Search

Dynamic analysis of structures is very important for structural design and health monitoring. Conventional numerical or experimental methods often suffer from the great challenges of analyzing the responses of linear and nonlinear structures, such as high cost, poor accuracy, and low efficiency. In this study, the recurrent neural network (RNN) and long short-term memory (LSTM) models were used to predict the responses of structures with or without nonlinear components. The time series k-means (TSkmeans) algorithm was used to divide label data into different clusters to enhance the generalization of the models. The models were trained with different cluster acceleration records and the corresponding structural responses obtained by numerical methods, and then predicted the responses of nonlinear and linear structures under different seismic waves. The results showed that the two deep learning models had a good ability to predict the time history response of a linear system. The RNN and LSTM models could roughly predict the response trend of nonlinear structures, but the RNN model could not reproduce the response details of nonlinear structures (high-frequency characteristics and peak values). [ABSTRACT FROM AUTHOR]

Defect and damage identification is a crucial task in structural health monitoring (SHM) systems. Recent advances in deep neural networks (DNNs) show success in identification from data for a wide range of SHM systems. However, this approach faces challenges in terms of robustness and scalability with respect to data scarcity. Data collection for the training of DNNs from both the field and laboratory experiments is costly. To address this issue, we employ transfer learning (TL) through the use of deep convolutional neural networks (CNNs) for defect identification in the context of a sensor network's vibration data. A deep TL (DTL) paradigm is used herein so that a pretrained CNN, primarily trained for generalized defect identification tasks where sufficient training data exist (source domain), can be re-trained partially (fine-tuned) as a later secondary process that targets this application domain (target domain) specifically. Different DTL cases are compared, and training data are enhanced with numerical simulation data. The efficacy and robustness of this method are demonstrated on defect identification for full-scale prefabricated concrete shear wall structures with different levels of data scarcity. This method utilizes dynamic responses collected using a sensor network. This is an extension of deep learning vision for non-vision tasks. Defect features are extracted from the dataset of dynamic responses using this DTL frame. Experimental results show that this approach can improve identification models on datasets with few samples. [ABSTRACT FROM AUTHOR]

Generalization remains a significant challenge in the application of deep learning methods, particularly in predicting mooring line tension for moored floating structures. The dynamic and complex nature of marine environments makes accurate prediction of mooring line tension a formidable task. Existing deep learning models, constrained by the data distribution of their training sets, typically perform well within the scope of these datasets. However, their limited generalization capability becomes apparent when confronted with unknown and changing marine conditions, resulting in a decline in prediction accuracy. To address this issue, this paper proposes a Domain Generalization (DG) model approach, which incorporates a distribution matching regularization term to the model's loss function, aimed at reducing distributional disparities between data from different domains. The proposed method was compared with existing Deep Neural Network (DNN) model method through validation with engineering case studies. The results indicate that the DG model method significantly surpasses the DNN model in terms of generalization capabilities. The mooring line tension prediction model developed herein holds potential for constructing digital twins, offering near real-time data support for moored floating engineering projects. This facilitates timely and accurate decision-making in response to environmental changes. • Develops a Domain Generalization model to predict mooring tension in unknown sea conditions. • Demonstrates improvements in model generalization, surpassing traditional Deep Neural Network (DNN) methods. • The effectiveness of the model was validated through a practical engineering case study of a floating wind turbine. [ABSTRACT FROM AUTHOR]

Data scarcity is a major challenge for the application of deep learning methods in detecting mooring line failure. Existing deep learning methods necessitate significant quantities of labeled data, which can be both costly and difficult to obtain in real-world scenarios. Moreover, a trained deep learning model may fail when faced with a shift in data distribution due to changes in the application scenario. To address this issue, this paper proposes a low-supervised cross-domain approach that utilizes gated recurrent unit (GRU) as the basic unit of the deep neural network for mooring line failure detection. The proposed approach employs domain-adversarial training paradigm, enabling it to extract domain-invariant features. The proposed domain-adversarial neural network (DANN) approach is verified through several case studies. These cases consider single and combined changes in conditions of a moored floating system, e.g., failure type, sea state, vessel displacement, and mooring water depth. The results show that the model demonstrates exceptional accuracy in detecting mooring line failures in the target domain without labeled data. The proposed approach is compared with classical deep neural network (DNN) approach, and the accuracy variation curves on source domain and target domain samples during the training process are presented. [ABSTRACT FROM AUTHOR]

