Fault early warning system and method for excitation system of power generation power assembly

Abstract

This invention provides a fault early warning system and method for the excitation system of a power generation assembly, relating to the field of fault early warning technology. The method includes: constructing a directed fault propagation topology; decomposing excitation fault cases into multiple fault feature sample sets of multiple system components, and pre-training multiple fault identification models; mapping and embedding multiple component topology nodes, performing fault propagation linkage reasoning correction, and constructing a dynamically correlated fault identification topology; sensing real-time monitoring data of multiple multi-physics parameters, performing fault linkage state identification, and outputting a multi-node collaborative fault feature vector; predicting fault evolution trends, outputting potential fault time windows, and performing graded early warning. This invention solves the technical problem that existing power generation assembly excitation system fault prediction relies on models based on simple statistical analysis, which cannot handle complex multi-dimensional data and fault evolution trends, resulting in the inability to take effective intervention measures before faults occur.

Description

Technical Field

[0001] This invention relates to the field of fault early warning technology, and in particular to a fault early warning system and method for the excitation system of a power generation assembly. Background Technology

[0002] With the increasing demand for power generation systems in critical scenarios such as data centers, especially in the use of megawatt-level power generation systems, the need for efficient and reliable power supply systems has become particularly urgent. In order to ensure that the power generation system can cope with dynamic changes and sudden operating conditions, especially in complex fault scenarios involving the excitation system, traditional fault diagnosis methods face many challenges.

[0003] Traditional fault diagnosis often relies on static models. Even if some methods introduce some dynamic adjustments, they often lack the ability to effectively learn from multi-time series data and cannot make accurate fault predictions and adjustments based on real-time monitoring data. This limitation is particularly evident in the fault diagnosis of excitation systems. Faults in excitation systems can seriously affect the overall stability and power quality of the power generation system. Traditional methods are unable to quickly identify and respond to such complex fault modes, especially under conditions of large load fluctuations and operating disturbances. Traditional static control methods cannot effectively handle the dynamic adjustment needs of excitation systems, resulting in a lag in fault diagnosis and early warning processes and affecting the system's emergency response capabilities.

[0004] If a fault in the excitation system is not detected and effectively addressed in its early stages, it may trigger a chain reaction of faults, affecting the stability of the entire power generation system and leading to a shutdown of the power generation system. Especially in critical application scenarios, a fault in the excitation system can cause unpredictable business interruptions or safety risks, resulting in a wider range of shutdowns or accidents.

[0005] It should be noted that the information disclosed in this background section is intended only to enhance the understanding of the overall background of the present invention, and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a fault early warning system and method for the excitation system of a power generation assembly, which solves the technical problem that the fault prediction of the excitation system of the power generation assembly in the prior art relies on a model based on simple statistical analysis, which cannot handle complex multidimensional data and fault evolution trends, resulting in the inability to take effective measures to intervene before the fault occurs.

[0007] The specific technical solution is as follows:

[0008] In a first aspect, the present invention provides a fault early warning system for the excitation system of a power generation assembly, the fault early warning system for the excitation system of a power generation assembly comprising:

[0009] The system comprises the following modules: a directed topology construction module, used to retrieve the electrical schematic diagram, energy flow path, and excitation fault cases of the excitation system based on the equipment identification code of the power generation assembly, and construct a directed fault propagation topology; a model pre-training module, used to decompose the excitation fault cases into multiple fault feature sample sets of multiple system components in the excitation system, and pre-train multiple fault identification models; a linkage inference correction module, used to map and embed the multiple fault identification models into multiple component topology nodes of the directed fault propagation topology, perform fault propagation linkage inference correction, and construct a dynamically associated fault identification topology; a linkage state identification module, used to sense real-time monitoring data of multiple multi-physics field parameters of the multiple system components, input the dynamically associated fault identification topology to perform fault linkage state identification, and output a multi-node collaborative fault feature vector; and a graded early warning module, used to predict the fault evolution trend based on the multi-node collaborative fault feature vector, output a potential fault time window, and perform graded early warning for the excitation system.

[0010] In one implementation, the directed topology building module is used to perform the following steps:

[0011] Based on the component composition of the excitation system, multiple component topology nodes corresponding to multiple system components are set; after initializing the directed connections between the multiple component topology nodes according to the electrical schematic diagram, the connection relationship is dynamically calibrated in conjunction with the energy flow path to obtain the physical connection directed topology; based on the excitation fault cases, the fault propagation rules are sorted out, and the fault propagation probability weight of the physical connection directed topology is set to obtain the fault propagation directed topology.

[0012] In one implementation, the directed topology building module is used to perform the following steps:

[0013] Based on the physical connection directed topology, the reachability of fault propagation paths is sorted out, and multiple sets of downstream fault propagation nodes of the multiple component topology nodes are output; the excitation fault cases are decomposed in time series to obtain multiple component abnormal time series; after aligning the multiple component abnormal time series, the trigger frequency is statistically analyzed according to the multiple sets of downstream fault propagation nodes to obtain multiple sets of upstream and downstream trigger correlation; after superimposing the multiple sets of upstream and downstream trigger correlation onto the physical connection directed topology, a time decay factor is introduced to perform fault downstream propagation path probability normalization calibration, and the fault propagation directed topology is output.

[0014] In one implementation, the linked reasoning correction module is used to perform the following operation steps:

[0015] Based on the multiple sets of upstream and downstream trigger correlations, the abnormal time series of the multiple components are screened and reconstructed based on causal correlation to obtain multiple cross-device fault evolution time series. Based on the multiple cross-device fault evolution time series, multiple original cross-device fault data are extracted from the excitation fault case, and time series feature quantification is performed to output multiple cross-device fault evolution data. Multiple cross-device upstream and downstream propagation path features of the multiple component topology nodes are extracted from the fault propagation directed topology. A cross-device fault probability collaborative inference model is constructed based on a graph neural network, with the fault propagation directed topology as its skeleton. Using the multiple cross-device upstream and downstream propagation path features as neighborhood correlation constraints, the inference parameters of the cross-device fault probability collaborative inference model are optimized based on the multiple cross-device fault evolution data to output the dynamic correlation fault identification topology.

[0016] In one implementation, the cross-device upstream and downstream propagation path features include propagation direction, path length, and fault propagation probability weight.

[0017] In one implementation, the cross-device fault evolution data includes a fault type time-varying sequence, a fault probability time-varying sequence, and a multi-physics parameter time-varying sequence.

[0018] In one implementation, the linked reasoning correction module is used to perform the following operation steps:

[0019] The features of the multiple cross-device upstream and downstream propagation paths are quantified into multiple neighborhood association constraint vectors; multiple fault probability time-varying sequences are used as supervision labels, and multiple fault type time-varying sequences and multiple multi-physics parameter time-varying sequences are used as joint input features to construct multiple training sample pairs; the multiple training sample pairs are used as model training data, and the multiple neighborhood association constraint vectors are used as physical prior constraints. The gradient descent algorithm is used to iteratively optimize the graph convolution kernel weights and fault propagation probability parameters of the cross-device fault probability collaborative inference model until the objective function converges, and then the dynamic associated fault identification topology is output. The objective function includes a main loss function and a regularization loss function.

[0020] In one embodiment, the linkage status recognition module is used to perform the following operation steps:

[0021] The real-time monitoring data of the multiple multiphysics parameters are input into the multiple fault identification models in the dynamic correlation fault identification topology for local fault diagnosis, outputting multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels. Fault-sensitive features are extracted from the real-time monitoring data of the multiple multiphysics parameters to obtain multiple multiphysics anomaly feature vectors. Using the multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels, cross-node fault probability collaborative calibration is performed in the dynamic correlation fault identification topology to obtain multiple node-level collaborative fault features. The topology concatenates the multiple node-level collaborative fault features to output the multi-node collaborative fault feature vector.

