The challenges of class-imbalance and partially unknown training labels often arise in fault detection tasks. When these two problems occur simultaneously, existing imbalanced classification methods cannot be directly used due to the absence of the label, and the class-imbalance would lead to severe bias prediction. In this study, we proposed a novel deep imbalance separation network (deepImSN) framework that is capable of dealing with fault detection problems with both class imbalance and partially unknown labels. This framework integrates the one-class learning concept into the positive-unlabeled (PU) learning theory for the first time. It alleviates the bias of the class-imbalance while making full use of the limited label information in the PU set to optimize the feature space and guide model training. The proposed deepImSN is designed to be used in different scenarios. It can accurately complete the fault detection task whether only part of fault samples or normal samples are labeled, and the class-prior is known or unknown. Experimental results on real-world problems, such as high-speed rail wheels fault inspection and wafer map fault detection, demonstrate that deepImSN outperforms existing methods in various experimental conditions.

In practical industrial applications, the need for a large number of accurately labeled training samples is a significant challenge for fault detection tasks. However, labeling all training samples is expensive and prone to labeling errors, especially for early fault detection of wind turbine blades. This article proposes a labeling bias hypothesis. Assuming the labeler only needs to label parts of normal samples that are easy to judge, we design a probability ratio least-squares importance fitting (PRL-SIF) method based on variable homogeneity. Unlike other state-of-the-art positive unlabeled learning methods, PRL-SIF does not require knowledge of the class priors to achieve training. Furthermore, to better handle the multidimensional time-series data of wind turbines, we provide a data preprocessing method based on functional analysis to achieve time series feature extraction and dimensionality reduction. The effectiveness and robustness of the proposed method are verified on 23 real-world wind turbine datasets. Experimental results show that the proposed method can achieve nearly 90% accuracy while only labeling 20% of normal samples.

Class-imbalance is a prevalent and challenging problem in the field of fault detection. The undersampling ensemble framework is an effective method to deal with imbalance problems. However, designing a suitable sampling strategy to generate effective and divergent subsets is a major difficulty of this type of method. Hence, we propose a novel adaptive undersampling framework. It models the entire training process as a Markov decision process (MDP), thus, enabling dynamic decision-making for subsequent sampling strategies based on the current training performance of the ensemble framework. The sampler is optimized by the soft actor-critic reinforcement learning method. Considering the imbalance dataset's nature and the need for state definition, the clustering method is applied to the original training dataset. The state (training performance) and the action (sampling strategy) are determined according to the clustering results. The unique state definition and sampling decision mechanism are designed to ensure the convergence speed of MDP and improve the divergence of the subsets. We validate the effectiveness of the proposed framework on the real-world wind turbine blade cracking datasets and the high-speed train braking system dataset. The experimental results show that the classification performance and robustness of the proposed framework are significantly better than the 16 benchmark methods.

Intelligent fault detection methods based on deep learning have been developed rapidly in recent years. However, most of these methods are based on supervised learning which requires a fully labeled training set. It is difficult to obtain massive labeled samples in real applications incredibly accurately labeled fault samples from an operating system. The lack of labels and label noise becomes a great challenge for fault detection. To tackle this problem, in this article, we propose a positive-unlabeled learning based hybrid network (PUHN). It only needs part of the normal operating samples to be labeled. All other samples (including the rest of the normal samples and all fault samples) are unlabeled, which greatly reduces the labeling cost. PUHN consists of three modules: a nonnegative risk positive-unlabeled (PU) network for training the classifier, a feature extraction module, and a clustering layer for improving data separability and estimating the class priors of PU learning. The three are optimized as a whole and the corresponding optimization strategy is designed. The monitoring data of 24 wind turbines are used to verify the effectiveness and robustness of the proposed method. The experimental results indicate that the proposed method is superior to the benchmark methods, and the performance is significantly better than the supervised learning method when there exists label noise.

With the lack of failure data, class imbalance has become a common challenge in the fault diagnosis of industrial systems. The oversampling methods can tackle the class-imbalanced problem by generating the minority samples to balance the training set. However, one of the main challenges of the existing oversampling methods is how to generate high-quality minority samples. Traditional oversampling methods regard all synthetic samples as minority ones to be added to the training set without filtering. The low-quality synthetic samples would distort the distribution of the dataset and worsen the classification performance. In this article, we propose a weakly supervised oversampling method that treats all synthetic samples as unlabeled samples and develops a graph semisupervised learning algorithm to select high-quality synthetic samples, adding into the final training set as minority samples. To improve the quality of synthetic samples, we propose a cost-sensitive neighborhood component analysis dimensionality reduction method to enhance domain information validity in high-dimensional datasets. Finally, combining a boosting-based ensemble framework, we propose a new imbalanced learning framework suitable for high dimensionality and highly imbalanced fault diagnosis in industrial systems. The experimental validation is performed on five real-world wind turbine blade cracking failure datasets and compared to 15 benchmark methods. The experimental results show that average performances and robustness of the proposed framework are significantly better than those of the benchmark methods.