CN-122020004-A - Interpolation method of structural health monitoring missing data, electronic equipment and program product

CN122020004ACN 122020004 ACN122020004 ACN 122020004ACN-122020004-A

Abstract

The disclosure provides an interpolation method of structural health monitoring missing data, electronic equipment and a program product, and relates to the technical field of computers. The interpolation method of the structural health monitoring missing data comprises the steps of obtaining time sequence observation data which aims at a target structure and comprises a plurality of sensor channels, generating a mask matrix corresponding to the time sequence observation data, filling random noise in a data missing position in the time sequence observation data according to the mask matrix to obtain a target input matrix, inputting the target input matrix into a generating interpolation network, wherein the generating interpolation network comprises a generator and a discriminator, performing anti-learning combined training on the generator and the discriminator to obtain a target data interpolation model, wherein the combined training loss function at least comprises a physical consistency loss item in the process of the anti-learning combined training, and performing data interpolation on the time sequence observation data to be interpolated according to the target data interpolation model.

Inventors

GONG NAN
LEI YING
LIANG SIYU
LIU LIJUN

Assignees

厦门大学

Dates

Publication Date: 20260512
Application Date: 20260408

Claims (9)

1. An interpolation method for structural health monitoring missing data, comprising: acquiring time sequence observation data comprising a plurality of sensor channels aiming at a target structure, and generating a mask matrix corresponding to the time sequence observation data, wherein the time sequence observation data has continuous synchronous deletion of multi-sensor monitoring data, and mask values in the mask matrix are used for identifying data deletion positions in the time sequence observation data; Filling random noise in the data missing position in the time sequence observation data according to the mask matrix to obtain a target input matrix; Inputting the target input matrix into a generating type interpolation network, wherein the generating type interpolation network comprises a generator and a discriminator, the generator generates complete time sequence data after data interpolation based on the target input matrix, and the discriminator judges that a data value in the complete time sequence data is real data or interpolation data; Performing an anti-learning combined training on the generator and the discriminator to obtain a trained generator and taking the trained generator as a target data interpolation model, wherein in the process of the anti-learning combined training, a combined training loss function at least comprises a physical consistency loss term, and the physical consistency loss term is used for constraining low-order dynamics characteristics of structural response of interpolation data generated by the generator to be maintained, and According to the target data interpolation model, performing data interpolation on time sequence observation data to be interpolated, wherein the time sequence observation data to be interpolated has continuous synchronous loss of the monitoring data of a plurality of sensors; Wherein constructing a physical consistency loss term comprises: performing Hank transformation on the complete time sequence data to obtain a corresponding Hank matrix; performing singular value decomposition on the Hank matrix to obtain a plurality of singular values corresponding to the Hank matrix; Applying a non-convex penalty function to each of the singular values to obtain a penalty corresponding to each of the singular values, and And constructing a physical consistency loss term according to the penalty corresponding to the singular values.
2. The method of interpolation of structural health monitoring missing data of claim 1 wherein said generator includes a dilation time convolution network comprising a plurality of layers of one-dimensional causal dilation convolution layers; The generator generates complete time series data after data interpolation based on the target input matrix, and the complete time series data comprises: Extracting time sequence characteristics of the target input matrix through a plurality of layers of one-dimensional causal expansion convolution layers to obtain multichannel time sequence characteristics, and And generating complete time sequence data after data interpolation according to the multi-channel time sequence characteristics.
3. The method of interpolating structural health monitoring missing data of claim 2, wherein the expansion rate of the plurality of layers of one-dimensional causal expansion convolutional layers is increased layer by layer, and the receptive field of each layer of one-dimensional causal expansion convolutional layers is fixed.
4. The method of interpolation of structural health monitoring missing data of claim 2, wherein said generator further comprises a local sparse multi-headed attention module; Generating complete time sequence data after data interpolation according to the multi-channel time sequence characteristics, wherein the complete time sequence data comprises the following steps: Based on the local sparse multi-head attention module, determining a cross-channel correlation matrix among a plurality of sensor channels in a time region where real data exist in the target input matrix, wherein the cross-channel correlation matrix is used for describing an association relation existing among monitoring data of each sensor when a plurality of sensors monitor the same structure, and And generating complete time sequence data after data interpolation according to the multi-channel time sequence characteristics and the cross-channel correlation matrix.
5. The method for interpolating structural health monitoring missing data according to claim 1, wherein said determining that the data value in said complete time series data is true data or interpolated data, comprises: The discriminator carries out two-class judgment on each data value in the complete time sequence data according to the complete time sequence data and a corresponding prompt matrix, and outputs the probability that each data value is real data, wherein the prompt matrix is generated based on the mask matrix and is used for partially revealing whether each data value in the complete time sequence data is real data or interpolation data.
6. The method for interpolating structural health monitoring missing data of claim 1, wherein said joint training loss function further comprises an counterloss term for constraining the interpolated data generated by said generator to agree with the real data in data distribution and an observed consistency loss term for constraining the value at the existing real data position to agree with the original real data in the complete time series data generated by said generator.
7. The method of interpolation of structural health monitoring deficiency data according to any one of claims 1-6, wherein said time series observation data comprises one or more of multichannel vibration, displacement, acceleration or strain data for a target structure.
8. An electronic device, comprising: a memory storing execution instructions, and A processor executing the execution instructions stored in the memory, causing the processor to perform the method of interpolating structural health monitoring missing data according to any one of claims 1 to 7.
9. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the method of interpolation of structural health monitoring missing data according to any of claims 1 to 7.

