Deep learning bearing-fault diagnosis has shown strong vitality in recent years. In industrial practice, the running state of bearings is monitored by collecting data from multiple sensors, for instance, the drive end, the fan end, and the base. Given the complexity of the operating conditions and the limited number of bearing-fault samples, obtaining complementary fault features using the traditional fault-diagnosis method, which uses statistical characteristic in time or frequency, is difficult and relies heavily on prior knowledge. In addition, intelligent bearing-fault diagnosis based on a convolutional neural network (CNN) has several deficiencies, such as single-scale fixed convolutional kernels, excessive dependence on experts’ experience, and a limited capacity for learning a small training dataset. Considering these drawbacks, a novel intelligent bearing-fault-diagnosis method based on signal-to-RGB image mapping (STRIM) and multichannel multiscale CNN (MCMS-CNN) is proposed. First, the signals from three different sensors are converted into RGB images by the STRIM method to achieve feature fusion. To extract RGB image features effectively, the proposed MCMS-CNN is established, which can automatically learn complementary and abundant features at different scales. By increasing the width and decreasing the depth of the network, the overfitting caused by the complex network for a small dataset is eliminated, and the fault classification capability is guaranteed simultaneously. The performance of the method is verified through the Case Western Reserve University’s (CWRU) bearing dataset. Compared with different DL approaches, the proposed approach can effectively realize fault diagnosis and substantially outperform other methods.
13 April 2023 in Neural Computing and Applications
Industrial fault diagnosis faces unique challenges with high-dimensional data, long time-series, and complex couplings, which are characterized by significant information entropy and intricate information dependencies inherent in datasets. Traditional image processing methods are effective for local feature extraction but often miss global temporal patterns, crucial for accurate diagnosis. While deep learning models like Vision Transformer (ViT) capture broader temporal features, they struggle with varying fault causes and time dependencies inherent in industrial data, where adding encoder layers may even hinder performance. This paper proposes a novel global and local feature fusion sequence-aware ViT (GLF-ViT), modifying feature embedding to retain sampling point correlations and preserve more local information. By fusing global features from the classification token with local features from the encoder, the algorithm significantly enhances complex fault diagnosis. Experimental analyses on data segment length, network depth, feature fusion and attention head receptive field validate the approach, demonstrating that a shallower encoder network is better suited for high-dimensional time-series fault diagnosis in complex industrial processes compared to deeper networks. The proposed method outperforms state-of-the-art algorithms on the Tennessee Eastman (TE) dataset and demonstrates excellent performance when further validated on a power transmission fault dataset.
17 March 2025 in IEEE Transactions on Instrumentation and Measurement
In recent years, few-shot detection has become a popular research direction in the field of industrial defect detection, which aims to perform defect detection tasks accurately using a limited number of labeled samples. The dual-branch architecture, which utilizes class center features to aggregate query features is one of the most commonly used methods to solve the few-shot detection problem due to its simplicity and effectiveness. Previous few-shot detection algorithms based on dual-branch architecture typically aggregate features by directly using the features of labeled samples or by performing simple averaging operations on those features to compute the class center features. However, the acquisition of class center features is related to the sample distribution of novel classes, resulting in poor robustness of such methods. Additionally, the regression of bounding boxes and the classification of objects are two coupled problems that are difficult to optimize simultaneously. Therefore, this paper proposes a decoupled few-shot detection algorithm to address the challenges that exist in the current few-shot detection. More specifically, we propose a Vector Quantization Feature Aggregation (VQFA) method, which utilizes vector quantization to map the features of support images into discrete vector sets and each vector comes from a learnable codebook. By employing this operation, we can acquire a category representation that is both robust and accurate, thereby enabling enhanced interactions between support and query features. Besides, we propose a Decoupled Few-Shot (DeFS) module to decouple classification and localization tasks so that the two tasks can obtain different visual regions to achieve better detection results. Experimental results on the public dataset NEU-DET and GC10-DET demonstrate that our method significantly outperforms other methods on various few-shot scenes, which proves the effectiveness of our proposed method.
Reservoir computing (RC) is an efficient recurrent neural network (RNN) method. However, the performance and prediction results of traditional RCs are susceptible to several factors, such as their network structure, parameter setting, and selection of input features. In this study, we employ a kernel time-delayed RC (KT-RC) method for time series prediction. The KT-RC transforms input vectors linearly to obtain a high-dimensional set of time-delayed linear eigenvectors, which are then transformed by various kernel functions to represent the nonlinear characteristics of the input signal. Finally, the Bayesian optimization algorithm adjusts the few remaining weights and kernel parameters to minimize the manual adjustment process. The advantages of KT-RC can be summarized as follows: 1) KT-RC solves the problems of uncertainty in weight matrices and difficulty in large-scale parameter selection in the input and hidden layers of RCs. 2) The KT module can avoid massive reservoir hyperparameters and effectively reduce the hidden layer size of the traditional RC. 3) The proposed KT-RC shows good performance, strong stability, and robustness in several synthetic and real-world datasets for one-step-ahead and multistep-ahead time series prediction. The simulation results confirm that KT-RC not only outperforms some gate-structured RNNs, kernel vector regression models, and recently proposed prediction models but also requires fewer parameters to be initialized and can reduce the hidden layer size of the traditional RCs. The source code is available at https://github.com/whs7713578/RC.