Bayesian structure system identification method based on adaptive rotation element learning sampling

Abstract

The invention provides a Bayesian structure system identification method based on adaptive rotation element learning sampling. According to the method, a self-adaptive rotation element learning sampling method is provided, and on the basis of a detailed probability distribution identification result of a Bayesian structure system identification method, rotation invariance of posterior trend characteristics is utilized to adaptively rotate a sampling direction from each parameter direction related to a specific problem to each principal component direction consistent with a posterior trend, so that the probability distribution identification result of the Bayesian structure system identification method is identified. According to the method, the efficient sampler after neural network training in the method has wide universality irrelevant to problems, the re-training requirement during task change is avoided, the method is suitable for the recognition problem of a complex structure automatic refined system which is difficult to train, and therefore the method better serves the field of structure health detection.

Description

Technical Field

[0001] This invention relates to the field of structural system identification and structural health monitoring technology, and in particular to a Bayesian structural system identification method based on adaptive rotational element learning sampling. The method is applicable to Bayesian structural system identification. Background Technology

[0002] In the field of structural health monitoring, Bayesian inference is often used for structural system identification, while Markov chain Monte Carlo (MCMC) sampling method, as a powerful computational tool, has been widely studied in the numerical simulation of non-normalized posterior distributions in Bayesian inference to avoid the problem of high-dimensional integral analytical computation.

[0003] The Markov Chain Motion (MCMC) method generates posterior distribution samples by simulating a specific Markov chain. Researchers have continuously improved the design of the Markov chain, i.e., the sampling strategy, in the MCMC method, and have proposed several classic algorithms. With the development of neural networks, some algorithms have further improved the flexibility of sampling strategy design by embedding neural networks into the MCMC algorithm, and have also achieved automated design for specific problems through neural network training. However, design optimization for specific problems often sacrifices the generality of the sampler, resulting in the need for retraining and optimization in new tasks. The training time actually offsets the efficiency improvement brought by design optimization.

[0004] Recent sampling methods, such as AM-SGHMC, propose a meta-learning technique that improves neural network embedding, reducing the differences in neural networks required for different tasks. This achieves efficient sampling for specific problems while improving the generality of the sampler after training and reducing training time. However, to achieve generality across sampling tasks of different dimensions, it restricts the neural network to process input-output relationships only according to the dimension, failing to consider the posterior trend of parameters—that is, the influence of the overall trend of high-probability regions of the posterior distribution in the high-dimensional parameter space—thus limiting its generality and sampling efficiency. Furthermore, the parameter category encoding in its network input limits its applicability only to Bayesian inference of models with similar structures. In system recognition practice, the posterior trends of parameters vary greatly for different problems, and some structural types are large and difficult to train. Therefore, there is an urgent need to research new, efficient sampling methods that overcome the limitations of posterior trend constraints and are applicable across structural types. Summary of the Invention

[0005] The purpose of this invention is to address the problems in existing technologies by proposing a Bayesian structural system recognition method based on adaptive rotational meta-learning sampling. This method establishes an adaptive principal component direction estimation method, iteratively updating the principal component directions of the current samples. By adaptively rotating the sampling strategy processing direction in AM-SGHMC from the directions of parameters relevant to the specific problem to the directions of principal components consistent with the posterior trend, and eliminating parameter category encodings in the network input, this method overcomes the limitations of posterior trend and structural type, simultaneously improving the efficiency and versatility of structural system recognition.

[0006] This invention is achieved through the following technical solution: This invention proposes a Bayesian structure system recognition method based on adaptive rotation element learning sampling, the method comprising: Step 1: For the target structure, deploy a structural health monitoring system. Based on the structural system model and structural health monitoring data, obtain the unnormalized posterior distribution of the model parameters through Bayesian inference. Step 2: Initialize the adaptive rotation meta-learning sampling method by importing the default or locally trained neural network parameters into the adaptive rotation meta-learning sampler. Step 3: Select whether to train the adaptive rotation meta-learning sampler. If yes, proceed to Step 4; otherwise, proceed to Step 5. Step 4: Based on the unnormalized posterior distribution, run the adaptive rotation meta-learning sampling method in training mode to train the adaptive rotation meta-learning sampler, and obtain and save the trained neural network parameters. Step 5: Based on the unnormalized posterior distribution, turn off the training mode and run the adaptive rotation meta-learning sampling method to obtain the parameter posterior distribution samples, thereby realizing the recognition of Bayesian structure systems.

