Search

The intelligent fault diagnosis powered deep learning (DL) is widely applied in various practical industries, but the conventional intelligent fault diagnosis methods cannot fully juggle the manifold structure information with multiple-order similarity from the massive unlabeled industrial data. Thus, a new Multiple-Order Graphical Deep Extreme Learning Machine (MGDELM) algorithm for unsupervised fault diagnosis (UFD) of rolling bearing is proposed in this study. Specifically, the developed MGDELM algorithm mainly contains two parts: 1) one is unsupervised multiple-order feature extraction, the first-order proximity with Cauchy graph embedded is applied to extract the local structural information, and the second-order proximity is simultaneously employed for mining global structural information and 2) the other used is the unsupervised Fuzzy-C-Mean (FCM) into fault clustering built on the extracted multiple-order graph embedding features. Empirically, two cases of rolling bearing failure data validate the effectiveness of the proposed algorithm and fault diagnosis method. By jointly optimizing the multiple-order objective function, the proposed MGDELM algorithm can synchronously extract local and global structural information from the raw industrial data. This study also provides a novel promising approach for UFD.

Geomagnetic data forecasting plays a critical role for natural disaster institutions to respond promptly or make decisions for the magnetic storm warning, earthquake early warning, and so on. However, the forecasting accuracy is not always reliable due to the spatial correlations among different sites and the temporal correlations from each site measurement update. To address the correlation problems, a novel sparse geomagnetic data forecasting (SGCast) matrix framework is proposed in this study. To be specific, a coupled matrix factorization is proposed to model the sparse multilocation geomagnetic data and spatial correlation, which is demonstrated with an example from seven Chinese cities. After the factorization, two subspaces, location subspace, and temporal subspace are derived. Finally, the future geomagnetic signals are reconstructed based on forecasting temporal subspace and matrix reconstruction. The experimental results from the extensive comparison studies demonstrate the superiority of the proposed SGCast approach compared to the state-of-the-art approaches, with an approximate improvement of the forecasting accuracy as 10%~15%.

The property value assessment in the real estate market still remains as a challenges due to incomplete and insufficient information, as well as the lack of efficient algorithms. House attributes, such as size and number of bedrooms, are currently being employed to perform the estimation by professional appraisers and researchers. Numerous algorithms have been proposed; however, a better assessment performance is still expected by the market. Nowadays, there are more available relevant data from various sources in urban areas, which have a potential impact on the house value. In this paper, we propose to fuse urban data, i.e., metadata and imagery data, with house attributes to unveil the market value of the property in Philadelphia. Specifically, two deep neural networks, i.e., metadata fusion network and image appraiser, are proposed to extract the representations, i.e., expected levels, from metadata and street-view images, respectively. A boosted regression tree (BRT) is adapted to estimate the market values of houses with the fused metadata and expected levels. The experimental results with the data collected from the city of Philadelphia demonstrate the effectiveness of the proposed model. The research presented in this paper also provides the real estate industry a new reference to the property value assessment with the data fusion methodology.

A miniature magnetic gradient tensor system may be used in small spacecraft for mineral prospecting and detection of unexploded explosive ordnance. The system consists of a magneto-inductive (MI) sensors array, a signal extraction circuit, and a microcontroller. A magnetic gradient full-tensor configuration utilizes four MI sensors arranged in a planar cross structure and is employed to minimize orientation and orthogonality errors. The system's performance was checked by data validation and magnetic anomaly detection. Numerous experimental results demonstrate that the system is able to detect a low-amplitude magnetic field as low as 8.28 nT at 10 Hz, with a typical noise floor of 0.235 nT/Hz $^{1/2}$ at 1 Hz. The system can identify nearby ferromagnetic objects clearly, which indicates its feasibility for weak magnetic anomaly detection.

Infrared to Visible (IR2VIS) image registration suffers from the challenge of cross-modal feature extraction and matching. Conventional methods usually design the same keypoint detector for both Infrared (IR) and Visible (VIS) images. The VIS images are even converted to gray-scale images before the keypoint detection. IR and VIS gray-scale images have different properties which might not be applicable for the same feature detector. Therefore, this paper proposes an IR2VIS image registration method, namely, Image Translation for Image Enhanced Registration (ITIER). The IR images are first translated to realistic VIS images by Wavelet-Guided Generative Adversarial Network (WGGAN) for the convenience of cross-modal feature detection. Then the keypoint detection and matching and the homography transformation, which have been integrated into our ITIER, are conducted on the translated and original VIS images. Experimental results demonstrate that the IR2VIS image registration accuracy is greatly enhanced by the image-to-image translation procedure, which transfers IR images to realistic VIS images.

