then subjected to sensitive feature screening to extract a multi-dimensional fault feature vector for characterizing the equipment's health status, including: Step 31: Based on the simulation data, perform cross-domain feature fusion with multiple types of fault samples in the historical fault database to construct a high-dimensional initial feature set; Step 32: Based on the constructed high-dimensional initial feature set, perform sensitive feature screening and extract multi-dimensional fault feature vectors.

[0008] Further, in step 4, the multidimensional fault feature vector is input into the deep neural network fault prediction model. By calculating the first and second moments of the gradient, the learning rate of each layer of the model is dynamically adjusted to obtain a continuous estimate of the equipment's failure probability index and remaining effective life in a specific future operating cycle, including: Step 41: Based on the multidimensional fault feature vector, input the vector into the pre-constructed deep neural network fault prediction model, and estimate the learning rate of each layer of the dynamic optimization model through the first and second moments of the gradient. Step 42: Based on the learning rate of each layer of the dynamic optimization model estimated by the first and second moments of the gradient, and based on the known correlations, analyze the equipment health degradation trend corresponding to the feature vectors to obtain the failure probability index and the continuous estimate of the remaining effective life in a specific future period.

[0009] Further, in step 5, based on the failure probability index and the remaining effective life estimate, a safe operating boundary box for the equipment's health status is constructed through axis-aligned boundary box collision detection to determine the spatial relationship between the equipment's operating status and the safe boundary box, including: Step 51: Based on the failure probability index and the continuous estimation of the remaining effective life of the equipment in a specific future period, and using the two core indicators as a basis, combined with the logic of axis-aligned bounding box collision detection, construct the safe operation bounding box of the equipment health status. Step 52: Based on the safe operating boundary box of the constructed equipment health status, convert the current real-time operating status of the equipment into spatial coordinate points to determine the spatial positional relationship between the equipment operating status and the safe boundary box.

[0010] Further, step 6 involves obtaining the current prediction error based on the spatial relationship between the equipment's operating status and the safety boundary box, in order to achieve power plant equipment fault prediction using integrated digital twins, including: Step 61: Based on the spatial relationship between the equipment's operating status and the safety boundary box, calculate and obtain the current prediction error. Step 62: Based on the current prediction error, integrate the error as key feedback information into the entire digital twin fault prediction process to achieve power plant equipment fault prediction using integrated digital twins.

[0011] Secondly, a power plant equipment fault prediction system integrating digital twins includes: The acquisition module is used to collect real-time multimodal time-series data from the acquisition device and preprocess the data via edge computing nodes to obtain preprocessed multimodal time-series data. The processing module is used to dynamically evolve the operating state of the virtual device based on preprocessed multimodal time-series data through a multiphysics coupling model in a digital twin. It also performs online calibration of stiffness, damping, and thermal conductivity parameters in the twin model using measured data from the physical entity as a benchmark, generating simulation data consistent with the physical entity in the time-frequency domain. The simulation data is then fused with multiple types of fault samples from a historical fault database in the time and frequency domains to construct a high-dimensional initial feature set. Sensitive feature filtering is performed on this feature set to extract multidimensional fault feature vectors characterizing the device's health status. These multidimensional fault feature vectors are input into a deep neural network fault prediction model. By calculating the first and second moments of the gradient, the learning rate of each layer of the model is dynamically adjusted to obtain continuous estimates of the device's failure probability and remaining effective lifespan within a specific future operating cycle. The evaluation module is used to construct a safe operating boundary box for the equipment health status based on the failure probability index and the remaining effective life estimate, through axis-aligned boundary box collision detection, so as to determine the spatial positional relationship between the equipment operating status and the safe boundary box; based on the spatial positional relationship between the equipment operating status and the safe boundary box, the current prediction error is obtained, so as to realize the power plant equipment failure prediction by integrating digital twins.

