LMDFFN: A Lightweight Neural Framework for Multidimensional Feature Fusion in Fault Diagnosis

Mechanical core components are vital to modern industrial systems, where operational stability and accurate fault diagnosis directly govern productivity and safety. Although vibration-based fault diagnosis has advanced significantly, it still faces two key challenges:incomplete feature representation and high computational cost from complex fusion strategies. To address these issues, this paper proposes a Lightweight Multi-Dimensional Feature Fusion Network (LMDFFN). The framework employs a dual-branch design: a time–frequency branch uses Continuous Wavelet Transform (CWT) and Depthwise Separable Convolution (DSC) for efficient spectral feature extraction, while a temporal branch utilizes a bidirectional gated recurrent unit (BiGRU) to capture sequential dependencies. A novel DualSE-Dynamic module adaptively fuses these complementary features through a gating mechanism that dynamically recalibrates their importance and enriches representation dimensions. Experiments on a plunger pump dataset show that LMDFFN achieves an average accuracy of 99.96% with a standard deviation of ±0.03 over five independent trials, confirming its stability and robustness. Importantly, it attains this high precision while maintaining low computational complexity, demonstrating an excellent balance between diagnostic performance and efficiency for mechanical fault diagnosis, as validated on plunger pump and rolling bearing datasets.