[0022] In one implementation, the model pre-training module is used to perform the following steps:

[0023] The first fault feature sample set is structured and analyzed to obtain multiple component-level fault feature sample tuples, wherein each component-level fault feature sample tuple includes the sample fault type, the sample time-domain fault parameter sequence, and the sample fault severity level. Based on the time-varying characteristics of the first component fault, a first set of identification model architectures is matched and called from the model architecture selection library. Using the multiple component-level fault feature sample tuples as input variables, the parameters of the first set of identification model architectures are iteratively optimized and trained to output a first set of fault identification channels. The first set of adaptation correlation of the first set of identification model architectures and the first set of identification accuracy of the first set of fault identification channels are weighted to obtain a first set of identification performance confidence. The first set of identification performance confidence is used to perform weighted, optimal, and parallel connection of the first set of fault identification channels to complete the construction of the first fault identification model.

[0024] Secondly, this invention provides a fault early warning method for the excitation system of a power generation assembly. The method includes: retrieving the electrical schematic diagram, energy flow path, and excitation fault cases of the excitation system based on the equipment identification code of the power generation assembly, and constructing a directed fault propagation topology; decomposing the excitation fault cases into multiple fault feature sample sets of multiple system components in the excitation system, and pre-training multiple fault identification models; mapping and embedding the multiple fault identification models into multiple component topology nodes of the directed fault propagation topology, performing fault propagation linkage reasoning correction, and constructing a dynamically associated fault identification topology; sensing real-time monitoring data of multiple multi-physics field parameters of the multiple system components, inputting them into the dynamically associated fault identification topology for fault linkage state identification, and outputting a multi-node collaborative fault feature vector; predicting the fault evolution trend based on the multi-node collaborative fault feature vector, outputting a potential fault time window, and performing a graded early warning for the excitation system.

[0025] In one implementation, the electrical schematic diagram, energy flow path, and excitation fault cases of the excitation system are retrieved based on the equipment identification code of the power generation system, and a directed topology for fault propagation is constructed, including:

[0026] Based on the component composition of the excitation system, multiple component topology nodes corresponding to multiple system components are set; after initializing the directed connections between the multiple component topology nodes according to the electrical schematic diagram, the connection relationship is dynamically calibrated in conjunction with the energy flow path to obtain the physical connection directed topology; based on the excitation fault cases, the fault propagation rules are sorted out, and the fault propagation probability weight of the physical connection directed topology is set to obtain the fault propagation directed topology.

[0027] In one implementation, fault propagation rules are sorted out based on the excitation fault cases, and fault propagation probability weights are set for the directed topology of the physical connection to obtain the directed topology of fault propagation, including:

[0028] Based on the physical connection directed topology, the reachability of fault propagation paths is sorted out, and multiple sets of downstream fault propagation nodes of the multiple component topology nodes are output; the excitation fault cases are decomposed in time series to obtain multiple component abnormal time series; after aligning the multiple component abnormal time series, the trigger frequency is statistically analyzed according to the multiple sets of downstream fault propagation nodes to obtain multiple sets of upstream and downstream trigger correlation; after superimposing the multiple sets of upstream and downstream trigger correlation onto the physical connection directed topology, a time decay factor is introduced to perform fault downstream propagation path probability normalization calibration, and the fault propagation directed topology is output.

[0029] In one implementation, the plurality of fault identification models are mapped and embedded into multiple component topology nodes of the fault propagation directed topology, and fault propagation linkage reasoning correction is performed to construct a dynamically associated fault identification topology, including:

[0030] Based on the multiple sets of upstream and downstream trigger correlations, the abnormal time series of the multiple components are screened and reconstructed based on causal correlation to obtain multiple cross-device fault evolution time series. Based on the multiple cross-device fault evolution time series, multiple original cross-device fault data are extracted from the excitation fault case, and time series feature quantification is performed to output multiple cross-device fault evolution data. Multiple cross-device upstream and downstream propagation path features of the multiple component topology nodes are extracted from the fault propagation directed topology. A cross-device fault probability collaborative inference model is constructed based on a graph neural network, with the fault propagation directed topology as its skeleton. Using the multiple cross-device upstream and downstream propagation path features as neighborhood correlation constraints, the inference parameters of the cross-device fault probability collaborative inference model are optimized based on the multiple cross-device fault evolution data to output the dynamic correlation fault identification topology.

[0031] In one implementation, the cross-device upstream and downstream propagation path features include propagation direction, path length, and fault propagation probability weight.

[0032] In one implementation, the cross-device fault evolution data includes a fault type time-varying sequence, a fault probability time-varying sequence, and a multi-physics parameter time-varying sequence.

[0033] In one implementation, using the upstream and downstream propagation path features of the multiple cross-device systems as neighborhood association constraints, the inference parameters of the cross-device fault probability collaborative inference model are optimized based on the fault evolution data of the multiple cross-device systems, and the dynamic associated fault identification topology is output, including:

[0034] The features of the multiple cross-device upstream and downstream propagation paths are quantified into multiple neighborhood association constraint vectors; multiple fault probability time-varying sequences are used as supervision labels, and multiple fault type time-varying sequences and multiple multi-physics parameter time-varying sequences are used as joint input features to construct multiple training sample pairs; the multiple training sample pairs are used as model training data, and the multiple neighborhood association constraint vectors are used as physical prior constraints. The gradient descent algorithm is used to iteratively optimize the graph convolution kernel weights and fault propagation probability parameters of the cross-device fault probability collaborative inference model until the objective function converges, and then the dynamic associated fault identification topology is output. The objective function includes a main loss function and a regularization loss function.

[0035] In one implementation, real-time monitoring data of multiple multiphysics parameters of the multiple system components are sensed, input into the dynamically correlated fault identification topology for fault linkage state identification, and output a multi-node collaborative fault feature vector, including:

[0036] The real-time monitoring data of the multiple multiphysics parameters are input into the multiple fault identification models in the dynamic correlation fault identification topology for local fault diagnosis, outputting multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels. Fault-sensitive features are extracted from the real-time monitoring data of the multiple multiphysics parameters to obtain multiple multiphysics anomaly feature vectors. Using the multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels, cross-node fault probability collaborative calibration is performed in the dynamic correlation fault identification topology to obtain multiple node-level collaborative fault features. The topology concatenates the multiple node-level collaborative fault features to output the multi-node collaborative fault feature vector.

[0037] In one implementation, the excitation fault case is decomposed into multiple fault feature sample sets of multiple system components in the excitation system, and multiple fault identification models are pre-trained, including:

[0038] The first fault feature sample set is structured and analyzed to obtain multiple component-level fault feature sample tuples, wherein each component-level fault feature sample tuple includes the sample fault type, the sample time-domain fault parameter sequence, and the sample fault severity level. Based on the time-varying characteristics of the first component fault, a first set of identification model architectures is matched and called from the model architecture selection library. Using the multiple component-level fault feature sample tuples as input variables, the parameters of the first set of identification model architectures are iteratively optimized and trained to output a first set of fault identification channels. The first set of adaptation correlation of the first set of identification model architectures and the first set of identification accuracy of the first set of fault identification channels are weighted to obtain a first set of identification performance confidence. The first set of identification performance confidence is used to perform weighted, optimal, and parallel connection of the first set of fault identification channels to complete the construction of the first fault identification model.

[0039] Beneficial effects of the embodiments of the present invention:

[0040] By retrieving the electrical schematic diagram, energy flow path, and historical fault cases of the excitation system, a directed fault propagation topology is constructed. This topology accurately reflects the relationships between components of the generator powertrain's excitation system and the fault propagation path. This structure allows for a more comprehensive understanding of the origin, propagation, and impact of faults within the system, thereby improving the accuracy and timeliness of fault diagnosis. Furthermore, by constructing a dynamically correlated fault identification topology, the fault characteristics of multiple components are linked and corrected through inference. This enables multi-dimensional fusion of multi-physics parameters, real-time identification of fault states, and the acquisition of multi-node-level collaborative fault characteristics. This method comprehensively considers changes in multiple physical fields, including electrical, mechanical, and thermal factors, improving the response to multi-source interference and... The system exhibits strong fault tolerance to load fluctuations. Based on multi-node collaborative fault feature vectors, it predicts the evolution trend of faults and accurately forecasts the occurrence window of potential faults. This not only identifies potential faults in advance and prevents sudden faults, but also provides sufficient early warning time for subsequent fault recovery. By predicting the evolution trend of faults, it can provide graded early warnings based on different fault severity levels. Combined with adaptive adjustment strategies, it optimizes the excitation system, improving the system's stability and operational capabilities under significant load changes. The combination of graded early warnings and adaptive adjustment allows the system to flexibly adjust according to actual operating conditions, which not only improves fault response speed but also effectively avoids over-adjustment or resource waste.