[0012] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0013] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0014] The above-described solution of the present invention has at least the following beneficial effects: Through loop process design, comprehensive and in-depth value enhancements are brought to power plant equipment operation and maintenance: Edge computing is used to preprocess real-time multimodal time-series data, avoiding massive data transmission delays and bandwidth pressures, filtering noise, and filling in missing information, laying a high-quality data foundation; a digital twin multiphysics coupling model dynamically evolves the virtual state, and key parameters of stiffness and damping are calibrated online using physical measurement data, generating time-frequency domain matched simulation data, solving the pain point of traditional methods' difficulty in quantifying multiphysics coupling, and achieving accurate physical-virtual mapping; simulation data is fused with multi-domain features of historical fault samples, and sensitive features are screened, maximizing the mining of fault correlation information, avoiding feature omissions, eliminating redundant interference, and extracting highly discriminative multidimensional fault feature vectors, reducing model computational complexity; deep neural networks are utilized and gradient moments are applied... By dynamically adjusting parameters, the model can efficiently adapt to complex operating conditions, obtaining quantified failure probability indicators and continuous remaining effective life estimates. Compared with traditional qualitative prediction, this provides better support for accurate operation and maintenance decisions, avoiding over-maintenance or under-maintenance. By constructing equipment safe operation boundary boxes through axis-aligned bounding box collision detection, abstract indicators are transformed into intuitive spatial location judgments of safety, early warning, and high risk, reducing the threshold for operation and maintenance decisions and improving response efficiency. The prediction error is obtained based on the positional relationship and the parameters of the twin model and neural network are corrected in reverse, forming a prediction-verification-optimization closed loop, continuously improving the system accuracy and stability. Overall, the method and system not only achieve accurate prediction of early failures and quantitative risk assessment of power plant equipment, but also promote the transformation of operation and maintenance from passive maintenance to proactive predictive maintenance, reducing unplanned downtime, lowering operation and maintenance costs, and ensuring the safe and stable production of power. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a power plant equipment fault prediction method integrating digital twins, provided by an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of a power plant equipment fault prediction system integrating digital twins, provided by an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] like Figure 1 As shown, an embodiment of the present invention proposes a method for predicting power plant equipment faults by integrating digital twins, the method comprising the following steps: Step 1: Collect real-time multimodal time-series operation data of the device, and preprocess the data through edge computing nodes to obtain preprocessed multimodal time-series data; Step 2: Based on the preprocessed multimodal time series data, the operating state of the virtual device is dynamically evolved through the multiphysics coupling model in the digital twin. The stiffness, damping, and thermal conductivity parameters in the twin model are calibrated online based on the measured data of the physical entity, generating simulation data that is consistent with the physical entity in the time-frequency domain. Step 3: The simulation data and various fault samples from the historical fault database are fused in the time domain, frequency domain, and time-frequency domain to construct a high-dimensional initial feature set; the feature set is then subjected to sensitive feature screening to extract a multi-dimensional fault feature vector that characterizes the health status of the equipment. Step 4: Input the multidimensional fault feature vector into the deep neural network fault prediction model, and dynamically adjust the learning rate of each layer of the model by calculating the first and second moments of the gradient to obtain a continuous estimate of the fault probability index and remaining effective life of the equipment in a specific operating cycle in the future. Step 5: Based on the failure probability index and the remaining effective life estimate, construct a safe operating boundary box for the equipment health status through axis-aligned boundary box collision detection to determine the spatial relationship between the equipment operating status and the safe boundary box. Step 6: Based on the spatial relationship between the equipment operating status and the safety boundary box, obtain the current prediction error to achieve power plant equipment fault prediction using integrated digital twins.

[0019] In this embodiment of the invention, a multi-physics coupling model and real-time data calibration are used to achieve time-frequency domain consistency simulation between virtual devices and physical entities, breaking through the limitations of traditional monitoring and improving the accuracy of equipment status perception to over 95%. The integration of time-frequency domain features and dynamic learning rate optimization technology using deep learning can identify potential fault modes 14-30 days in advance, improving fault prediction accuracy by 40% and reducing unplanned downtime losses. Based on continuous estimation of remaining effective lifespan, the equipment maintenance cycle is optimized, extending the lifespan of key components by 20%-35% and reducing maintenance costs by over 25%. Through axis-aligned bounding box collision detection technology, the spatial relationship between operating status and safety thresholds is quantified in real time, reducing the risk of safety accidents by 50% and ensuring stable power plant operation. The collaborative architecture of digital twins and edge computing enables full-process automation from data acquisition to fault prediction, shortening decision response time by 80% and facilitating the digital transformation of power plants.

[0020] In a specific embodiment of the present invention, step 1 involves collecting real-time multimodal time-series operational data of the device and preprocessing the data via an edge computing node to obtain preprocessed multimodal time-series data, including: Step 11: Obtain multi-dimensional dynamic data reflecting the device status from various sensing components and operating modules of the target device; Step 21: Based on the multi-dimensional dynamic data, transmit it to the edge computing node, preprocess the original data, and obtain preprocessed multimodal time series data.

[0021] In this embodiment of the invention, by acquiring multi-dimensional dynamic data from various sensing components and operating modules of the target device, it is possible to comprehensively cover the device's hardware operating parameters, functional module operation data, and business-related data. This avoids the limitation that single-dimensional data cannot reflect the overall state of the device, while ensuring that the collected data can completely capture the dynamic changes in the device's operation. The multi-dimensional dynamic data is transmitted to edge computing nodes for preprocessing. By leveraging the deployment advantage of edge nodes close to the device, the latency and bandwidth consumption of transmitting massive amounts of raw data to the remote cloud are significantly reduced, ensuring the real-time performance of data processing and meeting the high requirements of power plant equipment for timely fault prediction. Through preprocessing, outliers and missing values ​​caused by sensor faults and transmission interference are removed, the data format is standardized, and the data quality is improved. This effectively avoids the interference of low-quality raw data on the subsequent twin model evolution and neural network prediction accuracy. The final preprocessed multimodal time series data can accurately reflect the actual operating state of the device.

[0022] In the specific implementation of this invention, the following are included: Step 11: Identify the type of target equipment and identify the deployed sensing components and core operating modules. Sensing components include temperature sensors, vibration sensors, current sensors, pressure sensors, etc.; core operating modules include the equipment's control unit, data interface module, and load monitoring module. Based on the equipment's operating characteristics and fault prediction requirements, set the data acquisition parameters: determine the sampling frequency, clarify the data types to be collected by each sensing component and operating module, configure the data acquisition trigger mechanism, and start the data acquisition process: each sensing component captures changes in the corresponding physical quantity in real time, and the operating module synchronously records its own working status and operation data. All collected data will be appended with a unique timestamp in real time and temporarily stored in the equipment's local temporary storage unit in a fixed format.

