Search

CN-122020373-A - Bayesian structure system identification method based on self-adaptive rotation element learning sampling

CN122020373ACN 122020373 ACN122020373 ACN 122020373ACN-122020373-A

Abstract

The invention provides a Bayesian structure system identification method based on self-adaptive rotation element learning sampling. According to the invention, the self-adaptive rotation element learning sampling method is provided, on the basis of the detailed probability distribution recognition result of the Bayesian structure system recognition method, the rotation invariance of the posterior trend characteristics is utilized, the sampling direction is self-adaptively rotated from each parameter direction related to a specific problem to each principal component direction consistent with the posterior trend, so that the high-efficiency sampler after neural network training in the method has wide universality, the problem independence is realized, the retraining requirement during task change is avoided, and the method is suitable for the complex structure automation refinement system recognition problem which is difficult to train, thereby better serving the structural health detection field.

Inventors

  • HUANG YONG
  • MENG XIANGHAO
  • LI HUI

Assignees

  • 哈尔滨工业大学

Dates

Publication Date
20260512
Application Date
20260122

Claims (10)

  1. 1. A bayesian structure system identification method based on adaptive rotation element learning sampling, the method comprising: Firstly, laying out a structural health monitoring system for a target structure, and obtaining non-normalized posterior distribution of model parameters through Bayesian inference based on a structural system model and structural health monitoring data; initializing a self-adaptive rotating element learning sampling method, and introducing algorithm default or locally trained neural network parameters into a self-adaptive rotating element learning sampler; step three, selecting whether to train the self-adaptive rotation element learning sampler, if so, entering step four, otherwise, entering step five; step four, based on the unconditional posterior distribution, running the self-adaptive rotation element learning sampling method in a training mode, training the self-adaptive rotation element learning sampler, and obtaining and storing the trained neural network parameters; And fifthly, based on the unconditional posterior distribution, closing the training mode and running the self-adaptive rotation element learning sampling method to obtain a parameter posterior distribution sample, so as to realize the recognition of the Bayesian structure system.
  2. 2. The method of claim 1, wherein the adaptive rotation element learning sampling in step four and step five is based on a non-normalized parameter posterior distribution Direction estimation method and rotation processing by utilizing fusion self-adaptive principal component , , AM-SGHMC algorithm of (a) a process for obtaining a parameter posterior distribution sample, wherein For the structural health monitoring data, Is a structural model parameter vector, its dimension Namely the number of the parameters is the number, For an adaptive principal component representation of the parameter vector, for AM-SGHMC samples, Selecting parallelism based on computing device performance constraints and computing efficiency requirements A strip markov chain.
  3. 3. The method according to claim 2, wherein the adaptive rotation element learning sampling process is specifically: Step 4.1, initializing a rotation matrix ; Step 4.2, for each chain, randomly generating around the nominal value of the parameter Setting auxiliary variables Thereby obtaining an initial augmentation vector sample Order-making ; Step 4.3, when If not, enter step 4.4 to generate a new sample, otherwise enter step 4.11 to simulate the number of steps ; Step 4.4, when in the self-adaptive adjustment stage, entering step 4.5, otherwise, maintaining the rotation matrix Unchanged, go to step 4.7, formally sampling the starting point The time, the self-adaptive adjustment stage is correspondingly taken as the formal sampling starting point If the section is in the training mode, the self-adaptive adjustment stage is taken as A section; Step 4.5, calculating the current parameter sample And based on the sample Performing adaptive principal component direction estimation, and performing current rotation matrix Updated to ; Step 4.6, updating the sample to the corresponding direction of the new rotation matrix, , ; Step 4.7 at present In the corresponding direction, the current sample Performing one-step AM-SGHMC algorithm sampling step, and simulating to obtain a new sample ; Step 4.8, order ; Step 4.9, if the training mode is adopted, entering step 4.10, otherwise, directly returning to step 4.3; Step 4.10, executing an AM-SGHMC algorithm neural network updating step, and returning to the step 4.3; Step 4.11, if the training mode is adopted, the neural network parameters are saved, otherwise, the formal sampling stage is output Corresponding all parameter samples 。
  4. 4. A method according to claim 3, characterized in that the parameter samples entered each time in step 4.5 are noted as Adaptive adjustment phase Wherein Representing that it comes from the first A strip Markov chain for developing the current rotation matrix of each input as the direction vector of each principal component Establishing a main component direction self-adaptive estimator, setting the index attenuation rate array of the estimation process as In which is arranged Is a broken line type array Initializing the statistics of the built-in principal components to be when the call is initially made Wherein Is the dimension statistics built in the AM-SGHMC algorithm.
  5. 5. The method according to claim 4, wherein the adaptive principal component direction estimation method in step 4.5 specifically comprises: Step 4.5.1, calling the parameter mean statistics built in the AM-SGHMC algorithm Centering parameter samples ; Step 4.5.2 press Rearrangement in descending order And corresponding to And ; Step 4.5.3, pairing Sequentially performing steps 4.5.4 to 4.5.9; step 4.5.4 calculating a centralized parameter sample Is the first of (2) Dimension orthogonal margin Convention of ; Step 4.5.5 orthogonalization update Principal component statistics of vitamin origin ; Step 4.5.6 orthogonalizing update Direction of principal component of vitamin ; Step 4.5.7, calculate the first Dimension principal component input item ; Step 4.5.8, update the first Principal component statistics of dimensions ; Step 4.5.9, calculate the first Direction of main component ; Step 4.5.10, merging the direction vectors of the main components to obtain a new rotation matrix 。
  6. 6. The method according to claim 5, wherein the first step The formula of the dimension principal component statistics is: 。
  7. 7. The method according to claim 6, wherein the first The formula for maintaining the direction of the main component is as follows: 。
  8. 8. the method of claim 7, wherein the new rotation matrix formula is: 。
  9. 9. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1-8 when the computer program is executed.
  10. 10. A computer readable storage medium storing computer instructions which, when executed by a processor, implement the steps of the method of any one of claims 1-8.

