Continuous Learning Based Condition Monitoring for Crucial Components

Deep learning for bearing fault diagnosis often requires a large quantity of comprehensive data to give support in the field of rotating machinery fault diagnosis. However, large-quantity datasets for model training are difficult to obtain in actual working environments. Therefore, bearing fault diagnosis problems under practical working conditions are often considered few-shot problems. Meta-learning can be adopted to solve these few-shot problems. Traditional meta-learning methods, however, can lead to model overfitting, and shallow neural networks are usually used to avoid overfitting. As a result, the features extracted by the shallow neural network are insufficiently rich to exploit the optimal performance of the model. A few-shot fault diagnosis method based on meta-learning, named meta-transfer learning with freezing operation (MTLFO), is proposed in this study to solve these problems. MTLFO can learn new knowledge rapidly through a small number of samples. The hyperparameter self-regulation ability of meta-learning is adopted by MTLFO, and a freezing operation is used to deal with the neuronal nature of meta-learning to ensure that the neurons from different tasks are transferred by utilizing scaling and shifting. 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.

Owing to the nonlinearity and nonstationarity of the bearing fault signal, it is difficult to identify fault characteristics under the influence of a strong noise environment. The extraction of early weak fault features is critical for the reliability of bearing operations. Therefore, an urgent problem is reasonable noise reduction and feature enhancement in weak-signal processing. Traditional variational modal decomposition (VMD) and stochastic resonance (SR) are commonly applied to detect weak signals in fault diagnosis. The VMD method can decompose the signal into several intrinsic mode functions (IMFs) to effectively reduce the modal aliasing problem. However, uniform standards for the key parameters of decomposition and the selection of the optimal IMF after decomposition are lacking. Meanwhile, some disadvantages of SR still exist; for example, the interference of multiscale noise may lead to false detection by incorrect selection of high-pass filter parameters, and the system parameters are not adaptive to different signals to achieve the best response output. To better address the weak signal feature enhancement, a novel rolling bearing fault diagnosis method combining adaptive VMD and SR by improved differential search (IDS) optimization is proposed. First, the bearing fault signal is decomposed into several IMFs using the IDS-VMD. Second, the feature information of the fault signal is retained and reconstructed using the correlation kurtosis for sensitive modal extraction. Furthermore, the fault features of the reconstructed signal are effectively enhanced by the variable-step IDS-SR, which can reasonably transfer the noise energy of the input components to the fault characteristic frequency. Finally, the periodic pulse can be observed in the corresponding envelope spectrum. The simulated and experimental data show that the proposed method can not only effectively extract the signal feature information in the actual fault but also realize early weak fault diagnosis of rolling bearings more accurately.

Due to the complex working environment, effective fault data from wind turbine gears are often difficult to obtain. Aiming at this practical issue, a generative adversarial network (GAN)-based oversampling method is proposed in this paper, which can achieve fault classification with a small dataset. In the initial stage, wavelet packet transform is applied to generate and extract features. Then, the optimal discriminator and generator trained by GAN are used to generate data to compensate for the imbalanced fault dataset. Random forest, eXtreme gradient boosting and support vector machines are chosen to classify a real dataset, imbalanced dataset and generated dataset, respectively. Experiments indicate that the data generated by the proposed method stay at the same distribution as the real data. Therefore, for small or imbalanced dataset situations, the proposed method could be a solution to compensate for the dataset.

To address the issue of weak fault features and difficulty in feature extraction for planetary gearbox bearings of the circulation water pump (CRF pump) unit in the conventional island of a nuclear power plant, a deconvolution method, named the multipoint optimal minimum entropy deconvolution adjusted (MOMEDA), is optimised by a mixing operator improved quantum behaviour particle swarm optimisation (QPSO) algorithm, to extract the periodic fault impulse of planetary gearbox bearings. In the mixing operator, a differential evolution operator is introduced to improve particle swarms' diversity and enhance the algorithm's global optimisation capability. Meanwhile, a proposed adaptive CrossOver operator is incorporated into the algorithm to increase its convergence speed. Then, combining these two types of operators can construct a parameter optimisation algorithm displaying both global optimisation capability and high algorithm execution efficiency. Consider capturing the periodic impulsive features relating to bearing faults. 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 optimisation on MOMEDA. Thereby, it enhances the optimisation capability of our feature extraction method. Experiments were conducted by adopting the proposed method on the signals collected from the inner- and outer-race faulty planetary bearing in a test bed, in which favourable feature extraction results are obtained. After the comparative analysis, we observed that both significances of features extracted by this method and the execution efficiency of the algorithm are superior compared to other methods. The on-site feature extraction results of the input shaft bearing in the planetary gearbox of a CRF pump unit in a nuclear power plant also demonstrated the engineering practicability of the method proposed in this work.

