An incipient fault diagnosis method based on multi-scale energy transform and two-stream network

To address the challenge of incipient fault diagnosis in rotating machinery under complex operating conditions, this paper proposes a novel methodology integrating Multi-scale Energy Transform (MSET) with a dual-stream neural network architecture. The proposed framework consists of three key components: First, a new MSET technique is introduced for preprocessing weak fault signals. By applying Energy Transform (ET) iteratively with optimized parameters, where ET is defined as the reconstruction of a signal multiplied by its filtered Teager–Kaiser Energy Operator (TKEO), the proposed MSET effectively enhances fault impulse components while preserving their structural characteristics. This process yields a high signal-to-noise ratio (SNR) output suitable for subsequent sparse decomposition. Second, a sparse decomposition approach combining MSET with time–frequency spectrum optimization (TFSO) is developed to extract sparse atoms representing fault impulses. After MSET preprocessing, the maximum energy trajectory of the time–frequency representation across scales exhibits convex properties. This enables the reformulation of optimal atom matching pursuit into a convex optimization problem, which is efficiently solved using the Golden Jackal Optimization (GJO) algorithm to rapidly and accurately identify the optimal time–frequency spectrum. Third, a two-stream diagnostic network incorporating a time-series Transformer and an Inception-ResNet-v2 architecture is designed to comprehensively exploit multi-dimensional features in fault signals. This model effectively captures both global temporal dependencies and local pattern characteristics, enabling accurate fault mode identification. Simulation and experimental case studies confirm that the proposed method achieves high diagnostic accuracy and computational efficiency, particularly under strong noise conditions.