Special features on continuous learning based condition monitoring for crucial components

Fault diagnosis of crucial components has become an indispensable technology in modern industrial complex systems due to the rapid development of high-speed heavy load, and complex mechanical equipment. The condition monitoring tasks are generally submitted in a sequence when addressing a series of fault diagnosis tasks with increments of working conditions, fault types, or machines that often occur in real-world scenarios. Compared with transfer learning- and meta-learning-based fault diagnosis models that focus only on the model performance of the target task and perform poorly on previous tasks due to catastrophic forgetting, the continual learning-based fault diagnosis model requires good performance on all learned tasks and does not require all historical fault data to retrain the model. Continual learning-based fault diagnosis models can continuously learn information on new fault diagnosis tasks to reduce training costs and accumulate this diagnosis information to improve the reliability and generalization capabilities of the diagnosis model.

The special feature discusses the latest advances in measurement science and technology enabling engineering condition-based decisions. The feature includes fault mechanism analysis based on dynamics. The article [1] takes the aero-engine dual-rotor system as the research object and separately studies the vibration behavior of the dual-rotor bearing system in the absence of fault or misalignment, collision, misalignment and collision coupling faults, and provides structural optimization design for the aero-engine rotor system. A surrogate-based model correction method was proposed for multi-condition and multi-measurement point finite element models to address the large uncertainty issue in the material parameters of real objects and the difficulty of constructing finite element simulation models. This method displays strong environmental adaptability, high accuracy, and fast iteration [2].

Condition monitoring is an important way to understand and master the running state of the equipment, which is indispensable in industrial production. Article [3] proposed a health indicator (HI) construction method based on statistical learning modeling for machine condition monitoring. This method aims to address the lack of inherent threshold, robustness, monotonicity, and interpretability of existing virtual and physical HIs. In article [4], the ResNet101 model is used to monitor the state of the molten pool during direct energy deposition. In addition, a new combined model based on Transformer and GRU is proposed in article [5]. This model is used to detect the bearing state of offshore wind turbine gearboxes, providing a valuable solution for the offshore wind power industry.

Signal processing and recognition classification are crucial contents of mechanical fault diagnosis. Article [6] proposes a fault diagnosis method based on improved variational mode decomposition (VMD) and stochastic resonance for diagnosing nonlinear and nonstationary signals of bearings. Article [7] proposes a fault diagnosis method based on improved VMD and maximum correlated peak kurtosis deconvolution for weak vibration signals of electronic equipment. In article [8], an adaptive selective ensemble classification algorithm is introduced based on the idea of ensemble learning to improve the classification performance of a single extreme learning machine (ELM) for fault diagnosis. The proposed algorithm achieves better classification performance than a single model, and the selected ELM subnetworks are chosen adaptively. The ensemble weights of selected ELM subnetworks are also calculated adaptively, which avoids the complex process of manually selecting subnetworks and calculating ensemble weights. Article [9] introduces an improved Fast Newman algorithm based on reliability judgment for complex network community clustering in mechanical fault diagnosis. This method has high clustering accuracy and stability under multiload and multispeed conditions. Thus, the method is expected to have broad application prospects. Article [10] uses machine learning technology combined with sensor signal fusion and feature selection algorithm (mRMR) to improve the fault detection accuracy. In article [11], a fitness function based on the characteristic information of the deconvolution signal in the time and frequency domain is proposed. This function serves as the objective function to facilitate the parameter optimization on adjusted multipoint optimal minimum entropy deconvolution.

Deep learning is one of the most popular technologies in the field of fault diagnosis. In [12], a novel impact feature extraction deep neural network is proposed for intelligent gear and bearing fault diagnosis. The analysis results of gear and bearing faults of different types and defect sizes show that the IFE-DN present remarkable accuracy and satisfactory generalization capability. In [13], a bearing fault diagnosis method based on mixed attention mechanism and deep separable dilated convolution neural network is proposed. To a certain extent, this method reflects the time correlation between signals, alleviates the problems of low recognition accuracy, and addresses excessive parameters under complex working conditions and noise interference. In [14], a fault diagnosis model was proposed on the basis of a dynamic model of a coal mill and deep belief network. Model in this paper can significantly improve the capability to simulate and analyze the coal mill. In [15], a generative adversarial network based on the oversampling method is proposed to realize fault classification with a small dataset. For small or imbalanced dataset conditions, the proposed method could be used to compensate for the dataset. In [16], a bearing fault diagnosis method based on adaptive deep residual shrinkage network and global parametric rectified linear unit is proposed, which can effectively extract the weak fault pulse of CRF pump planetary gearbox bearing, thereby displaying significant engineering application value. In [17], a multisource information fusion classification model based on neural network is proposed to diagnose pump mechanical faults. The model uses a multihead attention mechanism (multihead attention) and an uncertain value throwing mechanism to enhance the adaptability of the model to complex information fusion environments.

In [18], a two-stage multiscale domain adversarial transfer learning method is proposed, demonstrating an effective performance in fault diagnosis. This method can effectively extract domain-invariant features and achieve high accuracy diagnosis in multiple transfer tasks. In [19], a transfer condition monitoring method across wind turbines using feature alignment and parameter finetuning is introduced. This method can accurately monitor and identify the health states of wind turbines. In [20], a few-shot fault diagnosis method based on meta-learning, namely meta-transfer learning with freezing operation (MTLFO), is proposed to solve the aforementioned problems. MTLFO avoids the overfitting problem in traditional meta-learning methods and presents more advantages in solving few-shot problems in fault diagnosis compared with other types of methods.

In [21], a spatiotemporal feature extraction network based on the sensor dynamic graph, namely the DST-GT model, is proposed. The method extracts the unidirectional relationship between sensors from monitoring data to construct a dynamic spatiotemporal graph reflecting the sensor relationship. In [22], a fusion network with two subnets is introduced to adapt to the prediction problem in different scenarios. Meanwhile, a method of cross-weighted joint analysis of the two features is proposed to compensate for the shortcomings of feature analysis and achieve complementarity between time–domain and time–frequency features.

No new data were created or analysed in this study.