Taking the dual-rotor system in an aero-engine as a research object, the vibration behaviors of a dual-rotor bearing system that is fault-free or has misalignment, rub-impact or misalignment rub-impact coupling faults are studied, respectively. First, the three-dimensional model of the dual rotor bearing system is established, and then the first six modes and critical speeds of the dual rotor bearing system are obtained by the finite element method. Then, the dynamic equations of the dual rotor bearing system with and without faults are derived based on the Lagrange method. The Runge–Kutta method is used to solve the dynamic equation, and the nonlinear dynamic response of the system is obtained. The vibrational behaviors of the dual rotor system with misalignment, rub-impact and misalignment rub-impact coupling faults are analyzed and discussed through the time history, phase diagram and frequency spectrum of the rotor. The vibrational behaviors of the first six modes of the dual rotor system with and without faults are analyzed, respectively. The results provide theoretical guidance for the structural optimization design of an aero-engine rotor system and the prior decision information for misalignment and rub-impact coupling fault diagnosis.

The complex network, as a method for the analysis of nonlinear and non-stationary signals, overcomes the shortcomings of traditional time-frequency analysis methods and proves its effectiveness in mechanical fault diagnosis. Community clustering, a type of complex network, has made great progress in recent years. However, the existing community clustering algorithms have disadvantages in that they lack significant global extreme value and huge search spaces. Therefore, a Fast Newman algorithm based on reliability judgment is proposed. Starting from the community structure characteristics of the complex network, with the fault sample as a network node, the relationship between the samples as a connected edge and a complex network model of fault data is established. Clusters in troubleshooting are transformed into community structure discovery in the network. Firstly, the initial division of the community is obtained by measuring the distance between the samples. Then, the modularity index of the network is used as a standard function of the community division, and the bottom-up community merger is performed. The local edge density index is used for reliability determination before each combination to achieve global optimization, and the network block structure is the most obvious. Finally, with all the data merged into one community, the optimal division of the community structure is obtained, while accurate community clustering and fault diagnosis is realized. The benchmark graphs for testing community detection (Lancichinetti–Fortunato–Radicchi benchmark standard test network, LFR) and different fault data of rolling bearings under multiple operating conditions are applied to verify the effectiveness of this method; the results prove that the modified Fast Newman algorithm has better clustering effects and a higher accuracy rate than the original algorithm. Compared with K-means clustering and fuzzy clustering, the modified Fast Newman algorithm achieves higher performance in fault diagnosis of rolling bearings under multiple operating conditions.

Machine condition monitoring aims to evaluate machine health conditions by analyzing machine vibration signals, which is helpful to make timely maintenance decisions and prevent unexpected accidents. Currently, constructions of virtual and physical health indicators (HIs) are commonly used methods for machine condition monitoring. However, most classic physical and virtual HIs lack inherent thresholds, robustness, monotonicity, and interpretability for machine condition monitoring. In this paper, a statistical learning modeling based HI construction method for machine condition monitoring is proposed to solve these problems. Firstly, a statistical decision theory is suggested to clearly describe a machine condition monitoring objective, and subsequently shapes of square envelope spectra are robustly modeled by using a parametric statistical model called a penalized B-spline approximation. Further, an interpretable HI named B-spline weight HI (BSWHI) as well as an inherent statistical threshold is accordingly constructed based on the Mahalanobis distance between B-spline weights of testing samples and a healthy sample. Experiments on bearing and gear run-to-failure datasets are studied to show that the proposed BSWHI and its inherent statistical threshold can effectively detect early machine faults and simultaneously provide monotonic degradation assessment trends. The proposed interpretable BSWHI has achieved a substantial improvement over existing classic HIs.

