In the field of fault diagnosis for rotating machinery, where available fault data are limited, numerous studies have employed a generative adversarial network (GAN) for data generation. However, the limited fault data for training GAN exacerbate GAN’s inherent training instability and mode collapse issues, which are induced by adversarial training. Moreover, the stochastic nature of random sampling for latent vectors sampling often results in low-fidelity and poor diversity generation, which negatively affects the fault diagnosis models. To address these issues, this paper presents two novel approaches: a spectrum-guided GAN (SGAN) and density-directionality sampling (DDS). SGAN mitigates training instability and mode collapse through combinatorial data utilization, adversarial spectral loss, and a tailored model structure. DDS ensures the high-fidelity and high-diversity of the generated data by selectively sampling the latent vectors through two steps: density-based filtering and directionality-based sampling in the feature space. Validation on both rotor and rolling element bearing datasets demonstrates that SGAN-DDS considerably improves classification results under the limited fault data. Furthermore, fidelity and diversity analyses are conducted to validate DDS, which increase the credibility of the proposed method; and offer advancement toward the application of deep-learning and GAN in industrial fields.

Though existing cross-domain fault-diagnosis methods have shown promising results under domain shift conditions, existing approaches are only valid for class-balanced data. However, situations of class imbalance are inevitable in industrial fields, due to the difficulty in acquiring fault data from in-use machinery. Thus, if existing cross-domain fault diagnosis approaches are directly applied when domain shift and class imbalance coexist, performance degradation can occur. Thus, this research develops a domain adversarial learning network for class-imbalanced data to address situations where class imbalance and domain shift coexist; this situation is referred to as the problem of class imbalance domain adaptation (CIDA). In the proposed method, domain adversarial training is implemented for learning domain-invariant features by reducing the domain shift, and a label-aligned sampling strategy is utilized to deal with the class imbalance. In addition, for further performance enhancement of label-aligned sampling by increasing the accuracy of pseudo labels, metric learning is introduced to enhance the feature distinctiveness by expanding the distance of samples from different classes while decreasing the distance of samples from the same class. The efficiency of the proposed method is validated by applying it to various circumstances using two bearing datasets. The proposed method demonstrates superior performance compared to conventional algorithms in addressing the CIDA problem, according to quantitative and qualitative evaluations.

This paper proposes a new generative model to produce signals of variable lengths. The proposed frequency-learning generative network (FLGN), which is designed and trained based on signal processing knowledge, can generate signals in a desired time range. The frequency information of the training data can be directly learned by the proposed method. A frequency is assumed to be decomposed to include deterministic and stochastic frequency parts. In the proposed approach, the deterministic frequency is learned in the form of a trainable parameter and the stochastic frequency is determined by the output of a frequency extractor. First, a phase extractor outputs a feature that corresponds to the phase of each frequency component. Then, a sine-basis is defined using the phase feature and the final frequency which is the summation of the deterministic and stochastic frequencies. Next, a magnitude extractor produces the magnitude feature from the sine-basis. Finally, the final output becomes the dot product of the sine-basis and the magnitude features. In the work described here, the proposed method is evaluated both quantitatively and qualitatively using three datasets: one simulation dataset and two experimental testbed datasets. The validation results indicate that the generated signal is similar to the true signal, when comparing them in the time-domain and frequency-domain. The results of the quantitative evaluation show that the signal generated by the proposed method has statistical characteristics that are similar to the true signal. Finally, the evaluation shows that the proposed model focuses on the characteristic frequencies while learning the frequency components.

This paper proposes a supervised learning with a class-balancing loss function (SL-CBL) approach for fault detection and feature-similarity-based recipe optimization (FSRO) for a plastic injection molding process. SL-CBL is a novel method that can accurately classify an input sample as a normal or fault condition, even when the training data are severely class-imbalanced. The proposed class-balancing loss function consists of the weighted focal loss and the loss of the F1 score; together, these are used to correctly classify even a small number of faulty samples. SL-CBL is investigated with four classifiers of different structures; the classifiers consist of several fully connected and batch normalization layers. FSRO is an optimization scheme that finds the optimal recipe whose feature is similar to the features of normal samples. The optimal solution is obtained by minimizing the Euclidean distance to the centroid of the normal features. In this research, the proposed SL-CBL and FSRO methods are validated by applying them to an industrial plastic injection molding dataset. The validation results show that the proposed SL-CBL approach achieves the highest F1 score with the lowest misclassification rate, as compared to the alternative methods. When visualizing the feature space, the optimal recipe found by the FSRO scheme was found to be close to the centroid of the normal features, even if the initial recipe is classified as a fault. Furthermore, each variable of the optimized recipe lies within the confidence interval of 3σ for the normal condition. This indicates that the optimal recipe is statistically similar to the normal samples.