[0041] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A schematic diagram of the fault early warning system for the excitation system of the power generation assembly provided by the present invention is shown.

[0044] Figure 2 A schematic flowchart of the fault early warning method for the excitation system of a power generation assembly provided by the present invention is shown.

[0045] Figure 3 The diagram illustrates the process of constructing a directed topology for fault propagation in the fault early warning method for the excitation system of a power generation assembly provided by the present invention.

[0046] Figure labeling: Directed topology construction module 10, model pre-training module 20, linkage inference correction module 30, linkage state recognition module 40, hierarchical early warning module 50. Detailed Implementation

[0047] To facilitate understanding of the present invention, a more complete description of the invention will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein; rather, these embodiments are provided so that the disclosure of the invention will be more thorough and complete.

[0048] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0049] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0050] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0051] The present invention provides a fault early warning system and method for the excitation system of a power generation assembly, which solves the technical problem that the fault prediction of the excitation system of the power generation assembly in the prior art relies on a model based on simple statistical analysis, which cannot handle complex multidimensional data and fault evolution trends, resulting in the inability to take effective measures to intervene before the fault occurs.

[0052] Example 1, see Figure 1 The present invention provides a fault early warning system for the excitation system of a power generation assembly, comprising:

[0053] The directed topology construction module 10 is used to retrieve the electrical schematic diagram, energy flow path, and excitation fault cases of the excitation system based on the equipment identification code of the power generation assembly, and construct a directed topology for fault propagation. The model pre-training module 20 is used to decompose the excitation fault cases into multiple fault feature sample sets of multiple system components in the excitation system, and to pre-train multiple fault identification models. The linkage reasoning correction module 30 is used to map and embed the multiple fault identification models into multiple component topology nodes of the directed topology for fault propagation, perform fault propagation linkage reasoning correction, and construct a dynamically associated fault identification topology. The linkage state identification module 40 is used to sense real-time monitoring data of multiple multi-physics field parameters of the multiple system components, input the dynamically associated fault identification topology to perform fault linkage state identification, and output a multi-node collaborative fault feature vector. The hierarchical early warning module 50 is used to predict the fault evolution trend based on the multi-node collaborative fault feature vector, output a potential fault time window, and perform hierarchical early warning for the excitation system.

[0054] In one implementation, the directed topology construction module 10 is used to perform the following operation steps:

[0055] Based on the component composition of the excitation system, multiple component topology nodes corresponding to multiple system components are set; after initializing the directed connections between the multiple component topology nodes according to the electrical schematic diagram, the connection relationship is dynamically calibrated in conjunction with the energy flow path to obtain the physical connection directed topology; based on the excitation fault cases, the fault propagation rules are sorted out, and the fault propagation probability weight of the physical connection directed topology is set to obtain the fault propagation directed topology.

[0056] In one implementation, the directed topology construction module 10 is used to perform the following operation steps:

[0057] Based on the physical connection directed topology, the reachability of fault propagation paths is sorted out, and multiple sets of downstream fault propagation nodes of the multiple component topology nodes are output; the excitation fault cases are decomposed in time series to obtain multiple component abnormal time series; after aligning the multiple component abnormal time series, the trigger frequency is statistically analyzed according to the multiple sets of downstream fault propagation nodes to obtain multiple sets of upstream and downstream trigger correlation; after superimposing the multiple sets of upstream and downstream trigger correlation onto the physical connection directed topology, a time decay factor is introduced to perform fault downstream propagation path probability normalization calibration, and the fault propagation directed topology is output.

[0058] In one implementation, the linked reasoning correction module 30 is used to perform the following operation steps:

[0059] Based on the multiple sets of upstream and downstream trigger correlations, the abnormal time series of the multiple components are screened and reconstructed based on causal correlation to obtain multiple cross-device fault evolution time series. Based on the multiple cross-device fault evolution time series, multiple original cross-device fault data are extracted from the excitation fault case, and time series feature quantification is performed to output multiple cross-device fault evolution data. Multiple cross-device upstream and downstream propagation path features of the multiple component topology nodes are extracted from the fault propagation directed topology. A cross-device fault probability collaborative inference model is constructed based on a graph neural network, with the fault propagation directed topology as its skeleton. Using the multiple cross-device upstream and downstream propagation path features as neighborhood correlation constraints, the inference parameters of the cross-device fault probability collaborative inference model are optimized based on the multiple cross-device fault evolution data to output the dynamic correlation fault identification topology.

[0060] In one implementation, the cross-device upstream and downstream propagation path features include propagation direction, path length, and fault propagation probability weight.

[0061] In one implementation, the cross-device fault evolution data includes a time-varying sequence of fault types, a time-varying sequence of fault probabilities, and a time-varying sequence of multiphysics parameters.

[0062] In one implementation, the linked reasoning correction module 30 is used to perform the following operation steps:

[0063] The features of the multiple cross-device upstream and downstream propagation paths are quantified into multiple neighborhood association constraint vectors; multiple fault probability time-varying sequences are used as supervision labels, and multiple fault type time-varying sequences and multiple multi-physics parameter time-varying sequences are used as joint input features to construct multiple training sample pairs; the multiple training sample pairs are used as model training data, and the multiple neighborhood association constraint vectors are used as physical prior constraints. The gradient descent algorithm is used to iteratively optimize the graph convolution kernel weights and fault propagation probability parameters of the cross-device fault probability collaborative inference model until the objective function converges, and then the dynamic associated fault identification topology is output. The objective function includes a main loss function and a regularization loss function.

[0064] In one implementation, the linkage state recognition module 40 is used to perform the following operation steps:

[0065] The real-time monitoring data of the multiple multiphysics parameters are input into the multiple fault identification models in the dynamic correlation fault identification topology for local fault diagnosis, outputting multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels. Fault-sensitive features are extracted from the real-time monitoring data of the multiple multiphysics parameters to obtain multiple multiphysics anomaly feature vectors. Using the multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels, cross-node fault probability collaborative calibration is performed in the dynamic correlation fault identification topology to obtain multiple node-level collaborative fault features. The topology concatenates the multiple node-level collaborative fault features to output the multi-node collaborative fault feature vector.

[0066] In one implementation, the model pre-training module 20 is used to perform the following steps:

[0067] The first fault feature sample set is structured and analyzed to obtain multiple component-level fault feature sample tuples, wherein each component-level fault feature sample tuple includes the sample fault type, the sample time-domain fault parameter sequence, and the sample fault severity level. Based on the time-varying characteristics of the first component fault, a first set of identification model architectures is matched and called from the model architecture selection library. Using the multiple component-level fault feature sample tuples as input variables, the parameters of the first set of identification model architectures are iteratively optimized and trained to output a first set of fault identification channels. The first set of adaptation correlation of the first set of identification model architectures and the first set of identification accuracy of the first set of fault identification channels are weighted to obtain a first set of identification performance confidence. The first set of identification performance confidence is used to perform weighted, optimal, and parallel connection of the first set of fault identification channels to complete the construction of the first fault identification model.

[0068] Example 2, see Figure 2 The present invention provides a method for early warning of faults in the excitation system of a power generation assembly, the method comprising:

[0069] Y100: Based on the equipment identification code of the power generation system, retrieve the electrical schematic diagram, energy flow path and excitation fault cases of the excitation system, and construct a directed topology for fault propagation.

