Time–frequency ensemble network for wind turbine mechanical fault diagnosis

Abstract

Wind turbines typically operate under variable speed conditions, so the collected vibration signals are affected by non-linearity and information mixing, while also containing a large amount of noise interference. However, most existing methods extract fault features from a single domain, failing to capture the signals’ diverse and complex characteristics. To fully exploit multi-domain discriminative features under variable speed conditions, this paper proposes a time–frequency ensemble network (TFNet). First, the feature representation is improved by constructing an adaptive spectral block (ASB) using Fourier analysis, while an adaptive threshold is introduced to reduce noise interference. Second, the Transformer and Graph Convolutional Network (GCN) are combined to extract the time–frequency discriminative features of defects. Specifically. In the time domain module, the global time domain features of faults are extracted by the Transformer encoder block. In the frequency domain module, a mixhop graph convolutional network is used to extract the multi-scale frequency domain features of different neighbours, and a Multi Head Attention (MHA) mechanism is introduced to capture the intra-feature dependencies. To achieve better diagnostic results under variable speed conditions, a label smoothing algorithm is used to assist the training of the model. A case study is conducted using the WT-Planetary gearbox dataset and the XJTUSuprgear variable speed gearbox dataset as well as the CWRU Bearing dataset. The experimental results show that the proposed model has high diagnostic accuracy and strong generalisation ability compared to other fault diagnosis models.

Keywords

• Wind turbine
• Fault diagnosis
• Transformer-GCN
• Time–frequency features
• Label smoothing