One critical challenge of directed energy deposition (DED) in additive manufacturing (AM) is the lack of comprehension of the relationship between the part parameters and the formation quality. Components fabricated by the inappropriate manufacturing parameters will be too unreliable to satisfy the strict requirements of industrial applications. To address these issues, the present study established an experiment with a coaxial high-speed charge coupled device (CCD) camera to monitor the 316L deposition process and developed a data-driven model with ResNet101 to identify different melt pool states. We adopted the t-distributed stochastic neighbor embedding clustering algorithm, accuracy rate, and normalized confusion matrix to evaluate the performance of ResNet101. Furthermore, the visualization technique class activation mapping was used to analyze the morphological characteristics of the melt pool formed under different experimental parameters, explained the classification behavior of the ResNet101 model. The result indicated that ResNet101 gains better feature extraction and higher capability to classify different melt pool states with an average accuracy of 99.07%, compared with other CNNs (LeNet, GoogLeNet, AlexNet, ResNet34, and ResNet50) models. This demonstrated the effectiveness of ResNet101 in monitoring the DED process and the potential to reduce fabrication costs in DED.

Recently, deep learning has been widely used for intelligent fault diagnosis of rolling bearings due to its no-mankind feature extraction capability. The majority of intelligent diagnosis methods are based on the assumption that the data collected is from constant working conditions. However, rolling bearings often operate under variable working conditions in the real diagnosis scenario, which reduces the generalization capability of the diagnosis model. To solve this problem, a self-adaptive deep residual shrinkage network with a global parametric rectifier linear unit (DRSN-GPReLU) is proposed in this paper. First, the DRSN is used as the basic architecture to improve the anti-noise ability of the proposed method. Then, a novel activation function—the GPReLU—is developed, which can achieve better intra-class compactness for vibration signals, and the inter-class samples are better mapped into remote areas. Finally, a sub-network based on the attention mechanism is designed to automatically infer the slope of the GPReLU. Various experimental results demonstrate that the DRSN-GPReLU can realize better performance compared with traditional methods under variable working conditions, and has better robustness under noise interference.

Gears and bearings are important components in rotating machinery and are crucial for the safety and operation of the whole mechanical system. Intelligent fault diagnosis methods based on deep-learning algorithms have undergone rapid development in recent years. Despite this, integrating fault features in a deep network construction remains a challenge for intelligent fault diagnosis of rotating machinery. In this paper, a novel impact feature extraction deep neural network (IFE-DN) is proposed for intelligent gear and bearing fault diagnosis. An improved three-layer Laplace wavelet kernel convolutional neural network (LW-CNN), where the Laplace wavelet kernel is designed in the first convolutional layer, is constructed to extract and enhance the impact features in the vibration signal. Using a visualized heat map, the physical meaning of the LW-CNN's extracted features is explained and the interpretability of the network model is enhanced. The wavelet function selection in the deep neural network is also discussed. The extracted features are transferred to a primary capsule layer and a digital capsule layer. With a feature vector converting process and dynamic routing algorithm, more detailed features are optimized and the fault types are classified. Four experimental data sets from different laboratories are used to verify the performance of the proposed model, and t-distributed stochastic neighbour embedding is carried out to visually analyze the extracted features in different layers. The results of the analysis of gear and bearing faults of different types and defect sizes show that the IFE-DN presents significant accuracy and satisfactory generalization ability.

The finite element model inversion method has been widely used in recent years for iterative adjustment of finite element model parameters. However, the models constructed in the existing literature are weak and time consuming to adapt to the environment, which makes it difficult to adapt to the current needs of numerical simulations. To address the problem of large uncertainty in the material parameters of real objects and the difficulty of constructing finite element simulation models, a surrogate-based model correction method was proposed for multi-condition and multi-measurement point finite element models. The innovative use of the working condition parameter as one of the training parameters of the surrogate model to construct the optimal mathematical model for parameter correction of the finite element model to variable working conditions. To reduce the number of finite element model calls and speed up the convergence process, an Minimizing Prediction-CV-Voronoi parallel infill sampling method for the surrogate model was proposed to overcome the problems of easily falling into local optima and slow convergence when solving after constructing the surrogate model. The proposed parallel infill sampling method was tested using the test functions. The finite element model correction method with multiple working conditions and multiple measurement points was applied for material parameter correction and identification of aluminum alloys. The superiority of the proposed parallel point addition method in terms of the solution accuracy and speed was demonstrated. The results show that the multi-measurement points have a significant effect on improving the model correction effect, and the constructed multi-condition surrogate model can make fast predictions for arbitrary conditions and has strong environmental adaptability. The finite element model correction method proposed in this paper, with strong environmental adaptability, high accuracy and fast iteration, has been tested to be very effective.