Description

Bayesian structure system identification method based on self-adaptive rotation element learning sampling Technical Field The invention relates to the technical field of structural system identification and structural health monitoring, in particular to a Bayesian structural system identification method based on self-adaptive rotation element learning sampling. The method is suitable for Bayesian structure system identification. Background In the field of structural health monitoring, bayesian inference is commonly used for structural system identification, while Markov Chain Monte Carlo (MCMC) sampling methods are widely studied as powerful calculation tools in the aspect of numerical simulation of unconcerned posterior distribution in bayesian inference to avoid the difficulty of high-dimensional integral analysis and calculation. The MCMC method generates posterior distribution samples by simulating a specific Markov chain, researchers continuously improve the design of the Markov chain, namely a sampling strategy, in the MCMC method, and sequentially propose various classical algorithms. With the development of the neural network, some algorithms further improve the flexibility of sampling strategy design by embedding the neural network in the MCMC algorithm, and simultaneously realize the automatic design aiming at specific problems through the training of the neural network. However, design optimization for specific problems often loses the versatility of the sampler, resulting in the need for retraining optimization in new tasks, which takes time to counteract the efficiency improvement caused by design optimization. The latest sampling method such as AM-SGHMC is used for reducing the difference of the neural networks required by different tasks by providing a meta-learning technology for improving the embedding mode of the neural networks, realizing high-efficiency sampling aiming at specific problems, improving the universality of the sampler after training and reducing the training time. However, in order to have the universality among sampling tasks with different dimensions, the neural network is limited to process the input-output relationship according to the dimensions, and the parameter posterior trend cannot be considered, namely the influence of the overall trend of a high-probability area of parameter posterior distribution in a high-dimensional parameter space is not considered, so that the universality and the sampling efficiency are limited. Meanwhile, the parameter class coding in the network input also enables the network input to be only applicable to Bayesian inference of the similar structure model. In the system identification practice, the posterior trend of different problem parameters is changeable, and part of structural types are large in volume and difficult to train, so that a new efficient sampling method which breaks through the limit of the posterior trend of the parameters and is universal across the structural types is urgently needed to be researched. Disclosure of Invention The invention aims to solve the problems in the prior art and provides a Bayesian structure system identification method based on self-adaptive rotation element learning sampling. According to the method, the current sample principal component direction is iteratively updated by establishing a self-adaptive principal component direction estimation method, so that the sampling strategy processing direction in the AM-SGHMC is self-adaptively rotated from each parameter direction related to a specific problem to each principal component direction consistent with the posterior trend, and parameter class codes in network input are removed, the limitation of the parameter posterior trend and the structure type is broken through, and the identification efficiency and the universality of the structural system are synchronously improved. The invention is realized by the following technical scheme, and provides a Bayesian structure system identification method based on self-adaptive rotation element learning sampling, which comprises the following steps: Firstly, laying out a structural health monitoring system for a target structure, and obtaining non-normalized posterior distribution of model parameters through Bayesian inference based on a structural system model and structural health monitoring data; initializing a self-adaptive rotating element learning sampling method, and introducing algorithm default or locally trained neural network parameters into a self-adaptive rotating element learning sampler; step three, selecting whether to train the self-adaptive rotation element learning sampler, if so, entering step four, otherwise, entering step five; step four, based on the unconditional posterior distribution, running the self-adaptive rotation element learning sampling method in a training mode, training the self-adaptive rotation element learning sampler, and obtaining and storing the trained neural