Structural health monitoring multivariate data anomaly diagnosis method based on convolutional neural network and transfer learning

Abstract

The invention provides a structural health monitoring multivariate data anomaly diagnosis method based on a convolutional neural network and transfer learning, and the method comprises the steps: carrying out the time series segmentation data visualization processing of the multivariate monitoring data of a certain large-scale structure A, converting the data into a time domain response image, carrying out the manual marking according to the time domain response image data corresponding to a data segment, selecting samples of various abnormal types with manual marks to form a data set A; inputting the data set A into a convolutional neural network model A for anomaly detection, and training the model A; visualizing and manually marking multivariate monitoring data of a certain large structure B to form a data set B; and adding a data set B on the basis of the model A, carrying out transfer learning training, so that the generalization performance of a classification model is improved, a convolutional neural network model can adapt to data of different distributions, and the model trained through transfer learning serves as a multivariate data anomaly detector; the method can solve the problem that there is no detection method for structural health monitoring multivariate data at present.

Description

technical field [0001] The invention belongs to the technical field of transfer learning, convolutional neural network, and civil engineering structural health monitoring, and specifically relates to a structural health monitoring multivariate data abnormality diagnosis method based on convolutional neural network and transfer learning. Background technique [0002] At present, there are more and more large-scale infrastructures such as long-span bridges, super high-rise buildings, long-span space structures, and offshore platforms. These infrastructure structures will inevitably be affected by disasters during their long-term use, resulting in a decline in their reliability. Even if the structure is not affected by the disaster, the structure itself will inevitably be damaged due to the comprehensive influence of long-term environmental erosion, load effect and fatigue effect, resulting in material aging, structural deformation, and amplification of its own defects, resultin...

AI Technical Summary

Problems solved by technology

[0003] The structural health monitoring system works outdoors for a long time. Due to factors such as sensor failure and data transmission failure, the data obtained by the structural health monitoring system will inevitably contain various types of abnormal data. These abnormal data cannot fully represent the real situation of the structure. , and will seriously interfere with the automatic data analysis and early warning functions of the monitoring system. In order to ensure the accuracy of the data analysis results, abnormal data must be detected and cleaned.

Due to the huge amount of structural health monitoring data (considering efficiency, precision, and scale, manual expert detection methods are no longer applicable), there are many types of abnormal data (traditional single-objective and binary classification methods are not applicable), and the characteristics of abnormal data are uncertain (based on Modulus threshold method is difficult to apply) and other characteristics, even with the participation of professionals, it is prone to over-processing or under-processing, and it is time-consuming and labor-intensive, which cannot meet the real-time requirements of monitoring system data analysis

[0004] At present, the detection of abnormal data at home and abroad is mostly focused on the detection of dynamic data such as acceleration, and the data of other types of sensors (such as strain, temperature, humidity, inclination, displacement, GPS, etc.) that are more diverse and complex lack effective abnormal data. Detection method

The difference between these multi-type sensor monitoring data and vibration acceleration data is as follows. The sampling frequency of vibration acceleration is generally 20HZ, and the data quality is relatively good. The sampling frequency range of other types of sensor data is 1HZ-20HZ. Data sampling frequency is generally 1HZ, temperature, humidity, inclination, displacement and other data sampling frequency is 5HZ-10HZ, strain data sampling frequency is generally 20HZ; vibration acceleration data graph has obvious symmetry, and generally has a mean value of 0; other types of sensor data have a trend of random fluctuations, and the data images are generally asymmetrical, without the characteristics of 0; there are many studies related to vibration acceleration data, and there are many studies on anomaly detection work, while other types of sensors Data-related research is relatively vacant, and there is no corresponding abnormal data detection research for various data

Images