# Reconstruction of missing data in transferred generative adversarial networks with small sample data

## TL;DR

This paper introduces a new method using generative adversarial networks to reconstruct missing data in heavy-duty train systems when only small samples are available.

## Contribution

The novel VAE-FGAN model with GRU and SE-NET mechanisms enables effective missing data reconstruction in small sample scenarios.

## Key findings

- The VAE-FGAN model achieves reconstruction accuracy with MAE and MAPE below 1.5 for missing data.
- The reconstructed data aligns well with the distribution trend of original measured data.
- Migration learning and pre-training help overcome challenges from small sample sizes.

## Abstract

Under special working conditions, data collection systems of heavy-duty trains may be faced with a small sample size and missing data when executing measurement, operation, or maintenance tasks. Existing generative modeling methods are ineffective in reconstructing missing data with such a small sample. Hence, we set up a frame of migration learning generative adversarial network for small data samples, in which a new variational autoencoder semantic fusion generative adversarial network (VAE-FGAN) is developed to reconstruct missing data. First a GRU module is introduced in the encoder to fuse the underlying features of the data with higher-level features, which enables the VAE-FGAN to learn the correlation between the measured data through unsupervised training. Second, an SE-NET attention mechanism is introduced into the whole generative network to enhance the expression of the feature extraction network on data features. Finally, parameters are shared through migration learning and pre-training, thereby eliminating the difficulty in training the model due to the small size of certain operation and maintenance data. Experimental results show that the reconstruction accuracy indices MAE and MAPE can be kept below 1.5 when measured data is missing; the reconstructed data also fits well to the distribution trend of the measured data.