Search

Data-driven fault detection techniques have attracted extensive attention in engineering, industry and many other areas in recent years. In many real applications, the following situation often occurs: data for certain types of faults (unseen faults) are not available to train models that are used for fault detection. Such a scenario can occur when data collection becomes highly time-consuming or destructive. To address this challenging problem, a novel fault detection method using zero-shot learning (ZSL) is proposed in this paper, which contains three phases: feature extraction, label embedding, and feature embedding. The method first extracts features from raw signals by applying a one-dimensional convolutional neural network (1D CNN), then builds semantic descriptions (human-defined) as fault attributes shared between seen faults and unseen faults, and finally uses a bi-linear compatibility function to find the highest-ranking fault type. The proposed semantic space based zero-shot learning with 1D CNN is called SSB-ZSL-1DCNN. The cosine distance is used to measure the similarity between feature embeddings and fault attributes. An important characteristic of SSB-ZSL-1DCNN is that the model, trained using only samples of seen faults, can be used to detect unseen defects. To evaluate the proposed method, two case studies are designed based on two well-known benchmarks (the Tennessee-Eastman chemical control process and the rolling bearing experiments at the Case Western Reserve University, respectively). The results demonstrate that the proposed method shows remarkable performance in detecting unseen faults.

Rotating machinery is important to industrial production. Any failure of rotating machinery, especially the failure of rolling bearings, can lead to equipment shutdown and even more serious incidents. Therefore, accurate residual life prediction plays a crucial role in guaranteeing machine operation safety and reliability and reducing maintenance cost. In order to increase the forecasting precision of the remaining useful life (RUL) of the rolling bearing, an advanced approach combining elastic net with long short-time memory network (LSTM) is proposed, and the new approach is referred to as E-LSTM. The E-LSTM algorithm consists of an elastic mesh and LSTM, taking temporal-spatial correlation into consideration to forecast the RUL through the LSTM. To solve the over-fitting problem of the LSTM neural network during the training process, the elastic net based regularization term is introduced to the LSTM structure. In this way, the change of the output can be well characterized to express the bearing degradation mode. Experimental results from the real-world data demonstrate that the proposed E-LSTM method can obtain higher stability and relevant values that are useful for the RUL forecasting of bearing. Furthermore, these results also indicate that E-LSTM can achieve better performance.

Electroencephalogram (EEG) signal analysis plays an essential role in detecting and understanding epileptic seizures. It is\ud known that seizure processes are nonlinear and nonstationary, discriminating between rhythmic discharges and dynamic\ud change is a challenging task in EEG based seizure detection. In this paper, a new time-varying (TV) modeling framework,\ud based on an autoregressive (AR) model structure, is proposed to characterize and analyze EEG signals. The TV parameters\ud of the AR model are approximated through a multi-wavelet basis function expansion (MWBF) approach. An effective ultraregularized orthogonal forward regression (UROFR) algorithm is employed to significantly reduce and refine the resulting\ud expanded model. Given a time-varying process, the proposed TVAR-MWBF-UROFR method can generate a parsimonious\ud TVAR model, based on which a high-resolution power spectrum density (PSD) estimation can be obtained. Informative features are then defined and extracted from the PSD estimation. The TVAR-MWBF-UROFR method is applied to a number of\ud real EEG datasets; features obtained from these datasets are then used for seizure detection and classification. To make the\ud results more accurate and reliable, a PCA algorithm is adopted to select the optimal feature subset, and a Bayesian optimization technique based on the Gaussian process (GP) is performed to determine the coefficients associated with each of the\ud classifiers. Experimental results of the proposed approach outperform the compared state-of-the-art classifiers on two\ud benchmark datasets. Moreover, the results produced by the proposed time-frequency analysis scheme are more reliable for\ud seizure detection based on the noisy EEG datasets used in our case studies.

