Deep transfer-learning-based diagnosis models are promising to apply diagnosis knowledge across related machines, but from which the collected data follow different distribution. To reduce the distribution discrepancy, Gaussian kernel induced maximum mean discrepancy (GK-MMD) is a widely used distance metric to impose constraints on the training of diagnosis models. However, the models using GK-MMD have three weaknesses: 1) GK-MMD may not accurately estimate distribution discrepancy because it ignores the high-order moment distances of data; 2) the time complexity of GK-MMD is high to require much computation cost; 3) the transfer performance of GK-MMD-based diagnosis models is sensitive to the selected kernel parameters. In order to overcome the weaknesses, a distance metric named polynomial kernel induced MMD (PK-MMD) is proposed in this paper. Combined with PK-MMD, a diagnosis model is constructed to reuse diagnosis knowledge from one machine to the other. The proposed methods are verified by two transfer learning cases, in which the health states of locomotive bearings are identified with the help of data respectively from motor bearings and gearbox bearings in laboratories. The results show that PK-MMD enables to improve the inefficient computation of GK-MMD, and the PK-MMD-based diagnosis model presents better transfer results than other methods.

Due to the data distribution discrepancy, fault diagnosis models, trained with labeled data in one scene, likely fails in classifying by unlabeled data acquired from the other scenes. Transfer learning is capable to generalize successful application trained in one scene to the fault diagnosis in the other scenes. However, the existing transfer methods do not pay much attention to reduce adaptively marginal and conditional distribution biases, and also ignore the degree of contribution between both biases and among network layers, which limit classification performance and generalization in reality. To overcome these weaknesses, we establish a new fault diagnosis model, called subdomain adaptation transfer learning network (SATLN). First, two convolutional building blocks were stacked to extract transferable features from raw data. Then, the pseudo label learning is amended to construct target subdomain of each class. Furthermore, a subdomain adaptation is combined with domain adaptation to reduce both marginal and conditional distribution biases simultaneously. Finally, a dynamic weight term is applied for adaptive adjustment of the contributions from both discrepancies and each network layers. The SATLN method is tested with six transfer tasks. The results demonstrate the effectiveness and superiority of the SATLN in the cross-domain fault diagnosis field.

Event-based cameras are the emerging bioinspired technology in vision sensing. Different from the traditional standard cameras, the event-based cameras asynchronously record the brightness change per pixel, and have the great merits of high temporal resolution, low energy consumption, high dynamic range, etc. While the event-based cameras have been initially exploited in several common vision-based tasks in the recent years, the investigation on machine condition monitoring problem is quite limited. This article offers the first attempt in the current literature on exploring the contactless event vision data for machine fault diagnosis. A vibration event representation is proposed to transform the event records into typical data samples, and a deep convolutional neural network model is used for processing the event information. To enhance the model robustness against environmental noisy vision events, an event data augmentation method is proposed to introduce variations of the event patterns. A deep representation clustering method is further proposed to improve the pattern recognition performance with respect to different machine health conditions. Experiments on the event vision-based rotating machine fault diagnosis problem are carried out. It is extensively validated that high fault diagnosis accuracies can be obtained using the vision data from the event-based cameras, which are competitive with the popular accelerometer data. Considering the properties of flexibility, portability, and data recognizability, the event-based cameras thus provide a promising new tool for contactless machine health condition monitoring and fault diagnosis.

Intelligent machinery fault diagnosis methods have been popularly and successfully developed in the past decades, and the vibration acceleration data collected by contact accelerometers have been widely investigated. In many industrial scenarios, contactless sensors are more preferred. The event camera is an emerging bio-inspired technology for vision sensing, which asynchronously records per-pixel brightness change polarity with high temporal resolution and low latency. It offers a promising tool for contactless machine vibration sensing and fault diagnosis. However, the dynamic vision-based methods suffer from variations of practical factors such as camera position, machine operating condition, etc. Furthermore, as a new sensing technology, the labeled dynamic vision data are limited, which generally cannot cover a wide range of machine fault modes. Aiming at these challenges, a novel dynamic vision-based machinery fault diagnosis method is proposed in this paper. It is motivated to explore the abundant vibration acceleration data for enhancing the dynamic vision-based model performance. A cross-modality feature alignment method is thus proposed with deep adversarial neural networks to achieve fault diagnosis knowledge transfer. An event erasing method is further proposed for improving model robustness against variations. The proposed method can effectively identify unseen fault mode with dynamic vision data. Experiments on two rotating machine monitoring datasets are carried out for validations, and the results suggest the proposed method is promising for generalized contactless machinery fault diagnosis.

Deep transfer learning-based fault diagnosis has been developed to correct the data distribution shift, promoting a diagnosis knowledge transfer across related machines. However, there are two weaknesses: first, the assumption that all the target domain data are unlabeled is strict for robust applications of deep transfer learning to diagnosis across different machines; and second, the successes of existing methods are mostly achieved under the same conditional label distribution. For the weaknesses, this article relaxes a reasonable assumption that one-shot target domain samples called anchors are labeled, and further presents a deep targeted transfer learning (DTTL) method for tasks with different conditional label distributions. DTTL includes three parts. First, a domain-shared residual network is constructed to represent features from cross-domain data. Second, a target-domain clustering module gathers unlabeled target domain samples toward anchors. Third, a targeted adaptation module designs adaptation trajectories of target domain samples according to the associated labels of anchors and source domain data, and then corrects the joint distribution shift. The DTTL is demonstrated on transfer diagnosis tasks across different bearings. The results show that cross-domain data can be aligned by following the designable adaptation trajectories. Compared with other methods, the DTTL achieves higher diagnosis accuracy.