A Dual-Channel Feature Fusion Network of MACNN and Informer for Gearbox Fault Diagnosis under Variable Operating Conditions

Abstract

To address the difficulty of extracting and identifying gearbox fault features under varying operating conditions in industrial environments, a gearbox fault diagnosis method based on a dual-channel feature extraction and fusion network combining a multi-scale convolutional neural network (MACNN) and an Informer encoder is proposed. First, variational mode decomposition (VMD) and the fast Fourier transform (FFT) are applied to preprocess the raw vibration signals to explore multi-scale characteristics of fault information. A local pyramid attention mechanism is then introduced, and the MACNN architecture is designed to extract local features, while the Informer encoder is employed to capture global features from gearbox vibration signals. Subsequently, a frequency-domain self-attention feature fusion module based on adaptive spectral filtering (FMFF) is constructed to integrate local and global features, and the fault classification results are obtained through attention-weighted pooling and a linear layer. The proposed method is validated on the MCC5-THU gearbox fault dataset, and experimental results demonstrate that, under identical preprocessing conditions, the proposed MACNN-Informer dual-channel feature fusion network consistently outperforms MACNN, CNN-LSTM, and 1D-CNN-Informer models. The average fault classification accuracy reaches 99.16%. These results verify the effectiveness and feasibility of the proposed method under variable operating conditions, indicating strong practical applicability.