Thermal power plant fault early warning diagnosis method and system based on nebula system

Abstract

The invention relates to a thermal power plant fault early warning and diagnosis method based on a nebula system, and the method comprises the steps: collecting the multi-dimensional operation time sequence data of a thermal power plant, carrying out the feature extraction and lexical element processing of the multi-dimensional operation time sequence data through an encoder, and obtaining a unified equipment state lexical element sequence; based on the equipment state lexical element sequence, constructing a dynamic star map representing the operation state of the whole power plant; inputting the dynamic star map into a space-time fusion backbone network; the space-time fusion backbone network performs iterative processing on the dynamic star map and generates a health degree attenuation trajectory; when the slope of the health degree attenuation trajectory exceeds a preset threshold value, dynamic early warning and system diagnosis are triggered, and a natural language diagnosis report containing a causal reasoning chain is generated; new multi-dimensional operation time sequence data are collected in real time, and an incremental learning algorithm is used to update the encoder and the space-time fusion backbone network online; compared with the prior art, the system can continuously adapt to working condition changes and has high self-optimization capacity.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for thermal power plants, and in particular to a fault early warning and diagnosis method and system for thermal power plants based on the nebula system. Background Technology

[0002] As the cornerstone of national energy supply, the safe, stable, and efficient operation of thermal power plants is of paramount importance. With the deepening of the concepts of "Industry 4.0" and "intelligent operation and maintenance," fault early warning and diagnosis technologies have become crucial for ensuring the smooth operation of this complex system engineering project. Currently, the technological development in this field mainly follows the following directions, which in turn exposes a series of core problems that urgently need to be solved: 1. Data-driven early warning methods are increasingly becoming mainstream. These methods typically rely on sensors deployed throughout the unit to collect massive amounts of time-series data (such as temperature, pressure, and vibration), and use deep learning models (such as LSTM and CNN) for anomaly detection. However, these methods have significant limitations: First, they are mostly "data silo" analyses, focusing only on data from a single device or a single modality, making it difficult to characterize the complex physical connections and dynamic interactions between equipment such as boilers, turbines, and generators, resulting in one-sided early warnings and an inability to predict system-level cascading failures. Second, these purely data-driven models are like "black boxes," and their predictions sometimes violate basic physical laws (such as the law of conservation of energy and the law of thermodynamics), producing unrealistic "false warnings" that are difficult for maintenance personnel to trust and make decisions based on.

[0003] 2. In the fault diagnosis stage, existing systems generally suffer from a disconnect between "early warning" and "diagnosis". The system may be able to issue an alarm, but locating the root cause of the fault still relies heavily on expert experience. Operation and maintenance engineers need to manually review historical logs, compare equipment images, and cross-analyze data from different systems. This process is time-consuming and labor-intensive, and the accuracy and efficiency of diagnosis are limited by the engineer's personal skill level and state. Although some studies have attempted to introduce knowledge graphs to store expert rules, traditional knowledge graphs are costly to build, difficult to update, and lack flexible causal reasoning and natural language interaction capabilities, making it impossible to automatically generate clear and easy-to-understand diagnostic reports.

[0004] 3. Existing models are generally static and rigid. A model trained under specific operating conditions is difficult to adapt to dynamic changes such as frequent load adjustments in thermal power plants and the natural aging of equipment performance over time. This results in the model performing well in the early stages of operation, but its early warning accuracy will continue to decline over time. It lacks self-learning and self-evolution capabilities and ultimately requires expensive periodic retraining and calibration.

[0005] In summary, the existing technical framework for fault early warning and diagnosis in thermal power plants faces several challenges: it cannot achieve forward-looking and accurate early warning across modes and equipment; it cannot provide automated and interpretable causal diagnosis and tracing after early warning; and it cannot build an intelligent system that can continuously adapt to changes in operating conditions and has self-optimization capabilities. Therefore, developing a new method that can solve the above problems has become an urgent need to promote the intelligent operation and maintenance upgrade of thermal power plants.

