Variable scale multilayer perceptron for helicopter transmission system vibration data abnormity beyond efficient recovery

In the task of health monitoring of helicopter transmission systems, data transmission is often abnormal due to harsh working environments, sparse wire connections, and other factors. This paper addresses the problem of recovering abnormal vibration data, focusing on common Bias anomalies and precision degradation anomalies in the collected vibration data. The Transformer model has gained popularity to address time series for its ability to capture strong long-range correlations. However, it also poses intensive demands on memory and computation resources. To address this issue, an efficient model called Variable Scale Multilayer Perceptron is proposed for recovering abnormal vibration data. Initially, the input data is divided into patches of varying scales to alter the model's input view and reduce computational complexity. Subsequently, compute the relative variations in the input sequence to capture more motion information in the vibration data. Then, a simple Multilayer Perceptron (MLP) layer is employed to learn the latent relationships among the vibration data, followed by a linear mapping layer to reconstruct normal vibration data from the latent representations. Experimental evaluations were conducted on vibration datasets of bearings and gears, and the results demonstrate that our proposed method outperforms the Transformer -based and MLP-based models in terms of both performance and efficiency.