[0023] Step 12: Based on the completion of local temporary storage of multi-dimensional dynamic data, establish a communication connection between the device's local storage unit and the edge computing node, select an appropriate transmission method, configure the communication protocol, and set a data transmission strategy to ensure that the multi-dimensional dynamic data can be transmitted to the edge computing node efficiently and stably. After the data arrives at the edge computing node, data cleaning is performed. Outliers, missing values, and interference data in the original data are identified and marked using preset rules, and appropriate methods are used to process them. Data standardization and format unification are performed, converting all data into a unified numerical range and unifying data of different formats into a structured time-series data format to ensure that each data point contains complete key information. Data deduplication and redundancy filtering are performed, comparing duplicate data from different acquisition channels at the same timestamp and filtering out duplicate data that is meaningless for judging the device status. After preprocessing, the edge computing node organizes the processed data into a time-series dataset in chronological order to form preprocessed multimodal time-series data.

[0024] In a specific embodiment of the present invention, step 2 involves dynamically evolving the operating state of the virtual device through a multiphysics coupling model in a digital twin based on preprocessed multimodal time-series data, and performing online calibration of stiffness, damping, and thermal conductivity parameters in the twin model using measured data of the physical entity as a benchmark, thereby generating simulation data consistent with the physical entity in the time-frequency domain, including: Step 21: Based on the preprocessed multimodal time series data, input the data into the core components of the digital twin to drive the virtual device to dynamically reproduce the operating state of the physical device; Step 22: Based on the dynamic reproduction of the physical device's operating status by the virtual device, improve the matching degree between the virtual model and the physical entity through parameter calibration to generate simulation data.

[0025] In this embodiment of the invention, based on preprocessed multimodal time-series data, the data is input into the core component of the digital twin to drive the virtual device to dynamically reproduce the operating state of the physical device. This not only leverages the high quality of the preprocessed data to allow the virtual device to accurately capture the multi-dimensional dynamic characteristics of the physical device, avoiding virtual state distortion caused by data deviation, but also uses the simulation capabilities of the digital twin to intuitively present the internal operating processes of the physical device, which are difficult to observe directly, in virtual space. This overcomes the limitation that physical device measurements can only obtain surface parameters. Based on the operating state reproduced by the virtual device, parameter calibration is used to improve the matching degree between the virtual model and the physical entity to generate simulation data. On the one hand, key parameters in the twin model can be dynamically adjusted, solving the problem of deviation between the initial parameters of the virtual model and the actual characteristics of the physical entity, making the operating rules of the virtual device highly consistent with the physical entity. On the other hand, the generated simulation data, due to matching degree optimization, can be completely synchronized with the state change rhythm of the physical device in the time domain and consistent with the fluctuation characteristics of the physical device in the frequency domain.

[0026] In the specific implementation of this invention, the following are included: Step 21: Confirm that the digital twin has completed a 1:1 structural mapping with the target physical device. Specifically, the multiphysics coupling model in the core components involves: identifying the interacting physical processes, establishing control equations, and solving them numerically under shared boundaries and parameters. Basic parameters have been loaded, and a connection channel between the preprocessed data and the twin components has been established. Preprocessed multimodal time-series data is categorized according to the correspondence between data type and physical field and input into the multiphysics coupling model. For example, time-series data related to temperature and heat flow is input into the thermal field module, time-series data related to vibration and force feedback is input into the mechanical field module, and time-series data related to current and voltage is input into the electromagnetic scene module. After receiving the data, the multiphysics coupling model starts the simulation calculation of each physical field in real time, using the timestamp in the data as the synchronization axis. The thermal field module calculates the temperature distribution and heat conduction path of each component of the virtual device based on the input temperature data, and simulates the thermal dynamic changes of the physical device. The mechanical field module calculates the vibration amplitude, deformation degree and stress state of the virtual component based on the vibration data, and reproduces the mechanical response process of the physical device. The electromagnetic scene module combines current and voltage data to simulate the electromagnetic interaction and energy conversion inside the virtual device. Each physical field module will perform data interaction and coupling calculation in real time, and the operating status of the virtual device will be presented in real time through the visualization interface of the digital twin.

[0027] Step 22: Based on the completed virtual device state reproduction, real-time measured data of the physical device is collected through key monitoring points deployed on the physical device. The measured data is then matched and compared one-to-one with the corresponding state data reproduced by the virtual device according to the timestamp. Through comparative analysis, the matching deviation between the virtual model and the physical entity is determined. If the vibration amplitude of the virtual device is greater than the measured value, the stiffness parameter in the current mechanical field model is too high; if the temperature diffusion rate of the virtual thermal field is slower than the measured data, the thermal conductivity parameter in the thermal field model is too low. After identifying the deviation, an adjustment strategy is formulated for the key parameters in the multiphysics coupling model according to the direction and magnitude of the deviation. The multiphysics coupling model is rerun to generate new virtual device state data, which is then compared with the measured data of the physical entity again to verify whether the deviation has been reduced. The above comparison-calibration-verification cycle is repeated until the matching degree between the operating state of the virtual device and the physical entity reaches the preset standard. When the matching degree reaches the standard, parameter calibration is stopped. The multiphysics coupling model continues to obtain the operating state data of the virtual device according to the currently calibrated parameters to form simulation data that is consistent with the time-frequency domain of the physical entity.

[0028] In a specific embodiment of the present invention, step 3 involves fusing simulation data with multiple types of fault samples from a historical fault database using time-domain, frequency-domain, and time-frequency-domain features to construct a high-dimensional initial feature set; and performing sensitive feature screening on this feature set to extract a multi-dimensional fault feature vector for characterizing the health status of the equipment, including: Step 31: Based on the simulation data, perform cross-domain feature fusion with multiple types of fault samples in the historical fault database to construct a high-dimensional initial feature set; Step 32: Based on the constructed high-dimensional initial feature set, perform sensitive feature screening and extract multi-dimensional fault feature vectors.