[0006] Patent CN120449035A discloses a method and system for intelligent monitoring and early warning of equipment defects in thermal power plants based on a multimodal large model. The method includes: intelligent monitoring and early warning of equipment defects in thermal power plants; real-time collection of equipment operation data, i.e., multimodal data; transmitting the multimodal data to a multimodal Transformer model; constructing a multimodal Transformer model; encoding different types of multimodal data to generate a unified multimodal representation; performing cross-modal data fusion and analysis on the unified multimodal representation to obtain unified fusion features; calculating the anomaly degree of the equipment based on the unified fusion features using contrastive learning and an autoencoder to determine whether the equipment is abnormal; and predicting the future state of the equipment based on the unified fusion features, predicting the probability of potential faults, and providing operation and maintenance suggestions and early warnings. This intelligent monitoring and early warning method for equipment defects in thermal power plants improves monitoring accuracy, reduces false alarm and false negative rates, reduces the probability of sudden faults, and has strong adaptability. However, the system's early warning and diagnosis modules are not strongly coupled, resulting in low efficiency and accuracy in locating the root cause of faults, and it cannot achieve automated, high-precision tracking of the fault source equipment. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a fault early warning and diagnosis method and system for thermal power plants based on the nebula system, so as to promote the intelligent operation and maintenance upgrade of thermal power plants.

[0008] The objective of this invention can be achieved through the following technical solutions: This invention provides a fault early warning and diagnosis method for thermal power plants based on nebula systems, comprising the following steps: Multi-dimensional runtime sequence data of a thermal power plant is collected, and an encoder is used to extract features and lexicalize the multi-dimensional runtime sequence data to obtain a unified equipment status lexical sequence. Based on the equipment status lexical sequence, a dynamic star map representing the operating status of the entire power plant is constructed. The dynamic star map is input into the spatiotemporal fusion backbone network; the spatiotemporal fusion backbone network iteratively processes the dynamic star map and generates a health decay trajectory; when the slope of the health decay trajectory exceeds a preset threshold and fails to pass the multi-source data fusion verification, dynamic early warning and system diagnosis are triggered and a natural language diagnostic report containing a causal inference chain is generated; a digital twin model of the power plant is constructed, and the fault propagation path is dynamically simulated; New multi-dimensional runtime sequence data are collected in real time, and the encoder and spatiotemporal fusion backbone network are updated online using an incremental learning algorithm.

[0009] Furthermore, the encoder includes a time-series encoder, an image encoder, and a text encoder; the multi-dimensional runtime time-series data includes sensor time-series data, equipment monitoring image data, and maintenance log text data; the time-series encoder performs feature extraction and lexicalization on the time-series data, specifically: The time-series encoder, image encoder, and text encoder are used to extract features and lexicalize the sensor time-series data, equipment monitoring image data, and maintenance log text data, respectively, to generate the equipment status lexical sequence. The Informer model is used as the time-series encoder, and its attention window size is set to 96 to 192. The time-series data with more than 1000 data points are feature-encoded. Then, the K-Means clustering algorithm is used on the encoded feature vectors, and the cluster centers with a silhouette coefficient higher than 0.6 are defined as different operating condition lexicals.

[0010] Furthermore, the spatiotemporal fusion backbone network is a physical information spatiotemporal Transformer; during training, the loss function of the physical information spatiotemporal Transformer consists of data loss and physical constraint loss in a ratio of 1:0.5 to 1:2; the physical constraint loss is calculated jointly from the residual obtained after discretization of the heat conduction equation and the residual of the continuity equation in the Navier-Stokes equations; the heat conduction equation is used to constrain temperature to ensure that the predicted temperature satisfies the laws of thermodynamics; the Navier-Stokes equations are used to constrain fluid flow to ensure that the predicted fluid flow satisfies the law of conservation of mass.

