Search

Fault diagnosis of rotor systems is important to prevent unexpected failures. Recently, deep learning (DL) methods, such as a convolutional neural network (CNN), have been utilized in many research areas, including fault diagnosis. DL has gained significant attention thanks to its ability to efficiently learn proper features from input data. It is possible to learn enriched hierarchical features by making the DL architectures deeper; therefore, many studies have been conducted to stack the neural networks, which are the basic building blocks of DL, deeper. However, it becomes difficult to comprehensively train neural network architectures as they become deeper, due to problems in the flow of gradient information during the training phase. In this paper, a direct connection based CNN (DC-CNN) method is proposed to significantly improve training efficiency and diagnosis performance. DC-CNN connects feature maps of different layers within a CNN to improve the gradient information flow over the layers. These additional connections, however, can increase the number of trainable parameters within the network. To prevent problems that might be caused by an increased number of parameters, dimension reduction modules are also developed. Moreover, to consider the anisotropic characteristics inherent in rotor systems, the vibration images containing both spatial and temporal information are generated and utilized. The effectiveness of DC-CNN is validated using experimental data from a rotor testbed. The experimental results indicate that the proposed method outperforms other conventional approaches with a smaller number of parameters. Also, visualizations of the learned features indicate that the proposed method can learn much more effective and significant features. Furthermore, the proposed method outperforms other approaches under conditions of insufficient or noisy data.

Recently, substantial research has explored the development of deep-learning-based methods to diagnose faults in rotating machinery. For these diagnosis methods, it is difficult to obtain high target diagnosis accuracy when the amount of labeled data obtained pertaining to the rotating machinery under study is insufficient or in cases involving a discrepancy in the distribution types found in the training and test data sets. To deal with this research need, the paper outlines a new method, a domain adaptation with semantic clustering (DASC), capable of diagnosing faults in rotating machinery. The method outlined in this research learns both domain-invariant and discriminative features. The method reduces the domain discrepancy by minimizing the domain-related loss. In addition, by defining an additional loss, which is called semantic clustering loss, and reducing it at multiple feature levels, the DASC method learns features that make samples better semantically clustered, according to their health conditions. Consequently, fault diagnosis performance for target rotating machinery can be enhanced through the use of the DASC approach. The effectiveness of the DASC approach is confirmed by examining various fault diagnosis situations with domain discrepancies across the source and target domains, using experimental data from three bearing systems. Also, various analyses are explored to better understand the advantages of the DASC method.

Some anomaly states of journal bearing rotor systems are direction-oriented (e.g., rubbing, misalignment). In these situations, vibration signals vary according to the direction of the sensors and the health state. This makes diagnosis difficult with traditional diagnosis methods. This paper proposes an omnidirectional regeneration method to develop a robust diagnosis algorithm for rotor systems. The proposed method can generate vibration signals in arbitrary directions without using extra sensors. In this method, signals are generated around the entire circumference of the rotor to consider all possible directions. Then, the directionality of each state is proved by mathematically and is evaluated using a proposed metric. When a directional state is determined, the classification is carried out on all of the generated signals. When a non-directional state is found, the classification is performed on only one of the generated signals to minimize computational load without sacrificing accuracy. The proposed ODR method was validated using experimental data. The classification results show that the proposed method generally outperforms the conventional classification method. The results support the proposed concept of using ODR signals in diagnosis procedures for journal bearing systems.

Deep learning-based research has drawn much attention in the field of fault diagnosis of various mechanical systems due to its powerful performance. In deep learning-based methods, signals become the input for a deep learning algorithm. However, the performance of an algorithm can be diminished if there is significant noise in the data. To address the noise issue, this paper proposes a fault diagnosis method called multi-task learning of classification and denoising (MLCD). The proposed method is designed to make a fault diagnosis algorithm robust against the noise in vibration signals by learning the denoising task simultaneously with the classification. Given a noisy input, MLCD can improve test accuracy by implementing denoising as an auxiliary task, using hyperparameters chosen by Bayesian optimization. To validate the proposed MLCD method, it is integrated with the two most commonly used deep learning algorithms: long short-term memory and one-dimensional convolutional neural network. For a case study, these algorithms are tested to classify the states of a rotor testbed data. The results show that the proposed MLCD method extracts noise-robust and meaningful features; ultimately, this improves the fault diagnosis performance.