SNS optical fiber impact identification method based on automatic encoder deep learning

Abstract

The present invention relates to a flexible sheet structure impact load identification method based on combination of a distributed SNS (Single-mode-No core-Single mode) multiple-mode interference fiber optic sensor and an automatic encoder deep learning algorithm, belonging to the technical field of structure health monitoring impact monitoring. The method comprises the following steps of: the step 1: layout of an SNS fiber optic sensor; the step 2: construction of a distributed SNS fiber optic sensor sheet structure impact load monitoring system; the step 3: real-time monitoring and collection of impact response dynamic signals, and recording of an impact test data to generate a sample bank through impact test for different positions and different energy; the step 4: preprocessing for the sheet structure SNS fiber optic sensor impact sample bank data; the step 5: selection of an automatic encoder as a deep learning model, construction of a network structure, and training of a deep learning neural network; and the step 6: through adoption of a trained model obtained in the step 5, processing of the SNS fiber optic sensor impact response data to achieve identification for the impact load position and energy.

Description

technical field [0001] The invention belongs to the technical field of impact monitoring for structural health monitoring, in particular to a plate structure impact monitoring method combining SNS optical fiber sensing technology and automatic encoder deep learning technology. Background technique [0002] For some large structures with high cost and high reliability requirements in actual engineering, such as aerospace vehicles, civil engineering, offshore platforms, etc., they will be affected by complex environmental loads and various sudden external factors such as The impact of foreign objects, vibration, etc., resulting in different degrees of damage to the structure. If the damage of the structure is not detected in time and prevented and controlled in advance, with the accumulation of damage, the ability of the structure to bear the load will gradually decrease, and even cause catastrophic events. [0003] At present, in the research of shock monitoring technology, ...

AI Technical Summary

Problems solved by technology

[0006] To overcome the traditional shallow model neural network method, it is necessary to master a large number of shock signal processing techniques combined with rich engineering practice experience to extract features, and most of these features have not been verified by big data

For example, the shock monitoring based on wavelet packet decomposition has the problems of wavelet base selection and wavelet decomposition level determination; there is no definite theoretical guidance, and it can only be determined through experience; at the same time, the shock characteristics constructed by wavelet coefficients are not complete, which affects the structure. Adaptation and environmental changes are not adaptable; in terms of model training, the shallow model is used to represent the complex mapping relationship between the signal and the impact load, resulting in a significantly insufficient generalization performance of the model when faced with big data

Images