[0011] Furthermore, the process of constructing a dynamic star map includes: Each word in the device status word sequence is regarded as a graph node. Based on the prior topology of the physical connection relationship of the device and the dynamic time warping distance of the historical data sequence between nodes, the edge weight between nodes is calculated with a weight ratio of 1:1 to 1:3 to construct a dynamic star map representing the operating status of the entire power plant. The window size of the dynamic time warping distance is set to 24 to 168 data points.

[0012] Furthermore, the process for generating the health decay trajectory specifically includes: After the dynamic star map is input into the spatiotemporal fusion backbone network, the spatiotemporal fusion backbone network performs three alternating iterative processes through its internal spatial attention layer, temporal attention layer, and spatiotemporal fusion layer. In each iteration, the spatial attention layer updates the node features based on the graph attention mechanism and then inputs the result into the temporal attention layer. The temporal attention layer updates the temporal series features of the nodes based on the probabilistic sparse self-attention mechanism and then feeds the result back to the spatial attention layer to update the node features for the next round. Finally, the spatiotemporal fusion backbone network outputs the health decay trajectory of the system and key subsystems over the next 6 to 72 hours.

[0013] Furthermore, the process of dynamic early warning and system diagnosis is as follows: The current operating status of the power plant is determined in real time by a scenario-aware model. The operating status includes load level and equipment operating time. Among them, the load is higher than 85% of the rated load, which is a high load state, and the load is lower than 85% of the rated load, which is a low load state. The equipment operating time is more than 1,000 hours, which is a long-term operating state, and less than 1,000 hours, which is a short-term operating state. A meta-learner based on a model-independent meta-learning framework is invoked, and the meta-learner dynamically generates an early warning threshold that is adaptive to the running state based on the running state. The health decay trajectory is compared with the dynamically generated warning threshold. When the health trajectory is lower than the warning threshold for three consecutive time points, a warning is triggered.

[0014] When the slope of the health decay trajectory exceeds a preset threshold, the system diagnosis detects an anomaly in the system; the system anomaly is matched with a pre-built global causal knowledge graph by subgraph matching, and the global causal knowledge graph is queried to obtain the cross-power plant causal evidence chain of the fault; and a natural language diagnostic report containing the causal inference chain is generated using a large language model that has been fine-tuned based on thermal power field operation and maintenance manuals and fault reports.

[0015] Furthermore, the steps for constructing a global causal knowledge graph are as follows: A federated learning architecture is established among multiple thermal power plants; each local power plant uses an attention-weight-based causal discovery algorithm to mine causal relationships between variables from local data and generate local causal graphs; the federated server uses a graph structure fusion algorithm to aggregate all local causal graphs and generate a global causal knowledge graph. The dynamic simulation of the fault propagation path specifically includes: By combining the global causal knowledge graph and the power plant digital twin model, the dynamic process of the fault root cause propagating to related equipment along the causal path is simulated and visualized in real time on the power plant digital twin model, with the extrapolation time range being 1 to 24 hours in the future.

[0016] Furthermore, the multi-source data fusion verification process specifically includes: Time-series data, image data, and text data targeting the same observation target are collected from three independent data sources at the thermal power plant. Each data source is processed using an independent encoder and a separate dynamic early warning process, resulting in three independent processing results. The DS evidence theory is then used to fuse and verify these results into a final judgment result that includes an anomaly confidence level. When the anomaly confidence level is higher than 0.75, the dynamic early warning and system diagnosis are triggered, generating a natural language diagnostic report containing a causal inference chain.

[0017] Furthermore, the incremental learning algorithm is an elastic weight consolidation algorithm, and the Fisher information threshold for the model parameters is set to 0.1 to 0.3 during each online update; the Fisher information is used to measure the impact of the model parameters on the model's predictive performance.

[0018] The present invention also provides a fault early warning and diagnosis system for thermal power plants based on the nebula system, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of any of the methods described above.