[0007] Furthermore, the adaptive rotational meta-learning sampling process in steps four and five is based on the unnormalized parameter posterior distribution. The method utilizes a fusion adaptive principal component orientation estimation method and rotation processing. , , The AM-SGHMC algorithm is used to obtain the posterior distribution of the parameters in the sample. For structural health monitoring data, The structural model parameter vector has the following dimensions. That is, the number of parameters. An adaptive principal component representation of the parameter vector, used for AM-SGHMC sampling. An adaptive rotation matrix is ​​used; parallel processing is selected based on the performance limitations of the computing device and the computational efficiency requirements. A Markov chain.

[0008] Furthermore, the adaptive rotation element learning sampling process is specifically as follows: Step 4.1: Initialize the rotation matrix ; Step 4.2: For each chain, randomly generate parameters around their nominal values. Set auxiliary variables Thus, the initial augmented vector samples are obtained. ,make ; Step 4.3, when If the condition is met, proceed to step 4.4 to generate a new sample; otherwise, proceed to step 4.11. The total number of simulation steps is [number missing]. ; Step 4.4: If in the adaptive adjustment phase, proceed to step 4.5; otherwise, maintain the rotation matrix. No change, proceed to step 4.7, and set the starting point for formal sampling. At any given time, the adaptive adjustment phase will be set accordingly based on the starting point of the formal sampling. For intervals, if it is training mode, no formal sampling is required; the adaptive adjustment phase takes the interval as 1. interval; Step 4.5: Calculate the current parameter sample. and based on samples Perform adaptive principal component orientation estimation, and then use the current rotation matrix. Updated to ; Step 4.6: Update the sample to the corresponding direction of the new rotation matrix. , ; Step 4.7, in the current In the corresponding direction, for the current sample Perform one sampling step of the AM-SGHMC algorithm to simulate and obtain a new sample. ; Step 4.8, let ; Step 4.9: If it is in training mode, proceed to step 4.10; otherwise, return directly to step 4.3. Step 4.10: Execute the AM-SGHMC algorithm neural network update step, then return to step 4.3; Step 4.11: If in training mode, save the neural network parameters; otherwise, output the formal sampling phase. All corresponding parameter samples .

[0009] Furthermore, the parameter samples input in each step 4.5 are denoted as... Adaptive adjustment phase ,in This represents its origin from the first A Markov chain; the current rotation matrix of each input is expanded and denoted as the direction vector of each principal component. Establish a principal component orientation adaptive estimator, and set the exponential decay rate series of the estimation process as follows: , where settings A broken line sequence The built-in principal component statistics are initialized upon initial call. ,in These are the scale statistics built into the AM-SGHMC algorithm.

[0010] Furthermore, the adaptive principal component direction estimation method in step 4.5 is specifically as follows: Step 4.5.1: Call the built-in parameter mean statistic in the AM-SGHMC algorithm. Center the parameter samples ; Step 4.5.2, Press Rearrange in descending order and its corresponding and ; Step 4.5.3, for Perform steps 4.5.4 through 4.5.9 sequentially; Step 4.5.4: Calculate the centered parameter sample The Orthogonal residual Agreement ; Step 4.5.5, Orthogonalization Update Principal component statistics ; Step 4.5.6, Orthogonalization Update Principal component direction of vitamin ; Step 4.5.7, Calculate the first... Principal component input item ; Step 4.5.8, Update the first Principal component statistics ; Step 4.5.9, Calculate the first... Innovation principal component direction ; Step 4.5.10: Merge the direction vectors of each principal component to obtain a new rotation matrix. .

[0011] Furthermore, the first The formula for the principal component statistic is: .

[0012] Furthermore, the first The formula for the direction of the principal component in the reformation is: .

[0013] Furthermore, the formula for the new rotation matrix is: .

