Electromechanical system fault diagnosis method and device based on digital twinning
By employing a fault diagnosis method based on hierarchical collaborative modeling using digital twins and integrating graph neural networks, the problems of fault diagnosis accuracy and stability of electromechanical systems in harsh environments were solved, achieving higher fault diagnosis accuracy and stability.
Patent Information
- Application Number
- CN202511296140.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-11
- Publication Date
- 2026-01-16
AI Technical Summary
Existing fault diagnosis methods for electromechanical systems have low accuracy and stability in harsh environments, making it difficult to meet the requirements for safe and reliable operation.
A fault diagnosis method based on digital twins is adopted. A digital twin model is constructed through hierarchical collaborative modeling to generate virtual fault data and structured measurement knowledge. Combined with integrated graph neural networks and multi-task learning algorithms, a fault diagnosis model with composite fault modes is constructed.
It improved the accuracy and stability of fault diagnosis for electromechanical systems, increasing the fault diagnosis accuracy and stability by 5.94% and 22.32%, respectively.
Smart Images
Figure CN121350447A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and apparatus for fault diagnosis of electromechanical systems based on digital twins, belonging to the field of electrical data processing technology. Background Technology
[0002] Electromechanical systems (EMS) often operate in harsh environments. Due to these harsh conditions, EMS are prone to failure, which can lead to performance degradation or even catastrophic accidents. To maintain safe and reliable operation, there is an increasing need to develop fault diagnosis technologies for EMS. Recently, some scholars have conducted research on fault diagnosis methods for typical EMS components and faults. These methods can be broadly categorized into two types: knowledge / model-based fault diagnosis and data-driven fault diagnosis. The former primarily utilizes prior knowledge of the system to construct observers or identification models. They achieve fault diagnosis through fault reasoning, which relies almost entirely on accurate fault mechanism knowledge and is almost independent of fault data. The latter mainly extracts fault-sensitive information from sensor measurements and achieves fault diagnosis through fault classification. These data-driven methods do not require prior knowledge and are gaining increasing attention in the field of fault diagnosis due to their excellent adaptability. However, due to their data-driven paradigm, their performance largely depends on the quality and quantity of available fault data. Nevertheless, their fault diagnosis accuracy and stability are relatively low. Summary of the Invention
[0003] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and apparatus for fault diagnosis of electromechanical systems based on digital twins, which improves the accuracy and stability of fault diagnosis.
[0004] To achieve the aforementioned objective, this invention provides a method for fault diagnosis of electromechanical systems based on digital twins, comprising the following steps: S01: A digital twin model is constructed using a hierarchical collaborative modeling method to simulate the physical characteristics of an electromechanical system; fault information is generated through the digital twin model, the fault information including virtual fault data and structured interconnection knowledge; S02: Construct an integrated graph neural network to extract learnable heterogeneous graph topological representations from fault information, aiming to provide a complete representation of nonlinear and diverse fault behaviors; S03: A multi-task learning algorithm with bi-level gradient optimization is used to reconstruct and train an integrated graph neural network through multiple sub-tasks to generate a fault diagnosis model with composite fault modes.
[0005] To achieve the aforementioned objective, this invention also provides a fault diagnosis device for electromechanical systems based on digital twins, comprising: a digital twin model construction unit configured to construct a digital twin model to simulate the physical characteristics of an electromechanical system using a hierarchical collaborative modeling method; generating fault information through the digital twin model, the fault information including virtual fault data and structured interconnection knowledge; an integrated graph neural network construction unit configured to extract learnable heterogeneous graph topological representations from the fault information, aiming to provide a complete representation for nonlinear and diverse fault behaviors; and a fault diagnosis model generation unit for composite fault modes configured to reconstruct and train the integrated graph neural network through a multi-task learning algorithm with bi-level gradient optimization to generate a fault diagnosis model for composite fault modes.