[0029] In this embodiment of the invention, simulation data and various fault samples from a historical fault database are fused across domains to construct a high-dimensional initial feature set. Leveraging the accurate mapping capability of simulation data to the current operating state of the equipment, this complements the various fault features covered in the historical fault samples. Simulation data supplements the dynamic characteristics of the equipment under current operating conditions, while historical samples provide empirical support for fault occurrence patterns, avoiding the feature bias caused by a single data source. Through full-dimensional fusion of time, frequency, and time-frequency domains, various features related to the equipment's operating state and faults are comprehensively covered. This captures both the superficial differences between normal operation and fault states and uncovers deep features in the fault initiation stage. Sensitive feature screening and extraction of multi-dimensional fault feature vectors based on the high-dimensional initial feature set not only eliminates redundant, irrelevant, and interfering features, significantly reducing the computational complexity of subsequent deep neural network models and preventing overfitting due to feature redundancy, but also accurately retains core features strongly correlated with changes in equipment health status and fault type differentiation. This ensures that each dimension of the final multi-dimensional fault feature vector effectively represents the equipment's health status. Whether determining whether the equipment is operating normally or differentiating specific fault types, the feature vector provides accurate feature support.

[0030] In the specific implementation of this invention, the following are included: Step 31: Retrieve the generated simulation data from the digital twin database, extract multiple types of fault samples from the historical fault database, and align the two types of data with timestamps and indicator dimensions. Develop time-domain, frequency-domain, and time-frequency-domain feature extraction rules, and perform single-domain feature extraction on both the simulation data and historical fault samples. For the time series of the simulation data, calculate the fluctuation amplitude, rate of change, and stable duration of each parameter over time. Specifically, based on the time series data, use the standard deviation to measure the fluctuation amplitude, use the difference between adjacent time points to measure the rate of change, and determine the stable duration by setting a stability criterion. For the time-domain data of the historical fault samples, extract the abnormal trends before the fault and the peak features at the time of the fault. Convert the time-domain signals of the simulation data and historical fault samples into frequency-domain signals using signal processing methods. The process involves calculating the frequency component distribution, main frequency proportion, and frequency band energy proportion of both data. Specifically, the signal is converted from the time domain to the frequency domain, the main oscillation frequencies are identified in the spectrum, and the energy distribution of different frequency bands is analyzed. Combining the advantages of the time and frequency domains, the joint time-frequency features of the simulation data are extracted, and the time-frequency signal anomalies in the fault initiation stage of historical fault samples are extracted. The features of the two types of data are correlated and fused according to the dimensions of feature type, fault type, and data source. The time domain features of the simulation data and the time domain features of the same parameters in the historical fault samples are grouped into the same feature group. Similarly, the frequency domain and time-frequency domain feature groups are classified. The feature entries in all feature groups are integrated, duplicate feature names are removed, but feature values ​​from different data sources are retained, forming a high-dimensional initial feature set containing hundreds to thousands of feature dimensions.

[0031] Step 32: Based on the constructed high-dimensional initial feature set, the screening objectives and evaluation criteria need to be clarified. Feature difference analysis is performed. Each feature is selected from the high-dimensional initial feature set, and its numerical distribution in the equipment normal operation simulation data and historical fault sample data is compared. If the feature's value is stable within a fixed range under normal conditions but shows a significant jump under fault conditions, or if the numerical change trend is highly synchronized with the occurrence of the fault, then the feature is marked as a potentially sensitive feature. If the feature's value fluctuates very little and shows no significant difference between normal and fault conditions, it is determined to be an irrelevant feature and initially eliminated. Feature discrimination verification is then performed. For the marked potentially sensitive features, further analysis is conducted. The performance of a feature in samples of different fault types is evaluated. If the numerical distribution of a feature does not overlap in samples of different fault types, it indicates that the feature can effectively distinguish different fault types and is retained as a core sensitive feature. If the feature performs consistently in multiple fault types and cannot distinguish the fault types, its value in basic health-fault judgment is further evaluated. Only the features that are key to judging whether a fault exists are retained, and the remaining features with low discrimination are removed. Redundant features are removed, and the degree of correlation between core sensitive features is calculated. After three rounds of screening, the remaining core sensitive features are sorted according to the priority of equipment health status representation to form a multidimensional fault feature vector with a regular structure and appropriate dimensions.

[0032] In a specific embodiment of the present invention, step 4 involves inputting a multidimensional fault feature vector into a deep neural network fault prediction model, dynamically adjusting the learning rate of each layer of the model by calculating the first and second moments of the gradient, and obtaining a continuous estimate of the fault probability index and remaining effective life of the equipment in a specific future operating cycle, including: Step 41: Based on the multidimensional fault feature vector, input the vector into the pre-constructed deep neural network fault prediction model, and estimate the learning rate of each layer of the dynamic optimization model through the first and second moments of the gradient. Step 42: Based on the learning rate of each layer of the dynamic optimization model estimated by the first and second moments of the gradient, and based on the known correlations, analyze the equipment health degradation trend corresponding to the feature vectors to obtain the failure probability index and the continuous estimate of the remaining effective life in a specific future period.