[0014] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the Bayesian structure system recognition method based on adaptive rotation element learning sampling.

[0015] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the Bayesian structure system recognition method based on adaptive rotation element learning sampling.

[0016] The beneficial effects of this invention are: 1. This invention is a structural system identification method based on Bayesian inference. Compared with non-probabilistic methods, the identification result of this invention is a more detailed posterior probability distribution of parameters, which can be used to calculate reference indicators such as confidence intervals in practical applications. 2. This invention employs a neural network-enhanced sampling method. Compared with traditional sampling methods, the neural network can automatically and finely adjust the sampling strategy through training, significantly improving sampling efficiency. 3. This invention innovatively designs an adaptive rotation meta-learning sampling method, which has a wider distribution level versatility compared with existing neural networks or meta-learning methods, avoids the need for retraining when the task category changes, and is suitable for structural system identification problems that are difficult to train, such as large bridge finite element models. 4. This invention can adaptively handle the posterior distribution sampling problem in high-dimensional parameter space, and is suitable for automated multi-parameter fine-grained structural system identification of large structures. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1This is a flowchart of the Bayesian structure system recognition method based on adaptive rotation element learning sampling described in this invention; Figure 2 This is a flowchart of the adaptive rotation element learning sampling process involved in this invention; Figure 3 This is an overall implementation plan diagram of a building Bayesian structural system identification training task and two cross-category bridge Bayesian structural system identification generality test tasks executed in an embodiment of the present invention; Figure 4 The following are time history curves of structural health monitoring data for three Bayesian structural system identification tasks in this embodiment of the invention: (a) is the acceleration measurement data of the building task, and (b) is the displacement and acceleration measurement data shared by the 6-parameter model and the 17-parameter model bridge task. Figure 5 The sampling results of the building task in this embodiment of the invention are two-dimensional projection lines or scatter plots in several pairs of principal component directions. Each line plot contains 160 Markov chains, and one of them is highlighted. The color of the scatter points is the value of the two-dimensional marginal probability density function estimated based on the sample projection in the figure. Figure 6 The above is a two-dimensional projection line or scatter plot of the sampling results of the 6-parameter model bridge task in several principal component directions in this embodiment of the invention. The different colors of the line represent 8 different Markov chains, and the colors of the scatter plots are the two-dimensional marginal probability density function values ​​estimated based on the sample projection in the figure. Figure 7 The above is a two-dimensional projection line or scatter plot of the sampling results of the 17-parameter model bridge task in several principal component directions in this embodiment of the invention. The different colors of the line represent 8 different Markov chains, and the colors of the scatter plots are the two-dimensional marginal probability density function values ​​estimated based on the sample projection in the figure. Detailed Implementation

[0019] 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.

[0020] In sampling-based Bayesian structural system identification methods, classical sampling algorithms often suffer from low sampling efficiency, making it difficult to obtain sufficient effective samples. While neural network-enhanced sampling methods can improve sampling efficiency through network training, they require retraining when the system identification problem changes, making them unsuitable for complex structural models that are difficult to train within a limited time. To improve the practicality of Bayesian structural system identification methods by combining the versatility of classical sampling algorithms with the high sampling efficiency of neural network-enhanced algorithms, this invention designs a Bayesian structural system identification method based on adaptive rotational meta-learning sampling. First, this invention rotates the sampling direction from the directions of parameters related to the specific problem to the directions of principal components consistent with the posterior trend, ensuring the method's rotation invariance to the posterior distribution and its versatility across different problems. Second, the relationship between the samples and their probability density functions before and after rotation is processed to ensure the correctness of the sampling results under the rotation transformation. Finally, an adaptive principal component direction estimation method is designed, which adaptively updates the previously unknown principal component directions as the sampling process progresses, ensuring the feasibility of the algorithm.

