Fig 1.
End-to-end workflow of the proposed generative repair and RUL prediction framework under sensor failures.
The framework starts from the original dataset and proceeds through data preprocessing, including initial preprocessing, sensor fault simulation, and dataset pair construction. The corrupted sensor sequences are then repaired by the missing-parameter generator based on the WGAN-GP architecture, followed by RUL estimation using a CNN–LSTM predictor.
Table 1.
List of monitored sensor parameters serving as multivariate time-series inputs for the proposed model.
Fig 2.
Data preprocessing workflow including sensor fault simulation and dataset construction, including initial preprocessing, simulated sensor failure generation, ten-fold dataset partitioning, and the construction of corrupted–complete sequence pairs for model training.
Table 2.
Imputation performance of the missing parameter generator under different failure rates.
Table 3.
Comparison of different data repair methods for multivariate time-series imputation under a 20% simulated sensor failure conditions based on reconstruction accuracy (RMSE, SSIM) and statistical significance.
Table 4.
Comparison of RUL prediction performance between the proposed framework and representative deep learning models under sensor fault conditions, evaluated using RMSE, standard C-MAPSS scoring function and statistical significance.
Table 5.
Ablation study evaluating the impact of the missing-parameter generator on RUL prediction performance under sensor failure conditions.
Table 6.
Computational efficiency and deployment-related complexity of the proposed model, including model size, training time, and inference latency measured on an NVIDIA RTX 4090 GPU.