CN113743019A - Digital twin enhanced complex equipment health monitoring method - Google Patents

Digital twin enhanced complex equipment health monitoring method Download PDF

Info

Publication number
CN113743019A
CN113743019A CN202111084393.5A CN202111084393A CN113743019A CN 113743019 A CN113743019 A CN 113743019A CN 202111084393 A CN202111084393 A CN 202111084393A CN 113743019 A CN113743019 A CN 113743019A
Authority
CN
China
Prior art keywords
model
digital twin
data
physical system
physical
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111084393.5A
Other languages
Chinese (zh)
Inventor
于劲松
宋悦
周倜
唐卿
梁思远
唐荻音
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202111084393.5A priority Critical patent/CN113743019A/en
Publication of CN113743019A publication Critical patent/CN113743019A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

The invention discloses a health monitoring method for digital twin enhanced complex equipment. The method comprises three parts of physical system interface design, digital twin model design, data synchronization between a physical system and a digital twin model and model update. The physical system interface transmits various monitoring data of the complex equipment to the virtual model by acquiring. The digital twin model employs a nonparametric Bayesian network based representation of the dynamic degradation process of the state of health and the propagation of cognitive uncertainty. The data synchronization and model update between the physical system and the virtual model comprise two parts, on one hand, for those parameters in the nonparametric Bayesian network with prior models, the improved GPF is used for updating them in real time; on the other hand, for the parameters lacking the prior model, the DPMM learning hidden variable is provided, the model structure is updated in a self-adaptive mode, and the accuracy of health monitoring is effectively improved.