[0070] A powertrain is a comprehensive power system integrating power generation, energy storage, transmission, and related control systems. It is commonly used in electric vehicles, hybrid vehicles, and range extenders. In these systems, power sources include electric motors, internal combustion engines (such as range extenders), and batteries. All components work in coordination to provide stable power output. Each powertrain has a unique equipment identification code. This code allows retrieval of related documents, drawings, and fault cases. The equipment identification code ensures that the retrieved excitation system data corresponds to a specific powertrain, thus improving data accuracy. Electrical schematics illustrate the connections between the various electrical components of the powertrain, particularly the electrical connections of the excitation system. These diagrams help understand the working principle of the excitation system and how faults are transmitted through electrical loops. Energy flow path diagrams show the energy transmission route from the powertrain to the load, including the flow of parameters such as current and voltage. This process helps analyze the energy flow path and energy loss points in the excitation system, thereby identifying potential fault sources. By studying historical excitation system failure cases, we can learn about different failure modes and their propagation methods. These cases provide rich reference data for subsequent failure prediction and diagnosis.

[0071] Directed topology for fault propagation refers to establishing a topological structure encompassing all components of an excitation system based on electrical schematics, energy flow paths, and fault cases. In this topology, each component represents a node in the excitation system, and the connections between nodes represent their interactions or fault propagation paths. The process of constructing a directed topology is based on fault propagation rules, graphically representing the dependencies between different components and possible fault propagation paths to form a directed topological structure for fault propagation. This structure clearly demonstrates how a fault in one component affects other components and gradually propagates throughout the entire excitation system.

[0072] Y200: Decompose the excitation fault case into multiple fault feature sample sets of multiple system components in the excitation system, and pre-train multiple fault identification models.

[0073] Excitation failure cases include various types of failure modes, such as electrical short circuits, overheating, and sensor failure. Decomposing these cases into failure feature samples of multiple excitation system components means breaking down a complex failure mode into failure samples of multiple individual components. For example, the failure mode of the entire excitation system can be decomposed into individual failure events such as "exciter failure", "control module failure" and "current sensor failure".

[0074] Fault feature samples for each component are used to train independent fault identification models. These models can employ different machine learning or deep learning algorithms, such as decision trees, support vector machines, and convolutional neural networks. Pre-training allows these models to learn fault patterns from historical fault data, enabling them to identify future faults occurring under similar conditions. Through pre-training, these models can master the typical characteristics and abnormal patterns of various faults, thereby enabling fault detection and prediction in actual operation.

[0075] Y300: The multiple fault identification models are mapped and embedded into multiple component topology nodes of the fault propagation directed topology, and fault propagation linkage reasoning correction is performed to construct a dynamically associated fault identification topology.

[0076] In the constructed directed fault propagation topology, multiple fault identification models serve as the intelligent layer of the topology nodes. Each node is equipped with a corresponding fault identification model. When a component fails, the corresponding model can identify the fault based on its input data and output the fault type and severity. The linked reasoning process refers to inferring not only from the fault diagnosis results of a single component but also from the overall operating state of the excitation system. For example, when a fault is identified in one component, the states of other related components may be affected, requiring linked correction. Reasoning correction involves adjusting the fault prediction and diagnosis results of the excitation system by analyzing the interactions between components and the fault propagation path. This process, based on the directed fault propagation topology, pushes the fault from the fault source component to the affected components, thereby ensuring the overall accuracy of fault identification in the excitation system.

[0077] The dynamic correlation fault identification topology is an excitation system architecture that integrates multiple fault identification models and fault propagation reasoning and correction mechanisms. It can not only identify the independent faults of each component, but also dynamically update the fault status of the excitation system through fault propagation path analysis. As the fault propagates, it dynamically adjusts the topology structure based on real-time data to adapt to different fault scenarios.

[0078] Y400: Real-time monitoring data of multiple multi-physics parameters of multiple system components are sensed, input the dynamic correlation fault identification topology to identify fault linkage status, and output multi-node collaborative fault feature vector.

[0079] Each excitation system component is equipped with multiple sensors, such as Hall current sensors, temperature sensors, and pressure sensors. These sensors collect data from multiple physical domains, including electrical, mechanical, and thermal, based on the different components of the power generation assembly and the requirements of the excitation system. This real-time monitoring data is input into a previously constructed dynamically correlated fault identification topology. Because the components in the power generation assembly are interconnected, a fault in one component can affect the normal operation of others. Therefore, the topology integrates the states of each component and can be dynamically adjusted to generate accurate fault state predictions based on real-time data and fault propagation. Using the input multi-physics data, faults in multiple excitation system components are analyzed and identified through linked state identification. By correlating data from different physical domains, the initial signs of fault occurrence can be identified more accurately. After completing the fault linkage state identification, a multi-node collaborative fault feature vector is output. This feature vector contains information on the fault correlation and interaction between multiple components in the excitation system, providing a comprehensive description of the current fault state.

[0080] Y500: Based on the multi-node collaborative fault feature vector, the fault evolution trend is predicted, the potential fault time window is output, and the excitation system is given a graded early warning.

[0081] Fault evolution trend prediction refers to predicting how faults develop or propagate in the excitation system by analyzing the collaborative fault characteristics collected from each node. Through training on historical and real-time data, the future evolution trend of faults can be predicted. In this process, the multi-node collaborative fault feature vector provides the real-time fault characteristics of each component in the excitation system and the interaction between them. Through this data, it is possible to identify which faults will lead to further deterioration of the excitation system, and thus predict the fault evolution process.

[0082] Based on the predicted fault evolution, a potential fault time window is calculated. This time window refers to the period before a fault occurs and spreads to the entire excitation system. By accurately predicting the fault's occurrence time and impact range, a proactive response can be made, ensuring that the power generation system is effectively addressed before a fault occurs. According to the potential fault time window, a tiered early warning system is implemented for the excitation system. Tiered early warning means generating different levels of warning information based on the fault's severity, impact range, and evolution trend. For example: Level 1 warning indicates the initial stage of a fault, with a slight impact on the excitation system's operation; Level 2 warning indicates the fault has developed, having a certain impact on the excitation system, but still within a controllable range; Level 3 warning indicates the fault has expanded, seriously threatening the stability of the excitation system, requiring immediate action. Through tiered early warning, power generation system operators can clearly understand the severity of the fault and take different emergency measures according to different warning levels, ensuring the safe operation of the power generation system.

[0083] In one implementation, see Figure 3 Based on the equipment identification code of the power generation system, the electrical schematic diagram, energy flow path, and excitation fault cases of the excitation system are retrieved, and a directed topology for fault propagation is constructed, including:

[0084] Y110: Based on the component composition of the excitation system, set the multiple component topology nodes corresponding to the multiple system components; Y120: After initializing the directed connections between the multiple component topology nodes according to the electrical schematic diagram, perform dynamic calibration of the connection relationship in conjunction with the energy flow path to obtain the physical connection directed topology; Y130: Based on the excitation fault case, sort out the fault propagation rules, set the fault propagation probability weight of the physical connection directed topology, and obtain the fault propagation directed topology.

[0085] The excitation system consists of multiple components, such as the generator stator and rotor, exciter, control module, and current sensor. These components perform different functions in the power generation assembly, and their state and interactions directly affect the stability and operating efficiency of the excitation system. In the directed topology graph of fault propagation, each component is considered a node. Each node not only has an independent function but also has close electrical and physical connections with other components. The topology nodes are set by analyzing the excitation system to determine the function, role, and interrelationships of each component.

[0086] The electrical schematic diagram illustrates the electrical connections between various components of the excitation system, such as the transmission paths of current and voltage. In this step, based on the electrical schematic diagram, the connections between multiple component nodes are first initialized. During initialization, directed connections are established between component nodes according to the electrical schematic diagram, forming a preliminary topology. Here, "directed" means that each connection has a definite direction, indicating that energy or signals are transferred from one component to another.