[0019] Compared with the prior art, the present invention has the following advantages: (1) This invention uses a time encoder, an image encoder and a text encoder to extract features and lexicalize multi-dimensional runtime time data: sensor time data, equipment monitoring image data and operation and maintenance log text data, to obtain a unified equipment status lexical sequence that represents the dynamic star map of the entire power plant's operating status. This achieves a unified representation of multimodal data and depicts the complex physical connections and dynamic interactions between equipment such as boilers, steam turbines and generators. It also deeply integrates technologies such as unified representation of multimodal data, dynamic equipment star map and physical information spatiotemporal Transformer, avoiding the problem of data silos and improving the ability to predict and diagnose faults.

[0020] (2) By constructing a dynamic star map, this invention realizes the dynamic graph model and spatiotemporal fusion iteration, generating a system health decay trajectory that conforms to physical laws, significantly extending the warning window and improving the accuracy of the warning; at the same time, by using knowledge graphs and domain large language models, the diagnosis is upgraded from manual investigation to automatic tracing and reasoning, and can automatically generate natural language diagnostic reports containing causal chains and easy to understand; and can accurately search for fault-related equipment and sub-equipment based on the graph structure and graph neural network of the dynamic star map, greatly improving diagnostic efficiency, diagnostic accuracy and interpretability.

[0021] (3) By constructing a context-aware meta-learning model and an online incremental learning model, this invention enables the encoder and spatiotemporal fusion backbone network to adapt to equipment aging and changes in operating conditions, thereby achieving continuous self-optimization. By designing a multi-source data fusion verification mechanism, a security firewall for decision-making is constructed to ensure that alarms are triggered only when there is sufficient evidence, thereby achieving the accuracy and reliability of early warning and comprehensively improving the practicality of the system in complex industrial environments. Attached Figure Description

[0022] Figure 1 A flowchart of a fault early warning and diagnosis method for thermal power plants based on a nebula system, provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] Example 1 like Figure 1 As shown, this embodiment of the invention provides a fault early warning and diagnosis method for thermal power plants based on a nebula system, including the following steps: S1: Collect multi-dimensional runtime sequence data of thermal power plants, use an encoder to extract features and lexicalize the multi-dimensional runtime sequence data to obtain a unified equipment status lexical sequence; based on the equipment status lexical sequence, construct a dynamic star map representing the operating status of the entire power plant; S101: Collect time-series sensor data, equipment monitoring image data, and operation and maintenance log text data from thermal power plants; use a time-series encoder, image encoder, and text encoder to extract features and lexicalize the time-series data, image data, and text data respectively, generating a unified equipment status lexical sequence; when constructing the equipment status lexical sequence, the operating condition lexicals from the time-series encoder, the visual lexicals from the image encoder, and the text lexicals from the text encoder are aligned and spliced ​​according to their corresponding unified timestamps to form a unified, cross-modal sequence; Preferably, a temporal encoder is used to extract features and lexicalize the temporal data, specifically as follows: The Informer model was used as the temporal encoder, with its attention window size set to 96 to 192. Long-sequence temporal data with more than 1000 data points were used for feature encoding. Subsequently, the K-Means clustering algorithm was used on the encoded feature vectors, and the cluster centers with silhouette coefficients higher than 0.6 were defined as different working condition terms.