Description

Digital twin enhanced complex equipment health monitoring method
Technical Field
The invention relates to a method for solving the problem of complex equipment health monitoring by using a digital twinning method, in particular to a digital twinning method based on a nonparametric Bayesian network.
Background
In order to ensure safe and reliable operation of a complex system, fault Prediction and Health Management (PHM) technology has become a popular topic in recent years. PHM was originally proposed in the field of military aviation in the nineties of the twentieth century, along with the united states army combat Fighter (JSF) project. The PHM technology combines the inherent physical mechanism of the target system and the sensor data obtained in actual operation, and adopts machine learning, artificial intelligence algorithm, etc. to realize the health monitoring and the Remaining service Life prediction (RUL) of the complex system, and can actively take the pre-maintenance measures to avoid the occurrence of serious accidents. Therefore, the PHM technology is receiving attention in more and more fields.
Health monitoring is an important link in PHM technology. The health monitoring process models the system degradation process by considering current conditions, environments, etc. to estimate the current and future health of the system. The prediction methods proposed at present mainly include data-driven methods and physical model-based methods. Data-driven based methods extract useful features from the collected data to characterize the current state, thereby modeling the degradation trend. Physical model-based methods use mathematical formulas to generalize the relevant physics of system degradation behavior, which assumes that system behavior can be accurately analyzed and characterized. In contrast to physical model-based approaches, data-driven approaches do not require a complete understanding of the physical mechanisms of the system. However, the data-driven approach requires more data than the physical model-based approach.
The invention provides a health monitoring method based on digital twin, which can combine the advantages of a method based on a physical model and a method based on data driving and combine the physical mechanism of complex equipment with data, thereby improving the accuracy of health monitoring and having stronger adaptability under different working conditions. The first concept of the digital twin derives from the presentation of the Air Force Research Laboratory (AFRL) with the aim of achieving comprehensive diagnosis and preliminary maintenance to ensure safe operation throughout the life cycle of the flight system. The digital twin was originally dedicated to the modeling, simulation and visualization of complex systems. However, with the development of CPS and internet of things (IoT), the digital twin architecture is more concerned about real-time interaction and data communication between the physical world and the information world. A digital twin model based on a nonparametric Bayesian network is established to carry out health monitoring on complex equipment, and a data synchronization and model updating method between a physical system and the digital twin model is provided.
Disclosure of Invention
The invention provides a health monitoring method of complex equipment based on digital twins, aiming at solving the problem of health management of the complex equipment.
The digital twin-based complex equipment health monitoring algorithm has the main innovation points that: the method comprises three parts of physical system interface design, digital twin model design, data synchronization between a physical system and a virtual model and model update.
The physical system can be any industrial complex equipment, and the physical system interface design provides a data interface between the physical system and the digital twin model and is used for collecting various types of data capable of reflecting the health condition of the equipment. The data acquisition board card on the physical system sends a data packet to the upper computer through the serial port, and the upper computer unpacks the data packet to obtain information capable of completely reflecting the state of the physical system and sends the information to the digital twin model. Note that the entire physical system needs to use a common serial protocol and data resolution protocol.
The digital twin model is built through a nonparametric Bayesian network according to the physical mechanism of the complex device. For complex devices, on the one hand, part of their physical mechanisms (e.g., electromechanical, control, etc.) can be described by certain mathematical formulas, which can be converted into bayesian network models that determine the structure; on the other hand, uncertain influence factors in the complex equipment can be regarded as implicit variables, and are expressed as nodes with unknown prior structures in the Bayesian network. Thus, the digital twin model of the entire physical system can be expressed in the form of a non-parametric bayesian network.
The data synchronization and Model update between the physical system and the virtual Model comprise two Model parameter updating modes, namely Gaussian Particle Filter (GPF) Model reasoning based on kernel smoothing and Model self-learning based on a Dirichlet Process Mix Model (DPMM). The non-parametric Bayesian network established by the invention is unknown in local model, and two types of nodes exist in the network and need to be updated along with real-time data. One is a node with a known prior model structure, and the node realizes real-time reasoning through a Gaussian particle filter algorithm based on kernel smoothing; the other is a node with an unknown prior model structure, and the hidden variables of the node are estimated by a Dirichlet process mixed model algorithm, so that the model has the structure self-learning capability.
The invention can be applied to any complex equipment, wherein a digital twin model is established in an upper computer, the health state of a physical system can be monitored in real time by combining a physical mechanism and data, and uncertain factors in the physical mechanism can be taken into consideration, so that the accuracy of health monitoring is greatly improved.
The invention has the advantages that:
(1) according to the invention, the health monitoring of complex equipment is carried out by constructing a digital twin model based on a nonparametric Bayesian network, so that the dynamic degradation process of the health state and the propagation of the uncertainty of the prior knowledge can be represented.
(2) The invention provides a novel data synchronization and model updating method between a physical system and a digital twin model. Real-time updating of the parameters of the digital twin model is achieved by using the improved GPF, so that synchronous evolution with a real physical system is achieved. For the parameters lacking the prior model, the DPMM learning hidden variable is provided, the model structure is updated in a self-adaptive mode, the uncertainty is greatly reduced, and the accuracy of health monitoring is improved.
Drawings
FIG. 1 is a digital twinning reference frame diagram provided by the present invention;
FIG. 2 is a diagram of a non-parametric Bayesian network architecture in accordance with the present invention;
fig. 3 is a schematic diagram of a DPMM provided by the present invention.
Detailed Description
The digital twin-based complex equipment health monitoring method provided by the invention is described in detail below with reference to the accompanying drawings.
The digital twin reference frame provided by the invention is shown in figure 1. The digital twin framework is composed of three parts, namely a physical system, a digital twin model and interaction between the digital twin model and the physical system. The physical system is used for collecting information which can reflect the actual working state of the equipment, such as environment, operation, working condition, perception and the like. The digital twin model comprises three parts of digital simulation, a physical model and data fusion, wherein the digital simulation is used for simulating the machinery, the structure and the like of the equipment and providing a visual display function; the physical model refers to a nonparametric Bayesian network model; data fusion refers to the combination of various types of data from physical systems that can represent health status. The interaction between the digital twin model and the physical system comprises three parts of data interaction, model updating and service, wherein the data interaction transmits data from the physical system to the digital twin model through a communication interface and a communication protocol; the model updating is to realize the real-time updating of the digital twin model by means of 3D visualization, big data, machine learning, statistical model and the like on the real-time data; the services refer to services such as fault diagnosis, health monitoring, optimized maintenance and the like provided for a physical system by a digital twin model.
