Accurate prediction of remaining useful life (RUL) has been a critical and challenging problem in the field of prognostics and health management (PHM), which aims to make decisions on which component and when should be replaced. In this paper, a novel deep neural network named convolution based long short-term memory (CLSTM) network is proposed to predict the RUL of rotating machineries mining the in-situ vibration data. Different from previous research that simply connects convolutional neural network (CNN) to a long short-term memory (LSTM) network serially, the proposed network conducts convolutional operation on both the input-to-state and state-to-state transitions of the LSTM, which contains both time-frequency and temporal information of signals, and not only preserving the advantages of LSTM but also incorporating time-frequency features. The convolutional structure in the LSTM has the ability to capture long-term dependencies and extract features from time-frequency domain at the same time. By stacking the multiple CLSTM layer-by-layer and forming an encoding-forecasting architecture, the deep-learning model is established for RUL prediction in this paper. Run-to-failure tests on bearings are conducted and vibration responses are collected. Using the proposed algorithm, RUL is predicted and as a comparison, the performance from other methods including deep CNNs and deep LSTM are evaluated using the same data set. The comparative study indicates that the proposed CLSTM network outperforms the current deep learning algorithms in URL prediction and system prognosis with respect to better accuracy and computation efficiency.

Effective fault diagnosis of rotating machinery has multifarious benefits such as improved safety, enhanced reliability and reduced maintenance cost for complex engineered systems. With many kinds of installed sensors for conducting fault diagnosis, one of the key tasks is to develop data fusion strategies which can effectively handle multimodal sensory signals. Most traditional methods use hand-crafted statistical features and then combine these multimodal features simply by concatenating them into a long vector to achieve data fusion. The present study proposes a deep coupling autoencoder (DCAE) model which handles the multimodal sensory signals not residing in a commensurate space, such as vibration and acoustic data, and integrates feature extraction of multimodal data seamlessly into data fusion for fault diagnosis. Specifically, a coupling autoencoder (CAE) is constructed to capture the joint information between different multimodal sensory data, and then a DCAE model is devised for learning the joint feature at a higher level. The CAE is developed by coupling hidden representations of two single-modal autoencoders, which can capture the joint information from multimodal data. The performance of the proposed method is evaluated by two experiments, which shows that the DCAE model succeeds in efficiently utilizing multi-source sensory data to perform accurate fault diagnosis. Compared with other methods, the proposed method exhibits better performance.

In this communication, a multi-task deep learning-driven homogenization scheme is proposed for predicting the effective thermomechanical response of unidirectional composites consisting of a random array of inhomogeneity. Toward this end, 40 000 repeating unit cells (RUCs) comprising an arbitrary number of locally irregular inclusions are generated over a wide range of fiber volume fractions. The finite-volume direct averaging micromechanics is then employed to evaluate the homogenized thermo-mechanical moduli of each RUC. Subsequently, a two-dimensional deep convolution neural network (CNN) is constructed as a surrogate model to extract the statistical correlations between the RUC geometrical information and the corresponding homogenized response. The RUC images together with their homogenized moduli are divided into two datasets in a ratio of 9:1 with the former part used for training the CNN model and the latter part used for verification. The results presented in this contribution demonstrate that the deep CNN predictions exhibit remarkable correlations with the theoretical values generated by the finite-volume micromechanics, with a maximum relative prediction error of less than 8%, providing good support for the data-based homogenization approach.

Fault diagnosis is essential for high energy systems such as liquid rocket engines (LREs) due to harsh thermal and mechanical working environment. In this study, a novel method based on one-dimension Convolutional Neural Network (1D-CNN) and interpretable bidirectional Long Short-term Memory (LSTM) is proposed for intelligent fault diagnosis of LREs. 1D-CNN is responsible for extracting sequential signals collected from multi sensors. Then the interpretable LSTM is developed to model the extracted features, which contributes to modeling the temporal information. The proposed method was executed for fault diagnosis using the simulated measurement data of the LRE mathematical model. The results demonstrate the proposed algorithm outperforms other methods in terms of accuracy of fault diagnosis. Through experimental verification, the method proposed in this paper was compared with CNN, 1DCNN-SVM and CNN-LSTM in terms of LRE startup transient fault recognition performance. The model proposed in this paper had the highest fault recognition accuracy (97.39%).

Anomaly detection plays an important role in ensuring the safety and reliability of complex Internet of Things (IOT) systems through analysis of time-series data. Generally, real-world signals exhibit intrinsic non-stationary characteristics, resulting in a significant challenge the context of anomaly detection for time-series data. In this study, a novel Wavelet-Koopman predictor is proposed for time-series data modeling and anomaly detection underlying non-stationary dynamics of complex systems. The proposed wavelet-Koopman decomposes non-stationary data into time variant and time invariant components and processes them separately using different neural koopman operators. The operation of Wavelet decomposition avoids the impact of non-stationary temporal data on prediction results so the proposed Wavelet-Koopman is able to model and predict non-stationary time series data, which lays the foundation for the implementation of anomaly detection. Several experiments are conducted on common datasets and anomaly detection of liquid rocket engines are realized. The experimental results demonstrate that the proposed method outperforms several recent methods.

Accurate prediction of remaining useful life (RUL) plays a significant role in ensuring the safe flight of aircraft. With the recent rapid development of deep learning, there has been a growing trend towards more precise RUL prediction. However, while many current deep learning methods are capable of extracting spatial features—those along the sensor dimension—through convolutional kernels or fully connected layers, their extraction capacity is often limited due to the small scale of kernels and the high uncertainty associated with linear weights. Graph neural networks (GNNs), emerging as effective approaches for processing graph-structured data, explicitly consider the relationships between sensors. This is akin to imposing a constraint on the training process, thereby allowing the learned results to better approximate real-world situations. In order to address the challenge of GNNs in extracting temporal features, we augment our proposed framework for RUL prediction with a Transformer encoder, resulting in the adaptive graph convolutional transformer encoder (AGCTE). A case study using the C-MAPSS dataset is conducted to validate the effectiveness of our proposed model.