[0026] S102: Treat each word in the equipment status word sequence as a graph node. Based on the prior topology of the physical connection relationship of the equipment and the dynamic time warping distance of the historical data sequence between nodes, calculate the edge weight between nodes with a weight ratio of 1:1 to 1:3 to construct a dynamic star map representing the operating status of the entire power plant. The window size of the dynamic time warping distance is set to 24 to 168 data points. S2: Input the dynamic star map into the spatiotemporal fusion backbone network; the spatiotemporal fusion backbone network iteratively processes the dynamic star map and generates a health decay trajectory; when the slope of the health decay trajectory exceeds the preset threshold and fails to pass the multi-source data fusion verification, a dynamic early warning and system diagnosis are triggered and a natural language diagnostic report containing a causal inference chain is generated; a digital twin model of the power plant is constructed and the fault propagation path is dynamically simulated. S201: The dynamic star map is input into the spatiotemporal fusion backbone network. The spatiotemporal fusion backbone network performs at least three alternating iterations through its internal spatial attention layer, temporal attention layer, and spatiotemporal fusion module. In a single iteration, the spatial attention layer updates the node features based on the graph attention mechanism and then inputs the result into the temporal attention layer. The temporal attention layer updates the time series features of the nodes based on the ProbSparse self-attention mechanism and then feeds the result back to the spatial attention layer for recalculating the spatial attention weights for the next round. Finally, the network outputs the health decay trajectory of the system and key subsystems over the next 6 to 72 hours. Preferably, the spatiotemporal fusion backbone network is a physical information spatiotemporal Transformer; During training, the loss function L of the physical information spatiotemporal Transformer is composed of the data loss L. data and physical constraint loss L physics They are composed of each other in a ratio of 1:0.5 to 1:2, that is, L = L data +λ* L physics Where λ is the weighting coefficient; physical constraint loss L physics The residuals calculated from the discretized heat conduction equation are obtained together with the residuals from the continuity equation in the Navier-Stokes equations.

[0027] S202: When the slope of the health decay trajectory exceeds the preset threshold, the diagnosis module is triggered; the diagnosis module performs subgraph matching between the abnormal pattern and the pre-built equipment knowledge graph, and uses a large language model that has been fine-tuned by the thermal power field operation and maintenance manual and fault reports to generate a natural language diagnosis report containing causal reasoning chains. Preferably, it also includes a dynamic early warning step: The scenario perception module determines the current operating scenario of the power plant in real time. The operating scenario includes high and low load scenarios and equipment operating time scenarios. Among them, the load is higher than 85% of the rated load, which is a high load scenario, and lower than 85% is a low load scenario. The equipment operating time is more than 1,000 hours, which is a long-term operating scenario, and less than 1,000 hours is a short-term operating scenario. The meta-learner, based on a model-independent meta-learning framework, is invoked. The meta-learner dynamically generates an early warning threshold that is adaptive to the current scenario based on the operating scenario and the current device state characteristics. The health decay trajectory is compared with the dynamically generated warning threshold. When the health trajectory is lower than the warning threshold for three consecutive time points, a warning is triggered.

[0028] Preferably, it also includes the step of constructing a global causal knowledge graph, specifically: A federated learning architecture is established among multiple thermal power plants. Each local power plant uses an attention-weighted causal discovery algorithm to mine causal relationships between variables from local data and generate a local causal graph. The federated server uses a graph structure fusion algorithm to aggregate all local causal graphs and generate a global causal knowledge graph. The diagnostic module queries the global causal knowledge graph to obtain cross-power plant causal evidence chains for faults.

[0029] Preferably, after generating the diagnostic report, a dynamic simulation of the fault propagation path is also performed: By combining a global causal knowledge graph with a power plant digital twin model, the dynamic process of a fault propagating from its source along a causal path to related equipment can be simulated and visualized in real time on the digital twin, with a time range of 1 to 24 hours in the future.

[0030] Preferably, it also includes a multi-source data fusion verification step: Time-series data, image data, and text data for the same observation target were collected from at least three independent data sources in the thermal power plant. Each data source was processed using an independent encoder and early warning diagnosis process. The DS evidence theory was used to fuse and verify the multiple processing results. A diagnostic report is triggered by the diagnostic module only if the anomaly confidence level obtained from the multi-source data fusion verification step is higher than 0.75.

[0031] S3: During system operation, new time-series data, image data, and text data are collected in real time. Incremental learning algorithms are used to update the time-series encoder, image encoder, text encoder, and spatiotemporal fusion backbone network online, enabling the model to adapt to the aging of power plant equipment and changes in operating conditions.

[0032] Preferably, the incremental learning algorithm is an elastic weight consolidation algorithm, and the Fisher information threshold for important parameters of the model is set to 0.1 to 0.3 for each online update.