[0087] Energy flow path refers to the route along which energy (such as current and power) flows within the power generation assembly. By analyzing the energy flow path, the direction of energy transmission, the location of energy loss, and the impact on various components can be clearly identified. Dynamic calibration refers to adjusting and optimizing the initially established topological connections based on the actual energy flow conditions. Through dynamic calibration, the actual energy flow state during operation and its dependencies on various components can be reflected more accurately, thereby improving the accuracy of fault propagation analysis.

[0088] Through directed connection initialization and dynamic calibration of energy flow paths, a directed topology of physical connections that reflects the actual physical connections and energy transfer is finally obtained. This topology helps subsequent steps identify fault propagation paths and ensures the accuracy of fault identification and early warning.

[0089] Excitation fault cases include historical excitation system faults. Each case demonstrates the path and propagation method of the fault from one component to another. By analyzing these cases, we can identify different types of fault propagation rules. For example, a certain type of fault first affects the voltage regulator, then the exciter, and finally the generator fails to operate normally. These rules help to build a fault propagation model.

[0090] Fault propagation probability weight refers to the probability of a fault occurring along each propagation path. By analyzing fault cases, a weight is assigned to each connection path, representing the likelihood of a fault propagating along that path. For example, if historical fault data indicates that voltage instability frequently leads to exciter failure, then the propagation probability weight of the connection path between the voltage and the exciter will be higher. The weighting is based on factors such as the frequency of fault occurrence, the speed of fault propagation, and the scope of impact, enabling more accurate prediction of the fault's expansion within the excitation system.

[0091] By combining fault propagation rules and fault propagation probability weights, a directed fault propagation topology is obtained. This topology diagram illustrates the fault propagation paths between various components in the excitation system and the propagation probability of each path. In this topology, the weight of each connection path represents the probability of fault propagation along that path, enabling rapid identification of fault sources and prediction of fault expansion during real-time monitoring.

[0092] In one implementation, fault propagation rules are sorted out based on the excitation fault cases, and fault propagation probability weights are set for the directed topology of the physical connection to obtain the directed topology of fault propagation, including:

[0093] Y131: Based on the directed topology of the physical connection, perform fault propagation path reachability analysis and output multiple sets of downstream fault propagation nodes for the multiple component topology nodes; Y132: Decompose the excitation fault case into time series to obtain multiple component abnormal time series; Y133: After aligning the multiple component abnormal time series, perform trigger frequency statistics based on the multiple sets of downstream fault propagation nodes to obtain multiple sets of upstream and downstream trigger correlation; Y134: After superimposing the multiple sets of upstream and downstream trigger correlation onto the directed topology of the physical connection, introduce a time decay factor to perform fault downstream propagation path probability normalization calibration and output the fault propagation directed topology.

[0094] In a physically connected directed topology, each component node is interconnected with other nodes via connection paths. Reachability analysis determines whether a fault can propagate from one component node to other component nodes, and the path of fault propagation. For each component node, one or more downstream propagation nodes are output. These downstream propagation nodes represent other components that will be affected after the fault propagates from this component. For example, if the exciter fails, it will affect other components such as the control system and sensors; these components are identified as downstream propagation nodes.

[0095] The excitation fault case studies record the operating status and fault type of each component in the excitation system when a fault occurs. Time series decomposition extracts the state changes of each component in these fault cases, forming different abnormal time series. An abnormal time series refers to the abnormal fluctuations in physical quantities such as current, voltage, and temperature of a component over time when a fault occurs. For example, when a component fails, its voltage may experience sudden and drastic fluctuations, or its temperature may rise. These fluctuation time series will reflect the fault characteristics of that component. Through time series decomposition, the data in the fault cases are classified by component, and an abnormal time series is generated for each component. These time series help analyze the timing of the fault occurrence and the sequence of fault propagation.

[0096] Since fault occurrence and propagation is a dynamic process, aligning the anomaly time series of each component in chronological order allows the aligned time series to reflect the synchronicity and propagation patterns of different components during fault occurrence. The purpose of time alignment is to ensure that the fault occurrence times of different components can be accurately compared. For example, when one component fails, the abnormal responses of other components may be delayed. By aligning the time series, these delays and correlations can be accurately captured.

[0097] After alignment, the triggering frequency of each component failure is analyzed. Specifically, the occurrence frequency of each component failure is statistically analyzed, and it is determined which downstream components are more frequently affected by upstream component failures. By calculating the triggering frequency, the triggering correlation degree between each pair of components is obtained, that is, whether the failure of a certain component will lead to the failure of the downstream component. This correlation degree reflects the strength and frequency of failure propagation.

[0098] Based on the calculated trigger correlation, this information is superimposed on the directed topology of physical connections. This means that the weight of each connection path in the topology will reflect the correlation strength of fault propagation between upstream and downstream components. The superposition of these correlations helps to further improve the accuracy of the topology and ensure that the probability of fault propagation can truly reflect the influence relationship between the components.

[0099] Fault propagation involves a time delay. To more accurately reflect the actual situation of fault propagation, a time decay factor is introduced. This means that as the propagation time increases, the propagation probability gradually decreases. Specifically, the propagation of a fault from the source component to downstream components is affected by the time delay; the further the component is, the lower the probability of being affected by the fault. The time decay factor normalizes and calibrates the probability of the fault propagation path, making the probability distribution of the propagation path more consistent with the actual fault propagation process. The final output is a calibrated directed fault propagation topology. This topology contains the probabilistic information of the fault propagation path and considers the time decay effect, enabling it to more accurately reflect the propagation patterns of faults in various components of the excitation system.

[0100] In one implementation, the multiple fault identification models are mapped and embedded into multiple component topology nodes of the fault propagation directed topology, and fault propagation linkage reasoning correction is performed to construct a dynamically associated fault identification topology, including:

[0101] Y310: Based on the multiple sets of upstream and downstream trigger correlations, the abnormal time series of the multiple components are screened and reconstructed according to causal correlation to obtain multiple cross-device fault evolution time series; Y320: Based on the multiple cross-device fault evolution time series, multiple original cross-device fault data are extracted from the excitation fault case, and time series feature quantification is performed to output multiple cross-device fault evolution data; Y330: Multiple cross-device upstream and downstream propagation path features of the multiple component topology nodes are extracted from the fault propagation directed topology; Y340: A cross-device fault probability collaborative reasoning model with the fault propagation directed topology as the skeleton is constructed based on a graph neural network; Y350: Using the multiple cross-device upstream and downstream propagation path features as neighborhood correlation constraints, the inference parameters of the cross-device fault probability collaborative reasoning model are optimized based on the multiple cross-device fault evolution data to output the dynamic correlation fault identification topology.

[0102] Causal correlation refers to whether a fault occurring in one component will have a direct or indirect impact on other components. By understanding the causal relationships between components, we can identify which component anomaly time series have a direct causal correlation. Time series reconstruction refers to reorganizing and synchronizing the anomaly time series of multiple components based on causal relationships. Through reconstruction, we obtain the time series information of fault evolution in different components, showing the occurrence and propagation process of the fault. Through the screening and reconstruction process, we obtain multiple cross-device fault evolution time series, which describe how a fault starts from one component, affects other components, and the fault evolution process of the entire excitation system.

[0103] From the cross-device fault evolution time series, raw cross-device fault data related to the fault are extracted. This data includes the time-series changes of various physical quantities such as current, voltage, and temperature. By analyzing this raw data, the impact of the fault on different components and the specific process of fault propagation can be captured. Time series feature quantization refers to the analysis of the raw time series data to extract meaningful features. These features include parameters such as peak value, volatility, and frequency, reflecting the dynamic behavior of the excitation system when a fault occurs. Through time series feature quantization, complex time series data can be transformed into feature data that is easier to analyze and compare. Based on time series feature quantization, multiple cross-device fault evolution data are finally output. This data provides detailed information on the fault evolution of each component in the excitation system, helping to further analyze fault modes and trends.