Mooring lines are a crucial component of offshore mooring systems, and long-term fatigue damage assessment is a challenging task in mooring line design and structural health monitoring. Traditional methods require a large number of structural time-domain simulations. Because the sea state varies over time, the variations in wave parameters (e.g., wave heights, periods, and directions) must be considered in long-term fatigue damage assessment. Wave directions have usually been ignored in previous studies, while the wave direction sensitive analysis conducted in this study showed that the wave directions could not be neglected in the long-term fatigue damage assessment of mooring lines. Recent deep neural network (DNN) methods have shown great potential in mooring line/riser response prediction for offshore structures. However, the utilization of float responses as input data in these DNNs restricts their applicability during the structure design phase. To address this issue, a gated recurrent unit (GRU) network is proposed for mooring line tension modeling. The proposed GRU network surrogate models are capable of accurately predicting the mooring line tension using wave elevation components or float motion responses. The rain-flow counting algorithm, T-N approach, and Miner's rule were employed to estimate the structural fatigue damage. The results showed that the proposed method could replace the heavy cost numerical computation, assess the long-term fatigue damage with high accuracy and alleviate the data dependency of DNN methods. • Deep learning models using the wave components or vessel motion as input were proposed to predict mooring line response. • An approach for long-term fatigue damage assessment of mooring lines considering wave directionality was proposed. • For the model using vessel motion as input, a small number of data collected from several sea states is sufficient. [ABSTRACT FROM AUTHOR]

This work proposes a solution to improve transfer learning from laboratory environment to real-world hydraulic machinery (centrifugal pump) for the effective identification of defects. The proposed method involves collecting vibration data, normalizing and computing Fast Fourier Transform (FFT) of the data, and adding attributes by combining the FFT of healthy real-world machinery with the FFT of laboratory machinery. The attribute addition is performed in the frequency domain to address phase lag issues in time-domain data. The resulting signal is transformed back to the time domain, and the envelope spectrum is obtained. An approximation model is constructed using the envelope spectrum and refined using data from industrial machinery. The refined model is then employed to identify defects in real machinery, specifically the centrifugal pump. The proposed knowledge addition-based transfer learning achieves an accuracy of 93%, which is 51% higher than the accuracy attained by the domain-adversarial neural network method. [ABSTRACT FROM AUTHOR]

Automatic identification of bearing and rotor defects, when operated at varying speed is challenging. To make this challenging task possible, a tacho-less deep learning model is developed which can effectively learn, even from small data set. For accurate learning from small data set, existing CNN is made sparse. Sparsity is incorporated in the CNN by adding newly developed sparsity cost in the existing cost function of CNN to enhance the learning capability of CNN. The method works in the following steps. First, vibration signals are processed with Fourier synchro squeezed transform (FSST) to obtain tachometer information. The extracted tachometer information is used to change the time domain signal to angular domain signal. Second, wavelet transform of angular domain signals is carried out to produce time–frequency images. Third, time–frequency images of angular domain signals are applied to the improved version of CNN. After learning, time–frequency images obtained from angular domain signals of defective bearings and rotor are applied to detect defects. The defect identification accuracy attained by the proposed method is 96.6 %. This accuracy is higher as compared to the accuracy achieved by the methods used in existing works. This has been made possible due to sparsity cost functions assimilated in the cost function of CNN that evade avoidable activation of neurons in the feature extraction layers of CNN, which makes the learning of modified CNN becomes deeper in comparison to existing CNN. [ABSTRACT FROM AUTHOR]

• A cost function is developed to make CNN sparse. • A transfer learning methodology based on sparse learning is proposed. • The developed transfer learning methodology is applied to identify defects in rotor and gear. • The accuracy of developed transfer learning is better than the existing transfer learning. It is incredibly difficult to build a data-driven machine learning model for the automatic detection of defects in rotating machinery. The existing techniques, based on machine learning models, work satisfactory for one machinery but unsatisfactorily for others. The problem of identifying defects becomes sever when the enough training data is not available. Here, a sparse deep learning model is put forward that can efficiently learn from the limited training data. The existing cost function of CNN has been improved by adding sparsity cost for the purpose of improving the performance of deep learning. To assimilate the sparsity, unnecessary activation of neurons is averted in the feature extraction layer of CNN. A trigonometric sparsity cross entropy (TSCE) function is built to obtain the sparsity cost. The updated CNN is trained with enough samples from the source domain. Thereafter, fine-tuning of the model is carried out from small data samples of the target domain for detecting defects. The testing of proposed deep learning model is carried out on two types of data set, one is gear and another is rotor. The concluding results obtained from the proposed work has been compared with current state-of-the-artwork. The comparison analysis shows the usefulness of the proposed methodology over the existing methods. [ABSTRACT FROM AUTHOR]