A digital twin-based method for assessing the health status of engineering operations and maintenance.

By constructing a unified health semantic space and health assessment network across operating conditions, the problem of inconsistent health assessments under different operating conditions was solved, achieving stable and accurate assessments when operating conditions change, and ensuring the semantic consistency and physical accuracy of the assessment results.

CN123115996APending Publication Date: 2026-11-17HAOYI DECORATION (SHANGHAI) PROJECT MANAGEMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611211992.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing health assessment methods struggle to ensure that health scores correspond to the same failure risk level under different operating conditions, and lack a collaborative mechanism between digital twin simulation and assessment networks, resulting in inconsistent assessment results when operating conditions change.

Method used

A unified health semantic space is constructed across operating conditions. By setting health semantic anchor points and nonlinear mapping functions, the health semantic axes of each operating condition are aligned to a unified reference coordinate system. Adversarial training is carried out using graph attention networks and gradient inversion layers to establish a health assessment network. The assessment results are corrected using semantic offsets.

Benefits of technology

It achieves stability and consistency in health assessment under different working conditions, improves the accuracy of the projection position of the working condition boundary area and the semantic calibration of the assessment results, and ensures the accuracy of the physical meaning of the health score.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115996A_ABST
    Figure CN123115996A_ABST
Patent Text Reader

Abstract

This invention discloses a digital twin-based method for assessing the health status of engineering operation and maintenance, relating to the field of engineering operation and maintenance health management technology. It addresses the technical problems of existing health assessment methods, such as the inability to guarantee that the same health score corresponds to the same fault risk level under different operating conditions, and the lack of a collaborative mechanism between the digital twin model and the assessment network. The method includes the following steps: constructing a digital twin model to generate source and target domain data; setting benchmark health anchor points and critical fault anchor points to construct a cross-operating condition health semantic space; constructing a bi-branch assessment network based on a graph attention network and a bidirectional long short-term memory network, decoupling operating condition information through a gradient inversion layer; inputting the target domain data into the assessment network to obtain an initial health score, and simulating the final projection position of the target domain data in the health semantic space using the twin model; calculating the semantic offset based on the semantic distance between the projection position and the health semantic anchor point, and correcting the initial health score.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of engineering operation and maintenance health management technology, specifically involving a method for assessing the health status of engineering operation and maintenance based on digital twins. Background Technology

[0002] During the long-term service of engineering equipment, its health status is affected by multiple operating conditions such as load, ambient temperature and operating speed, exhibiting complex time-varying degradation characteristics. Traditional health assessment methods mainly rely on single sensor threshold discrimination or statistical models based on historical data, which are difficult to cope with the data distribution drift caused by frequent switching of equipment operating conditions, and also lack a set of universal health measurement benchmarks that do not depend on operating conditions.

[0003] The emergence of digital twin technology has provided new solutions for engineering operation and maintenance. In recent years, patents have proposed using digital twin models for equipment condition monitoring and multi-source data fusion assessment. For example, by constructing a twin to simulate and extrapolate the equipment degradation process, or by fusing multi-source sensor signals such as vibration, temperature, and pressure to score the health status. However, existing methods generally have the following shortcomings: First, the generalization ability of the health assessment model under different operating conditions is insufficient. After the model is trained under a certain operating condition, the assessment accuracy drops significantly when switching to a new operating condition. Second, existing domain adaptive transfer methods focus on the alignment at the data feature distribution level, only reducing the data distribution difference between the source domain and the target domain from a statistical perspective. It is difficult to ensure that the same health score has a consistent fault risk semantic meaning under different operating conditions. As a result, when operation and maintenance personnel obtain a score of "80 points" at different times and under different operating conditions, they cannot determine whether the actual corresponding equipment risk level is the same. Third, there is a lack of effective collaboration mechanism between the digital twin model and the assessment network. The simulation capability of the twin model is only used to expand the training data and fails to participate in the calibration of the assessment results during the inference stage.

[0004] Therefore, a method for assessing the health status of engineering operations and maintenance based on digital twins has emerged. Summary of the Invention

[0005] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a health assessment method for engineering operation and maintenance status based on digital twins. This method addresses the technical problems of existing health assessment methods, which struggle to ensure that the same health score corresponds to the same fault risk level under different operating conditions, and also struggle to establish an effective collaborative mechanism between the data-driven model and digital twin simulation during the health assessment process. Furthermore, it lacks the technical ability to embed the degraded physical semantics of digital twin simulation into the health assessment network and use this semantics as a reference to perform cross-condition consistency calibration of the assessment score.

[0006] To address the above problems, this invention provides a method for assessing the health status of engineering operations and maintenance based on digital twins, comprising the following steps: S1: Collect multi-source operating data of engineering equipment under historical operating conditions, construct a digital twin model, and generate source domain data and target domain data through the digital twin model; S2: Set health semantic anchor points in the digital twin model, the health semantic anchor points include the baseline health state and critical fault state under each operating condition, and construct a cross-operating condition health semantic space based on the health semantic anchor points; S3: Construct a health assessment network and train the health assessment network using the source domain data; S4: Input the target domain data into the health assessment network to obtain an initial health score, and simultaneously calculate the projection position of the target domain data in the health semantic space through the digital twin model; S5: Based on the semantic distance between the projection position and the health semantic anchor point, calculate the semantic offset, and use the semantic offset to correct the initial health score to obtain the corrected health assessment result.