[0006] Compared with existing technologies, the electromechanical system fault diagnosis method and apparatus based on digital twins provided by this invention firstly constructs a full-system digital twin model through hierarchical collaborative modeling technology, generating two information forms: virtual fault data and structured measurement-related knowledge, enriching the type and scale of the dataset. Secondly, starting from the two dimensions of digital twin model data and digital twin model knowledge, an integrated graph neural network (IGNN) is constructed to build a learnable heterogeneous graph representation for nonlinear and diverse fault behaviors. The IGNN is trained using a bi-level gradient multi-task learning algorithm to improve the model's ability to handle complex fault modes. This improves the accuracy and stability of fault diagnosis, demonstrating the effectiveness and superiority of the electromechanical system fault diagnosis method considering coupled fault characteristics. Attached Figure Description
[0007] Figure 1 This is a flowchart of the first embodiment of the present invention, which provides a fault diagnosis method for electromechanical systems based on digital twins. Detailed Implementation
[0008] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0009] First Embodiment
[0010] Figure 1 This is a flowchart of the electromechanical system fault diagnosis method based on digital twin provided in the first embodiment of the present invention; as follows: Figure 1 As shown, the fault diagnosis method for electromechanical systems based on digital twins provided by this invention includes the following steps: S01: A digital twin model (DT) is constructed using a hierarchical collaborative modeling method to simulate the physical characteristics of an electromechanical system; fault information is generated through the digital twin model, the fault information including virtual fault data and structured interconnection knowledge; S02: Construct an integrated graph neural network (IGNN) to extract learnable heterogeneous graph topological representations from fault information, aiming to provide a complete representation of nonlinear and diverse fault behaviors; S03: The multi-task learning (MTL) algorithm with bi-level gradient optimization is used to reconstruct and train the ensemble graph neural network through multiple sub-tasks to generate a fault diagnosis model with composite fault modes.
[0011] In the first embodiment, the entity description of the electromechanical system (EMS) is divided into three types: element description (ED), data description (DD), and relation description (RD).
[0012] The element descriptions (EDs) of an EMS entity include: structural dimension elements. Physical dimension elements Operating condition dimension elements and fault dimension elements The intrinsic mechanism elements describing the characteristics of EMS entities, including structural dimension elements. Includes: structural components and structural details Physical dimension elements Including control principles Electromagnetic principles Mechanical principles and fluid principles Working condition dimension elements Including the power generation process Power transmission process and the process of running under load Fault dimension elements Including fault location Cause of the malfunction Impact of the fault and symptoms of failure .
[0013] The data description (DD) of an EMS entity includes: dynamic data description. and static data description The description provides a quantitative description of EMS entity sensor measurements, including dynamic data description. Includes: rapidly changing data and slowly changing data Static data description Includes: sensor name Sensor position Sensor sampling frequency Sensor base and signal preprocessing methods .
[0014] EMS entity data relationship description (RD) includes: inline relationships and interconnection Among them, internal relationships Includes: structural dimension elements and physical dimension elements Coupling between Describes the intrinsic relationship between spatial structure and physics; structural dimension elements Physical dimension elements and working condition dimension elements Coupling between This describes how to integrate the physical behaviors of some unit-level entities into functional domain-level operating conditions; structural dimension elements. Physical dimension elements Operating condition dimension elements and fault dimension elements Coupling between This describes how a normally functioning entity evolves into a faulty entity through coupling with faulty elements. Interconnection relationships. This includes the mutual coupling relationship between the power functional domain and the transmission functional domain. The mutual coupling relationship between the transmission functional domain and the load functional domain The mutual coupling relationship between the load functional domain and the power functional domain .
[0015] In the first embodiment, constructing a digital twin model using a hierarchical collaborative modeling method includes:
[0016] S01-1: Construct digital twin models sequentially for the functional domains of the electromechanical system. , respectively representing the spatial level sub-model, behavioral level sub-model, process level sub-model, and state level sub-model; for the i-th level sub-model First, inherit from the previous level. As a baseline model.