During the manufacturing process of electronic equipment, objects such as tin beads and glue blocks may be left in the electronic equipment, causing failure of the electronic equipment. This paper uses experimental equipment to collect weak vibration signals on the surface of electronic equipment. In view of the nonlinear and non-stationary characteristics of the vibration signal and its easy to be masked by strong background noise, a fault diagnosis method of weak vibration signal based on improved variational mode decomposition (VMD) and maximum correlation kurtosis deconvolution (MCKD) is proposed. Cosine factors and adaptive weights are introduced to improve the convergence accuracy of the Whale Optimization Algorithm (WOA). The envelope spectrum peak factor is used as the adaptability function of the improved whale algorithm (IWOA) to optimize the parameters of VMD and MCKD. Firstly, based on the decomposition results of weak fault signals by IWOA-VMD, the optimal modal components are selected. Secondly, the fault impact component in the optimal modal component is enhanced based on the IWOA-MCKD algorithm. Finally, the fault characteristic frequency is extracted through the envelope spectrum. The feasibility and superiority of the proposed optimization method are verified through simulation signal analysis and actual case study.

Through the analysis of data from the SCADA system of a wind turbine unit in a specific offshore wind farm located in Zhanjiang, it was observed that the most prevalent type of fault is bearing alarms on the gearbox's generator side. Considering the growing need for intelligent offshore wind turbine maintenance, this study employed GRA on SCADA data collected over a significant duration from a representative wind turbine unit. Relevant features were extracted, with temperature serving as the target parameter. To address the challenge of long-term dependencies in long-term time series forecasting tasks, this study uniquely combined the GRU with the advanced Transformer neural network which incorporates attention mechanisms, to predict the temperature trend of the gearbox's generator-side bearing. Based on the prediction residuals obtained during normal operation and their subsequent analysis, the study devised an effective anomaly detection process to identify early abnormal states of the gearbox's generator-side bearing. Comparative performance evaluations were conducted, comparing the combined model with its individual component models, as well as the traditional LightGBM, in terms of temperature time series prediction and their application in anomaly detection. The results unequivocally demonstrate that the combined model outperforms both the individual models and LightGBM in terms of time series prediction accuracy and anomaly detection effectiveness, indicating an enhanced ability to handle long-term memory challenges. Furthermore, the combined model exhibits great potential of practical application for the early warning of gearbox bearing anomalies during actual wind turbine daily operation and maintenance, providing a valuable solution for the offshore wind turbine industry.

In machine intelligence fault diagnostic and health status decision-making systems, rich, complex, and fuzzy feature information cannot facilitate fault decision-making merely on a single data source. This requires utilizing the heterogeneity of information gathered from multiple sources to diminish the system's uncertainty and improve the accuracy of decision-making. In this work, a novel neural network-based multi-source fusion classification model is proposed to diagnose the pump mechanical faults. The multi-head attention Dempster–Shafer (D–S) evidence fusion (MADS) system extends the model's ability to focus on rich features. Furthermore, the uncertain values throwing mechanism can effectively eliminate samples from uncertain categories and increase the model's ability to distinguish diagnostic results with low confidence. Compared with a single sensor, our multi-sensor joint decision based on seven sensors considerably improved the fault diagnostic accuracy of MADS system, which has increased by at least 12.34%. Experimental validation demonstrates that utilizing the improved combination rules provided for multi-source evidence fusion fault diagnosis can significantly improve the efficacy of conventional D–S fusion and reduce the probability of misjudgment; combining the multi-head attention mechanism can dramatically increase the precision of model fault diagnosis. The proposed method has the potential to substantially accelerate research in the field of multi-source sensor joint fault diagnosis.