[0104] Based on the constructed directed topology of fault propagation, the upstream and downstream propagation path features of each component node are extracted. These features describe the process of fault propagation from one component to another, including the direction of propagation, path length, and propagation probability. Focusing on cross-device fault propagation paths, that is, those propagation paths that affect multiple devices or components, these features can identify which components play a key role in the fault propagation process and the interaction relationships between these components.

[0105] Graph Neural Networks (GNNs) are deep learning models that learn through graph structures, making them particularly suitable for processing data with dependencies between nodes. GNNs can learn the relationships between nodes in a directed fault propagation topology. GNNs treat the fault propagation path as a graph, with nodes representing components and edges representing connections between them. Through GNNs, effective fault reasoning can be performed on the graph structure, capturing complex interactions and fault propagation patterns between components. GNNs can also be used for collaborative reasoning of cross-device fault propagation probabilities; that is, based on fault information from multiple components, the probability of fault propagation between them is jointly inferred to obtain the collaborative fault relationships between components. Through collaborative reasoning, complex multi-point fault propagation can be effectively analyzed, key fault paths can be identified, and fault warning and repair strategies can be optimized.

[0106] The cross-device upstream and downstream propagation path features include the path features of the fault propagation from one component to another in the excitation system, including the propagation direction, path length, and fault propagation probability weight. Neighborhood association constraints refer to using these propagation path features as constraints and applying them to the subsequent model optimization process. These features limit the propagation mode of the fault in the network and ensure that the model follows these propagation characteristics during inference.

[0107] Cross-device fault evolution data includes fault change information among multiple components at different time scales. This data helps analyze how faults propagate across multiple components and evolve at different physical levels. Using this cross-device fault evolution data, inference parameters are optimized based on a graph neural network. The optimization process adjusts the model's parameters to enable it to more accurately predict the fault propagation process. The goal of inference parameter optimization is to update the weights and connection strengths in the model by learning from historical and real-time monitoring data, thereby improving the model's accuracy. The optimized inference model can better simulate the fault propagation dynamics among components in the excitation system.

[0108] Based on the optimized inference model, a dynamic correlation fault identification topology is output. This topology describes the fault propagation paths and their relationships among the components in the entire excitation system, enabling the excitation system to identify and provide early warnings of potential faults in real time. The dynamic topology means that the topology structure can be updated in real time as faults occur and propagate to adapt to new fault scenarios. This flexible topology structure allows the system to efficiently identify different types of faults and accurately predict their impact range.

[0109] In one implementation, the cross-device upstream and downstream propagation path features include propagation direction, path length, and fault propagation probability weight.

[0110] The propagation direction indicates the direction of fault propagation, such as from the generator exciter to the control module, or from the sensor to the actuator. This ensures that the model can identify the path direction when a fault is transmitted from one component to another. The path length represents the number of components the fault propagates through, i.e., the number of excitation system components that the fault passes through from the fault source to the target component. The longer the path length, the greater the propagation delay and the complexity of the impact. This allows the model to better understand the propagation delay of the fault in the network and the possibility of affecting more distant components. The fault propagation probability weight represents the probability of the fault propagating on a specific path. This weight is calculated based on factors such as historical fault data, physical connection strength, and the stability of the excitation system.

[0111] In one implementation, the cross-device fault evolution data includes a time-varying sequence of fault types, a time-varying sequence of fault probabilities, and a time-varying sequence of multiphysics parameters.

[0112] The time-varying sequence of fault types describes the changes in fault type of various components of the excitation system at different points in time when a fault occurs. For example, a component may initially exhibit abnormal voltage, then become overheated, and finally cause the excitation system to shut down, helping the model understand the possible evolution path of the fault. The time-varying sequence of fault probability records how the probability of fault occurrence fluctuates over time. It reflects the speed of fault expansion and the scope of its impact. As time goes on, the probability of fault occurrence may gradually increase, which helps to predict the timing of fault occurrence and the speed of fault expansion. The time-varying sequence of multiphysics parameters includes time-series data from multiple physical domains, such as electrical, mechanical, and thermal, reflecting the changes of various physical quantities over time during fault propagation, helping the model identify the impact of complex faults on different physical levels.

[0113] In one implementation, the upstream and downstream propagation path features of the multiple cross-device systems are used as neighborhood association constraints. Based on the fault evolution data of the multiple cross-device systems, the inference parameters of the cross-device fault probability collaborative inference model are optimized to output the dynamic associated fault identification topology, including:

[0114] Y351: Quantize the features of the multiple cross-device upstream and downstream propagation paths into multiple neighborhood association constraint vectors; Y352: Construct multiple training sample pairs using multiple fault probability time-varying sequences as supervision labels, multiple fault type time-varying sequences and multiple multi-physics parameter time-varying sequences as joint input features; Y353: Use the multiple training sample pairs as model training data, use the multiple neighborhood association constraint vectors as physical prior constraints, and use the gradient descent algorithm to iteratively optimize the graph convolution kernel weights and fault propagation probability parameters of the cross-device fault probability collaborative inference model until the objective function converges, and output the dynamic associated fault identification topology, wherein the objective function includes a main loss function and a regularization loss function.

[0115] The characteristics of cross-device upstream and downstream propagation paths, including propagation direction, path length, and fault propagation probability weights, are quantified into neighborhood association constraint vectors. These vectors represent the fault propagation relationships between components and their degree of influence. Different characteristics of the propagation path are converted into numerical representations, specifically in the following ways: different propagation directions are represented using binary or numerical encoding, such as 1 for forward propagation and -1 for backward propagation; the path length is converted into a numerical value, reflecting the complexity of propagation and the hierarchy of influence; and the propagation probability is standardized within a certain range, representing the propagation probability of each path as a numerical value, such as a floating value between 0 and 1. These neighborhood association constraint vectors define the dependencies between components and guide the model during inference, ensuring that fault propagation conforms to actual physical and causal relationships.

[0116] The time-varying sequence of fault probability serves as a supervision label, representing the probability of fault occurrence in the excitation system at different time points. These labels reflect the occurrence and development trends of faults, and the model will learn how to predict future fault probabilities. The time-varying sequence of fault type represents the change of fault type of each component in the excitation system over time; it is discrete time-series data. For example, a component may exhibit a short circuit at a certain moment, which may later transform into overheating. The time-varying sequence of multiphysics parameters includes time-varying data from multiple physical domains such as electrical, mechanical, and thermal, reflecting the impact of faults on various physical levels of the excitation system. This data provides the model with multi-dimensional perception, enabling it to comprehensively analyze the multifaceted impact of faults on the excitation system. Each time-varying sequence of fault probability is combined with the corresponding time-varying sequence of fault type and time-varying sequence of multiphysics parameters to form a training sample pair. Each training data pair contains input features and corresponding supervision labels for the model to learn and optimize.

[0117] Multiple training sample pairs are fed as input data into the cross-device fault probability collaborative inference model, which is then trained based on this input data. Neighborhood association constraint vectors serve as physical prior constraints, helping the model follow the actual laws of fault propagation during inference, ensuring that the model's predictions conform to physical and causal relationships. Through these constraints, the model can pay more attention to the relevance and reachability of fault propagation paths during fault inference. In the graph neural network, the graph convolution kernel is responsible for propagating information in the graph structure, obtaining the final node representation by calculating the weighted information transfer between nodes. In this step, the weights of the graph convolution kernel are optimized to ensure that it can effectively capture the relationships of fault propagation and perform accurate inference based on input features. During the training of the graph neural network, fault propagation probability parameters are also included in the optimization objective. These parameters describe the probability of fault propagation in the network; optimizing these parameters can improve the accuracy of the model in fault propagation inference.

[0118] The objective function includes the main loss function and the regularization loss function. The main loss function measures the error of the model on the training data, which is the difference between the predicted failure probability and the actual failure occurrence. For example, the mean squared error loss function is used to control the model complexity and prevent overfitting. Regularization techniques include L2 regularization, which maintains the simplicity of the model by penalizing excessively large weight values.