The use of the multiscale generalized radial basis function (MSRBF) neural networks for image feature extraction and medical image analysis and classification is proposed for the first time in this work. The MSRBF networks hold a simple and flexible architecture that has been successfully used in forecasting and model structure detection of input-output nonlinear systems. In this work instead, MSRBF networks are part of an integrated computer-aided diagnosis (CAD) framework for breast cancer detection, which holds three stages: an input-output model is obtained from the image, followed by a high-level image feature extraction from the model and a classification module aimed at predicting breast cancer. In the first stage, the image data is rendered into a multiple-input-single-output (MISO) system. In order to improve the characterisation, the nonlinear autoregressive with exogenous inputs (NARX) model is introduced to rearrange the available input-output data in a nonlinear way. The forward regression orthogonal least squares (FROLS) algorithm is then used to take advantage of the previous arrangement by solving the system as a model structure detection problem and finding the output layer weights of the NARX-MSRBF network. In the second stage, once the network model is available, the feature extraction takes place by stimulating the input to produce output signals to be compressed by the discrete cosine transform (DCT). In the third stage, we leverage the extracted features by using a clustering algorithm for classification to integrate a CAD system for breast cancer detection. To test the method performance, three different and well-known public image repositories were used: the mini-MIAS and the MMSD for mammography, and the BreaKHis for histopathology images. A comparison exercise was also made between different database partitions to understand the mammogram breast density effect in the performance since there are few remarks in the literature on this factor. Classification results show that the new CAD method reached an accuracy of 93.5% in mini-Mammo graphic image analysis society (mini-MIAS), 93.99% in digital database for screening mammography (DDSM) and 86.7% in the BreaKHis. We found that the MSRBF networks are able to build tailored and precise image models and, combined with the DCT, to extract high-quality features from both black and white and coloured images.

The Atlantic meridional overturning circulation (AMOC), an important component of the climate system, has only been directly measured since the RAPID array’s installation across the Atlantic at 26°N in 2004. This has shown that the AMOC strength is highly variable on monthly timescales; however, after an abrupt, short-lived, halving of the strength of the AMOC early in 2010, its mean has remained ~ 15% below its pre-2010 level. To attempt to understand the reasons for this variability, we use a control systems identification approach to model the AMOC, with the RAPID data of 2004–2017 providing a trial and test data set. After testing to find the environmental variables, and systems model, that allow us to best match the RAPID observations, we reconstruct AMOC variation back to 1980. Our reconstruction suggests that there is inter-decadal variability in the strength of the AMOC, with periods of both weaker flow than recently, and flow strengths similar to the late 2000s, since 1980. Recent signs of weakening may therefore not reflect the beginning of a sustained decline. It is also shown that there may be predictive power for AMOC variability of around 6 months, as ocean density contrasts between the source and sink regions for the North Atlantic Drift, with lags up to 6 months, are found to be important components of the systems model.

System identification and data-driven modeling techniques have seen ubiquitous applications in the past decades. In particular, parametric modeling methodologies such as linear and nonlinear autoregressive with exogenous input models (ARX and NARX) and other similar and related model types have been preferably applied to handle diverse data-driven modeling problems due to their easy-to-compute linear-in-the-parameter structure, which allows the resultant models to be easily interpreted. In recent years, several variations of the NARX methodology have been proposed that improve the performance of the original algorithm. Nevertheless, in most cases, NARX models are applied to regression problems where all output variables involve continuous or discrete-time sequences sampled from a continuous process, and little attention has been paid to classification problems where the output signal is a binary sequence. Therefore, we developed a novel classification algorithm that combines the NARX methodology with logistic regression and the proposed method is referred to as logistic-NARX model. Such a combination is advantageous since the NARX methodology helps to deal with the multicollinearity problem while the logistic regression produces a model that predicts categorical outcomes. Furthermore, the NARX approach allows for the inclusion of lagged terms and interactions between them in a straight forward manner resulting in interpretable models where users can identify which input variables play an important role individually and/or interactively in the classification process, something that is not achievable using other classification techniques like random forests, support vector machines, and k-nearest neighbors. The efficiency of the proposed method is tested with five case studies.

System identification has been applied in diverse areas over past decades. In particular, parametric modelling approaches such as linear and nonlinear autoregressive with exogenous inputs models have been extensively used due to the transparency of the model structure. Model structure detection aims to identify parsimonious models by ranking a set of candidate model terms using some dependency metrics, which evaluate how the inclusion of an individual candidate model term affects the prediction of the desired output signal. The commonly used dependency metrics such as correlation function and mutual information may not work well in some cases, and therefore, there are always uncertainties in model parameter estimates. Thus, there is a need to introduce a new model structure detection scheme to deal with uncertainties in parameter estimation. In this work, a distance correlation metric is implemented and incorporated with a bagging method. The combination of these two implementations enhances the performance of existing forward selection approaches in that it provides the interpretability of nonlinear dependency and an insightful uncertainty analysis for model parameter estimates. The new scheme is referred as bagging forward orthogonal regression using distance correlation (BFOR-dCor) algorithm. A comparison of the performance of the new BFOR-dCor algorithm with benchmark algorithms using metrics like error reduction ratio, mutual information, or the Reversible Jump Markov Chain Monte Carlo method has been carried out in dealing with several numerical case studies. For ease of analysis, the discussion is restricted to polynomial models that can be expressed in a linear-in-the-parameters form.