In modern industries, context enhancement is critical to capture objects, especially in a low illuminance environment. Deploying thermal or infrared (IR) sensors is straightforward to perceive the objects beyond the visible spectrum. Through IR and visual red-green-blue (RGB) image fusion, the images of thermal objects are enhanced for improved detection. Unfortunately, IR/RGB image fusion is limited to dark night, where the RGB camera is not capable of capturing the visual scene. Generative adversarial networks (GANs) have been implemented to convert IR images into RGB images for enriching semantic information. Nonetheless, contemporary methods are limited due to spatial distortion in translated results. Therefore, in this chapter, a wavelet-guided generative adversarial network (WGGAN) is proposed to address the problem. The proposed WGGAN model adopts the discrete wavelet transform to a variational autoencoder for preserving the structural information during image translation. The qualitative and quantitative analyses show considerable improvements with respect to the state-of-the-art GANs in image translation.

Speech emotion recognition (SER) is a task that cannot be accomplished solely depending on linguistic models due to the presence of figures of speech. For a more accurate prediction of emotions, researchers adopted acoustic modelling. The complexity of SER can be attributed to a variety of acoustic features, the similarities among certain emotions, etc. In this paper, we proposed a framework named Cross Languages One-Versus-All Speech Emotion Classifier (CLOVASEC) that identifies speeches’ emotions for both Chinese and English. Acoustic features were preprocessed by Synthetic Minority Oversampling Technique (SMOTE) to diminish the impact of an imbalanced dataset then by Principal component analysis (PCA) to reduce the dimension. The features were fed into a classifier that was made up of eight sub-classifiers and each sub-classifier was tasked to differentiate one class from the other seven classes. The framework outperformed regular classifiers significantly on The Chinese Natural Audio-Visual Emotion Database (CHEAVD) and an English dataset from Deng.

The intelligent fault diagnosis powered deep learning (DL) is widely applied in various real industrial fields. However, the classical intelligent fault diagnosis methods cannot fully juggle the manifold structure information with multiple-order similarity from the massive industrial data. At the same time, the scarcity of the labeled information can also result in inferior generalization performance. To this end, a new multiple-order graphical deep extreme learning machine (MGDELM) algorithm for unsupervised fault diagnosis (UFD) of rolling bearing is designed in this study. By jointly optimizing the multiple-order objective function, the proposed MGDELM algorithm can synchronously extract local and global structural information from the raw industrial data. Empirically, rolling bearing failure data validates the effectiveness of the designed algorithm and fault diagnosis method.

Closed-circuit television (CCTV) is being widely adopted in water pipeline inspection. The inspector needs to spend a long time to watch the recorded video during the office-based survey and can get fatigue easily. An automated process can release the inspector’s work load and ensure the consistent quality of the survey. However, a fully automated survey of varied structural discontinuities still remains as a challenge. This study aims to first identify the anomaly frames of the CCTV video, which contain the major anomalies captured from the internal surface of the pipe. Thus, the inspector can focus more on these anomaly frames. In this paper, an anomaly frame detection framework based on steerable pyramid autoencoder (SPAE) is proposed. The SPAE can generate discriminative representations to be used in the prediction. Both the parameter optimization and comparative studies for the proposed SPAE were carried out in this research. The experimental results demonstrate that this novel SPAE algorithm can achieve 0.984 accuracy and 0.984 F1-score, which outperforms other state-of-the-art methods selected for comparison. Thus, the proposed framework can significantly improve the accuracy and efficiency for anomaly frame detection, which will highly facilitate the pipeline condition assessment through the CCTV inspection.

Deep learning recently attracts considerable attention thanks to its powerful computational capacities in image processing and natural language processing. More and more real estate brokers provide online “Deep” expert systems to help clients with their inquiry of targeted properties before deciding on the transaction of properties. The real estate appraisal is one of the most significant concerns for the clients. In the appraisal process, the estimation of house price depends not only on its attributes but also their neighbors. The influence from neighbors is known as peer-dependence, which is not directly measurable. Thus, real estate appraisal can be improved if the valuation includes the peer-dependence measurement. In this paper, we propose a peer-dependence valuation model (PDVM), which is capable of converting the peer-dependence-based valuation problem into a sequence prediction problem. In the proposed model, we first develop a method, K-nearest similar house sampling (KNSHS), to generate sequences from the to-be-value house and nearby houses. Secondly, the bidirectional long short-term memory (B-LSTM) layers extract the features of sequences. Finally, the fully connected (FC) layer estimates the house price based on the features. The experimental results indicate that our model outperforms the other state-of-the-art machine learning models being used for real estate appraisal.

In-service bridge structural performance analysis and prediction are usually complicated and challenging because of many unknown and uncertain factors. Contrary to the traditional structural appearance inspections and load tests, structural health monitoring (SHM) can provide a perspective for online analysis, prediction, and early warning. So far, SHM has been widely used in many bridge structures, and a lot of bridge SHM data have also been collected. However, the existing studies usually focus on some independent and unsystematic analysis methods, which are hard to use widely in engineering applications to reveal the overall structural performance. This study focuses on the structural performance analysis and prediction of the highway in-service bridge. The dynamic problems in bridge SHM are pointed out firstly, followed by a detailed analysis about the characteristics of bridge SHM data. With the consideration of different characteristics, three targeted analysis methods are proposed. An urban concrete-filled steel tube (CFST) truss girder bridge (opened to traffic in 1995) is also presented, which once experienced some prominent vibration problems. The bridge SHM system is designed and stalled after several appearance inspections, load tests, and some reinforcement measures. The data mining methods proposed (distribution function, association analysis, and time-series analysis) are employed for the analysis and prediction of structural response and deterioration extent. This study can provide some references for maintenance and management and can also build a foundation for further online analysis and early warning.

