A self-growth convolution network for thermal and mechanical fault detection with very limited engine data

Severe faults occur infrequently but are critical for the prognostics and health management (PHM) of power machinery. Due to the scarcity of fault data, diagnostic models are always facing a very limited data problem. Basic convolutional neural networks require a large number of samples to train, and widely used data augmentation methods are influenced by data quality, which can exacerbate overfitting. To address this issue, a self-growth convolution network (SGNet) is proposed to make the deep learning process a self-growing scheme in both depth and width dimensions. The direct similarity measurement is utilized to supervise the depth-growth in the layer-by-layer training process. The feature redundancy metric is employed to control the width expansion. The self-growth scheme is proposed to disrupt the coadaptation between layers and that between kernels in order to mitigate the overfitting issue of small-sample cases. The SGNet is verified and implemented in the PHM of a heavy-duty diesel engine. It exhibits remarkable diagnostic capabilities in extremely sample-limited scenarios. With only three training samples per faulty type, the recognition rates of SGNet for the misfire fault and the gear tooth fracture fault are 88.44% and 98.11%, respectively. Further, the feature contrast, the information transmission, the noise resistance, and the frequency domain activation heat of SGNet are discussed by the ablation experiment in detail. The results indicate a novel path to solve the data-limitation problem in the PHM of important power machinery.