[0033] In summary, the technical solution of this embodiment enhances fault early warning and diagnosis capabilities through the deep integration of technologies such as unified multimodal data representation, dynamic equipment star map, and physical information spatiotemporal Transformer. It elevates early warning from post-event alarms to precise pre-event prediction. By iteratively integrating dynamic graph models with spatiotemporal data, it generates a system health decay trajectory that conforms to physical laws, significantly extending the early warning window and improving accuracy. Simultaneously, it upgrades diagnosis from manual investigation to automatic tracing and reasoning. Leveraging knowledge graphs and domain-specific large language models, it can automatically generate natural language reports containing causal chains, greatly improving diagnostic efficiency and interpretability. Furthermore, through context-aware meta-learning and online incremental learning, the model can adapt to equipment aging and changes in operating conditions, achieving continuous self-optimization. The multi-source data fusion verification mechanism constructs a "safety firewall" for decision-making, ensuring that alarms are triggered only when evidence is sufficient, thereby comprehensively improving the system's practical value and reliability in complex industrial environments.

[0034] Example 2 This embodiment provides a fault early warning and diagnosis system for thermal power plants based on the Nebula system, including a memory and a processor. The memory stores a computer program, and the processor calls the computer program to execute the steps of the method described in any of the embodiments in 1.

[0035] Definitions: Heat conduction equation: The heat equation is a partial differential equation describing the change of temperature distribution over time. It belongs to the field of thermodynamics and is also known as the thermal equation. This equation is derived from Fourier's law of cooling, and the uniqueness of the solution needs to be determined in conjunction with the boundary conditions of the medium.

[0036] The solution to the equation has the property of smoothing the initial temperature distribution and tending towards thermal equilibrium. Formally, it uses the Laplace operator to represent the spatial derivative and belongs to the parabolic partial differential equation. The thermal equation technically violates special relativity because its solution allows disturbances to propagate instantaneously to all parts of space, but the actual effect is usually negligible. Therefore, a hyperbolic equation is needed to describe a reasonable heat conduction rate.

[0037] This equation is applicable not only to heat conduction and particle diffusion, but also to the Black-Scholes option pricing model in financial mathematics. The Crank-Nicolson method is widely used in numerical solutions.

[0038] The heat conduction equation is a partial differential equation describing the diffusion and transfer of heat in a medium over time. It establishes the relationship between the change of temperature distribution over time and its spatial distribution. Its standard form states that the rate of change of temperature over time is proportional to the second derivative of the spatial temperature distribution (i.e., the curvature of the temperature distribution). This equation is widely used in thermodynamics, engineering, and materials science, and is a fundamental mathematical model for analyzing heat transfer processes.

[0039] Navier-Stokes equations: The Navier-Stokes equations are a set of classical partial differential equations describing the motion of viscous fluids. Based on Newton's second law, these equations express the conservation of fluid momentum, and their core lies in characterizing the dynamic equilibrium between inertia, pressure gradient, viscous forces, and external volume forces during fluid motion. As the fundamental governing equations of fluid mechanics, they have wide applications in aerospace, weather forecasting, shipbuilding, and energy and power engineering. The existence and smoothness of their solutions remain one of the unsolved millennium problems in mathematics.

[0040] Graph attention mechanism: Graph attention is a graph neural network algorithm based on an attention model. It dynamically calculates the association weights between nodes and their neighbors to aggregate differentiated information about node features. This mechanism allows each node to be assigned different attention coefficients based on the importance of its neighbors, thus overcoming the limitations of fixed-weight aggregation in traditional graph convolutional networks. It significantly improves the ability to model complex graph structures and is widely used in fields such as social network analysis, recommender systems, and biochemical molecular structure modeling.

[0041] Probabilistic sparse self-attention mechanism: Probabilistic sparse self-attention is an improved attention algorithm for the Transformer model. It uses probabilistic statistical methods to select key attention weights for each query vector, significantly reducing computational complexity. Based on the sparsity assumption of key vector distribution, this mechanism uses a random measurement strategy to approximate the similarity distribution between the query and all keys, performing refined attention calculations only on significantly relevant key-value pairs. This reduces the computational load from O(n^2) to O(n^2) while maintaining model performance. 2 The time complexity is reduced to O(n log n), making it particularly suitable for long sequence data processing scenarios.

