The non-identical and non-independent distribution problem, caused by various working conditions of gears, inevitably makes effectiveness of gear fault diagnosis (GFD) degrade because of insufficient or low-quality data. Although some domain adaptation learning (DAL) models exist, most of them only aims at minimizing the global distributional mean discrepancy between the source domain (SD) and the target domain (TD), while the contribution of individual data is neglected. To address this issue, a new DAL, aiming to reduce the discrepancy on extracted features under a least square support vector machine (LSSVM) framework, is studied to exploit SD signals from another working conditions or adjacent mechanical parts to assist and boost target gear fault diagnostic performance in this paper. In addition, the white cosine similarity criterion is adopted to quantize the distributional weights of SD and TD feature data, and then the weights are added into a regularization function that measures the projected distributional discrepancy in LSSVM framework. Related experiments prove that the proposed method has higher diagnostic accuracies than other models. So, this strategy is expected to be a useful tool to transfer from SD to TD, and to boost the GFD performance under various working conditions.

Gears are often operated under various working conditions, which may cause the training and testing data have different but related distributions when conducting gear fault diagnosis. To address this issue, a latent feature-based transfer learning (TL) strategy is proposed in this paper. Firstly, the Bag-of-Fault-Words (BOFW) model combined with continuous wavelet transform (CWT) method is developed to extract and represent every fault feature parameter as a histogram. Before identifying the gear fault, the latent feature-based TL strategy is carried out, which adopts the joint dual-probabilistic latent semantic analysis (JD-PLSA) to model the shared and domain specific latent features. After that, a mapping matrix between two domains can be constructed by using Pearson’s Correlation Coefficients (PCCs) to effectively transfer shared and mapped domain specific latent knowledge, and to reduce the gap between two domains. Then, a Fisher kernel-based support vector machine (FSVM) is used to identify the gear fault types. To verify the effectiveness of the proposed approach, gear data sets gathered from the Spectra Quest’s Drivetrain Dynamics Simulator (DDS) are analyzed. Experimental results prove that the proposed approach has a significant advantage on gear fault diagnosis especially under the varying working conditions.

The various structures and working conditions make gearbox fault recognition (GFR) more challenging. This paper presents a new penalty domain selection machine (PDSM) enabled transfer learning for GFR study. The domain selection rules are designed using the band-selective independent component analysis (BS-ICA) to obtain the relation between different sensor locations and fault components for signal separation. Meanwhile, the initial penalty factors are calculated to speed up the PDSM process. For PDSM, the domain/signal penalty factors are added to the objective error function of original domain selection machine (DSM) to adapt varying working conditions and different sensor locations simultaneously. To solve the mixed optimization problem involved in PDSM, the Karush Kuhn Tucker (KKT) conditions are utilized to transform it to a two-layer single problem. Experiments using drivetrain dynamics simulator (DDS) prove that PDSM has higher diagnostic accuracy than other domain adaptation models. Meanwhile, it indicates faster convergence and stronger clustering capability than DSM.

This paper presents a transfer learning‐based approach for induction motor fault diagnosis, where the Transfer principal component analysis (TPCA) is proposed to improve diagnostic performance of the induction motors under various working conditions. TPCA is developed to minimize the distribution difference between training and testing data by mapping cross‐domain data into a shared latent space in which domain difference can be reduced. The trained model can achieve a good performance in testing data by using the learned features consisting of common latent principal components. Experimental results show that the proposed approach outperforms traditional machine learning techniques and can diagnose induction motor fault under various working conditions effectively.