[0021] Specifically, in combination Figures 1-7 This invention proposes a Bayesian structure system identification method based on adaptive rotation element learning sampling, the method comprising the following steps: Step 1: For the target structure, deploy a structural health monitoring system. Based on the structural system model and structural health monitoring data, obtain the unnormalized posterior distribution of the model parameters through Bayesian inference. Step 2: Initialize the adaptive rotation meta-learning sampling method by importing the default or locally trained neural network parameters into the adaptive rotation meta-learning sampler. Step 3: Select whether to train the adaptive rotation meta-learning sampler. If yes, proceed to Step 4; otherwise, proceed to Step 5. Step 4: Based on the unnormalized posterior distribution, run the adaptive rotation meta-learning sampling method in training mode to train the adaptive rotation meta-learning sampler, and obtain and save the trained neural network parameters. Step 5: Based on the unnormalized posterior distribution, turn off the training mode and run the adaptive rotation meta-learning sampling method to obtain the parameter posterior distribution samples, thereby realizing the recognition of Bayesian structure systems.

[0022] Furthermore, the adaptive rotational meta-learning sampling process in steps four and five is based on the unnormalized parameter posterior distribution. The method utilizes a fusion adaptive principal component orientation estimation method and rotation processing. , , The AM-SGHMC algorithm is used to obtain the posterior distribution of the parameters in the sample. For structural health monitoring data, The structural model parameter vector has the following dimensions. That is, the number of parameters. An adaptive principal component representation of the parameter vector, used for AM-SGHMC sampling. It is an adaptive rotation matrix; parallelism can be selected based on the performance limitations of the computing device and the computational efficiency requirements. A Markov chain; the simulation process of the sampling method is as follows: Step 4.1: Initialize the rotation matrix ; Step 4.2: For each chain, randomly generate parameters around their nominal values. Set auxiliary variables Thus, the initial augmented vector samples are obtained. ,make ; Step 4.3, when If the condition is met, proceed to step 4.4 to generate a new sample; otherwise, proceed to step 4.11. The total number of simulation steps is [number missing]. Can be taken ; Step 4.4: If in the adaptive adjustment phase, proceed to step 4.5; otherwise, maintain the rotation matrix. The process remains unchanged. Proceed to step 4.7, where the starting point for the formal sampling can be set. At any given time, the adaptive adjustment phase can be set accordingly based on the starting point of the formal sampling. If it's a training mode, no formal sampling is needed; the adaptive adjustment phase can use a range of [range name missing]. interval; Step 4.5: Calculate the current parameter sample. and based on samples Perform adaptive principal component orientation estimation, and then use the current rotation matrix. Updated to ; Step 4.6: Update the sample to the corresponding direction of the new rotation matrix. , ; Step 4.7, in the current In the corresponding direction, for the current sample Perform one sampling step of the AM-SGHMC algorithm to simulate and obtain a new sample. ; Step 4.8, let ; Step 4.9: If it is in training mode, proceed to step 4.10; otherwise, return directly to step 4.3. Step 4.10: Execute the AM-SGHMC algorithm neural network update step, then return to step 4.3; Step 4.11: If in training mode, save the neural network parameters; otherwise, output the formal sampling phase. All corresponding parameter samples .

[0023] Furthermore, the adaptive principal component direction estimation method in step 4.5 is specifically as follows: The parameter samples input in each step 4.5 are denoted as... Adaptive adjustment phase ,in This represents its origin from the first A Markov chain; the current rotation matrix of each input is expanded and denoted as the direction vector of each principal component. Establish a principal component orientation adaptive estimator, and set the exponential decay rate series of the estimation process as follows: , which can be set A broken line sequence The built-in principal component statistics are initialized upon initial call. ,in The scale statistics are built into the AM-SGHMC algorithm; therefore, the adaptive principal component direction estimation method in step 4.5 is as follows: Step 4.5.1: Call the built-in parameter mean statistic in the AM-SGHMC algorithm. Centering the parameter samples ; Step 4.5.2, Press Rearrange in descending order and its corresponding and ; Step 4.5.3, for Perform steps 4.5.4 through 4.5.9 sequentially; Step 4.5.4: Calculate the centered parameter sample The Orthogonal residual Agreement ; Step 4.5.5, Orthogonalization Update Principal component statistics ; Step 4.5.6, Orthogonalization Update Principal component direction of vitamin ; Step 4.5.7, Calculate the first... Principal component input item ; Step 4.5.8, Update the first Principal component statistics ; Step 4.5.9, Calculate the first... Innovation principal component direction ; Step 4.5.10: Merge the direction vectors of each principal component to obtain a new rotation matrix. .