[0007] Preferably, in step S1, constructing the digital twin model includes the following steps: The geometric parameters, material properties, and historical operation and maintenance records of the engineering equipment are acquired to construct a three-dimensional physical model. Based on the operating condition parameters and sensor signals in the multi-source operation data, a behavior model is constructed using multi-physics coupling simulation. The behavior model is associated with the three-dimensional physical model to form an initial twin. The preset operating condition parameter space is traversed, and several degradation levels are set along the full degradation path of the equipment at each operating condition node. The initial twin is driven to simulate and run at each degradation level to generate a multi-level source domain dataset containing operating condition annotations, degradation level annotations, and health semantic anchor point annotations. The real-time collected variable operating condition operation data is used as the target domain dataset.

[0008] Preferably, in step S2, setting health semantic anchor points and constructing a cross-condition health semantic space includes the following steps: For each operating node, the steady-state operation of the digital twin model under rated parameters is defined as the baseline health anchor point; the state of the digital twin model when the key parameters reach the design critical threshold is defined as the critical failure anchor point. Using the baseline health anchor point and critical fault anchor point as endpoints and the degree of equipment degradation as a metric, a health semantic axis is constructed under the aforementioned operating conditions. Using the health semantic anchor point annotations of each sample in the multi-level source domain dataset as supervision signals, a nonlinear mapping function from the health semantic axis of each working condition to the unified reference coordinate system is learned. The health semantic axis under each working condition is projected onto the unified reference coordinate system to form a unified health semantic space across working conditions.

[0009] Preferably, in step S3, constructing the health assessment network includes the following steps: A graph attention network is constructed, with each sensor signal in the multi-source operational data as a graph node. The spatial coupling weight between each node is calculated through the graph attention mechanism. Each sensor signal is sliced ​​according to a time window, and temporal features are extracted through a bidirectional long short-term memory network. The temporal features of each node are weighted and aggregated with the spatial coupling weight to form a health status feature vector. A dual-branch output structure connected to a graph attention network is constructed. The first branch of the dual-branch output structure is a multilayer perceptron regression network, which takes the health state feature vector as input and outputs the projection position coordinates in the health semantic space. The second branch is a fully connected mapping layer, which takes the health state feature vector as input and outputs a one-dimensional health score. The graph attention network and the dual-branch output structure together constitute a health assessment network; The degradation level labels in the multi-level source domain dataset are used as the supervision labels for the first branch, and the health semantic anchor labels in the multi-level source domain dataset are used as the supervision labels for the second branch. The health assessment network is jointly trained. During training, the health status feature vector is input to the working condition discriminator. The working condition discriminator is connected to the graph attention network through a gradient inversion layer. Adversarial training is performed with the working condition category as the prediction target. The gradient inversion layer transmits the feature vector identically during forward propagation and inverts the error gradient of the working condition discriminator before backpropagating it to the graph attention network during backward propagation.

[0010] Preferably, in step S4, calculating the projection position of the target domain data in the health semantic space using a digital twin model includes the following steps: Input the target domain data into the digital twin model for simulation and obtain the simulation running status corresponding to the target domain data; In the health semantic space, the membership degree of the simulation running state relative to the health semantic axis under each working condition node is calculated, and the working condition where the health semantic axis with the highest membership degree is located is taken as the current working condition belonging to the target domain data; at the same time, the membership degree value with the highest membership degree is used as the confidence weight. Obtain the coordinate transformation parameters between the local coordinate system and the unified reference coordinate system corresponding to the health semantic axis to which the current working condition belongs. The coordinate transformation parameters are determined by a nonlinear mapping function. The projection position coordinates output by the first branch in the health assessment network are transformed from the local coordinate system of the current working condition to the unified reference coordinate system using the coordinate transformation parameters to obtain the transformed coordinates. The simulation running state is mapped to a known coordinate position in the unified reference coordinate system corresponding to the degradation level of the simulation running state to obtain direct projected coordinates; The transformed coordinates and the direct projected coordinates are weighted and averaged using the confidence weights to obtain the final projection position of the target domain data in the cross-condition unified health semantic space.

[0011] Preferably, in step S5, calculating the semantic offset and using the semantic offset to correct the initial health score includes the following steps: Obtain the coordinates of the final projected position in the unified reference coordinate system, calculate the Euclidean distance between the final projected position and the reference health anchor point, and define it as the actual semantic distance. ; Obtain the initial health score from the output of the second branch in the health assessment network. The theoretical semantic distance corresponding to the initial health score is obtained by querying a preset mapping table. ; The difference between the actual semantic distance and the theoretical semantic distance is calculated as the semantic offset. The semantic offset is input into the nonlinear correction function to obtain the correction coefficient. ,in, For preset amplitude coefficient, The preset steepness coefficient; with the correction coefficient Multiply by the initial health score to obtain the corrected health assessment result. ,in, This is the corrected health score.

[0012] Preferably, the construction method of the preset mapping table includes the following steps: In the digital twin model, a baseline working condition is selected, which is the working condition node with the highest coverage density in the preset working condition parameter space; Under the baseline operating conditions, the digital twin model is driven to run in a step-by-step manner along the full degradation path of the equipment according to a preset step size, generating several degradation states covering the entire range of health scores. The projection position coordinates of each degradation state in the health semantic space are recorded, and the Euclidean distance between each projection position and the baseline health anchor point is calculated to construct a one-to-one correspondence mapping table between health scores and theoretical semantic distances.

[0013] Preferably, it also includes a feedback update step: The corrected health assessment results are associated with the corresponding target domain data, final projection position, and semantic offset and stored in the historical database of the digital twin model; After each storage, the cumulative average of the semantic offset in the historical database is calculated. When the cumulative number of storage times that the cumulative average is continuously greater than the preset error threshold reaches the preset trigger number, the target domain data stored in the historical database and its corresponding corrected health score are used as new supervision samples. After the incremental fine-tuning is completed, the output of the first branch in the fine-tuned health assessment network is used to recalculate the projection position of the health semantic axis in the unified reference coordinate system under each working condition node, update the nonlinear mapping function, and recalibrate the position of the health semantic anchor point under each working condition node with the updated nonlinear mapping function, and update the health semantic space.

