Modified BiLSTM network for interval prediction based on Aerospace Load System

Abstract

In structural health monitoring, load identification significantly impacts the safety and reliability of structures, particularly in aerospace, maritime, and other industrial systems. To address the issues of low accuracy and unstable signal identification in airborne load recognition, this paper proposes a novel neural network-based load prediction model called Encoding Meta Attention-Adaptive Snow Ablation Optimizer-BiLSTM (EMA-ABiLSTM). First, a novel meta-attention mechanism is designed to directly train the source data and output weights to the prediction model. The features generated by positional encoding and meta-network are used as inputs to the prediction model. Second, an Adaptive Improved Snow Ablation Optimizer (ASAO) algorithm is proposed to optimize the hyperparameters of the BiLSTM network, forming the Adaptive Snow Ablation Optimizer-BiLSTM (ABiLSTM) model. Finally, both feature input and raw input data are used to train the ABiLSTM model for high-precision load identification. An interval prediction method is introduced to quantify the uncertainty of predictions, thereby enhancing the reliability of the results. Additionally, a novel meta-network feature loss function is designed to improve the model’s identification efficiency. In experiments conducted using real landing gear data from an aircraft, the proposed method is compared with several advanced prediction models. The results demonstrate that the EMA-ABiLSTM model achieves excellent accuracy and reliability in interval prediction and load identification tasks.