[0017] In the first embodiment, the spatial sub-model ,in, Representation of spatial level submodel The main body, for a given structural dimension element Using inter-geometric modeling algorithms Mechanistic models for constructing spatial-level sub-models to statically reflect the structural characteristics of entities, such as Boolean modeling and spline curve modeling; data-driven reverse engineering algorithms. It involves quantitatively constructing numerical space models, such as point cloud scanning and raster projection; and utilizing fusion algorithms. , Will be The geometric dimensions are updated and optimized to achieve a more objective digital representation of the physical spatial structure.
[0018] In the first embodiment, the behavior-level sub-model Inherited spatial submodel : , Represents behavioral sub-models The main body, for physical elements Employing multidisciplinary modeling algorithms Achieving physical virtualization, such as finite element analysis and differential equations; modifying data-driven algorithms using parameter optimization techniques. Parameters, such as least squares, are used; a fusion algorithm based on steady-state analysis is employed. right and Integrate.
[0019] In the first embodiment, the process-level sub-model is the third level of the hierarchical sub-model. It enables the interaction between local behavioral-level sub-models to virtualize the operational conditions of entities. Process-level sub-model for: It describes multiple local behavior-level sub-models. The logical interaction between them. Given logical operating condition elements, the mechanism-based element-driven process algorithm... Different disciplines These are connected to form operating conditions, such as bond graphs and Petri networks. Given transient data elements across different operating conditions, Used to update different The transfer parameters are used to achieve more accurate process transfer. Unlike steady-state analysis, The key focus is and The transient fusion between them focuses on solving and optimizing across working conditions. Its time-varying characteristics.
[0020] In the first embodiment, the state-level sub-model M4 is the last level of the hierarchical sub-model, which virtualizes functional domain entities under different fault modes through fault physics and deep generation models: ,in, Represents the main body of the state-level sub-model; fault physics-driven algorithm According to the fault element Fault injection can be performed, such as by using a slotting deformation method on the process-level sub-model driven by the finite element method. Fault injection is performed. (Deep generation algorithm) Intrinsic fault characteristics are learned directly from the entity's measurement data, such as through generative adversarial networks and variational encoders. Fusion algorithms... Utilizing various pre-training fine-tuning and transfer learning techniques to the above and Optimize.
[0021] In the first embodiment, constructing a digital twin model using a hierarchical collaborative modeling method further includes: S01-2: Using the functional modeling interface to integrate digital twin models Each is encapsulated into a compatible model functional unit. In the middle; and the model functional units Integrate into the digital twin model. In the first embodiment, guided by the description of the interconnection relationships between the three core functional domains—power functional domain, transmission functional domain, and load functional domain—[the following is done]: Integrating into the digital twin model, where, The mathematical description is as follows: , in, , These are sub-models representing the electromechanical system entities in the power functional domain, transmission functional domain, and load functional domain, respectively. express The set of states of the running state; They represent The set of input variables and the set of output variables; Represent the initial state; using extrapolation equations , For driving from arrive The equation is performed. For time intervals; Let be the state variable at time t. Let be the input variable at time t. This is the equation for generating virtual data output at time t; It is the solver that drives the equations.
[0022] In the first embodiment, each sub-model is encapsulated as a model functional unit (FMU), which includes a sub-model, a sub-environment, and a solver. Guided by interconnection, the digital twin model of the entire system is integrated from the FMUs of three core functional domain entities, which include interactive variables and collaborative environments. With the help of timing orchestration algorithms (such as Gauss-Seidel), the entire system can be iteratively driven to generate virtual data in timing fault states and operating states.
[0023] The fault diagnosis method for electromechanical systems based on digital twins provided by this invention also includes: S04: Consistency between the electromechanical system entity and the digital twin model is ensured through a consistency metric. This consistency metric is achieved by comparing the entity data of the electromechanical system measured by sensors with the virtual data generated by the digital twin model. It can be implemented through distance measurement, similarity measurement, etc. If the consistency metric result exceeds a predefined threshold, a consistency update of the digital twin model will be initiated.