[0014] The beneficial effects of this invention are: This invention constructs a unified health semantic space across operating conditions, defines the health semantic axis under each operating condition with the benchmark health anchor point and the critical fault anchor point as endpoints, and uses the degradation level label in the digital twin simulation data as the supervision signal to learn the nonlinear mapping function, thereby realizing the coordinate alignment of the health semantic axis of each operating condition to the unified reference coordinate system, and forming a unified health measurement benchmark that does not depend on specific operating conditions. This invention introduces a gradient inversion layer connected to a condition discriminator in the health assessment network for adversarial training. This prevents the health status feature vector extracted by the graph attention network from retaining discriminable condition information, thereby decoupling health features from condition information. This makes the assessment network insensitive to changes in condition, thus ensuring the stability and consistency of health assessment under different conditions. This invention overcomes the shortcomings of single data-driven methods in the boundary region of working conditions by dynamically calculating confidence weights through membership values ​​during the inference stage and weightedly fusing the coordinate transformation results output by the neural network with the direct projection results of digital twin simulation. It achieves collaborative decision-making between data-driven and model-driven approaches and improves the accuracy of the projection position in the boundary region of working conditions. This invention calculates the semantic offset between the actual semantic distance and the theoretical semantic distance, and maps it to a correction coefficient using an S-shaped nonlinear function to correct the initial health score. This achieves semantic calibration of the evaluation results and ensures the accuracy and interpretability of the final health score at the level of its physical meaning. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, this invention is a method for assessing the health status of engineering operations and maintenance based on digital twins, including the following steps: S1: Collect multi-source operating data of engineering equipment under historical operating conditions, construct a digital twin model, and generate source domain data and target domain data through the digital twin model; S2: Set health semantic anchor points in the digital twin model, the health semantic anchor points include the baseline health state and critical fault state under each operating condition, and construct a cross-operating condition health semantic space based on the health semantic anchor points; S3: Construct a health assessment network and train the health assessment network using the source domain data; S4: Input the target domain data into the health assessment network to obtain an initial health score, and simultaneously calculate the projection position of the target domain data in the health semantic space through the digital twin model; S5: Based on the semantic distance between the projection position and the health semantic anchor point, calculate the semantic offset, and use the semantic offset to correct the initial health score to obtain the corrected health assessment result.

[0018] In one embodiment of the present invention, step S1, constructing a digital twin model, includes the following steps: The geometric parameters, material properties, and historical operation and maintenance records of the engineering equipment are acquired to construct a three-dimensional physical model. Based on the operating condition parameters and sensor signals in the multi-source operation data, a behavior model is constructed using multi-physics coupling simulation. The behavior model is associated with the three-dimensional physical model to form an initial twin. The preset operating condition parameter space is traversed, and several degradation levels are set along the full degradation path of the equipment at each operating condition node. The initial twin is driven to simulate and run at each degradation level to generate a multi-level source domain dataset containing operating condition annotations, degradation level annotations, and health semantic anchor point annotations. The real-time collected variable operating condition operation data is used as the target domain dataset.

[0019] Specifically, the geometric parameters, material properties, and historical operation and maintenance records of the engineering equipment are obtained, and a three-dimensional physical model is constructed using finite element modeling software. This model includes the geometric structure of the equipment, the constitutive relationship of the materials, and the initial boundary conditions. Based on operating parameters and sensor signals from multi-source operational data, a behavioral model is constructed using multi-physics coupled simulation. Specifically, load rate, ambient temperature, and operating speed are used as input boundary conditions, and vibration, temperature, and pressure sensor signals are used as model verification targets. The temperature field, stress field, and structural dynamic response equations of the equipment are solved using the finite element method. The calculation results of the behavioral model are compared with the measured sensor signals, and the model parameters are adjusted so that the root mean square error between the simulation output and the measured data is less than a preset error threshold (the preset error threshold is determined based on the sensor measurement accuracy, usually taken as 1% of the sensor range). The calibrated behavioral model is then associated with the three-dimensional physical model to form an initial twin. The preset operating parameter space is traversed, which consists of a load rate range [20%, 100%], an ambient temperature range [-20℃, 50℃], and an operating speed range [80%, 120% of rated speed]. The grid nodes are divided into several dimensions at intervals of 10%, 10°C, and 5%, forming several operating condition nodes. Under each operating condition node, several degradation levels are set along the full degradation path of the equipment. The degradation levels are uniformly divided from 0 to 1, where degradation level 0 corresponds to a brand new state and degradation level 1 corresponds to a critical fault state. The initial twin is driven to run in simulation at each degradation level. That is, the degradation process is simulated by gradually adding material degradation parameters (such as fatigue cumulative damage coefficient, wear thickness, etc.) of key components of the equipment to the twin model. The simulation sensor data corresponding to each degradation level is obtained. The simulation sensor data is associated and stored with the corresponding operating condition parameters, degradation level, and health semantic anchor points to generate a multi-level source domain dataset containing operating condition labels, degradation level labels, and health semantic anchor point labels. The real-time collected variable operating condition operation data is used as the target domain dataset. The target domain data only contains sensor signals and operating condition parameters, and does not contain health status labels.

