Search

Highlights • A CNN-based end-to-end classifier for detection of freezing of gait is proposed. • The discriminative features are learned directly from multiple 1D time series. • Two different feature fusion schemes are designed and compared. • More than 99% accuracy was obtained in the patient-dependent setting. • An average of 80.70% accuracy was obtained in the patient -independent setting. Abstract Background and objective Freezing of gait (FOG) is a symptom that manifests as an episodic inability to move. It happens typically in patients with advanced Parkinson’s disease (PD), and it is a common cause of falls in PD patients. The management of FOG is extremely difficult due to its sudden and transient property. Methods In this study, we implemented a novel FOG detection system that was based on deep convolutional neural network (CNN). By taking data segments from 1-dimensional (1D) acceleration signals as its inputs, the proposed CNN-based approach can realize automatic feature learning and discrimination of FOG events from normal walking in a streamline manner. By this way, it can remove the need for extracting hand-crafted features and the time-consuming feature selection. The proposed method was tested on a dataset comprised of more than eight hours of recorded lab data from 10 PD patients that experience FOG in their daily life. Results The final system achieved more than 99% classification accuracy in a patient-dependent setting, and an average of 80.70% accuracy in the patient -independent setting. The time for classification of a 4 s data segment is only 3.6 ms without the acceleration of graphics processing unit (GPU). Conclusions These results indicate that the proposed CNN-based system can provide satisfactory effectivity and efficiency in detecting FOG gaits if used suitably and can be beneficial to realize an accurate monitoring and gait assistance during daily living and rehabilitation therapy. [ABSTRACT FROM AUTHOR]

A new denoising algorithm based on low-rank matrix approximation (LRMA) with regularization of weighted nuclear norm minimization (WNNM) is proposed to remove Rician noise of magnetic resonance (MR) images. This technique simply groups similar non-local cubic blocks from noisy 3D MR data into a patch matrix with each block lexicographically vectorizing to be as a column, calculates the singular value decomposition (SVD) on this matrix, then the closed-form solution of LRMA is achieved by hard-thresholding different singular values with a different threshold. The denoised blocks are obtained from this estimate of the low-rank matrix, and the final estimate of the whole noise-free MR data is built up by aggregating all the denoised exemplar blocks that are overlapped each other. To further improve the denoising performance of the WNNM algorithm, we first realize the above denoising procedure in a two-iteration regularization framework, and then a simple non local means (NLM) filter based on single-pixel patch is utilized to reduce the intensity jumping at the homogeneous area. The proposed denoising algorithm was compared with related state-of-the-art methods and produced very competitive results over synthetic and real 3D MR data. [ABSTRACT FROM AUTHOR]

The reduction of multiplicative speckle noise which always complicates the human and automatic interpretation of objects is very significant for the practical applications of synthetic aperture radar (SAR) image. In this paper, a new maximum a posteriori (MAP) despeckling method based on directionlet transform is proposed. To convert the multiplicative noise into an additive one, the logarithmic transform is first applied to the SAR images. Then, the directionlet coefficients of the noise-free (or underlying backscatter) image and of the speckle noise are modeled as Laplace mixture distribution with zero-mean and Gaussian distribution, respectively. Within Bayesian framework, a MAP estimator is constructed using these assumed prior distributions. After obtaining the parameter estimates using expectation–maximization algorithm, the noise-free coefficients are estimated by a non-linear shrinkage function based on the average version of Bayesian estimator. To improve the denoising performance, we combine the intra-scale dependency in terms of mutual information with the MAP estimator to refine the estimated results. Finally, we compare the proposed algorithm with several other speckle filters applied on synthetic and actual SAR images. Experimental results show that the proposed method outperforms other filters in terms of signal-to-noise ratio, edge preservation and equivalent number of looks measures in most cases. [ABSTRACT FROM AUTHOR]

Abstract: A new denoising algorithm based on directionlet transform using a Cauchy probability density function (PDF) is proposed to remove speckle noise. First, an anisotropic directionlet transform is taken on the logarithmically transformed SAR images. The directionlet transform coefficients of reflectance image are modeled as a zero-location Cauchy PDF, while the distribution of speckle noise is modeled as an additive Gaussian distribution with zero-mean. Then a maximum a posteriori (MAP) estimator is designed using the assumed priori models. And a regression-based method is proposed to estimate the parameters from the noisy observations. Finally, the performance of the proposed algorithm is compared with those of existing despeckling methods applied on both synthetic speckled images and actual SAR images. Experimental results show that the proposed scheme efficiently removes speckle noise from SAR images. [Copyright &y& Elsevier]