[0024] The fault diagnosis method for electromechanical systems based on digital twins provided by this invention also includes: S05: Updating the digital twin model through consistent updates: for spatial sub-models Reverse engineering methods are used to reconstruct the spatial geometric parameters of the computer electrical system entity, generating a new spatial-level sub-model. : , In the formula, and These are the parameters before and after the spatial-level sub-model update, respectively.
[0025] This invention updates the behavioral-level sub-model using a constrained parameter estimation algorithm based on steady-state data. Generate new behavioral sub-models : , In the formula, , and yes The upper and lower limits; and These are the steady-state data and behavioral sub-models of the electromechanical system entity. The generated data.
[0026] This invention updates the process-level sub-model using a constrained parameter estimation algorithm based on transient data. Generate new process-level sub-models : , In the formula, , and yes The upper and lower limits; and These are entity data and process-level sub-models of electromechanical systems under fault conditions. The generated data.
[0027] This invention employs an implicit deep generative model and uses the gradient descent algorithm to update the state-level sub-models. Generate new state-level sub-models .
[0028] Furthermore, the system-wide digital twin model itself is a highly structured knowledge container from which knowledge related to fault propagation can be extracted, such as spatial structure knowledge, physical mechanism knowledge, state process knowledge, and mutual coupling knowledge. Digital twin model knowledge describes the potential impact of fault propagation within the electromechanical system from multiple dimensions.
[0029] In the first embodiment, constructing an integrated graph neural network to extract a learnable heterogeneous graph topological representation from fault information includes: S02-1 Constructing a prior knowledge image subnetwork for digital twin model knowledge; S02-2: Construct a fault characteristic data graph subnetwork for digital twin model data; S02-3: Combine the prior knowledge image subnetwork and the fault characteristic data graph subnetwork to generate an integrated graph neural network.
[0030] In the first embodiment, the prior knowledge image subnetwork for constructing the digital twin model knowledge includes: S02-1-1: Define basic nodes of the knowledge graph topology from the digital twin model. These basic nodes represent the location and number of sensors. Based on spatial knowledge, this invention abstracts the sensors of an electromechanical system into a set of basic nodes in the graph topology, categorizing them into three types: power nodes, transmission nodes, and load nodes. , i=1,2,3 In the formula, N i This represents the number of sensors in the r-th basic node. The signal length of the sensor is indicated. The power node mainly represents a three-phase current sensor, the transmission node mainly represents a vibration sensor, and the load node mainly represents a pressure and flow sensor.
[0031] In the first embodiment, the prior knowledge image subnetwork for constructing the digital twin model knowledge includes: S02-1-2: Based on the physical characteristics of the basic nodes, corresponding augmented nodes are obtained. These augmented nodes represent the physical characteristics of the sensor. Considering that sensor measurement data contains both fault-sensitive and interference components, physical behavior knowledge is used to transform the sensor measurement data into features, thereby highlighting the inherent fault-sensitive components of the sensor. Therefore, each augmented node represents the measurement characteristics of a basic node: , i=1,2,3 In the formula, This represents the number of the r-th expansion node. This represents the dimension of the amplified node. Power-based nodes (such as motor current) exhibit rapid, periodic changes in the time domain driven by the law of electromagnetic induction, and fault-sensitive information is reflected in both the time and frequency domains. Therefore, power-based amplified nodes consist of time-domain characteristics (such as mean characteristics, peak characteristics, etc.) and frequency-domain characteristics (such as peak frequency characteristics). Transmission-based nodes (such as bearing vibration measurement) exhibit rapid, periodic changes and oscillating behavior driven by motion / dynamic laws, where fault-sensitive information is reflected more in the frequency domain than the time domain. Therefore, transmission-based amplified nodes consist of frequency-domain characteristics (such as rational frequencies). Load-based nodes exhibit slow, time-domain changes driven by fluid dynamics, and fault-sensitive information is reflected in the time domain. Therefore, load-based amplified nodes are primarily composed of time-domain characteristics.
