A vibration mapped RGB-GADF images driven semi-supervised fault diagnosis method for helical gear

Abstract Aiming at the problems of small number of labeled data and low utilization rate of unlabeled data in mechanical fault diagnosis, a vibration mapped RGB-Gramian angular difference field (RGB-GADF) images driven semi-supervised fault diagnosis method for helical gear is proposed. Firstly, the 1D vibration signal collected by one three-channel acceleration sensor is subjected to piecewise aggregation approximation (PAA) to obtain three GADF images. Based on the characteristics of RGB images, three single-channel GADF images are fused into one RGB-GADF image to enhance the operation state characterization of helical gear. Then, based on the FixMatch model, the attention of the model to unlabeled data is increased, so a new multi-attention FixMatch (MA-FixMatch) semi-supervised model is constructed. Next, the weight parameters of the teacher model are subjected to exponential moving average (EMA) to improve the quality and stability of pseudo labels. The final MA-FixMatch with EMA semi-supervised model is obtained. Afterwards, based on inverted triangular channel distribution-ConvNeXt (ITCD-ConvNeXt), an improved ConvNeXt feature extraction model using adaptive maximum pooling (AMP-ConvNeXt) is proposed. The stacking times and the number of input channels of each ConvNeXt Block are finely designed. The data with three label rates are studied under four combined conditions of speed and load. The results show that the semi-supervised fault diagnosis method proposed has the fault diagnosis rate of 99.6% and above on the data sets under four working conditions when the label rate reaches the threshold.