[0020] In one embodiment of the present invention, step S2, setting health semantic anchor points and constructing a cross-condition health semantic space, includes the following steps: For each operating node, the steady-state operation of the digital twin model under rated parameters is defined as the baseline health anchor point; the state of the digital twin model when the key parameters reach the design critical threshold is defined as the critical failure anchor point. Using the baseline health anchor point and critical fault anchor point as endpoints and the degree of equipment degradation as a metric, a health semantic axis is constructed under the aforementioned operating conditions. Using the health semantic anchor point annotations of each sample in the multi-level source domain dataset as supervision signals, a nonlinear mapping function from the health semantic axis of each working condition to the unified reference coordinate system is learned. The health semantic axis under each working condition is projected onto the unified reference coordinate system to form a unified health semantic space across working conditions.

[0021] Specifically, for each operating condition node, the steady-state operation of the digital twin model under rated parameters (load rate of 100% of rated load, ambient temperature of 25℃, and operating speed of rated speed) is defined as the baseline health anchor point, which corresponds to degradation level 0 in the health semantic space. The state of the digital twin model when key parameters (such as stress, temperature, and vibration amplitude) reach the critical threshold specified in the design specifications is defined as the critical failure anchor point, which corresponds to degradation level 1. For specific equipment types, the critical threshold is determined based on the ultimate stress value or alarm threshold in the equipment design manual, with the specific value based on the technical specifications provided by the equipment manufacturer. Using the baseline health anchor point and the critical failure anchor point as endpoints, and the degree of equipment degradation as a metric, a health semantic axis is constructed for this operating condition: specifically, the line segment starting from the baseline health anchor point and ending at the critical failure anchor point is used as the health semantic axis for this operating condition. Each position on this axis corresponds to a degradation level, and the degree of degradation at each position on the health semantic axis is uniformly divided from 0 to 1, i.e., any degradation level... The position on the health semantic axis is the product of the direction vector from the baseline health anchor point to the critical failure anchor point and... Using the health semantic anchor points of each sample in the multi-level source domain dataset as supervision signals, a nonlinear mapping function from the health semantic axes of each working condition to a unified reference coordinate system is learned. Each working condition's health semantic axis is located in its own local coordinate system, with the origin of each local coordinate system set at the baseline health anchor point for that working condition. The coordinate axis direction is the direction of increasing degradation level under that working condition. The nonlinear mapping function is implemented by a multilayer perceptron, which includes an input layer, two hidden layers (each containing 64 neurons with ReLU activation function), and an output layer (the output dimension is the same as the unified reference coordinate system dimension, which is a three-dimensional space). A training sample set is constructed using the degradation level labels and health semantic anchor point labels of each sample in the multi-level source domain dataset. Each training sample contains the working condition category, degradation level label, and local coordinate system coordinates of that sample in the source domain data. Each training sample is input into the multilayer perceptron, which outputs the projected coordinates of that sample in the unified reference coordinate system. The loss function of the nonlinear mapping function is: ,in, The total number of nodes in the operating condition. The total number of degradation levels. For the working condition node index, For degradation level index, For the first Under the first working condition The coordinates of the sample corresponding to each degradation level in the local coordinate system. Let be the nonlinear mapping function represented by the multilayer perceptron. These are the projected coordinates of the sample in a unified reference coordinate system. This is the variance operator, used to calculate the same degradation level. The variance between the projected coordinates of all working condition nodes is calculated. With minimizing this loss function as the optimization objective, the multilayer perceptron is trained using the Adam optimizer (learning rate set to 0.001) until the loss function value decreases by less than 0.001 for 10 consecutive iterations. After training, the health semantic axes under each working condition are substituted into the trained multilayer perceptron to obtain the projection of the health semantic axes of each working condition in a unified reference coordinate system. This ensures that the coordinate positions of samples from each working condition with the same degradation level coincide in the unified reference coordinate system, forming a unified health semantic space across working conditions.

[0022] In one embodiment of the present invention, step S3, constructing a health assessment network, includes the following steps: A graph attention network is constructed, with each sensor signal in the multi-source operational data as a graph node. The spatial coupling weight between each node is calculated through the graph attention mechanism. Each sensor signal is sliced ​​according to a time window, and temporal features are extracted through a bidirectional long short-term memory network. The temporal features of each node are weighted and aggregated with the spatial coupling weight to form a health status feature vector. A dual-branch output structure connected to a graph attention network is constructed. The first branch of the dual-branch output structure is a multilayer perceptron regression network, which takes the health state feature vector as input and outputs the projection position coordinates in the health semantic space. The second branch is a fully connected mapping layer, which takes the health state feature vector as input and outputs a one-dimensional health score. The graph attention network and the dual-branch output structure together constitute a health assessment network; The degradation level labels in the multi-level source domain dataset are used as the supervision labels for the first branch, and the health semantic anchor labels in the multi-level source domain dataset are used as the supervision labels for the second branch. The health assessment network is jointly trained. During training, the health status feature vector is input to the working condition discriminator. The working condition discriminator is connected to the graph attention network through a gradient inversion layer. Adversarial training is performed with the working condition category as the prediction target. The gradient inversion layer transmits the feature vector identically during forward propagation and inverts the error gradient of the working condition discriminator before backpropagating it to the graph attention network during backward propagation.