Shipping is one of the major transportation approaches around the world. With the growing demands for global shipping service, vessel destination prediction has shown its significant role in improving the efficiency of decision making in industry and ensuring a safe and efficient maritime traffic environment. Currently, most vessel destination prediction methods focus on regional destination prediction, which has restrictions on destinations and regions. Thus, this paper proposes a general AIS (Automatic Identification System) data-driven model for vessel destination prediction. In this random forest-based model, the similarity between the vessel’s traveling and historical trajectories are measured and utilized to predict the destination. The destination of the historical trajectory, which shares the highest similarity with the traveling trajectory, is predicted as the vessel’s destination. The method is different from previous work which used maritime records as input to predict the destination. In our method, a historical trajectory database was generated from more than 141 million AIS records, which covers 534,824 traveling patterns between ports and more than 5.9 million historical trajectories. Comparative studies were carried out to validate the performance of the proposed model, where eleven state-of-the-art trajectories similarity measurement methods combined with two different decision strategies were implemented and compared. The experimental results demonstrate that the proposed model combined with the port frequency-based decision strategy achieves the best prediction accuracy on 35,937 testing trajectories.

Concealed metallic object detection is one of the critical tasks for any security system. It has been proved that different objects have their own magnetic fingerprints, which are a series of magnetic anomalies determined by shape, size, physical composition, etc. This study addresses the design of a low-cost power security system for the detection of metallic objects according to their response to the magnetic field. The system consists of three anisotropic magnetoresistance (AMR) sensor arrays, detection circuits, and a microcontroller. A magnetic gradient full-tensor configuration, utilizing four AMR sensors arranged on a planar cross structure, was employed to construct a two-dimensional image from the obtained data, which can further suppress the background noise and reduce the orientation and orthogonality errors. The performance of the system is validated by data validation and multiple object feature segmentation. Numerous magnetic fingerprinting results demonstrate that the system can configure metallic objects more than 50cm clearly and identify multiple objects separated by less than 20 cm, which indicates the feasibility of using this magnetic gradient tensor fingerprint method for metallic object detection.

The free induction decay (FID) transversal signal is always employed by a proton precession magnetometer (PPM) to evaluate the time-domain geomagnetic field. Nevertheless, the signal-to-noise ratio (SNR) is an important factor that severely affects the detection accuracy of the magnetic field due to uncontrollable interference sources, including random noise and power frequency noise. In this study, aiming to boost the SNR of the FID transversal signal, a novel filtering algorithm based on a prewhiten (PW) strategy is proposed and the PW filtering was combined with singular value decomposition (SVD) for further noise reduction. This method aims to generate adaptive PW input data before filtering, further decorrelating the noise to reduce the impact of varying noise levels in the received FID signals. The efficiency of the proposed joint filtering framework, dubbed PW-SVD, was evaluated by comparing with two state-of-the-art methods, i.e., SVD and principal component analysis and decomposition, using the same data. The results demonstrated that the proposed PW-SVD method obtained the smallest root mean square error and the highest signal-to-noise ratio improvement among all the compared methods, especially for the strong-noisy scenario, which enhances the environmental adaptability of a PPM.The free induction decay (FID) transversal signal is always employed by a proton precession magnetometer (PPM) to evaluate the time-domain geomagnetic field. Nevertheless, the signal-to-noise ratio (SNR) is an important factor that severely affects the detection accuracy of the magnetic field due to uncontrollable interference sources, including random noise and power frequency noise. In this study, aiming to boost the SNR of the FID transversal signal, a novel filtering algorithm based on a prewhiten (PW) strategy is proposed and the PW filtering was combined with singular value decomposition (SVD) for further noise reduction. This method aims to generate adaptive PW input data before filtering, further decorrelating the noise to reduce the impact of varying noise levels in the received FID signals. The efficiency of the proposed joint filtering framework, dubbed PW-SVD, was evaluated by comparing with two state-of-the-art methods, i.e., SVD and principal component analysis and decomposition, using the s...

Wind power is one of the majority sustainable power sources in the world. However, wind turbine is facing the highly possible blade frozen event on the blade. While numerous ice detection methods have been reported, the detection with the data acquired by the SCADA (supervisory control and data acquisition) system has not been investigated yet. Thus, in this study we propose a wind turbine ice assessment framework with SCADA data by using an ITL (inductive transfer learning) approach. We applied different machine learning methods to the ice detection problem, in which the fully-connected neural networks (FNN) outperformed the others for accuracy and stability. The experimental results demonstrate that the knowledge gained from one wind turbine can be transferred to others, where insufficient labeled data are available to build a prediction model.