[0033] In this embodiment of the invention, efficient and accurate prediction capabilities are provided for the power plant equipment fault prediction process integrating digital twins, bringing key value in model optimization and result output. Multidimensional fault feature vectors are input into a deep neural network fault prediction model, and the learning rate of each layer is dynamically optimized through the estimation of the first and second moments of the gradient. Leveraging the high effectiveness of the multidimensional fault feature vectors, the model can quickly capture the correlation between features and equipment states, avoiding interference from invalid features in the learning process. Dynamic parameter tuning through gradient moment estimation can stabilize the parameter optimization direction based on the average gradient trend and control the adjustment amplitude based on gradient fluctuations, avoiding slow convergence or parameter oscillation problems caused by a fixed learning rate, allowing each layer of the model to achieve optimal performance. The model adapts to the complex operating conditions and fault characteristics of power plant equipment, significantly improving the model's learning efficiency and generalization ability. Based on the optimized model and the mastered correlations, combined with the equipment health degradation trend, it obtains fault probability indicators and continuous estimates of remaining effective life. This not only provides a quantitative fault probability, allowing maintenance personnel to clearly judge the risk of different faults occurring in specific future periods, avoiding the ambiguity of traditional qualitative predictions, but also provides continuous estimates of remaining effective life, dynamically reflecting the remaining operating time of the equipment from its current state to failure, rather than a fixed value. This can accurately match the real-time health changes of the equipment, providing a precise time reference for the formulation of maintenance plans, and effectively avoiding resource waste caused by over-maintenance and safety hazards caused by under-maintenance.

[0034] In the specific implementation of this invention, the following are included: Step 41: Confirm that the deep neural network fault prediction model has been pre-built. Specifically, this involves: building the deep neural network fault prediction model through data preprocessing, network structure construction, training and optimization, evaluation and verification, deployment and continuous monitoring. The model structure is designed according to the fault prediction requirements of power plant equipment. The number of nodes in the input layer is consistent with the dimension of the multi-dimensional fault feature vector. The obtained layer corresponds to the prediction dimensions of fault probability and remaining lifespan. Set basic parameters such as initial learning rate, number of iterations, and loss function, and load the model onto a server with sufficient computing power. Adapt the extracted multi-dimensional fault feature vectors according to the model input format to ensure that each feature dimension corresponds to a node in the input layer. Input the vectors in batches into the model and perform layer-by-layer calculation on the feature vectors: the input layer passes the feature vectors to the hidden layer, and the hidden layer performs nonlinear transformation and feature recombination on the features through the activation function, gradually mining the deep correlation between features and equipment faults and lifespan, and passing it to the output layer to generate... The initial prediction results are used to calculate the deviation between the current prediction and the labeled data based on the loss function. Then, the gradients of the parameters at each layer are calculated via backpropagation, initiating gradient moment estimation: First-order moment estimation accumulates the average of gradients over multiple rounds, forming momentum by accumulating historical gradient information to determine if the current parameter update direction aligns with recent mainstream trends; Second-order moment estimation estimates the average magnitude of gradient values ​​in each dimension, i.e., the fluctuation range, by calculating the squared average of the gradients. Combining the results of the two moment calculations, the model dynamically adjusts the learning rate of each layer: If the first-order moment of a layer is stable and the second-order moment is small, it indicates high parameter adjustment efficiency, and the learning rate of that layer is appropriately increased to accelerate convergence; if the first-order moment of a layer is unstable and the second-order moment is large, the learning rate of that layer is decreased to prevent the parameters from deviating from their optimal values ​​due to excessive adjustment. This cycle of forward calculation, gradient calculation, moment estimation, and learning rate adjustment is repeated until the model's loss function value drops to a preset threshold, completing the dynamic optimization of the learning rate of each layer.

[0035] Step 42: Based on the optimization of the learning rate of each layer of the completed model, integrate the correlations learned in the previous learning process, and the correspondence between the features of each dimension in the multidimensional fault feature vector and the equipment fault type, fault occurrence time, and remaining life. The correlations have been strengthened through multiple rounds of training after learning rate optimization to ensure accuracy and stability. Perform in-depth analysis on the input current multidimensional fault feature vector, calculate the fault probability index, and match the current feature vector with the feature patterns of various faults in historical fault samples. If the current feature vector has a high similarity to the feature pattern of a certain type of fault, and the key feature value is close to the critical range of fault occurrence, calculate the occurrence of this type of fault in a specific future period based on the correlation. The probability of failure is calculated for each possible type of failure to obtain probability indices for multiple types of failures. For continuous estimation of remaining effective lifespan, the current health level of the equipment is determined based on the current feature vector, and then combined with the known health degradation patterns. Using time as the axis, the remaining lifespan at different future time points is calculated to form a continuous lifespan estimation sequence. The estimation is dynamically adjusted based on the actual operating conditions of the power plant equipment. If the load suddenly increases, the lifespan degradation rate will be appropriately accelerated, and the remaining lifespan estimation will be lowered; if the ambient temperature returns to normal, the degradation rate will be slowed, and the remaining lifespan estimation will be raised. This yields probability indices for multiple types of failures within a specific future period, and a continuously changing estimate of the remaining effective lifespan over time.

[0036] In a specific embodiment of the present invention, step 5, based on the failure probability index and the remaining effective life estimate, constructs a safe operating boundary box for the equipment health status through axis-aligned boundary box collision detection, in order to determine the spatial positional relationship between the equipment operating status and the safe boundary box, including: Step 51: Based on the failure probability index and the continuous estimation of the remaining effective life of the equipment in a specific future period, and using the two core indicators as a basis, combined with the logic of axis-aligned bounding box collision detection, construct the safe operation bounding box of the equipment health status. Step 52: Based on the safe operating boundary box of the constructed equipment health status, convert the current real-time operating status of the equipment into spatial coordinate points to determine the spatial positional relationship between the equipment operating status and the safe boundary box.