Description

Interpolation method of structural health monitoring missing data, electronic equipment and program product Technical Field The disclosure relates to the field of computer technology, and in particular, to an interpolation method, electronic device and program product for structural health monitoring missing data. Background The structural health monitoring (Structural Health Monitoring, SHM) system provides data support for structural safety assessment and damage identification by distributing multiple types of sensors at key locations of an engineered structure (e.g., bridge, building, tunnel, etc.), continuously collecting dynamic response data of the engineered structure. In the actual operation of the structural health monitoring system, the continuous synchronization loss of multi-sensor data is often caused by the factors of sensor faults, communication interruption and the like, and the subsequent structural state evaluation is seriously influenced. In the prior art, one is a discriminant method based on a convolution or cyclic neural network, which relies on partial observation data at each moment and is difficult to process the situation that all sensors are continuously missed synchronously, the other is a generation type method based on a generation countermeasure network, which is capable of modeling data distribution but does not fully consider low-order characteristics of structural dynamics response, and generates data under high missing rate, which is easy to deviate from a physical rule, and the other is a method which relies on an accurate finite element model by introducing a definite physical model as constraint, and the traditional nuclear norm regularization is easy to cause excessive attenuation of singular values and difficult to accurately maintain dominant modal characteristics. Disclosure of Invention The disclosure provides an interpolation method of structural health monitoring missing data, electronic equipment and a program product. According to one aspect of the present disclosure, there is provided a method of interpolating structural health monitoring missing data, comprising: acquiring time sequence observation data comprising a plurality of sensor channels aiming at a target structure, and generating a mask matrix corresponding to the time sequence observation data, wherein the time sequence observation data has continuous synchronous deletion of multi-sensor monitoring data, and mask values in the mask matrix are used for identifying data deletion positions in the time sequence observation data; Filling random noise in the data missing position in the time sequence observation data according to the mask matrix to obtain a target input matrix; Inputting the target input matrix into a generating type interpolation network, wherein the generating type interpolation network comprises a generator and a discriminator, the generator generates complete time sequence data after data interpolation based on the target input matrix, and the discriminator judges that a data value in the complete time sequence data is real data or interpolation data; Performing counterlearning combined training on the generator and the discriminator to obtain a trained generator and taking the trained generator as a target data interpolation model, wherein in the process of the counterlearning combined training, a combined training loss function at least comprises a physical consistency loss term, and the physical consistency loss term is used for restraining low-order dynamics characteristics of structural response of interpolation data generated by the generator; and according to the target data interpolation model, performing data interpolation on the time sequence observation data to be interpolated, which have continuous synchronous loss of the multi-sensor monitoring data. According to the technical scheme of one aspect, the method specifically identifies the positions of continuous synchronous deletion of data in time sequence observation data by acquiring the time sequence observation data containing multiple sensor channels and generating a corresponding mask matrix, and provides structured deletion information for subsequent processing. Then, a target input matrix is constructed by filling random noise in missing positions and is input into a generating type interpolation network, wherein the generating type interpolation network comprises a generator and a discriminator, and the generator is driven by an antagonism learning mechanism of the generator and the discriminator to generate interpolation data conforming to real data distribution. In addition, in the process of the combined training of the countermeasure learning, constraint is applied to the output of the generator by introducing a physical consistency loss term, so that the interpolation result is forced to keep the low-order dynamics characteristic of the structural response, and the rationality and consistency of the reconstru