[0023] Specifically, a graph attention network is constructed, using vibration, temperature, pressure, and current signals from multi-source operational data as graph nodes, with a total of 4 nodes. The initial feature of each node is the original sequence of its corresponding sensor signal within a time window. The spatial coupling weights between nodes are calculated using a graph attention mechanism. The specific calculation method is as follows: for any node pair... Calculate the attention coefficient ,in, and They are nodes and nodes initial characteristics, It is a learnable linear transformation matrix. This is a learnable attention weight vector. This is a vector concatenation operation. The activation function is used to process the attention coefficients. After normalization, node pairs are obtained Spatial Coupling Weights ,in, The total number of nodes is given. Each sensor signal is sliced ​​into time windows, with a window length of 10 seconds and a sliding step of 2 seconds. The sequence data within each time window is input into a Bidirectional Long Short-Term Memory (BiLSTM) network. The hidden layer dimension of the BiLSTM is set to 128. The temporal features of each sensor signal are extracted. The temporal features of each node at the same time are weighted and aggregated using the aforementioned spatial coupling weights to obtain the health status feature vector. ,in, For nodes through Extracted temporal feature vector; Construct a two-branch output structure connected to the graph attention network. The first branch is a multilayer perceptron regression network, using the health state feature vector. The input is [value], the hidden layer dimensions are 64 and 32 respectively, and the activation function is [function]. The output layer has a dimension of 3, and outputs the coordinates of its projection position in the health semantic space. The second branch is a fully connected mapping layer, using health state feature vectors. The input is given, the hidden layer dimension is 32, and the activation function is... The output layer has a dimension of 1, and outputs a one-dimensional health score. The graph attention network and the dual-branch output structure together constitute the health assessment network; the health assessment network is jointly trained using degradation levels labeled from the multi-level source domain dataset. The supervision label for the first branch is the theoretical projection coordinates of each sample in the multi-level source domain dataset in the health semantic space. (That is, mapping the local coordinates of the sample to coordinates in a unified reference coordinate system using the nonlinear mapping function trained in step S2) serves as the supervision label for the second branch; the loss function for the first branch is... ,in, For batch size, The projection position coordinates are the output of the first branch. These are the theoretical projection coordinates of the sample. The squared Euclidean distance is given; the loss function for the second branch is... ,in, The second branch outputs a one-dimensional health score. Label the degradation level of this sample; the total loss function is... , In this embodiment, preset weighting coefficients are used to balance the loss contributions of the two branches. The value is set to 1; a working condition decoupling regularization term is introduced during the training process to change the health state feature vector. The input is fed into a working condition discriminator, which is a fully connected classification network with hidden layer dimensions of 64 and 32, and the activation function is... The output layer dimension is ( The output, after Softmax activation, is the probability distribution of the work condition categories (total number of work condition categories). Each work condition category is the node number corresponding to the work condition parameter space partitioning mentioned above. The work condition discriminator is connected to the graph attention network through a gradient inversion layer. During forward propagation, the gradient inversion layer identically transmits the feature vector, meaning the forward output equals the input. During backward propagation, it multiplies the error gradient of the work condition discriminator by... The data is then fed back to the graph attention network, where... The inversion coefficient is used in this embodiment. The value is set to 0.1; the loss function of the working condition discriminator is... ,in, For the first One-hot encoding of the working condition category label for each sample. The first output of the operating condition discriminator Each sample belongs to category The probability; during training, the graph attention network and dual-branch output structure are used to minimize To optimize the objective, the operating condition discriminator aims to minimize... To optimize the objective, both are jointly optimized. However, the gradient reversal layer causes the parameter update direction of the graph attention network to be opposite to the optimization direction of the working condition discriminator. As a result, the health status feature vector h cannot retain information that can be used to accurately determine the working condition category, thus achieving decoupling between health status features and working condition information. In this embodiment, the Adam optimizer is used with a learning rate of 0.001 and a batch size M of 64. The training is conducted for 200 rounds, and the parameters of the health evaluation network and the working condition discriminator are updated sequentially in each round of training.

[0024] In one embodiment of the present invention, step S4, calculating the projection position of the target domain data in the health semantic space using a digital twin model, includes the following steps: Input the target domain data into the digital twin model for simulation and obtain the simulation running status corresponding to the target domain data; In the health semantic space, the membership degree of the simulation running state relative to the health semantic axis under each working condition node is calculated, and the working condition where the health semantic axis with the highest membership degree is located is taken as the current working condition belonging to the target domain data; at the same time, the membership degree value with the highest membership degree is used as the confidence weight. Obtain the coordinate transformation parameters between the local coordinate system and the unified reference coordinate system corresponding to the health semantic axis to which the current working condition belongs. The coordinate transformation parameters are determined by a nonlinear mapping function. The projection position coordinates output by the first branch in the health assessment network are transformed from the local coordinate system of the current working condition to the unified reference coordinate system using the coordinate transformation parameters to obtain the transformed coordinates. The simulation running state is mapped to a known coordinate position in the unified reference coordinate system corresponding to the degradation level of the simulation running state to obtain direct projected coordinates; The transformed coordinates and the direct projected coordinates are weighted and averaged using the confidence weights to obtain the final projection position of the target domain data in the cross-condition unified health semantic space.