[0032] In the first embodiment, the prior knowledge image subnetwork for constructing the digital twin model knowledge includes: S02-1-3: Inline connections are made between expanded nodes within the same functional domain to obtain an inner-connection graph topology. The expanded nodes reflect the physical characteristics of local entities, and these physical characteristics, when connected internally, construct the operating conditions. The inlines between expanded nodes reflect the process-level sub-model. Knowledge; , i=1,2,3, E i Let i be the set of topological edges of the connection graph within the i-th functional domain. Let be the number of edges in the topology of the connected graph within the i-th functional domain. For example, in the variable frequency control process driven by the motor-based power functional domain, the power amplification nodes have equal amplitude and the same period (like the characteristics of three-phase current). Therefore, there is an intrinsic relationship between the current periodic frequency characteristic node of phase a and the current periodic frequency characteristic node of phase b. Similarly, in the shaft transmission process of the transmission functional domain, the bearings at both ends of the shaft have the same rotational speed. Therefore, there is an intrinsic relationship between the rotational frequency characteristic node of the bearing at end a and the rotational frequency characteristic node of the bearing at end b. Taking a centrifugal pump as an example, during the fluid transport process in the pump pipe, its flow rate and pressure conform to the pump's characteristic curve. Therefore, there is an intrinsic relationship between the average flow rate characteristic node and the average pressure characteristic node.
[0033] In the first embodiment, the prior knowledge image subnetwork for constructing the digital twin model knowledge includes: S02-1-4: Based on the mutual coupling knowledge between entities in different functional domains, the internal connection graph topology of the corresponding functional domains is interconnected. This invention constructs a node internal connection graph topology based on the internal coupling knowledge within each functional domain entity. On this basis, it applies node interconnection to the graph topology using the mutual coupling knowledge between functional domain entities: , Represents the set of edges connected to the network. , This represents the number of edges in the interconnection. The above interconnection relationships can also be generalized to entities in other functional domains; in summary, the prior knowledge image subnetwork representation is obtained: .
[0034] In the first embodiment, constructing the fault characteristic data graph sub-network of the digital twin model data includes: S02-2-1: Reconstruct the topology data of channel M based on the data of different fault modes generated by the digital twin model; the reconstructed topology data of channel m is represented as: In the formula The input is the fault data for the m-th channel. Let m be the true state label, where m = 1, ..., M; S02-2-2: M data characteristic graph sub-networks are constructed respectively based on the topology data of the M channels using an M-channel graph topology learner; S02-2-3: Obtain the edge weights between any two different augmentation nodes p and q in the m-th data characteristic graph subnetwork using the graph topology learner of the m-th channel. :
[0035] In the formula, Let be the parameter vector of the fully connected layer in the m-th channel graph topology learner; Indicates splicing; Indicates the activation function; and These represent the prior knowledge image subnetworks, respectively. The characteristics of the p-th and q-th amplification nodes, Image subnetwork representing prior knowledge The weights between the features of the p-th and q-th augmented nodes; S02-2-4: Threshold for given weighted values ,if Then the edge and edge weight between the expanded nodes p and q are retained; S02-2-5: Repeat steps S02-2-3 to S02-2-4 for all augmented nodes on the m-th data characteristic graph subnetwork to obtain the edge set E of the augmented nodes on the m-th data characteristic graph subnetwork. m And edge weight set A m ; S02-2-6: Will The graph convolutional layer of the topology learner input to the m-th channel yields graph features. ;Graph features Convert to state estimation labels Specifically, it includes: Graph features Max graph pooling is used to obtain features: , ; Graph features Features are obtained by performing average graph pooling: ; Features and characteristics By splicing the components together, we obtain the fused features: , Based on fusion characteristics State estimation label : , ; S02-2-7: Calculate State Estimation Labels and state reality label Binary cross-entropy loss between Through binary cross-entropy loss Update the parameters of the subnetwork of the fault characteristic data graph for channel m; S02-2-8: Repeat steps S02-2-3 to S02-2-7 to obtain the fault characteristic data graph subnetwork of the digital twin model data: .