[0119] Using the gradient descent algorithm, the model progressively adjusts the graph convolution kernel weights and fault propagation probability parameters based on the values ​​of the main loss function and the regularization loss function. After each iteration, the model's performance gradually improves until the objective function converges, meaning the model no longer experiences significant performance gains. After optimization, the final dynamically correlated fault identification topology is output. This topology includes the optimized graph convolution kernel weights and fault propagation probability parameters, accurately describing the process of fault propagation from the source component to other components, providing precise evidence for fault diagnosis and early warning.

[0120] In one implementation, real-time monitoring data of multiple multiphysics parameters of the multiple system components are sensed, the dynamically correlated fault identification topology is input for fault linkage state identification, and a multi-node collaborative fault feature vector is output, including:

[0121] Y410: Input the real-time monitoring data of the multiple multiphysics parameters into the multiple fault identification models in the dynamic correlation fault identification topology for local fault diagnosis, and output multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels; Y420: Extract fault-sensitive features from the real-time monitoring data of the multiple multiphysics parameters to obtain multiple multiphysics anomaly feature vectors; Y430: Use the multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels to perform cross-node fault probability collaborative calibration in the dynamic correlation fault identification topology to obtain multiple node-level collaborative fault features; Y440: Topology concatenate the multiple node-level collaborative fault features to output the multi-node collaborative fault feature vector.

[0122] Real-time monitoring data of multiple multiphysics parameters are fed into various fault identification models within a dynamically correlated fault identification topology. Each model corresponds to a component in the excitation system. Local fault diagnosis is performed based on real-time data and the previously established topology. The initial fault probability refers to the probability of a current fault calculated by each fault identification model based on the input multiphysics data, representing the likelihood of a fault occurring in any component of the excitation system. The initial fault type refers to the fault type identified by the model, such as electrical short circuit, mechanical wear, or overheating. The initial fault severity level refers to the severity level of the fault, such as minor, severe, or extreme, based on its impact, to aid in subsequent decision-making.

[0123] Fault-sensitive features refer to key information that reflects whether a component of the excitation system has failed. For example, voltage fluctuations reflect electrical faults, and temperature increases indicate mechanical or thermal excitation system faults. The feature extraction process analyzes real-time monitoring data to identify these sensitive features, such as sudden current changes, frequency fluctuations, and temperature variations, and converts them into quantifiable features. The extracted fault-sensitive features are organized into multiple multiphysics anomaly feature vectors. Each feature vector contains anomaly information of multiple physical parameters related to a specific component of the excitation system. These feature vectors reflect anomalies in different physical domains, which is helpful for subsequent fault diagnosis and prediction.

[0124] This method uses multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels as input. Based on a dynamically correlated fault identification topology, cross-node collaborative calibration of fault probabilities is performed. Cross-node calibration integrates fault information from various components in the excitation system, eliminating potential errors in local diagnosis and ensuring more accurate overall fault diagnosis results. By analyzing the propagation path and time delay of the fault between different components, the overall fault situation can be identified more precisely. After collaborative calibration, multiple node-level collaborative fault features are output, including collaboratively corrected fault probabilities, fault types, severity, anomaly feature vectors, and propagation path information.

[0125] Multiple node-level collaborative fault features are concatenated according to the topology to form a complete multi-node collaborative fault feature vector. This feature vector integrates fault information from multiple components, including the fault probability, type, severity, abnormal characteristics, and fault propagation path of each component, representing the fault state of the entire excitation system. The concatenated feature vector, as the final output, provides a comprehensive view of the faults in the entire excitation system. These features will provide a basis for subsequent fault diagnosis and early warning.

[0126] In one implementation, the excitation fault case is decomposed into multiple fault feature sample sets of multiple system components in the excitation system, and multiple fault identification models are pre-trained, including:

[0127] Y210: Structure the first fault feature sample set to obtain multiple component-level fault feature sample tuples, wherein the component-level fault feature sample tuples include sample fault type, sample time-domain fault parameter sequence, and sample fault severity level; Y220: Match and call the first set of identification model architectures in the model architecture selection library according to the time-varying characteristics of the first component fault; Y230: Use the multiple component-level fault feature sample tuples as input variables to perform iterative optimization training of the parameters of the first set of identification model architectures, and output the first set of fault identification channels; Y240: Weight the first set of adaptation correlation of the first set of identification model architectures and the first set of identification accuracy of the first set of fault identification channels to obtain the first set of identification performance confidence; Y250: Use the first set of identification performance confidence to perform weighted selection and parallel connection of the first set of fault identification channels to complete the construction of the first fault identification model.

[0128] The first fault feature sample set contains relevant data when the first system component fails. Through structured parsing, this data is classified and organized according to component and fault type, so that the data can be analyzed at the component level and organized into multiple component-level fault feature sample tuples, including multiple pieces of information related to the fault.

[0129] The sample fault type indicates the type of fault occurrence, such as short-circuit fault, overload fault, or overheating fault. The fault type of each component is a key factor in fault diagnosis, helping the model understand the nature of the fault. The sample time-domain fault parameter sequence includes the sequence of parameters in each physical domain that change over time when the fault occurs, such as records of changes in current, voltage, temperature, and pressure over time. Time-domain data helps analyze the change patterns of the fault at different time scales. The sample fault severity level indicates the severity of the fault, such as minor, major, or extreme. This feature helps to determine the impact of the fault on the stability of the excitation system and make corresponding decisions.

[0130] When a component fails, its characteristics change over time, reflecting the development and propagation of the fault. Examples include current fluctuations in electrical components during short-circuit faults and temperature changes in temperature sensors during overheating faults. By analyzing the time-varying characteristics of the first component's fault, its specific fault evolution patterns can be identified. These time-varying characteristics help in selecting an appropriate model architecture for fault identification. The model architecture selection library contains various model architectures specifically designed for identifying and analyzing different types of faults. This library includes deep learning model architectures such as neural networks, convolutional neural networks, and long short-term memory networks. Based on the time-varying characteristics of the first component's fault, the most suitable model architecture for the current fault type is selected from the selection library—the first set of identification model architectures. For example, if the fault characteristics exhibit time-series fluctuations, a long short-term memory network is selected for sequence modeling. This ensures that the model can accurately identify and analyze the fault based on its time-varying characteristics.

[0131] Multiple component-level fault feature sample tuples are used as input data for the model. These sample tuples include the fault type, time-domain fault parameter sequence, and fault severity level for each component. The model is trained using this input data. During training, iterative optimization methods, such as gradient descent, are used to continuously adjust the model's parameters, gradually optimizing the model's weights and parameters to more accurately identify fault features. During training, the model continuously adjusts its parameters to minimize the loss function. After each iteration, the model's performance improves until the optimal parameter settings are reached. Through this iterative process, the model continuously learns and improves its ability to identify fault features. After training is complete, the first set of fault identification channels is output.

[0132] The first set of fit correlation indicates the degree to which the model matches the actual situation under a specific fault scenario. Each model architecture has its own adaptability to fault characteristics, and different architectures may perform differently in different fault types or different components. The fit correlation is calculated based on the difference between the model's predicted output and the actual recorded fault data. A higher fit correlation indicates that the model is more suitable for the current fault type. The second set of identification accuracy refers to the model's ability to correctly identify faults, measured by accuracy. Each identification channel predicts the fault probability, type, and severity of the corresponding component based on the input real-time monitoring data. Identification accuracy reflects the model's accuracy in predicting fault type, probability, and severity in actual diagnosis.

[0133] The first set of recognition performance confidence is obtained by weighting the first set of adaptation correlation and the first set of recognition accuracy. It represents the reliability of the model in the current task. If the confidence is high, it means that the model can make a relatively accurate diagnosis under specific fault conditions. If the confidence is low, further optimization or selection of other models is required.

[0134] Weighted parallel selection integrates the outputs of multiple fault identification channels and selects the optimal channel combination based on the performance confidence of each channel. This means that it does not rely solely on the output of a single channel, but combines the diagnostic results of multiple channels. The output of each channel is weighted according to its performance confidence, and the channel with better performance will have a larger weight in the final result. Finally, a comprehensive fault identification result is obtained through weighted integration.