[0025] Specifically, target domain data is input into a digital twin model for simulation, obtaining the simulated operating state corresponding to the target domain data. Specifically, operating parameters (load rate, ambient temperature, operating speed) from the target domain data are used as input boundary conditions to drive the behavioral model, obtaining the simulated operating state of the equipment under that operating condition, including stress distribution, temperature distribution, and vibration response of key components. The membership degree of the simulated operating state relative to the health semantic axis at each operating condition node is calculated in the health semantic space. Each position on the health semantic axis at each operating condition node corresponds to a complete description of the equipment degradation state (including simulation output parameter vectors such as stress, temperature, and vibration at that degradation level). The simulated operating state is then compared with the... The similarity of the state descriptions corresponding to each degradation level on the health semantic axis of each working condition node is calculated. The similarity metric is the reciprocal of the Euclidean distance, meaning the smaller the Euclidean distance between the feature vector of the simulation running state and the state description of a certain degradation level, the higher the membership degree of that degradation level; The membership degree of the simulated running state under each working condition node is the value with the highest similarity among all degradation levels on the health semantic axis of that working condition node, denoted as . The node corresponding to the maximum membership degree is selected as the current working condition of the target domain data. At the same time, the maximum membership value is used as the confidence weight. ; Obtain the coordinate transformation parameters between the local coordinate system and the unified reference coordinate system corresponding to the health semantic axis to which the current working condition belongs. These coordinate transformation parameters are derived from a trained nonlinear mapping function. Specifically, this means: determining any local coordinates on the current working condition health semantic axis. Substitution The function outputs its corresponding coordinates in a unified reference coordinate system. It inherently contains a complete coordinate transformation relationship from a local coordinate system to a unified reference coordinate system; the projected position coordinates of the first branch output in the health assessment network are transformed using coordinate transformation parameters. Transform the local coordinate system of the current working condition to a unified reference coordinate system to obtain the transformed coordinates. The simulation running state is mapped to the known coordinate position in the unified reference coordinate system corresponding to the degradation level of that simulation running state, thus obtaining the direct projected coordinates. Specifically, after simulation running under each working condition and degradation level using a digital twin model, the known coordinate position of each simulation running state in the health semantic space is obtained. The degradation level with the highest membership degree of the current simulation running state on the health semantic axis of each working condition node is taken, and the coordinate position corresponding to this degradation level in the unified reference coordinate system is the direct projected coordinate. ; weighted by confidence level For the transformed coordinates With direct projected coordinates By performing a weighted average, the final projection position of the target domain data in the unified health semantic space across operating conditions is obtained. .

[0026] In one embodiment of the present invention, step S5, which involves calculating a semantic offset and using the semantic offset to correct the initial health score, includes the following steps: Obtain the coordinates of the final projected position in the unified reference coordinate system, calculate the Euclidean distance between the final projected position and the reference health anchor point, and define it as the actual semantic distance. ; Obtain the initial health score from the output of the second branch in the health assessment network. The theoretical semantic distance corresponding to the initial health score is obtained by querying a preset mapping table. ; The difference between the actual semantic distance and the theoretical semantic distance is calculated as the semantic offset. The semantic offset is input into the nonlinear correction function to obtain the correction coefficient. ,in, For preset amplitude coefficient, The preset steepness coefficient; with the correction coefficient Multiply by the initial health score to obtain the corrected health assessment result. ,in, This is the corrected health score.

[0027] Specifically, obtain the final projection position. Calculate the final projected position using the coordinate values ​​in the unified reference coordinate system. Compared with the baseline health anchor point The Euclidean distance between them is defined as the actual semantic distance. ,in, The Euclidean distance operator is used to obtain the initial health score output by the second branch in the health assessment network. Obtain by querying the preset mapping table The corresponding theoretical semantic distance A pre-defined mapping table records the one-to-one correspondence between each health score under baseline conditions and the theoretical semantic distance in the health semantic space; the difference between the actual semantic distance and the theoretical semantic distance is calculated as the semantic offset. ,when When the actual projected position is farther from the baseline health anchor point than the theoretical position, it indicates that the actual degradation of the current device is more severe than represented by the initial health score, and the initial health score is overestimated; when This indicates that the actual projected position is closer to the baseline health anchor point than the theoretical position, meaning that the actual degree of degradation of the current device is less than that represented by the initial health score, and the initial health score is underestimated. semantic offset Input to nonlinear correction function The correction coefficient is obtained. ,in, The preset amplitude coefficient has a value range of [value range missing]. This is used to control the maximum adjustment range of the correction factor. To preset the steepness coefficient, This is used to control the sensitivity of the correction coefficient to changes in semantic offset. The larger the value.

[0028] In one embodiment of the present invention, the construction method of the preset mapping table includes the following steps: In the digital twin model, a baseline working condition is selected, which is the working condition node with the highest coverage density in the preset working condition parameter space; Under the baseline operating conditions, the digital twin model is driven to run in a step-by-step manner along the full degradation path of the equipment according to a preset step size, generating several degradation states covering the entire range of health scores. The projection position coordinates of each degradation state in the health semantic space are recorded, and the Euclidean distance between each projection position and the baseline health anchor point is calculated to construct a one-to-one correspondence mapping table between health scores and theoretical semantic distances.

[0029] Specifically, a baseline operating condition is selected in the digital twin model. The frequency of each operating condition node in the historical operating data within the preset operating condition parameter space is statistically analyzed. The operating condition node with the highest frequency of occurrence is selected as the operating condition node with the highest coverage density (wherein, the highest coverage density is the operating condition node with the highest frequency of occurrence), and this node is selected as the baseline operating condition. In this embodiment, if the combination of 80% load rate, 25℃ ambient temperature, and rated speed occurs most frequently in the historical operating data, then this operating condition node is the baseline operating condition. Under the baseline operating condition, the digital twin model is driven to simulate and run along the equipment's full degradation path step by step according to a preset step size, generating several degradation states covering the entire range of health scores [0, 1]. The aforementioned degradation level is used as the preset step size, starting from degradation level 0 and increasing step by step to degradation level 1. The digital twin model is driven to simulate and run once for each level, and the simulation running state corresponding to each degradation level and the health score corresponding to each degradation level are recorded. The correspondence between degradation level and health score is as follows: health score H = 1 - degradation level, that is, degradation level 0 corresponds to health score 1.00, degradation level 0.05 corresponds to health score 0.95, ..., degradation level 1.00 corresponds to health score 0; the simulation running state under each degradation level is input to the trained nonlinear mapping function. The projection position coordinates of the simulation running state in the unified reference coordinate system are obtained. ,in, The local coordinates of this degradation level on the health semantic axis; the coordinates of the reference health anchor point in the unified reference coordinate system. Using the origin as the reference point, calculate the Euclidean distance between the projected location of each degradation level and the baseline healthy anchor point. The distance between the health score H corresponding to each degradation level and the projected position of that degradation level is calculated. Associative storage is used to construct a one-to-one mapping table between health scores and theoretical semantic distances. range Each value in the equation uniquely corresponds to a theoretical semantic distance. Regarding health score exist The theoretical semantic distance corresponding to non-integer step sizes within the interval is obtained using linear interpolation, i.e., if Between and Between, among them, and For two adjacent health scores in the mapping table, the corresponding theoretical semantic distance is... ,in, and They are respectively and The corresponding theoretical semantic distance in the mapping table.

[0030] In one embodiment of the present invention, a feedback update step is further included: The corrected health assessment results are associated with the corresponding target domain data, final projection position, and semantic offset and stored in the historical database of the digital twin model; After each storage, the cumulative average of semantic offsets in the historical database is calculated. When the cumulative average is continuously greater than the preset error threshold (wherein the preset error threshold is related to the sensor measurement accuracy, and the preset trigger number is used to prevent false triggering caused by single error fluctuations), the cumulative number of storages reaches the preset trigger number. The target domain data stored in the historical database and its corresponding corrected health score are used as new supervision samples. After the incremental fine-tuning is completed, the output of the first branch in the fine-tuned health assessment network is used to recalculate the projection position of the health semantic axis in the unified reference coordinate system under each working condition node, update the nonlinear mapping function, and recalibrate the position of the health semantic anchor point under each working condition node with the updated nonlinear mapping function, and update the health semantic space.

[0031] Specifically, the revised health assessment results With the corresponding target domain data and final projection position and semantic offset The data is associated and stored in the historical database of the digital twin model. The historical database is a time-series database, and each record contains a timestamp, sensor signals and operating parameters of the target domain data, and , and After each storage operation, calculate the cumulative average of the semantic offsets in the historical database. Assume the historical database stores a total of [number missing] semantic offsets. There are [number] records, and the semantic offset of each record is [value]. The cumulative average of semantic offsets is ,Will Compared with the preset error threshold In comparison, in this embodiment The value is 0.05, when consecutive greater than The cumulative number of storage times has reached the preset trigger count. When an incremental adjustment is triggered, continuous cumulative statistics are used instead of a single judgment, in order to avoid false triggering caused by a single extreme data. Specifically, the target domain data and its corresponding corrected health scores stored in the historical database are used as new supervised samples to incrementally fine-tune the health assessment network. Specifically, the most recent L records are extracted from the historical database as the fine-tuning training set. In each training sample, the sensor signals and operating parameters from the target domain data are used as inputs to the health assessment network, along with their corresponding corrected health scores. As the supervision label for the second branch (fully connected mapping layer), it represents the final projection position of the target domain data in the healthy semantic space. The supervising label serves as the first branch (multilayer perceptron regression network). During fine-tuning, only the parameters of the fully connected mapping layer and the first branch are updated, while the graph attention network parameters remain unchanged. This is to preserve the learned spatial coupling features and temporal feature extraction capabilities. The loss function used in the fine-tuning... ,in, ,in, To fine-tune the batch size; In this embodiment, the value is 32. The Adam optimizer is used with a learning rate of 0.0001 (1 / 10 of the initial training learning rate of 0.001 to prevent significant parameter drift during fine-tuning), and training is performed for 50 epochs. After incremental fine-tuning, the output of the first branch in the fine-tuned health assessment network is used to recalculate the projection position of the health semantic axis in the unified reference coordinate system for each working condition node, and the nonlinear mapping function is updated. Specifically: The simulation running state corresponding to each degradation level under each working condition node is input into the fine-tuned first branch to obtain its projection position coordinates in the unified reference coordinate system. The mapping relationship of each working condition health semantic axis in the unified reference coordinate system is reconstructed using these coordinates. The updated projection positions of each working condition health semantic axis in the unified reference coordinate system are used as the new nonlinear mapping function. and with Recalibrate the health semantic anchor point positions (including baseline health anchor point and critical fault anchor point) under each operating condition node, and update the health semantic space.

[0032] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for assessing the health status of engineering operation and maintenance based on digital twins, characterized in that, Includes the following steps: S1: Collect multi-source operating data of engineering equipment under historical operating conditions, construct a digital twin model, and generate source domain data and target domain data through the digital twin model; S2: Set health semantic anchor points in the digital twin model, the health semantic anchor points include the baseline health state and critical fault state under each operating condition, and construct a cross-operating condition health semantic space based on the health semantic anchor points. S3: Construct a health assessment network and train the health assessment network using the source domain data; S4: Input the target domain data into the health assessment network to obtain an initial health score, and simultaneously calculate the projection position of the target domain data in the health semantic space using a digital twin model; S5: Based on the semantic distance between the projection position and the health semantic anchor point, calculate the semantic offset, and use the semantic offset to correct the initial health score to obtain the corrected health assessment result.

2. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, In step S1, constructing the digital twin model includes the following steps: The geometric parameters, material properties, and historical operation and maintenance records of the engineering equipment are acquired to construct a three-dimensional physical model. Based on the operating condition parameters and sensor signals in the multi-source operation data, a behavior model is constructed using multi-physics coupling simulation. The behavior model is associated with the three-dimensional physical model to form an initial twin. The preset operating condition parameter space is traversed, and several degradation levels are set along the full degradation path of the equipment at each operating condition node. The initial twin is driven to simulate and run at each degradation level to generate a multi-level source domain dataset containing operating condition annotations, degradation level annotations, and health semantic anchor point annotations. The real-time collected variable operating condition operation data is used as the target domain dataset.

3. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, In step S2, setting health semantic anchor points and constructing a cross-condition health semantic space includes the following steps: For each operating node, the steady-state operation of the digital twin model under rated parameters is defined as the baseline health anchor point; the state of the digital twin model when the key parameters reach the design critical threshold is defined as the critical failure anchor point. Using the baseline health anchor point and critical fault anchor point as endpoints and the degree of equipment degradation as a metric, a health semantic axis is constructed under the aforementioned operating conditions. Using the health semantic anchor point annotations of each sample in the multi-level source domain dataset as supervision signals, a nonlinear mapping function from the health semantic axis of each working condition to the unified reference coordinate system is learned. The health semantic axis under each working condition is projected onto the unified reference coordinate system to form a unified health semantic space across working conditions.

4. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, In step S3, constructing the health assessment network includes the following steps: A graph attention network is constructed, with each sensor signal in the multi-source operational data as a graph node. The spatial coupling weight between each node is calculated through the graph attention mechanism. Each sensor signal is sliced ​​according to a time window, and temporal features are extracted through a bidirectional long short-term memory network. The temporal features of each node are weighted and aggregated with the spatial coupling weight to form a health status feature vector. A dual-branch output structure connected to a graph attention network is constructed. The first branch of the dual-branch output structure is a multilayer perceptron regression network, which takes the health state feature vector as input and outputs the projection position coordinates in the health semantic space. The second branch is a fully connected mapping layer, which takes the health state feature vector as input and outputs a one-dimensional health score. The graph attention network and the dual-branch output structure together constitute a health assessment network; The degradation level labels in the multi-level source domain dataset are used as the supervision labels for the first branch, and the health semantic anchor labels in the multi-level source domain dataset are used as the supervision labels for the second branch. The health assessment network is jointly trained. During training, the health status feature vector is input to the working condition discriminator. The working condition discriminator is connected to the graph attention network through a gradient inversion layer. Adversarial training is performed with the working condition category as the prediction target. The gradient inversion layer transmits the feature vector identically during forward propagation and inverts the error gradient of the working condition discriminator before backpropagating it to the graph attention network during backward propagation.

5. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, In step S4, the projection position of the target domain data in the health semantic space is calculated using a digital twin model, including the following steps: Input the target domain data into the digital twin model for simulation and obtain the simulation running status corresponding to the target domain data; In the health semantic space, the membership degree of the simulation running state relative to the health semantic axis under each working condition node is calculated, and the working condition where the health semantic axis with the highest membership degree is located is taken as the current working condition belonging to the target domain data; at the same time, the membership degree value with the highest membership degree is used as the confidence weight. Obtain the coordinate transformation parameters between the local coordinate system and the unified reference coordinate system corresponding to the health semantic axis to which the current working condition belongs. The coordinate transformation parameters are determined by a nonlinear mapping function. The projection position coordinates output by the first branch in the health assessment network are transformed from the local coordinate system of the current working condition to the unified reference coordinate system using the coordinate transformation parameters to obtain the transformed coordinates. The simulation running state is mapped to a known coordinate position in the unified reference coordinate system corresponding to the degradation level of the simulation running state to obtain direct projected coordinates; The transformed coordinates and the direct projected coordinates are weighted and averaged using the confidence weights to obtain the final projection position of the target domain data in the cross-condition unified health semantic space.

6. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, In step S5, the semantic offset is calculated, and the initial health score is corrected using the semantic offset, including the following steps: Obtain the coordinates of the final projected position in the unified reference coordinate system, calculate the Euclidean distance between the final projected position and the reference health anchor point, and define it as the actual semantic distance. ; Obtain the initial health score from the output of the second branch in the health assessment network. The theoretical semantic distance corresponding to the initial health score is obtained by querying a preset mapping table. ; The difference between the actual semantic distance and the theoretical semantic distance is calculated as the semantic offset. The semantic offset is input into the nonlinear correction function to obtain the correction coefficient. ,in, For preset amplitude coefficient, The preset steepness coefficient; with the correction coefficient Multiply by the initial health score to obtain the corrected health assessment result. ,in, This is the corrected health score.

7. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 6, characterized in that, The method for constructing the preset mapping table includes the following steps: In the digital twin model, a baseline working condition is selected, which is the working condition node with the highest coverage density in the preset working condition parameter space; Under the baseline operating conditions, the digital twin model is driven to run in a step-by-step manner along the full degradation path of the equipment according to a preset step size, generating several degradation states covering the entire range of health scores. The projection position coordinates of each degradation state in the health semantic space are recorded, and the Euclidean distance between each projection position and the baseline health anchor point is calculated to construct a one-to-one correspondence mapping table between health scores and theoretical semantic distances.

8. The method for assessing the health status of engineering operation and maintenance based on digital twins according to claim 1, characterized in that, It also includes feedback update steps: The corrected health assessment results are associated with the corresponding target domain data, final projection position, and semantic offset and stored in the historical database of the digital twin model; After each storage, the cumulative average of the semantic offset in the historical database is calculated. When the cumulative number of storage times that the cumulative average is continuously greater than the preset error threshold reaches the preset trigger number, the target domain data stored in the historical database and its corresponding corrected health score are used as new supervision samples. After the incremental fine-tuning is completed, the output of the first branch in the fine-tuned health assessment network is used to recalculate the projection position of the health semantic axis in the unified reference coordinate system under each working condition node, update the nonlinear mapping function, and recalibrate the position of the health semantic anchor point under each working condition node with the updated nonlinear mapping function, and update the health semantic space.