Compressor Flow Perception via Deep Learning Modeling with Multi-Source Dynamic Fusion of Temporal Features by Bio-Inspired Optimization.

This is of significant engineering importance for enhancing the operation stability and reliability of aeroengines. To ensure the precise identification of aerodynamic instability, it proposes a deep learning model for multi-source fusion based on cross-attention and bidirectional Long Short-Term Memory (CA_BiLSTM) network. From a high-speed multistage compressor, multi-dimensional feature extraction is performed in the time domain, frequency domain, and entropy value range. Based on dispersion entropy, feature cross-identification is constructed with a multi-level early warning method. In response to the nonlinear aerodynamic parameters, Variational Mode Decomposition (VMD) and Dung Beetle Optimizer (DBO) for global optimization are integrated to construct a VMD_DBO_LSTM-coupled prediction model for aerodynamic stability. To address the limitation of single-point detection, this paper proposes a dual-channel fusion model based on cross-attention mechanism. Through shared convolution and dynamic weighting mechanism, the CA_BiLSTM model can precisely characterize the nonlinear features of the complex flow. It can fully integrate the complementary information of inlet and outlet signals, achieving the collaborative signal characterization. Its anti-interference capability is significantly superior to that of the original single-point signal. Combined with the dispersion entropy threshold, it can detect instability 1580 r in advance, effectively overcoming the problems of information deficiency and incomplete representation caused by traditional single-point monitoring.