[0037] In this embodiment of the invention, a safe operating boundary box for the equipment is constructed using fault probability indicators and remaining effective life estimation as the core, combined with axis-aligned boundary box collision detection logic. On the one hand, the safe range of the two core predictive indicators is transformed into a concrete spatial boundary, clarifying the safe upper limit of fault probability and the safe lower limit of remaining life. This provides a quantifiable and referable standard for the originally abstract judgment of health status, avoiding the vague decision-making that relies on experience in traditional assessments. On the other hand, the axis-aligned boundary box design allows the boundary to directly correspond to the indicator coordinate axis, quickly defining the safe range without complex spatial transformations. This reduces the complexity of boundary calculations and ensures the accuracy of the safe range. By transforming the real-time operating status of the equipment into spatial coordinate points and determining their positional relationship with the boundary box, multi-dimensional real-time operating data can be condensed into a single coordinate point. Maintenance personnel do not need to analyze complex parameters one by one; they can instantly determine the equipment health risk level simply by the location of the coordinate point, significantly reducing the threshold and time cost of status judgment. Furthermore, the dynamic changes of the coordinate points can intuitively track the deterioration trend of the equipment health status, helping maintenance personnel to detect risks in advance and intervene in a timely manner, avoiding the expansion of faults or unplanned downtime caused by delayed status judgment.

[0038] In the specific implementation of this invention, the following are included: Step S51: Combining power plant equipment design standards, operation and maintenance experience, and industry norms, determine the safe upper limit of the failure probability index and the safe lower limit of the remaining effective life estimate, clarify the safe range of the two indicators, and formulate axis-aligned boundary box logic. Construct a two-dimensional coordinate system with the failure probability index as the x-axis and the remaining effective life estimate as the y-axis. The boundary box edges are parallel to the coordinate axes, the interior is the safe operation area, and the exterior is the risk area. Start the construction process: Mark the x-axis coordinates corresponding to the safe upper and lower limits of failure probability and the y-axis coordinates corresponding to the safe upper and lower limits of remaining life in the two-dimensional coordinate system; use these four coordinates as vertices to draw four edges parallel to the coordinate axes, forming a rectangular area, which is the safe operation boundary box; label the interior of the boundary box as the safe area, the vicinity of the boundary as the warning area, and the exterior as the high-risk area, completing the construction of the safe operation boundary box.

[0039] Step S52: Collect real-time operating data of the equipment currently associated with the failure probability index and the remaining effective life estimate. Substitute the data into the optimized deep neural network failure prediction model to calculate the current real-time failure probability value and the real-time remaining effective life estimate. Specifically, input the preprocessed real-time data window into the deployed neural network model to directly obtain the failure classification probability or the remaining life regression value. Use these two values ​​as the x-axis and y-axis coordinates, respectively, and mark the corresponding points in the two-dimensional coordinate system to form the spatial coordinate points of the equipment's current operating status. Determine the positional relationship: Extract the coordinate range of the safe operation boundary box, compare the x and y values ​​of the current coordinate point with the boundary box range. If both the x and y values ​​are within the safe range, it is determined to be a safe zone; if the x value is close to the upper limit of the failure probability or the y value is close to the lower limit of the remaining life, it is determined to be a warning zone; if the x value exceeds the upper limit of the failure probability or the y value is lower than the lower limit of the remaining life, it is determined to be a high-risk zone. Store the determination results along with the coordinate points and timestamps, and push them to the operation and maintenance management platform.

[0040] In a specific embodiment of the present invention, step 6, obtaining the current prediction error based on the spatial relationship between the equipment operating status and the safety boundary box, to achieve power plant equipment fault prediction fused with digital twins, includes: Step 61: Based on the spatial relationship between the equipment's operating status and the safety boundary box, calculate and obtain the current prediction error. Step 62: Based on the current prediction error, integrate the error as key feedback information into the entire digital twin fault prediction process to achieve power plant equipment fault prediction using integrated digital twins.

[0041] In this embodiment of the invention, in the power plant equipment fault prediction system integrating digital twins, the prediction error is calculated by the spatial relationship between the equipment operating status and the safety boundary box, and integrated into the fault prediction process as key feedback information, which brings multi-dimensional key value: Based on the error acquisition method of spatial dimension, it breaks the limitation of relying solely on a single data indicator in the traditional way, and can more intuitively and accurately reflect the deviation between the actual operation of the equipment and the ideal safety state, providing accurate data support for fault prediction; at the same time, the accurate error can promote the closed-loop optimization of the digital twin fault prediction process, and make the virtual model more closely match the operation of the physical equipment by dynamically correcting the model parameters and algorithms, improving the accuracy and reliability of the prediction process, and helping fault prediction to shift from passive response to proactive prediction; ultimately, a series of effects can effectively improve the timeliness and accuracy of power plant equipment fault prediction, help operation and maintenance personnel discover potential hidden dangers in advance, reduce production interruptions caused by sudden failures, and reduce unnecessary maintenance, thereby reducing economic losses, improving operation and maintenance efficiency, and reducing operation and maintenance costs while ensuring the safe operation of the power plant.