[0036] In the first embodiment, a fault diagnosis model for composite fault modes is generated by reconstructing and training an ensemble graph neural network using a two-level gradient-optimized multi-task learning algorithm, including:
[0037] S03-1: The heterogeneous graph topology G is obtained by integrating the fault characteristic data graph subnetwork and the prior knowledge image subnetwork. , In the formula, , and These are the edge sets of the prior knowledge image subnetwork and the fault characteristic data graph subnetwork, respectively. , and These are the edge weight sets for the prior knowledge image subnetwork and the fault characteristic data graph subnetwork, respectively. m=1,…,M, The number of faults in the m-th channel is input. This represents graph convolution operations; S03-2: Yes Features are obtained by performing convolution operations. and M features The concatenation yields the feature vector F:
[0038] In the formula, Indicates splicing; S03-3: F is determined by the state classifier Transformed to a size of State vector y: .
[0039] In the first embodiment, the fault diagnosis model for composite fault modes generated by reconstructing and training the ensemble graph neural network through a multi-task learning algorithm with bi-level gradient optimization further includes: S03-4: Calculate the weighted loss of the c-th shared feature layer of the shared feature block in a multi-task ensemble graph neural network according to the following formula: c=1,…,C; In the formula, Let c be the loss function of the c-th layer of the t-th subtask. It is a topological number; The c-th layer shares the feature layer parameters; Parameters related to the t-th subtask; Weighting coefficients for tasks; =1; S03-5: Update parameters using the following formula and : , , t=1,…,T; In the formula, The first learning coefficient; The second learning coefficient; Indicates to The gradient; Indicates to The gradient.
[0040] In the first embodiment, the fault diagnosis model for composite fault modes generated by reconstructing and training the ensemble graph neural network through a multi-task learning algorithm with bi-level gradient optimization further includes: S03-6: Obtain the total loss of the shared feature block using the following formula: , In the formula, C represents the total number of layers sharing feature blocks; Shared layer weights;
[0041] In the formula, Represents the norm, Indicates to The gradient; .
[0042] In the first embodiment, the fault diagnosis model for composite fault modes generated by reconstructing and training the ensemble graph neural network through a multi-task learning algorithm with bi-level gradient optimization further includes: S03-7: Update the task weights and shared layer weights based on the total loss of the shared feature blocks using the following formula: , , In the formula, The third learning coefficient; It is the fourth learning coefficient; Indicates to The gradient; Indicates to The gradient.
[0043] The first embodiment of this invention provides a digital twin-based fault diagnosis method for electromechanical systems. First, it constructs a full-system digital twin model using hierarchical collaborative modeling technology, generating two information formats: virtual fault data and structured measurement-related knowledge, enriching the type and scale of the dataset. Second, it constructs an ensemble graph neural network (IGNN) from the two dimensions of digital twin model data and digital twin model knowledge, creating a learnable heterogeneous graph representation for nonlinear and diverse fault behaviors. The IGNN is trained using a bilevel gradient multi-task learning (MTL) algorithm to improve the model's ability to handle complex fault modes. This improves the accuracy and stability of fault diagnosis. Experiments show that, compared with existing mainstream methods, the digital twin-based fault diagnosis method for electromechanical systems provided by this invention improves fault diagnosis accuracy and stability by an average of 5.94% and 22.32%, respectively.
[0044] Second Embodiment
[0045] The second embodiment of the present invention only scans content that is different from that in the first embodiment; the same content will not be described again.
[0046] The second embodiment of the present invention provides a fault diagnosis device for electromechanical systems based on digital twins. Its digital twin model construction unit is configured to construct a digital twin model to simulate the physical characteristics of the electromechanical system through a hierarchical collaborative modeling method and generate fault information, which includes virtual fault data and structured interconnection knowledge; an integrated graph neural network construction unit is configured to extract learnable heterogeneous graph topological representations from the fault information, aiming to provide a complete representation for nonlinear and diverse fault behaviors; and a fault diagnosis model generation unit for composite fault modes is configured to reconstruct and train the integrated graph neural network through a multi-task learning algorithm with bi-level gradient optimization to generate a fault diagnosis model for composite fault modes.
