Exploring Sample/Feature Hybrid Transfer for Gear Fault Diagnosis Under Varying Working Conditions

Abstract

Unknown environmental noise and varying operation conditions negatively affect gear fault diagnosis (GFD) performance. In this paper, the sample/feature hybrid transfer learning (TL) strategies are adopted for GFD under varying working conditions, where source working conditions are considered to help the learning of target working conditions. Here, a multiple domains-feature vector is extracted where certain insensitive features offset the adverse effects of varying working conditions on sensitive features, including time domain, frequency domain, noise domain, and torque domain. Before TL, the signed-rank and chi-square test-based similarity estimation frame is adopted to select source data sets, aiming to reduce the possibility of negative transfer. Then, the hybrid transfer model, including the fast TrAdaBoost and partial model-based transfer (PMT) algorithm, is carried out, whose weights are allocated in sample and feature, respectively. Related experiments were conducted on the drivetrain dynamics simulator, which proves that feature transfer is more suitable for low-quality source domains while sample transfer is more suitable for high-quality source domains. Compared with non-transfer strategy, transfer learning is a useful tool to solve a practical GFD problem when facing with multiple working conditions, thus enhancing the universality and application value in fault diagnosis field.