Rotating machinery is the common types of equipment on ship and serves as the fundamental component to ensure ship operation and functionality. Developing intelligent fault diagnosis methods for rotating machinery is critical for timely and accurate detection of potential faults, guiding effective maintenance and repair, and ultimately extending the service life of the equipment. However, the lack of sufficient fault data for rotating machinery makes it challenging to employ deep learning to build intelligent fault diagnosis model. To address this issue, this paper introduces the prior knowledge of time-series signals and the inherent differences in samples under various states of rotating machinery. By leveraging few-shot fault samples in combination with normal state data, a novel contrastive loss function is proposed. This loss function enables the effective optimization of deep learning-based model hyperparameters under the condition of extreme training sample imbalance. Additionally, a fault diagnosis algorithm is designed to achieve the diagnosis modeling and application under highly imbalanced sample conditions. The proposed method is validated through experiments on two common types of rotating machinery: gearboxes and bearings. Furthermore, the effectiveness and scalability are demonstrated on an industrial robot fault experimental platform.

Traditional methods of monitoring industrial products and processes tend to focus on single-model approaches and ignore the cost of monitoring systems in fault detection. To fill the gap, this paper presents a novel cost-optimal fault monitoring and prediction model using machine learning and statistical process control for industrial systems. First, a least squares support vector machine model is presented to estimate the degradation process and obtain degradation residuals of the industrial systems. Secondly, a multidimensional hidden Markov model is presented to describe the residual process of the industrial systems. The multidimensional hidden Markov model parameters are derived and estimated via an expectation-maximization algorithm with explicit expressions. Afterward, based on the built hidden Markov model, a fault prediction model utilizing a Bayesian statistical control is developed to dynamically update the probability estimate of equipment failure throughout the entire operation process of industrial systems. The decision variable of the control process for fault prediction is optimized by a semi-Markov decision process (SMDP) under minimization of the average operation costs. Finally, a case study with NASA's PHM08 dataset is performed to verify the proposed approach, which demonstrates the effectiveness and superiority of the presented model.

The manual maintenance mode of gearboxes suffers from low efficiency and high subjectivity, making it difficult to meet the requirements of intelligent production. Therefore, achieving accurate condition identification and diagnosis of gearboxes is essential for further innovating intelligent maintenance of gearboxes. However, existing intelligent fault diagnosis methods for gearboxes suffer from insufficient mining of deep state features, resulting in poor effectiveness and generalization in practical applications. To solve the above challenges, this paper designed a new feature learning module that is named multiscale dilated multihead attention network (MDCN‐MHA). It can effectively expand the receptive field and enhance the ability to capture key fault features from gearbox condition data. Experimental verification and ablation studies show that the model with three multihead attention outperforms other comparison model variants in terms of accuracy and stability. Even under various noise levels, it still maintains high classification accuracy.