[0042] DS Evidence Theory: Dempster's evidence theory is a mathematical reasoning framework for handling uncertainty and incomplete information. It models and fuses multi-source evidence through basic probability assignment functions, belief functions, and likelihood functions. This theory breaks through the limitation of traditional probability theory, which requires precise allocation of probabilities for single events, by introducing the concept of an "uncertainty interval," which clearly distinguishes between the cognitive states of "uncertainty" and "not knowing." Its core Dempster combination rule enables the effective synthesis of conflicting evidence and is widely used in fields such as information fusion, fault diagnosis, and decision support systems.

[0043] Fisher's information content: Fisher information content is an important statistical indicator for measuring the amount of information contained in observed data regarding the parameters to be estimated. It is defined as the expected value of the variance of the score function or the second moment of the log-likelihood function. This indicator reflects the sensitivity of the probability distribution function to changes in parameters: the higher the Fisher information content, the more comprehensive the parameter information provided by the data, and the smaller the lower bound of the variance of the parameter estimate (Cramér-Rao lower bound). It plays a fundamental role in maximum likelihood estimation, model selection, experimental design, and information geometry, providing a theoretical limit benchmark for the accuracy of parameter estimation.

[0044] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for early warning diagnosis of faults in a thermal power plant based on a nebula system, characterized in that, The method comprises the following steps: Collecting multi-dimensional operation time series data of a thermal power plant, using an encoder to extract features and tokenize the multi-dimensional operation time series data, and obtaining a unified device state token sequence; Based on the device state token sequence, a dynamic star map representing the overall power plant operation state is constructed; The dynamic star map is input into a space-time fusion backbone network; the space-time fusion backbone network iteratively processes the dynamic star map and generates a health degree decay trajectory; when the slope of the health degree decay trajectory exceeds a preset threshold and does not pass the multi-source data fusion verification, a dynamic early warning and system diagnosis are triggered, and a natural language diagnosis report containing a causal reasoning chain is generated; a digital twin model of the power plant is constructed, and a dynamic deduction of the fault propagation path is performed; Real-time collection of new multi-dimensional operation time series data, online updating of the encoder and the space-time fusion backbone network using an incremental learning algorithm.

2. The method according to claim 1, characterized in that, The encoder includes a time series encoder, an image encoder, and a text encoder; the multi-dimensional operation time series data includes sensor time series data, device monitoring image data, and operation and maintenance log text data; the time series encoder extracts features and tokenizes the time series data, specifically: The time series encoder, image encoder, and text encoder are used to extract features and tokenize the sensor time series data, device monitoring image data, and operation and maintenance log text data, respectively, to generate the device state token sequence; the Informer model is used as the time series encoder, with an attention window size of 96 to 192; feature encoding is performed on time series data with more than 1000 data points; then, the K-Means clustering algorithm is used on the encoded feature vectors, and clustering centers with a silhouette coefficient higher than 0.6 are defined as different working condition tokens.

3. A method for early warning diagnosis of faults in a thermal power plant based on the nebula system according to claim 1, characterized in that, The space-time fusion backbone network is a physical information space-time Transformer; during training, the loss function of the physical information space-time Transformer is composed of data loss and physical constraint loss in a ratio of 1:0.5 to 1:2; the physical constraint loss is calculated from the residual error of the discretized heat conduction equation and the continuity equation residual error in the Navier-Stokes equation; the heat conduction equation is used to constrain the temperature, ensuring that the predicted temperature satisfies the laws of thermodynamics; the Navier-Stokes equation is used to constrain fluid flow, ensuring that the predicted fluid flow satisfies the law of conservation of mass.