[0024] This invention proposes a Bayesian structural system identification method based on adaptive rotation element learning sampling. By proposing an adaptive rotation element learning sampling method, and building upon the detailed probability distribution identification results of the Bayesian structural system identification method, this method utilizes the rotation invariance of posterior trend features to adaptively rotate the sampling direction from the directions of parameters related to the specific problem to the directions of principal components consistent with the posterior trend. This gives the efficient sampler trained by the neural network in this method broad versatility independent of the problem, avoiding the need for retraining when the task changes. This makes the method applicable to the identification of complex, automated, and refined structural systems that are difficult to train, thus better serving the field of structural health detection.

[0025] Example Combination Figures 3-7 This invention addresses the displacement and acceleration responses of building and bridge structures under seismic excitation, utilizing a Bayesian structural system identification method based on adaptive rotating element learning sampling. For multi-story buildings, a frame-braced structure model is used as the training task; for arch bridges, 6-parameter and 17-parameter models with different levels of refinement are used as two general performance testing tasks, forming three tasks. The invention is then applied sequentially for structural system identification. The overall implementation plan is as follows: Figure 3 As shown.

[0026] The following section utilizes the Bayesian structural system identification method based on adaptive rotation element learning sampling from this invention to identify structural systems: Step one specifically involves: Installing horizontal acceleration sensors along the exterior walls at the ground floor, first-floor slab, and roof perimeter of the frame-supported structure building. The sampling frequency is 100Hz. Response data for a duration of 3 seconds under seismic excitation is captured. The measurement data is as follows: Figure 4 As shown in (a), the unnormalized posterior distribution of model parameters was obtained through Bayesian inference based on the frame-support structure model. Longitudinal displacement sensors were installed at the bottom of the piers at both ends of the arch bridge, longitudinal acceleration sensors were installed at the top of the piers at both ends, and longitudinal and vertical acceleration sensors were installed at the road surface at mid-span and the arch crown. The sampling frequency was 50Hz, and response data for 4 seconds under seismic excitation was extracted. The measurement data are as follows. Figure 4 As shown in (b), the unnormalized posterior distribution of model parameters for two bridge test tasks was obtained by Bayesian inference based on 6-parameter models and 17-parameter arch bridge finite element models with different levels of refinement. The second step specifically involves: initializing the adaptive rotation meta-learning sampling method; when executing the first building training task, importing the algorithm's default neural network parameters into the adaptive rotation meta-learning sampler; and when executing the latter two bridge testing tasks, importing the local neural network parameters trained in the first building task into the adaptive rotation meta-learning sampler. The specific steps of step three are as follows: when executing the first building training task, select to train the adaptive rotation meta-learning sampler and proceed to step four; when executing the next two bridge test tasks, select not to train the adaptive rotation meta-learning sampler and proceed directly to step five. Step four specifically involves: based on the unnormalized posterior distribution of the building model parameters obtained in step one, starting the training mode and running the adaptive rotation meta-learning sampling method in parallel. A Markov chain, take the total number of simulation steps. In training mode, no formal sampling is required; the adaptive adjustment phase takes [a certain value]. In the interval, the adaptive principal component direction estimation method sets... A broken line sequence Finally, the parameters of the trained neural network are obtained and saved; Step five specifically involves: based on the unnormalized posterior distribution of the model parameters obtained in step one, disabling the training mode and running the adaptive rotation meta-learning sampling method; during the execution of the first construction task, parallel processing is performed. A Markov chain, take the total number of simulation steps. The official sampling starting point is taken as At what time, the adaptive adjustment phase is taken as follows: In the interval, the adaptive principal component direction estimation method sets... A broken line sequence When executing the last two bridge test tasks, in parallel A Markov chain, take the total number of simulation steps. The official sampling starting point is taken as At what time, the adaptive adjustment phase is taken as follows: In the interval, the adaptive principal component direction estimation method sets... A broken line sequence Finally, posterior distribution samples of the parameters are obtained to achieve Bayesian structured system recognition. The two-dimensional projections of the samples on the principal component directions at the maximum and minimum scales are as follows: Figure 5 , Figure 6 , Figure 7 As shown in the column on the right side of the middle.