Accurate implementation of remaining useful life (RUL) prediction of equipment is essential for health management and maintenance decisions. Advances in sensing and big data technologies have provided the possibility for research on data-driven life prediction methods. However, the current RUL prediction methods still need to improve in utilizing multi-source information. The present techniques consider more the relationship between the temporal information and RUL prediction among the multi-source information and rarely target the research and exploitation of the potential connection between sensor networks and RUL. Therefore, this paper proposes a spatio-temporal feature extraction network based on the sensor dynamic graph: DST-GT model. The method extracts the unidirectional relationship between sensors from monitoring data to construct a dynamic spatio-temporal graph reflecting the sensor relationship. The DST-GT model uses graph convolution based on message selection (MSGCN) to model the spatial dependencies of sensors, and uses multi-scale gated temporal convolution module (MGTCN) to model the temporal dependencies in sensor state monitoring data. In this paper, graph learning, graph convolution and temporal convolution modules are jointly learned in an end-to-end framework. The results on two widely used datasets and comparisons with other methods demonstrate the accuracy and advancement of DST-GT networks for RUL prediction.

Nowadays, the methods of remaining useful life (RUL) prediction based on deep learning only use single model, or a simple superposition of two models, which makes it difficult for to maintain good generalization performance in various prediction scenarios, and ignores the dynamic sensitivity of features in the prediction, limiting the accuracy. This paper proposes a method of RUL prediction of bearing using fusion network through two-feature cross weighting (FNT-F). First, a fusion network with two subnets is proposed in this paper to adapt to the prediction problem in different scenarios. Meanwhile, a method of cross weighted joint analysis of the two features is proposed to make up for the shortcomings of feature analysis and achieve complementarity between time-domain and time–frequency features.

Aiming at the problems of traditional fault diagnosis methods that do not represent the time correlation between signals, low recognition accuracy under complex working conditions and noise interference and too many parameters, a bearing fault diagnosis method based on mixed attention mechanism (MAM) and deep separable dilated convolution neural network (DSDCNN) is proposed. Firstly, a Markov transfer field encoding method is used to transform the original one-dimensional vibration signal into a two-dimensional feature image with temporal correlation. Secondly, a deep separable convolution algorithm is presented by taking advantage of the low computational complexity of deep separable convolution and the ability of dilated convolution to expand the receptive field under the condition of invariable number of parameters. Then, the MAM is designed to make the model capture the feature dependency of the feature map in spatial and channel dimensions, and the MAM-DSDCNN model is constructed. Finally, the fault diagnosis performance of the proposed model is verified with two different data sets. The results show that the average recognition accuracy of MAM-DSDCNN reaches 99.63% under variable load conditions, 99.42% under variable speed conditions, 94.26% under noisy environment with the signal-to-noise of 0 dB, which prove that the model has higher recognition accuracy, stronger generalization and noise immunity performance than other deep learning algorithms.

With the aim of improving the classification performance of a single extreme learning machine (ELM) for fault diagnosis, an adaptive selective ensemble classification algorithm is proposed based on the idea of ensemble learning. The proposed algorithm achieves better classification performance than a single model and the selected ELM subnetworks are selected adaptively and the ensemble weights of selected ELM subnetworks are calculated adaptively, which avoids the complex process of manually selecting subnetworks and calculating ensemble weights. First, a regularized error weight adjustment ELM is constructed by introducing regularization parameters and error weight adjustment matrix to a standard ELM, where the error weight adjustment matrix is obtained by the method of adaptive Gaussian kernel density estimation. Then, discrepancy subnetworks are constructed using six different activation functionsand the ensemble weights of subnetworks are obtained adaptively according to the normalized ratio of mean and variance of subnetwork F-scores to achieve the ensemble of subnetworks. Finally, the adaptive selective ensemble classification algorithm is validated using the UCI dataset and experimental acoustic emission signals of gearbox faults. The results show that the adaptive selective ensemble method can improve the stability and accuracy of classification algorithms and the achieved classification accuracy for experimental acoustic emission was 0.9773.