4. The method according to claim 1, characterized in that, The construction process of the dynamic star map includes: Each token in the device state token sequence is regarded as a graph node, and the edge weight between nodes is calculated based on the prior topology of device physical connection relationship and the dynamic time warping distance between nodes, with a weight ratio of 1:1 to 1:3, to construct a dynamic star map representing the overall power plant operation state; the window size of the dynamic time warping distance is set to 24 to 168 data points.

5. The method according to claim 1, characterized in that, The generation process of the health degree decay trajectory includes: After the dynamic star map is input into the space-time fusion backbone network, the space-time fusion backbone network is subjected to three times of alternating iterative processing through a space attention layer, a time attention layer and a space-time intermingling layer inside the space-time fusion backbone network; in a single iteration processing, the space attention layer updates node features based on a graph attention mechanism, and then inputs the result into the time attention layer; the time attention layer updates time sequence features of the node based on a probabilistic sparse self-attention mechanism, and then feeds back the result to the space attention layer for updating node features in the next round; finally, the space-time fusion backbone network outputs a health degree attenuation trajectory of the system and key subsystems within 6 to 72 hours in the future.

6. The method according to claim 1, wherein, The process of the dynamic early warning and system diagnosis is specifically as follows: The current operation state of the power plant is determined in real time through a scenario perception model, and the operation state includes load level and equipment operation time length; wherein, the load higher than 85% of the rated load is a high load state, and the load lower than 85% is a low load state; the equipment operation time length exceeding 1000 hours is a long-term operation state, and the equipment operation time length lower than 1000 hours is a short-term operation state; A meta-learner based on a model-independent meta-learning framework is called, and the meta-learner dynamically generates a warning threshold adaptive to the operation state according to the operation state; The health degree attenuation trajectory is compared with the dynamically generated warning threshold, and when the health degree trajectory is lower than the warning threshold at three consecutive time points, the early warning is triggered; When the slope of the health degree attenuation trajectory exceeds a preset threshold, the system diagnosis captures that the system is abnormal; the system abnormality is subjected to subgraph matching with a pre-constructed global causal knowledge graph, the global causal knowledge graph is queried to obtain a cross-power-plant causal evidence chain of the fault; and a large language model fine-tuned through a thermal power field operation and maintenance manual and a fault report is used to generate a natural language diagnosis report containing a causal reasoning chain.

7. The method according to claim 6, characterized in that, The step of constructing the global causal knowledge graph is specifically as follows: A federated learning architecture is established among multiple thermal power plants; a causal discovery algorithm based on attention weights is used by each local power plant to mine causal relationships between variables from local data to generate a local causal graph; A global causal knowledge graph is generated by a graph structure fusion algorithm on a federated server side by aggregating all local causal graphs; The dynamic deduction of the fault propagation path specifically includes: In combination with the global causal knowledge graph and a power plant digital twin model, a dynamic process in which a fault root propagates along a causal path to related equipment is simulated and visualized in real time on the power plant digital twin model, and the deduction time range is 1 to 24 hours in the future.

8. The method according to claim 1, wherein, The process of the multi-source data fusion verification specifically includes: Time series data, image data and text data for the same observation target are collected from three independent data sources of a thermal power plant, and are processed using independent encoders and independent dynamic early warning processes to obtain three independent processing results, and the processing results are fused and verified using D-S evidence theory, which is used to fuse the three independent processing results into a final judgment result containing abnormal confidence; when the abnormal confidence is higher than 0.75, the dynamic early warning and system diagnosis are triggered, and a natural language diagnosis report containing a causal reasoning chain is generated.

9. The method according to claim 1, wherein, The incremental learning algorithm is a flexible weight consolidation algorithm, and the Fisher information threshold of the model parameters is set to 0.1-0.3 during each online update; the Fisher information is used to measure the influence of the model parameters on the model prediction performance.

10. A star cloud system-based fault early warning diagnosis system for a thermal power plant, characterized in that, The computer program is stored in the memory and executed by the processor to perform the steps of the method according to any one of claims 1-9.

Images