.png){kind=icon}
.png){kind=icon}
Return
An impact localization method for composite structures based on time series features and machine learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compstruct.2025.119242.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Aircraft composite structures are susceptible to visually undetectable internal damage from low-velocity impacts. However, their anisotropy and complex geometry lead to intricate impact signals, making localization highly challenging. In this paper, a two-step impact localization method based on time series features (TSF) and machine learning is proposed. The first stage of this methodology transforms impact response sequences from different zones into a spectrum of TSF sets, including recursive quantized features (RQF), recursive plot features (RPF) and gridded representation features (GPF). This is achieved using a time series image-based representation approach. Subsequently, three distinct convolutional neural networks (CNNs) are constructed, namely RQF-1DCNN, RPF-2DCNN and GPF-2DCNN. These networks are employed to mine and learn deep-level features of time-series data, thereby transforming the impact localization task into a time-series feature classification task, specifically impact zone identification. The second step aims to precisely identify the impact location within the zone by utilizing impact response data and geometric center-of-mass algorithms at known locations within the identified impact zone. The proposed method is validated through low-velocity impact tests on composite honeycomb panels and aircraft wing structures. Additionally, differences in impact localization accuracy among various network models are analyzed. This method offers a cost-effective solution, achieving high accuracy with fewer sensors and less training data. Test results demonstrate that the RPF-2DCNN and GPF-2DCNN models, which are based on image features, outperform the RQF-1DCNN for impact monitoring in composite structures, achieving reliable impact localization with data from a single sensor. Moreover, compared to GPF, RPF, with its deeper time response sequence, is more appropriate for impact monitoring on wing structures with complex structural characteristics.
Keywords:
Composite
Impact localization
Time series images
Convolutional neural network
Journal
IF:
7.1
Papers:
1.8W
Citations:
8.0W