The data-driven fault diagnostic methods have made significant advances and breakthroughs in the past decades. However, due to the huge time and labor costs, single user is hard to collect sufficient monitoring data for model training and multiple users are preferred for collaborative modeling with distributed data. Considering the concerns of data privacy or conflicts of interest, the local data of different users cannot be directly shared with others for centralized training. Moreover, they are usually sampled from different working environments, leading to data island and distribution shift issues. To solve the two issues, this paper proposes a novel cloud-edge collaborative transfer fault diagnosis approach for rotating machinery. The federated learning scheme is introduced for collaborative modeling of multiple users while ensuring the data privacy of participants. To bridge the distribution gap across clients, the source training module is constructed to conduct local model training on insufficient labeled data from different users, and a federated fine-tuning module is then developed to remove distribution discrepancies between different source users, thus building accurate source models. Finally, a target self-adaptation module is designed to reduce prediction differences of source client models on target samples and achieve effective adaptation of the target client to source clients. Comprehensive comparative analysis shows that the proposed approach achieves a performance gain of 0.45% to the traditional centralized transfer learning method, demonstrating its potential in addressing the data island and distribution shift issues in practical applications. [ABSTRACT FROM AUTHOR]

1540 nm lasers are widely used for excellent anti-interference capability and atmospheric transmission, whose corresponding suppressing materials are still lack of investigation Er3+ ions can effectively absorb 1540 nm lasers due to energy level transitions near this wavelength. And the absorption effect could be further enhanced by doping with heterogeneous metal ions. Herein, Co2+ has been doped into ErBO 3 with different concentrations, which could induce lattice contraction of Er3+ ion cells. The ErBO 3 samples prepared by solvothermal and calcination processes are characterized by "flower-like" porous microspheres at 800 °C. The influence of different calcination temperatures and molar ratios of raw materials on the reflection performance of 1540 nm laser light was examined. The laser reflectance of the sample was reduced to 0.707 % by doping with 20 mol% Co2+. This Co2+ doped ErBO 3 material exhibits significant potential in 1540 nm laser protection applications. [ABSTRACT FROM AUTHOR]

• A novel DGNet_MSAC is proposed for cross-domain generalization fault diagnosis. • Multiple domain-specific auxiliary classifiers are designed in the DGNet_MSAC. • Domain-specific features are removed by the convolutional auto-encoder module. • Domain-invariant features are learned by the inter-domain alignment strategy. The domain adaptation-based intelligent diagnosis approaches have achieved promising performance on diagnosis tasks under different working conditions. However, these methods rely on a premise that the target data are available in the model training phase. In real industries, collecting interest data from target machines in advance may be infeasible, which greatly restricts the practicality of intelligent diagnosis approaches in reality. To solve this issue, this study proposes a novel domain generalization network for machinery fault diagnosis where interest data are completely unavailable during model training. In the proposed network, multiple domain-specific auxiliary classifiers are firstly designed to effectively learn domain-specific features from each source domain, and then, a convolutional auto-encoder module is further constructed to map raw signals into a new feature space where the learned domain-specific features are removed. Meanwhile, with the features outputted by the convolutional auto-encoder, a domain-invariant classifier with inter-domain alignment strategy is designed to learn generalization diagnostic knowledge among different source domains, thereby performing diagnosis tasks under unseen conditions. Experiments on three practical rotary machinery datasets validate the effectiveness of the proposed network, showing that the proposed network is promising for fault diagnosis tasks in practical scenarios. [ABSTRACT FROM AUTHOR]

Abstract: Synthetic aperture radar (SAR) images are inherently affected by multiplicative speckle noise, which is due to the coherent nature of scattering phenomena. This paper presents a depeckling method for SAR images based on stationary wavelet transform combining with directional filter banks. In particular, the traditional thresholding method is substituted by a novel and efficient method, Bayesian maximum a posteriori (MAP) estimation, to achieve more satisfying results. The performances of stationary wavelet transform, contourlet transform and stationary wavelet combining with directional filter banks for despeckling SAR images are compared through an experiment. Experiment results show this method not only suppresses the speckle noise effectively, but also preserves as many target characteristics of original images as possible. Besides, the recovered SAR image’s visual effect is satisfying. [Copyright &y& Elsevier]