Domain adaptation (DA) based intelligent fault diagnosis (IFD) methods have attracted great attention in recent years. The key motivation of DA methods is to extract the domain-invariant features. In most cases, the one-dimensional convolutional neural networks are adopted as the feature extractor, in which the kernels are usually single and fixed. However, the monitoring data for IFD usually involve various scale information, and the feature representation extracted by the above models may be incomprehensive. Moreover, the target domain data is only used to narrow the distribution discrepancy in an unsupervised way, which may lead to ignorance of class information of the target domain. To address these issues, in this paper, a two-stage multi-scale domain adversarial fault diagnosis method is proposed. A multi-scale feature extractor with different kernel sizes is designed to acquire more discriminative domain-invariant features. Meanwhile, pseudo label learning is adopted to provide the transfer learning process with the pseudo labels of the target domain, which are generated by a pre-trained network in the first stage and then are optimized through iterations in the second stage. The maximum mean discrepancy is also adopted to enhance the ability of the model for marginal distribution alignment, which can make the model more robust. Thirty-eight transfer tasks from two different datasets were conducted to evaluate the effectiveness of the proposed method. The experimental results demonstrated that the proposed method achieved higher average diagnosis accuracy compared with several popular methods. The superiority of our proposed method was further explained by visualization of learned features.

As the significant ancillary equipment of coal-fired power plants, coal mills are the key to ensuring the steady operation of boilers. In this study, a fault diagnosis model was proposed on the basis of a dynamic model of a coal mill and deep belief network (DBN). First, a dynamic coal mill model that considered the joint influence of drying, ventilation and grinding forces was established. Parameters in the model were identified by designing a two-phase optimization method based on the genetic algorithm. Then, this model was used for simulating the common faults of coal mills under a variety of operating conditions and obtaining extensive data. On this basis, the DBN fault diagnosis model was established and the combination of parameters was optimized by use of an orthogonal experiment. Finally, the validity of the model was verified by using the actual operation data of the coal mill. Compared with the dynamic models built in previous studies, that constructed in this paper can significantly improve the capability to simulate and analyze the coal mill. The convergence rate of the designed two-phase optimization method was improved. The experimental results show that the proposed method of coal mill fault diagnosis based on the dynamic model and DBN has an accuracy of 95%, which proves that this method has excellent application potential.

The predictive maintenance approach in the fault diagnosis of rotating machines is becoming important in industries in order to minimize operational cost and to improve the reliability of machines. Recent studies have focused on developing an effective prediction model to detect machinery faults at various speeds using machine learning techniques. In this research study, machine-learning-based models are developed to detect multi-component faults such as gear faults, bearing faults, and unbalanced shafts, by training the four speed data sets using decision trees, support vector machines and artificial neural networks. The machine learning models are improvised by fusing the vibration signals (X, Y and Z) and sound signals with the feature selection algorithm, minimal redundancy-maximum relevance (mRMR). The research work is extended by training a model with any three speeds and testing the trained model with the remaining speed. The state-of-the-art algorithms used to train all four speeds yield poor performance in the prediction of faults at testing speeds. A linear and quadratic discriminant analysis (QDA) is chosen based on its multivariate discrimination capability for better fault prediction at testing speed. The proposed combination of QDA with mRMR selection of fused vibration and sound feature set perform well in the detection of multiple faults at variable speeds.

Transfer condition monitoring across wind turbines still exists two issues. (1) The model of single turbine trained is directly used for other ones in the wind farm to result in low monitoring accuracy because of individual features of each turbine; (2) if the model is trained separately for each turbine, it is not only time-consuming and inefficient, but also is difficult to transfer across turbines. For solving the two issues, a transfer condition monitoring method across wind turbines using feature alignment and parameter fine-tuning is proposed. First, the representative normal turbine in a wind farm is screened by calculating the correlation between the monitoring parameters of wind turbines; second, the input features are constructed by using the massive normal supervisory control and data acquisition data of the representative turbine. The intelligent condition monitoring model of wind turbines is designed by using convolutional neural networks for spatial features and long-term and short-term memory networks for time features. Finally, feature alignment is performed using the normal data of the representative turbine and a large amount of historical normal data from other turbines to reduce data distribution differences. The aligned data is inputted into pre-training monitoring models for parameter fine-tuning. The personalized features of each turbine are incorporated to achieve high-precision transfer condition monitoring across turbines. The proposed method is verified by the real-world turbine data from a wind farm cooperated with us. The results show that the proposed transfer condition monitoring method using feature alignment and parameter fine-tuning can accurately monitor and identify the health states of wind turbines.