[0135] Based on the weighted selection and parallel connection results, a final first fault identification model is constructed. This model integrates information from different identification channels, enabling more accurate diagnosis of faults in the excitation system. The model's input data consists of real-time monitored multi-physics parameter data, including real-time information collected by various sensors such as electrical, mechanical, and thermal sensors. The model's output includes the initial fault probability, fault type, and fault severity level of multiple components. These outputs help determine which components have failed and assess the extent of their impact.

[0136] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0137] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A fault early warning system for the excitation system of a power generation assembly, characterized in that, The fault early warning system for the power generation and powertrain excitation system includes: The directed topology construction module is used to retrieve the electrical schematic diagram, energy flow path and excitation fault cases of the excitation system based on the equipment identification code of the power generation system, and construct a directed topology for fault propagation. The model pre-training module is used to decompose the excitation fault case into multiple fault feature sample sets of multiple system components in the excitation system, and to pre-train multiple fault identification models. The linkage reasoning correction module is used to map and embed the multiple fault identification models into multiple component topology nodes of the fault propagation directed topology, perform fault propagation linkage reasoning correction, and construct a dynamically associated fault identification topology. The linkage status recognition module is used to sense real-time monitoring data of multiple multi-physics field parameters of the multiple system components, input the dynamic correlation fault recognition topology to perform fault linkage status recognition, and output a multi-node collaborative fault feature vector. The graded early warning module is used to predict the fault evolution trend based on the multi-node collaborative fault feature vector, output potential fault time windows, and perform graded early warning for the excitation system.

2. The fault early warning system for the excitation system of the power generation assembly as described in claim 1, characterized in that, The directed topology construction module is used to perform the following operation steps: Based on the component composition of the excitation system, the multiple component topology nodes corresponding to the multiple system components are set; After initializing the directed connections between the multiple component topology nodes according to the electrical schematic diagram, the connection relationships are dynamically calibrated in conjunction with the energy flow path to obtain the physical connection directed topology. Based on the excitation fault case, the fault propagation rules are sorted out, and the fault propagation probability weights of the physical connection directed topology are set to obtain the fault propagation directed topology.

3. The fault early warning system for the excitation system of the power generation assembly as described in claim 2, characterized in that, The directed topology construction module is used to perform the following operation steps: Based on the physical connection directed topology, the reachability of fault propagation paths is sorted out, and multiple sets of downstream fault propagation nodes of the multiple component topology nodes are output; The excitation fault cases were decomposed into time series to obtain multiple component anomaly time series. After aligning the abnormal time series of the multiple components, the trigger frequency is statistically analyzed based on the multiple sets of downstream fault propagation nodes to obtain multiple sets of upstream and downstream trigger correlation. After superimposing the multiple sets of upstream and downstream trigger correlations onto the physical connection directed topology, a time decay factor is introduced to normalize and calibrate the probability of downstream fault propagation paths, and the fault propagation directed topology is output.

4. The fault early warning system for the excitation system of the power generation assembly as described in claim 3, characterized in that, The linked reasoning correction module is used to perform the following operation steps: Based on the multiple sets of upstream and downstream trigger correlation, the abnormal time series of the multiple components are screened and reconstructed based on causal correlation to obtain multiple cross-device fault evolution time series; Based on the multiple cross-device fault evolution time series, multiple original cross-device fault data are extracted from the excitation fault case, and then time series feature quantization is performed to output multiple cross-device fault evolution data. Extract the cross-device upstream and downstream propagation path features of the multiple component topology nodes from the fault propagation directed topology; A cross-device fault probability collaborative reasoning model is constructed based on graph neural networks, with the fault propagation directed topology as the backbone. Using the upstream and downstream propagation path features of the multiple cross-devices as neighborhood association constraints, the inference parameters of the cross-device fault probability collaborative inference model are optimized based on the fault evolution data of the multiple cross-devices, and the dynamic association fault identification topology is output.

5. The fault early warning system for the excitation system of the power generation assembly as described in claim 4, characterized in that, The cross-device upstream and downstream propagation path features include propagation direction, path length, and fault propagation probability weight.

6. The fault early warning system for the excitation system of the power generation assembly as described in claim 5, characterized in that, The cross-device fault evolution data includes time-varying sequences of fault types, time-varying sequences of fault probabilities, and time-varying sequences of multi-physics parameters.

7. The fault early warning system for the excitation system of the power generation assembly as described in claim 6, characterized in that, The linked reasoning correction module is used to perform the following operation steps: The characteristics of the multiple cross-device upstream and downstream propagation paths are quantified into multiple neighborhood association constraint vectors; Multiple training sample pairs are constructed using multiple time-varying sequences of fault probabilities as supervision labels and multiple time-varying sequences of fault types and multiple time-varying sequences of multi-physics parameters as joint input features. The multiple training sample pairs are used as model training data, and the multiple neighborhood association constraint vectors are used as physical prior constraints. The gradient descent algorithm is used to iteratively optimize the graph convolution kernel weights and fault propagation probability parameters of the cross-device fault probability collaborative reasoning model until the objective function converges. Then, the dynamic associated fault identification topology is output. The objective function includes a main loss function and a regularization loss function.

8. The fault early warning system for the excitation system of the power generation assembly as described in claim 1, characterized in that, The linkage status recognition module is used to perform the following operation steps: The real-time monitoring data of the multiple multiphysics parameters are input into the multiple fault identification models in the dynamic associated fault identification topology for local fault diagnosis, and multiple initial fault probabilities, multiple initial fault types and multiple initial fault severity levels are output. Fault-sensitive features are extracted from the real-time monitoring data of the multiple multiphysics parameters to obtain multiple multiphysics anomaly feature vectors. By employing the multiple multiphysics anomaly feature vectors, multiple initial fault probabilities, multiple initial fault types, and multiple initial fault severity levels, cross-node fault probability collaborative calibration is performed on the dynamic correlation fault identification topology to obtain multiple node-level collaborative fault features. The topology is spliced ​​together to obtain the multi-node collaborative fault feature vector.

9. The fault early warning system for the excitation system of the power generation assembly as described in claim 1, characterized in that, The model pre-training module is used to perform the following operations: The first fault feature sample set is structured and parsed to obtain multiple component-level fault feature sample tuples, wherein the component-level fault feature sample tuples include sample fault type, sample time-domain fault parameter sequence and sample fault severity level; Based on the time-varying characteristics of the first component failure, the first group of identified model architectures is matched and called from the model architecture selection library; Using the multiple component-level fault feature sample tuples as input variables, perform iterative optimization training of the parameters of the first set of identification model architectures, and output the first set of fault identification channels. The first set of recognition performance confidence is obtained by weighting the first set of adaptation correlation of the first set of recognition model architecture and the first set of recognition accuracy of the first set of fault recognition channels. The first fault identification model is constructed by weighted and optimized parallel connection of the first set of fault identification channels using the first set of identification performance confidence scores.

10. A fault early warning method for the excitation system of a power generation assembly, characterized in that, The method, implemented based on any one of claims 1 to 9, comprises: Based on the equipment identification code of the power generation system, retrieve the electrical schematic diagram, energy flow path and excitation fault cases of the excitation system, and construct a directed topology for fault propagation. The excitation fault case is decomposed into multiple fault feature sample sets of multiple system components in the excitation system, and multiple fault identification models are pre-trained. The multiple fault identification models are mapped and embedded into multiple component topology nodes of the fault propagation directed topology, and fault propagation linkage reasoning correction is performed to construct a dynamically associated fault identification topology. The system senses real-time monitoring data of multiple multi-physics parameters of multiple system components, inputs the dynamic correlation fault identification topology to identify fault linkage status, and outputs a multi-node collaborative fault feature vector. Based on the multi-node collaborative fault feature vector, the fault evolution trend is predicted, and the potential fault time window is output to perform graded early warning of the excitation system.

Images