[0027] Using the classical Hamiltonian Monte Carlo (HMC) sampling method instead of the adaptive rotation meta-learning sampling method in this invention as a control, the above three tasks were performed on the same computing device, and the resulting two-dimensional projections of the samples in the corresponding directions are as follows: Figure 5, Figure 6 , Figure 7 As shown in the left column, it can be seen that the sample distributions obtained by the two methods in the direction of the smallest principal component are basically the same, but the present invention achieves a more thorough exploration in the direction of the largest principal component.

[0028] In the above embodiments, the number of valid samples obtained for each task can be counted based on the autocorrelation function of the obtained samples, thereby calculating the sampling efficiency. In the above-mentioned building, 6-parameter bridge, and 17-parameter bridge system identification tasks, the average number of valid samples obtained per chain per hour by the method proposed in this invention is 410.38, 12.30, and 5.19, respectively. Under the same computing equipment, the sampling efficiencies of the HMC sampling method are 2.35, 0.16, and 0.064, respectively. Therefore, the sampling efficiency of the structural system identification method proposed in this invention is 174, 76, and 81 times that of the HMC method in the three tasks, respectively. Moreover, it only requires one training for tasks with significantly different structures, effectively improving the practicality of the structural system identification method and making it suitable for complex structural models that are difficult to train in practical engineering applications.

[0029] The method first uses a structural system model and structural health monitoring data to obtain the unnormalized posterior distribution of model parameters through Bayesian inference as the identification target. Secondly, it loads a default or local sampler and allows fine-tuning training. Finally, it utilizes this sampler to execute an adaptive rotation meta-learning sampling method to achieve efficient sampling of the target distribution, completing the Bayesian structural system identification task. The adaptive rotation meta-learning sampling method first rotates the sampling direction from the directions of parameters related to the specific problem to the directions of principal components consistent with the posterior trend of the parameters, ensuring the method's rotation invariance to the posterior distribution and its general applicability to different problems. Secondly, it processes the relationship between the samples and their probability density functions before and after rotation to ensure the correctness of the sampling results under the rotation transformation. Finally, it designs an adaptive principal component direction estimation method that adaptively updates the previously unknown principal component directions as the sampling process progresses, ensuring the algorithm's feasibility. This invention proposes an adaptive rotational meta-learning sampling method. Based on the detailed probability distribution recognition results of Bayesian structural system recognition methods, it utilizes the rotation invariance of posterior trend features to adaptively rotate the sampling direction from the directions of parameters related to the specific problem to the directions of principal components consistent with the posterior trend. This gives the efficient sampler trained by the neural network in the method broad versatility independent of the problem, avoids the need for retraining when the task changes, and makes the method applicable to the recognition of complex, difficult-to-train, automated, and refined structural systems, thus better serving the field of structural health detection.

[0030] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the Bayesian structure system recognition method based on adaptive rotation element learning sampling.

[0031] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the Bayesian structure system recognition method based on adaptive rotation element learning sampling.

[0032] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0033] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0034] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0035] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0036] The above provides a detailed description of the Bayesian structure system identification method based on adaptive rotation element learning sampling proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for recognizing Bayesian structured systems based on adaptive rotational element learning sampling, characterized in that, The method includes: Step 1: For the target structure, deploy a structural health monitoring system. Based on the structural system model and structural health monitoring data, obtain the unnormalized posterior distribution of the model parameters through Bayesian inference. Step 2: Initialize the adaptive rotation meta-learning sampling method by importing the default or locally trained neural network parameters into the adaptive rotation meta-learning sampler. Step 3: Select whether to train the adaptive rotation meta-learning sampler. If yes, proceed to Step 4; otherwise, proceed to Step 5. Step 4: Based on the unnormalized posterior distribution, run the adaptive rotation meta-learning sampling method in training mode to train the adaptive rotation meta-learning sampler, and obtain and save the trained neural network parameters. Step 5: Based on the unnormalized posterior distribution, turn off the training mode and run the adaptive rotation meta-learning sampling method to obtain the parameter posterior distribution samples, thereby realizing the recognition of Bayesian structure systems.

