A fault diagnosis approach based on multi-sensor data fusion is a promising tool to deal with complicated damage detection problems of mechanical systems. Nevertheless, this approach suffers from two challenges, which are (1) the feature extraction from various types of sensory data and (2) the selection of a suitable fusion level. It is usually difficult to choose an optimal feature or fusion level for a specific fault diagnosis task, and extensive domain expertise and human labor are also highly required during these selections. To address these two challenges, we propose an adaptive multi-sensor data fusion method based on deep convolutional neural networks (DCNN) for fault diagnosis. The proposed method can learn features from raw data and optimize a combination of different fusion levels adaptively to satisfy the requirements of any fault diagnosis task. The proposed method is tested through a planetary gearbox test rig. Handcraft features, manual-selected fusion levels, single sensory data, and two traditional intelligent models, back-propagation neural networks (BPNN) and a support vector machine (SVM), are used as comparisons in the experiment. The results demonstrate that the proposed method is able to detect the conditions of the planetary gearbox effectively with the best diagnosis accuracy among all comparative methods in the experiment.

A group of kurtosis-guided-grams, such as Kurtogram, Protrugram and SKRgram, is designed to detect the resonance band excited by faults based on the sparsity index. However, a common issue associated with these methods is that they tend to choose the frequency band with individual impulses rather than the desired fault impulses. This may be attributed to the selection of the sparsity index, kurtosis, which is vulnerable to impulsive noise. In this paper, to solve the problem, a sparsity index, called the Gini index, is introduced as an alternative estimator for the selection of the resonance band. It has been found that the sparsity index is still able to provide guidelines for the selection of the fault band without prior information of the fault period. More importantly, the Gini index has unique performance in random-impulse resistance, which renders the improved methods using the index free from the random impulse caused by external knocks on the bearing housing, or electromagnetic interference. By virtue of these advantages, the improved methods using the Gini index not only overcome the shortcomings but are more effective under harsh working conditions, even in the complex structure. Finally, the comparison between the kurtosis-guided-grams and the improved methods using the Gini index is made using the simulated and experimental data. The results verify the effectiveness of the improvement by both the fixed-axis bearing and planetary bearing fault signals.

The data distribution shift is inevitable in practical fault diagnosis due to internal and outside changes of equipment. These obstacles will lead to performance degrade or even failure of diagnostic models. In light of these problems, a novel unsupervised intelligent diagnostic framework named Adversarial Adaptation network based on Classifier Discrepancy (AACD) is introduced for mechanical fault diagnosis where the training and test domains have different data distribution. Specifically, AACD mainly contains two parts, one is the shared feature generator built by the one-dimensional convolutional neural network and the other is double task-specific classifiers. They play an adversarial training game to learn class-separable and domain-invariant features for fault diagnosis. The proposed AACD is evaluated by fifteen transfer diagnosis tasks constructed on the planetary gearbox test-bed and the rolling bearing test-bed. Moreover, six popular algorithms are selected for comparison. The comprehensive results validate the effectiveness and superiority of the proposed approach.

As more and more data become available for machine prognostic analysis in the big data environment, effective data suitability assessment methods become highly desired to help locate data with sufficient quality for analysis. Driven by this purpose, this paper proposes a novel and systematic methodology for data suitability assessment based on the needs of prognostics and health management (PHM). In this study, the data suitability for PHM is assessed from the aspects of detectability, diagnosability, and trendability, which correspond to the three major tasks of PHM: fault detection, fault diagnosis, and degradation assessment. The proposed methodology is mainly built upon the recent research studies on maximum mean discrepancy in the field of machine learning, which include a family of test statistics that are used to test the difference between two data distributions. The effectiveness of the proposed methodology is demonstrated in diverse industrial applications, which include semiconductors, boring tool degradation, and sensorless drive diagnosis. The results in the case studies indicate that the proposed methodology can be a promising tool to evaluate whether the data under study or the extracted feature set is suitable for PHM tasks.