[0047] The beneficial effects of the electromechanical system fault diagnosis device based on digital twin provided in the second embodiment of the present invention are the same as those in the first embodiment, and will not be described again here.
[0048] Furthermore, all or part of the steps of the digital twin-based electromechanical system fault diagnosis method provided in the first embodiment can be programmed into a computer program using a computer language, stored in a storage medium, and called by one or more processors.
[0049] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.
[0050] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A fault diagnosis method for electromechanical systems based on digital twins, characterized in that, Includes the following steps: S01: A digital twin model is constructed using a hierarchical collaborative modeling method to simulate the physical characteristics of an electromechanical system; fault information is generated through the digital twin model, the fault information including virtual fault data and structured interconnection knowledge; S02: Construct an integrated graph neural network to extract learnable heterogeneous graph topological representations from fault information, aiming to provide a complete representation of nonlinear and diverse fault behaviors; S03: A multi-task learning algorithm with bi-level gradient optimization is used to reconstruct and train an integrated graph neural network through multiple sub-tasks to generate a fault diagnosis model with composite fault modes.
2. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 1, characterized in that, Constructing digital twin models using a hierarchical collaborative modeling approach includes: S01-1: Construct digital twin models sequentially for the functional domains of the electromechanical system. , respectively representing the spatial level sub-model, the behavioral level sub-model, the process level sub-model, and the state level sub-model; S01-2: Using the functional modeling interface to integrate digital twin models Each is encapsulated into a compatible model functional unit. In, and the model functional units Integrate into the digital twin model.
3. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 2, characterized in that, Also includes: S04: Consistency between the electromechanical system entity and the digital twin model is ensured through a consistency metric, which is achieved by comparing the entity data of the electromechanical system measured by sensors with the virtual data generated by the digital twin model; S05: Updating the digital twin model through consistent updates: for spatial sub-models Reverse engineering methods are used to reconstruct the spatial geometric parameters of the computer electrical system entity, generating a new spatial-level sub-model. The behavioral sub-model is updated using a constrained parameter estimation algorithm based on steady-state data. Generate new behavioral sub-models ; Update the process-level sub-model using a constrained parameter estimation algorithm based on transient data. Generate new process-level sub-models An implicit deep generative model is adopted, and the gradient descent algorithm is used to update the state-level sub-models. Generate new state-level sub-models .
4. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 3, characterized in that, Constructing an ensemble graph neural network to extract learnable heterogeneous graph topological representations from fault information includes: S02-1 Constructing a prior knowledge image subnetwork for digital twin model knowledge; S02-2: Construct a fault characteristic data graph subnetwork for digital twin model data; S02-3: Combine the prior knowledge image subnetwork and the fault characteristic data graph subnetwork to generate an integrated graph neural network.
5. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 4, characterized in that, The prior knowledge image subnetwork for constructing digital twin model knowledge includes: S02-1-1: Define the basic nodes of the knowledge graph topology from the digital twin model; S02-1-2: Obtain the corresponding augmented nodes of the basic nodes based on the physical characteristics of the basic nodes; S02-1-3: Connecting expansion nodes within the same functional domain to obtain an inner connection graph topology; S02-1-4: Based on the knowledge of mutual coupling between entities in different functional domains, the internal connection graphs of the corresponding functional domains are interconnected topologically.
6. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 5, characterized in that, The fault characteristic data graph subnetwork for constructing digital twin model data includes: S02-2-1: Reconstruct the topology data of channel M based on the data of different fault modes generated by the digital twin model; the reconstructed topology data of channel m is represented as: In the formula The input is the fault data for the m-th channel. Let m be the true state label, where m = 1, ..., M; S02-2-2: M data characteristic graph sub-networks are constructed respectively based on the topology data of the M channels using an M-channel graph topology learner; S02-2-3: Obtain the edge weights between any two different augmentation nodes p and q in the m-th data characteristic graph subnetwork using the graph topology learner of the m-th channel. ; S02-2-4: Threshold for given weighted values ,if Then the edge and edge weight between the expanded nodes p and q are retained; S02-2-5: Repeat steps S02-2-3 to S02-2-4 for all augmented nodes on the m-th data characteristic graph subnetwork to obtain the edge set E of the augmented nodes on the m-th data characteristic graph subnetwork. m And edge weight set A m ; S02-2-6: Will The graph convolutional layer of the topology learner input to the m-th channel yields graph features. ;Graph features Convert to state estimation labels ; S02-2-7: Calculate State Estimation Labels and state reality label The binary cross-entropy loss between the data features is used to update the parameters of the m-th data feature map subnetwork. S02-2-8: Repeat steps S02-2-3 to S02-2-7 to obtain the fault characteristic data graph subnetwork of the digital twin model data.
7. The electromechanical system fault diagnosis method based on digital twin according to claim 6, characterized in that, A multi-task learning algorithm with bi-level gradient optimization is used to reconstruct and train an ensemble graph neural network into multiple sub-tasks to generate a fault diagnosis model with composite fault modes. S03-1: Integrate the fault characteristic data graph subnetwork and the prior knowledge image subnetwork to obtain the heterogeneous graph topology G: , In the formula, , and These are the edge sets of the prior knowledge image subnetwork and the fault characteristic data graph subnetwork, respectively. , and These are the edge weight sets for the prior knowledge image subnetwork and the fault characteristic data graph subnetwork, respectively. m=1,…,M, The number of faults in the m-th channel is input. This represents graph convolution operations; S03-2: Yes Features are obtained by performing convolution operations. and M features The concatenation yields the feature vector F: ， In the formula, Indicates splicing; S03-3: F is determined by the state classifier Transformed to a size of State vector y: 。 8. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 7, characterized in that, The multi-task learning algorithm using bi-level gradient optimization for reconstructing and training ensemble graph neural networks to generate fault diagnosis models with composite fault modes also includes: S03-4: Calculate the weighted loss of the c-th shared feature layer of the shared feature block in a multi-task ensemble graph neural network according to the following formula: ,c=1,…,C; In the formula, Let c be the loss function of the c-th layer of the t-th subtask. Topology data; The c-th layer shares the feature layer parameters; Parameters related to the t-th subtask; Weighting coefficients for tasks; =1; S03-5: Update parameters using the following formula and : , , t=1,…,T; In the formula, The first learning coefficient; The second learning coefficient; Indicates to The gradient; Indicates to The gradient.
9. The method for fault diagnosis of electromechanical systems based on digital twins according to claim 8, characterized in that, The multi-task learning algorithm using bi-level gradient optimization for reconstructing and training ensemble graph neural networks to generate fault diagnosis models with composite fault modes also includes: S03-6: Obtain the total loss of the shared feature block using the following formula: , In the formula, C represents the total number of layers sharing feature blocks; Shared layer weights; ， In the formula, Represents the norm, Indicates to The gradient; ; S03-7: Update the task weights and shared layer weights based on the total loss of the shared feature blocks using the following formula: , , The third learning coefficient; It is the fourth learning coefficient; Indicates to The gradient; Indicates to The gradient.
10. A fault diagnosis device for electromechanical systems based on digital twins, characterized in that, include: A digital twin model building unit is configured to construct a digital twin model to simulate the physical characteristics of an electromechanical system through a hierarchical collaborative modeling method and generate fault information, which includes virtual fault data and structured interconnection knowledge. An integrated graph neural network building block is configured to extract learnable heterogeneous graph topological representations from fault information, aiming to provide a complete representation of nonlinear and diverse fault behaviors. The fault diagnosis model generation unit for composite fault modes is configured to reconstruct and train an ensemble graph neural network through a multi-task learning algorithm with bi-level gradient optimization to generate a fault diagnosis model for composite fault modes.