The digital twin model based on the nonparametric Bayesian network structure provided by the invention is shown in FIG. 2. The nonparametric Bayesian network comprises five node types including a degenerate node, a static node, a system state node, an observation node, an implicit node and the like. In addition, the subscript t-1 or t indicates the time point, and a dotted line is used when connecting nodes across two different time points, and a solid line is used when connecting nodes within one time point. . An elliptical node is a degenerate node, representing a node that degrades over time, with the arrow pointing to it representing a conditional probability distribution. Circular nodes represent static nodes, representing nodes that do not change over time but may have unknown values for specific parameters, which may introduce data uncertainty. A triangle node is a system state node, computed from the value of the parent node, and the arrow pointing to it represents a deterministic function. The rectangular nodes represent observation nodes and represent variables observed by the sensor. In addition, the dotted nodes represent implicit nodes whose probability distribution functions are unknown, i.e., implicit variables. The number of hidden variables is uncertain, may change with the change of actual data, and is the node used to describe the model uncertainty. Two types of nodes exist in the nonparametric Bayesian network and need to be updated along with real-time data, one type of node is a node with a known prior model structure, such as a system state node, and the node realizes real-time reasoning through a Gaussian particle filter algorithm; the other is an implicit node with an unknown prior model structure, and the implicit variable of the node is estimated through a Dirichlet process mixed model algorithm, so that the model has the structure self-learning capability.
The Gaussian particle filter model inference based on the kernel smoothing can realize real-time updating of model parameters. By adding a kernel smoothing process in Gaussian particle filtering, the kernel smoothing algorithm can greatly improve the convergence speed of state and parameter estimation, and the variance of the particle set is kept unchanged after multiple updating iterative computations. Kernel Smoothing (KS) involves two steps of contraction and vibration.
The method comprises the following steps:
shrinking: in order to accelerate the convergence speed of the system state and the model parameters, an adjustment term is added to each particle in the particle set so that the particle is closer to the expected value of the particle set, and the formula is as follows:
Figure BDA0003264913040000041
wherein the content of the first and second substances,
Figure BDA0003264913040000042
and
Figure BDA0003264913040000043
respectively before and after the shrinkage of the ith particle at time t,
Figure BDA0003264913040000044
is the expected value of the particle set, h is the contraction factor, a larger h represents a larger contraction forceThe faster the system state and model parameters converge, the more recent h ∈ [0,1 ]]。
Step two:
vibration: adding a bias term to each particle in the set of particles to ensure that the set of particles remains variance unchanged during the update iteration, the vibration process can be expressed as:
Figure BDA0003264913040000045
wherein the content of the first and second substances,
Figure BDA0003264913040000046
represents the variance before contraction, after which the variance becomes
Figure BDA0003264913040000047
Vibration by adding bias terms
Figure BDA0003264913040000048
The variance is automatically adjusted.
The model self-learning based on the Dirichlet process hybrid model provided by the invention is shown in FIG. 3. Where O is the observed value for each time step, H is the hidden variable, μ and σ are the mean and standard deviation of the hidden variable, and H is the hyperparameter. Assuming the parametric observation is x, x is from a distribution with M hidden variables, each hidden variable j follows a normal distribution
Figure BDA0003264913040000049
The probability of x can be calculated by:
Figure BDA00032649130400000410
wherein ω isjThe weight of the jth hidden variable is obtained, the value of the observation data at the moment is obtained after all the hidden variables are subjected to weighted summation, and the weight meets the requirement
Figure BDA00032649130400000411
The DPMM realizes the autonomous updating of the non-parametric Bayesian network model by setting the prior distribution of Dirichlet Process (DP) and approximating the posterior distribution of DP by using a Gibbs sampling method, so as to obtain the number of hidden variables of observed data and the parameter of each hidden variable j
Figure BDA00032649130400000412
The method comprises the following steps:
setting the prior distribution of DP:
given the previous N-1 time step data, a new observation xtThe probability from the hidden variable j can be written as:
Figure BDA00032649130400000413
where α is a concentration parameter, representing the probability that we expect to see a new hidden variable. n isjIs the number of observations, c, obtained from the hidden variable j in the previous observation of N-1 time steptDenotes xtFrom the hidden variable j.
To obtain an analytical solution for the posterior distribution and simplify the calculation, μjAnd
Figure BDA0003264913040000051
the conjugate distribution of (a) is given by:
Figure BDA0003264913040000052
where m is the prior average of the observations and h is the hyperparameter. Γ (a, b) is the gamma distribution with a shape parameter a and a scale parameter b.
Step two:
estimate the posterior distribution of DP using gibbs sampling:
about posterior distribution
Figure BDA0003264913040000053
The exact reasoning of (a) is difficult to achieve in practice, so we use an iterative process called Gibbs (Gibbs) sampling to sample the relevant parameters afterwards. To apply Gibbs sampling to inference, we need to iterate over each conditional posterior distribution in turn. The specific process is as follows:
sampling ct: at a given parameter
Figure BDA0003264913040000054
Re-sampling of ct. Deriving x from an existing hidden variable jtThe probability of (c) is:
Figure BDA00032649130400000513
xtthe probability from the new hidden variable θ is calculated by the following equation:
Figure BDA0003264913040000055
② resampling muj: given ctSum variance
Figure BDA0003264913040000056
Resampling the mean μ from the conditional posterior distributionj：
Figure BDA0003264913040000057
Wherein
Figure BDA0003264913040000058
③ resampling
Figure BDA0003264913040000059
According to the posterior distribution pair under the condition of giving other parameters
Figure BDA00032649130400000510
Resampling is carried out:
Figure BDA00032649130400000511
the parameters of each hidden variable j after setting the prior distribution of DP and estimating the posterior distribution of DP by Gibbs sampling
Figure BDA00032649130400000512
An estimate is obtained and after obtaining the observation at each time point it is calculated whether the observation is from a new hidden variable. Therefore, after the updating of each time point, the number of all hidden variables and the corresponding parameters can be obtained, and the autonomous learning and updating of the whole model are realized.
After the hidden variable estimation of the unknown model parameters is realized, the whole non-parametric Bayesian network is updated in real time through observation data, and the autonomous updating of the model structure ensures the reduction degree of the model uncertainty and the accuracy of the health evaluation result.
The algorithm of the invention is partially developed by matlab.