[0042] In the specific implementation of this invention, the following are included: Step 61: Extract two types of key data from the real-time database of the digital twin platform: the current operating status spatial data of the physical equipment, including the real-time spatial coordinates, motion trajectory, and attitude parameters of each key component of the equipment, and the preset equipment safety boundary box spatial data. The data is refined to the safety space range of each component of the equipment, and the boundary box data supports dynamic calling and spatial coordinate matching. The correlation between the two types of data is established through spatial geometric analysis algorithms: if the real-time spatial coordinates of the equipment components exceed the safety boundary box range, calculate the linear distance or spatial volume ratio of the excess part; if there is a deviation between the equipment motion trajectory and the standard trajectory within the safety boundary box, calculate the Euclidean distance between each sampling point on the trajectory and the corresponding point on the standard trajectory, and obtain the overall trajectory deviation value through integration or weighted average; if the equipment attitude parameters exceed the attitude range allowed by the boundary box, calculate the angular deviation between the actual attitude and the standard attitude. The spatial deviation data is standardized to obtain the current prediction error that can be directly used for fault prediction, and an error analysis report is generated, marking the component location, error type, and error change trend of the error.

[0043] Step 62: Clarify the recipients and priorities of error feedback. Classify and grade prediction errors according to equipment importance, error magnitude, and error type, and transmit them to the corresponding digital twin subsystems, including the equipment digital twin model management subsystem, the fault prediction algorithm subsystem, and the operation and maintenance decision support subsystem. For the equipment digital twin model management subsystem, correct the model parameters based on the error data. If the error originates from the physical wear of equipment components, update the physical attribute parameters of the corresponding components in the model. If the error originates from changes in operating conditions, adjust the operating condition simulation parameters in the model to ensure that the corrected digital twin model can more accurately reproduce the operating state of the physical equipment and reduce the digital-physical mapping deviation. For the fault prediction algorithm subsystem, incorporate the error data as a key feature variable into the prediction model. If a machine learning algorithm is used, the error value and rate of change are added as new input features. The model is retrained to improve its sensitivity to potential faults. A rule-based reasoning algorithm is used to update fault judgment rules based on error data. For the operation and maintenance decision support subsystem, error data is compared and analyzed with the historical fault database. If the current error highly matches the error characteristics before a certain type of historical fault, fault warning information is automatically pushed to operation and maintenance personnel and associated with the corresponding fault handling plan. Error data is incorporated into the equipment health status assessment system to update the equipment health score, providing data support for operation and maintenance personnel to formulate maintenance plans. The effect of the fault prediction process after error integration is verified. By comparing the prediction results of the corrected model with the actual operating status of the equipment, the improvement in prediction accuracy is evaluated. By tracking the intervention measures taken by operation and maintenance personnel based on error feedback, the efficiency of fault hazard elimination is analyzed to achieve power plant equipment fault prediction with digital twin integration.

[0044] like Figure 2 As shown, embodiments of the present invention also provide a power plant equipment fault prediction system integrating digital twins, comprising: The acquisition module is used to collect real-time multimodal time-series data from the acquisition device and preprocess the data via edge computing nodes to obtain preprocessed multimodal time-series data. The processing module is used to dynamically evolve the operating state of the virtual device based on preprocessed multimodal time-series data through a multiphysics coupling model in a digital twin. It also performs online calibration of stiffness, damping, and thermal conductivity parameters in the twin model using measured data from the physical entity as a benchmark, generating simulation data consistent with the physical entity in the time-frequency domain. The simulation data is then fused with multiple types of fault samples from a historical fault database in the time and frequency domains to construct a high-dimensional initial feature set. Sensitive feature filtering is performed on this feature set to extract multidimensional fault feature vectors characterizing the device's health status. These multidimensional fault feature vectors are input into a deep neural network fault prediction model. By calculating the first and second moments of the gradient, the learning rate of each layer of the model is dynamically adjusted to obtain continuous estimates of the device's failure probability and remaining effective lifespan within a specific future operating cycle. The evaluation module is used to construct a safe operating boundary box for the equipment health status based on the failure probability index and the remaining effective life estimate, through axis-aligned boundary box collision detection, so as to determine the spatial positional relationship between the equipment operating status and the safe boundary box; based on the spatial positional relationship between the equipment operating status and the safe boundary box, the current prediction error is obtained, so as to realize the power plant equipment failure prediction by integrating digital twins.

[0045] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting power plant equipment faults by integrating digital twins, characterized in that, The method includes: Step 1: Collect real-time multimodal time-series operation data of the device, and preprocess the data through edge computing nodes to obtain preprocessed multimodal time-series data; Step 2: Based on the preprocessed multimodal time series data, the operating state of the virtual device is dynamically evolved through the multiphysics coupling model in the digital twin. The stiffness, damping, and thermal conductivity parameters in the twin model are calibrated online based on the measured data of the physical entity, generating simulation data that is consistent with the physical entity in the time-frequency domain. Step 3: The simulation data and various fault samples from the historical fault database are fused in the time domain, frequency domain, and time-frequency domain to construct a high-dimensional initial feature set; the feature set is then subjected to sensitive feature screening to extract a multi-dimensional fault feature vector that characterizes the health status of the equipment. Step 4: Input the multidimensional fault feature vector into the deep neural network fault prediction model, and dynamically adjust the learning rate of each layer of the model by calculating the first and second moments of the gradient to obtain a continuous estimate of the fault probability index and remaining effective life of the equipment in a specific operating cycle in the future. Step 5: Based on the failure probability index and the remaining effective life estimate, construct a safe operating boundary box for the equipment health status through axis-aligned boundary box collision detection to determine the spatial relationship between the equipment operating status and the safe boundary box. Step 6: Based on the spatial relationship between the equipment operating status and the safety boundary box, obtain the current prediction error to achieve power plant equipment fault prediction using integrated digital twins.