2. The method according to claim 1, characterized in that, The adaptive rotational element learning sampling process in steps four and five is based on the unnormalized parameter posterior distribution. The method utilizes a fusion adaptive principal component orientation estimation method and rotation processing. , , The AM-SGHMC algorithm is used to obtain the posterior distribution of the parameters in the sample. For structural health monitoring data, The structural model parameter vector has the following dimensions. That is, the number of parameters. An adaptive principal component representation of the parameter vector, used for AM-SGHMC sampling. An adaptive rotation matrix is ​​used; parallel processing is selected based on the performance limitations of the computing device and the computational efficiency requirements. A Markov chain.

3. The method according to claim 2, characterized in that, The adaptive rotation element learning sampling process is specifically as follows: Step 4.1: Initialize the rotation matrix ; Step 4.2: For each chain, randomly generate parameters around their nominal values. Set auxiliary variables Thus, the initial augmented vector samples are obtained. ,make ; Step 4.3, when If the condition is met, proceed to step 4.4 to generate a new sample; otherwise, proceed to step 4.

11. The total number of simulation steps is [number missing]. ; Step 4.4: If in the adaptive adjustment phase, proceed to step 4.5; otherwise, maintain the rotation matrix. No change, proceed to step 4.7, and set the starting point for formal sampling. At any given time, the adaptive adjustment phase will be set accordingly based on the starting point of the formal sampling. For intervals, if it is training mode, no formal sampling is required; the adaptive adjustment phase takes the interval as 1. interval; Step 4.5: Calculate the current parameter sample. and based on samples Perform adaptive principal component orientation estimation, and then use the current rotation matrix. Updated to ; Step 4.6: Update the sample to the corresponding direction of the new rotation matrix. , ; Step 4.7, in the current In the corresponding direction, for the current sample Perform one sampling step of the AM-SGHMC algorithm to simulate and obtain a new sample. ; Step 4.8, let ; Step 4.9: If it is in training mode, proceed to step 4.10; otherwise, return directly to step 4.

3. Step 4.10: Execute the AM-SGHMC algorithm neural network update step, then return to step 4.3; Step 4.11: If in training mode, save the neural network parameters; otherwise, output the formal sampling phase. All corresponding parameter samples .

4. The method according to claim 3, characterized in that, The parameter samples input in each step 4.5 are denoted as... Adaptive adjustment phase ,in This represents its origin from the first A Markov chain; the current rotation matrix of each input is expanded and denoted as the direction vector of each principal component. Establish a principal component orientation adaptive estimator, and set the exponential decay rate series of the estimation process as follows: , where settings A broken line sequence The built-in principal component statistics are initialized upon initial call. ,in These are the scale statistics built into the AM-SGHMC algorithm.

5. The method according to claim 4, characterized in that, The adaptive principal component direction estimation method in step 4.5 is as follows: Step 4.5.1: Call the built-in parameter mean statistic in the AM-SGHMC algorithm. Center the parameter samples ; Step 4.5.2, Press Rearrange in descending order and its corresponding and ; Step 4.5.3, for Perform steps 4.5.4 through 4.5.9 sequentially; Step 4.5.4: Calculate the centered parameter sample The Orthogonal residual Agreement ; Step 4.5.5, Orthogonalization Update Principal component statistics ; Step 4.5.6, Orthogonalization Update Principal component direction of vitamin ; Step 4.5.7, Calculate the first... Principal component input item ; Step 4.5.8, Update the first Principal component statistics ; Step 4.5.9, Calculate the first... Innovation principal component direction ; Step 4.5.10: Merge the direction vectors of each principal component to obtain a new rotation matrix. .

6. The method according to claim 5, characterized in that, The first The formula for the principal component statistic is: 。 7. The method according to claim 6, characterized in that, The first The formula for the direction of the principal component in the reformation is: 。 8. The method according to claim 7, characterized in that, The formula for the new rotation matrix is: 。 9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images