Claims (4)

1. A health monitoring method for complex equipment with enhanced digital twins is characterized in that: physical system interface design, digital twin model design, data synchronization between the physical system and the digital twin model, model update and the like. The digital twin Model is established through a nonparametric Bayesian network, and data synchronization and Model updating between the physical system and the digital twin Model are realized through Gaussian Particle Filter (GPF) Model reasoning and a Dirichlet Process Mix Model (DPMM) algorithm.
2. The physical system interface design of claim 1 obtains various data of the complex device through intelligent perception for real-time updating of the digital twin model, mainly comprising data capable of reflecting the health condition of the device, such as physical modeling, working state, operation and environmental conditions.
3. A nonparametric Bayesian network graph model of a complex plant system is established for the digital twin model design of claim 1, with its internal symbols representing system states and parameters. The graph model is built based on the internal physical mechanism of a complex system, and nodes of the graph model are connected through arrows which represent the propagation of conditional probability distribution and uncertainty sources. In non-parametric modeling, the structure of the bayesian network is not completely determined, so the model complexity can be adaptively adjusted according to the data complexity.
4. Data synchronization and model update between a physical system and a digital twin model according to claim 1 comprises two parts. On one hand, the Gaussian particle filter model inference based on kernel smoothing is used for updating parameters with prior models in a nonparametric Bayesian network in real time, so that the system state and the model parameters are estimated simultaneously, and the uncertainty of data and models is continuously reduced; on the other hand, model self-learning based on the Dirichlet process mixed model is used for estimating unknown hidden parameters of the prior model, so that the number and the probability distribution function of hidden variables are obtained, and the autonomous learning and model updating of the health monitoring digital twin model are realized.
CN202111084393.5A 2021-09-16 2021-09-16 Digital twin enhanced complex equipment health monitoring method Pending CN113743019A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111084393.5A CN113743019A (en) 2021-09-16 2021-09-16 Digital twin enhanced complex equipment health monitoring method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111084393.5A CN113743019A (en) 2021-09-16 2021-09-16 Digital twin enhanced complex equipment health monitoring method