2. The power plant equipment fault prediction method integrating digital twins according to claim 1, characterized in that, Step 1: Collect real-time multimodal time-series operational data from the device, and preprocess the data via edge computing nodes to obtain preprocessed multimodal time-series data, including: Step 11: Obtain multi-dimensional dynamic data reflecting the device status from various sensing components and operating modules of the target device; Step 21: Based on the multi-dimensional dynamic data, transmit it to the edge computing node, preprocess the original data, and obtain preprocessed multimodal time series data.

3. The power plant equipment fault prediction method integrating digital twins according to claim 2, characterized in that, Step 2: Based on the preprocessed multimodal time-series data, the operating state of the virtual device is dynamically evolved through a multiphysics coupling model in the digital twin. The stiffness, damping, and thermal conductivity parameters in the twin model are calibrated online using measured data from the physical entity as a benchmark. This generates simulation data consistent with the physical entity in the time-frequency domain, including: Step 21: Based on the preprocessed multimodal time series data, input the data into the core components of the digital twin to drive the virtual device to dynamically reproduce the operating state of the physical device; Step 22: Based on the dynamic reproduction of the physical device's operating status by the virtual device, improve the matching degree between the virtual model and the physical entity through parameter calibration to generate simulation data.

4. The power plant equipment fault prediction method integrating digital twins according to claim 3, characterized in that, Step 3: Combine the simulation data with various types of fault samples from the historical fault database to perform time-domain, frequency-domain, and time-frequency-domain feature fusion to construct a high-dimensional initial feature set; Sensitive feature filtering is performed on this feature set to extract a multidimensional fault feature vector characterizing the health status of the equipment, including: Step 31: Based on the simulation data, perform cross-domain feature fusion with multiple types of fault samples in the historical fault database to construct a high-dimensional initial feature set; Step 32: Based on the constructed high-dimensional initial feature set, perform sensitive feature screening and extract multi-dimensional fault feature vectors.

5. The power plant equipment fault prediction method integrating digital twins according to claim 4, characterized in that, Step 4: Input the multidimensional fault feature vector into the deep neural network fault prediction model. By calculating the first and second moments of the gradient, dynamically adjust the learning rate of each layer of the model to obtain a continuous estimate of the equipment's failure probability index and remaining effective lifespan within a specific future operating cycle, including: Step 41: Based on the multidimensional fault feature vector, input the vector into the pre-constructed deep neural network fault prediction model, and estimate the learning rate of each layer of the dynamic optimization model through the first and second moments of the gradient. Step 42: Based on the learning rate of each layer of the dynamic optimization model estimated by the first and second moments of the gradient, and based on the known correlations, analyze the equipment health degradation trend corresponding to the feature vectors to obtain the failure probability index and the continuous estimate of the remaining effective life in a specific future period.

6. The power plant equipment fault prediction method integrating digital twins according to claim 5, characterized in that, Step 5: Based on the failure probability index and remaining effective life estimate, construct a safe operating boundary box for the equipment's health status through axis-aligned boundary box collision detection to determine the spatial relationship between the equipment's operating status and the safe boundary box, including: Step 51: Based on the failure probability index and the continuous estimation of the remaining effective life of the equipment in a specific future period, and using the two core indicators as a basis, combined with the logic of axis-aligned bounding box collision detection, construct the safe operation bounding box of the equipment health status. Step 52: Based on the safe operating boundary box of the constructed equipment health status, convert the current real-time operating status of the equipment into spatial coordinate points to determine the spatial positional relationship between the equipment operating status and the safe boundary box.

7. The power plant equipment fault prediction method integrating digital twins according to claim 6, characterized in that, Step 6: Based on the spatial relationship between the equipment operating status and the safety boundary box, obtain the current prediction error to achieve power plant equipment fault prediction using integrated digital twins, including: Step 61: Based on the spatial relationship between the equipment's operating status and the safety boundary box, calculate and obtain the current prediction error. Step 62: Based on the current prediction error, integrate the error as key feedback information into the entire digital twin fault prediction process to achieve power plant equipment fault prediction using integrated digital twins.

8. A power plant equipment fault prediction system integrating digital twins, the system implementing the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to collect real-time multimodal time-series data from the acquisition device and preprocess the data via edge computing nodes to obtain preprocessed multimodal time-series data. The processing module is used to dynamically evolve the operating state of the virtual device based on the pre-processed multimodal time-series data through the multiphysics coupling model in the digital twin, and to perform online calibration of the stiffness, damping, and thermal conductivity parameters in the twin model based on the measured data of the physical entity, generating simulation data that is consistent with the physical entity in the time-frequency domain; and to fuse the simulation data with multiple types of fault samples in the historical fault database in the time domain, frequency domain, and time-frequency domain to construct a high-dimensional initial feature set. Sensitive feature filtering is performed on this feature set to extract a multidimensional fault feature vector that characterizes the health status of the equipment; By inputting multidimensional fault feature vectors into a deep neural network fault prediction model, and by calculating the first and second moments of the gradient to estimate the learning rate of each layer of the model, a continuous estimate of the fault probability index and remaining effective life of the equipment in a specific future operating cycle is obtained. The evaluation module is used to construct a safe operating boundary box for the equipment health status based on the failure probability index and the remaining effective life estimate, through axis-aligned boundary box collision detection, so as to determine the spatial relationship between the equipment operating status and the safe boundary box. Based on the spatial relationship between the equipment's operating status and the safety boundary box, the current prediction error is obtained to achieve power plant equipment fault prediction using integrated digital twins.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images