Publications (1)

Publication Number Publication Date
CN113743019A true CN113743019A (en) 2021-12-03

Family

ID=78739385

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111084393.5A Pending CN113743019A (en) 2021-09-16 2021-09-16 Digital twin enhanced complex equipment health monitoring method

Country Status (1)

Country Link
CN (1) CN113743019A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114943185A (en) * 2022-07-13 2022-08-26 北京航空航天大学 Nondestructive acquisition method for twin data of complex equipment under extreme working conditions
CN114943304A (en) * 2022-06-17 2022-08-26 广东工业大学 Bayesian-based 3C electronic product assembly error transfer method
CN115292834A (en) * 2022-07-20 2022-11-04 北自所(北京)科技发展股份有限公司 Method, device and system for diagnosing faults of digital twin equipment

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114943304A (en) * 2022-06-17 2022-08-26 广东工业大学 Bayesian-based 3C electronic product assembly error transfer method
CN114943185A (en) * 2022-07-13 2022-08-26 北京航空航天大学 Nondestructive acquisition method for twin data of complex equipment under extreme working conditions
CN114943185B (en) * 2022-07-13 2022-09-30 北京航空航天大学 Nondestructive acquisition method for twin data of complex equipment under extreme working conditions
CN115292834A (en) * 2022-07-20 2022-11-04 北自所(北京)科技发展股份有限公司 Method, device and system for diagnosing faults of digital twin equipment

Similar Documents

Publication Publication Date Title
CN113743019A (en) Digital twin enhanced complex equipment health monitoring method
Lv et al. Artificial intelligence for securing industrial-based cyber–physical systems
US20200272905A1 (en) Artificial neural network compression via iterative hybrid reinforcement learning approach
CN105787563B (en) A kind of Rapid matching fuzzy reasoning method based on self-study mechanism
CN106953862A (en) The cognitive method and device and sensor model training method and device of network safety situation
Yesil et al. Fuzzy cognitive maps learning using artificial bee colony optimization
CN113313947A (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN111240350A (en) Unmanned aerial vehicle pilot dynamic behavior evaluation system
Zheng et al. Parameter identification of nonlinear dynamic systems using an improved particle swarm optimization
US20210158162A1 (en) Training reinforcement learning agents to learn farsighted behaviors by predicting in latent space
Wang et al. Event-triggered consensus control for second-order multi-agent system subject to saturation and time delay
CN110309537A (en) A kind of the intelligent health prediction technique and system of aircraft
CN112434813B (en) Multi-multidimensional language reasoning method based on attribute-oriented language concept lattice
CN111797815A (en) Complex equipment health assessment method and system based on deep learning
CN114919559A (en) System and method for predicting residual service life of braking system based on digital twinning
CN114819054A (en) Power electronic system state monitoring method based on physical information neural network
CN115809596A (en) Digital twin fault diagnosis method and device
He et al. Influence-augmented online planning for complex environments
Koutsoukos Optimal control of stochastic hybrid systems based on locally consistent Markov decision processes
CN111144572A (en) Power distribution network disaster situation inference method and system based on tree-shaped Bayesian network
Montana et al. Towards a unified framework for learning from observation
Li et al. Evolvehypergraph: Group-aware dynamic relational reasoning for trajectory prediction
Perreault et al. Deriving prognostic continuous time Bayesian networks from D-matrices
CN115212549A (en) Adversary model construction method under confrontation scene and storage medium
Liu et al. Imprecise fusion operators for collective learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination