Search

Electrocardiography (ECG) arrhythmia heartbeat classification is essential for automatic cardiovascular diagnosis system. However, the enormous differences of ECG signals among individuals and high price of labeled data have brought huge challenges for current classification algorithms based on deep neural networks and prevented these models from achieving satisfactory performance on new data. In order to build a classification system with better adaptability, we propose a novel Domain-Adaptative ECG Arrhythmia Classification (DAEAC) model based on convolutional network and unsupervised domain adaptation (UDA). Based on observation of clustering characteristics of data, we present two original objective functions to enhance the inter-patient performance. A Cluster-Aligning loss is presented to align the distributions of training data and test data. Simultaneously, a Cluster-Maintaining loss is proposed to reinforce the discriminability and structural information of features. The proposed method requires no expert annotations but a short period of unsupervised data in new records to make deep models more adaptive. Extensive experimental results on three public databases demonstrate that our method achieves competitive performance with other state-of-the-arts on the detection of ventricular ectopic beats (V), supraventricular ectopic beats (S) and fusion beats (F). The cross-dataset experimental results also verify the great generalization capability.

This work focuses on combining nonparametric topic models with Auto-Encoding Variational Bayes (AEVB). Specifically, we first propose iTM-VAE, where the topics are treated as trainable parameters and the document-specific topic proportions are obtained by a stick-breaking construction. The inference of iTM-VAE is modeled by neural networks such that it can be computed in a simple feed-forward manner. We also describe how to introduce a hyper-prior into iTM-VAE so as to model the uncertainty of the prior parameter. Actually, the hyper-prior technique is quite general and we show that it can be applied to other AEVB based models to alleviate the collapse-to-prior problem elegantly. Moreover, we also propose HiTM-VAE, where the document-specific topic distributions are generated in a hierarchical manner. HiTM-VAE is even more flexible and can generate topic representations with better variability and sparsity. Experimental results on 20News and Reuters RCV1-V2 datasets show that the proposed models outperform the state-of-the-art baselines significantly. The advantages of the hyper-prior technique and the hierarchical model construction are also confirmed by experiments.

Based on the measurement principle of the ultrasonic method, the prediction and evaluation of the contact stress at the rubber contact interface are experimentally and numerically studied in this paper. First, the relationship between the ultrasonic echo signal and the interface contact stress is calibrated based on the ultrasonic method. Second, both the propagation and the interface reflection of ultrasonic at the rubber interface under different displacements are simulated using the finite element method, and the effects of the contact stress and the interfacial roughness on the ultrasonic propagation are analysed. Finally, the empirical fitting formula of ultrasonic reflection coefficient at the rubber interface between rubber and different materials is established by considering the damping, the thickness, the contact stress and the interfacial roughness through dimensional analysis method and finite element simulation, which is well validated by the experiment.

The key challenges faced in the automatic diagnosis of arrhythmia by electrocardiogram (ECG) is enormous differences among individual patients and high cost of labeling clinical ECG records. In order to establish a system with an automatic feature learning scheme and an effective optimization mechanism, we propose a global and updatable classification scheme named Global Recurrent Neural Network (GRNN). Recurrent Neural Network (RNN) is adopted to explore the underlying features of ECG beats, based on morphological and temporal information. In order to improve system performance when new samples are obtained, active learning is applied to select the most informative beats and incorporate them into training set. The system is then updated as the training set grows. Our GRNN has three main innovations. Firstly, relying on the large capacity and fitting ability of GRNN, we can classify samples of multiple different patients with a single model. Secondly, the GRNN improves generalization performance when training samples and test samples are from distinct databases. This can be explained that the optimization mechanism finds the most informative samples to improve performance as training data. Finally, RNN automatically learns the underlying differences among the samples from different classes. Experimental results prove that the GRNN system achieves the state-of-the-art performance with a single model. In across-database experiments where the training data and test data are from different databases respectively, the GRNN achieves significant improvement compared with other algorithms. This study illustrates the feasibility of a global and updatable ECG beat classification system in practical applications.

A single clock cycle approximate adder (SCCA) with hybrid prediction and error compensation methods has been proposed, where the computing accuracy and energy efficiency are carefully balanced. Meanwhile, the addition could be performed in one clock cycle, which decreases the critical path delay and simplify the configuration when applied to a system. Additionally, the proposed adder is able to prevent larger error output in accumulation computing mode. The new approximate adder is applied into the algorithms of Gaussian filter and Convolutional Neural Network (CNN). The simulation results show that the proposed approximate adder achieves 2.8X speedup and 59.9% Power-Delay-Product (PDP) reduction compared with conventional ripple-carry adder (RCA), respectively compared with two outstanding representative approximate adders in [24,28], 19.3% and 2.4% PDP reduction are also achieved with extremely low absolute errors. Our adder shows negligible output quality reduction in image Gaussian filter, where PSNR decreases from 26.74 dB to 26.72 dB; and CNN is also trained successfully using proposed approximate adder with 0.37% and 9.23% accuracy loss in MNIST and CIFAR10 datasets, respectively.

An analog memory based on the high linearity source-follower buffer topology is proposed, which is applied to the emerging Analog Convolutional Neural Network (CNN) for buffering parameters and operation results. The proposed memory consists of a source-follower type buffer, which delivers an appreciably enhanced accuracy over that of the conventional buffer, and a storage capacitor to meet the Analog CNN processing demands of accurate short-term storage and multi-reading capability. The enhanced linearity of the proposed buffer is achieved by adopting high-threshold cascode (HTC) structure and low parasitic capacitance switch (LPCS). Moreover, a low leakage bootstrap (LLB) structure is integrated to enhance the turn-off performance of switch, which reduces the leakage and improves the accuracy of buffer significantly. Simulated with 180 nm CMOS mixed-signal process, the proposed analog memory unit achieves power consumption of 3.6 μW and output error of 0.4%, with the input voltage swing of 1.1 V .

Focus stacking is a computational technique to extend the Depth of Field (DOF) through combining multiple images taken at various focus distances. However, existing focus stacking methods could not cope with false edges produced by propagation of blur kernels, and are affected by colored texture in the stack. In this work, we propose a novel all-in-focus method based on directional-max-gradient flow (DMGF) and labeled iterative depth propagation. Firstly, we present a novel directional-max-gradient flow to describe gradient propagation along different directions in the stack to remove false edges and preserve accurate depth values of both strong and weak edges(also called source points). Then the source points are further labeled as in-plane edges and off-plane edges by unsupervised classification technique. Finally in our proposed labeled iterative Laplacian optimization, these edges are utilized to remove artifacts produced by colored texture in the stack and refine the all-in-focus image. Extensive experiments on both synthesized data and real data show that our method has achieved superior performance to state-of-the-art methods.

In this paper, a new leakage characterization method for the sealing interface is proposed to detect the leakage location and measure the leakage rate based on the temperature field and the temperature gradient field. First, the formulas of the leakage rate are theoretically established based on the two-dimensional and the three-dimensional heat conduction model, respectively. Second, the finite element models of the general leakage area and the concentrated leakage area are built, in which the leakage rate calculated by means of heat conduction model fits the simulated leakage rate well. Finally, the interface leakage experiment is conducted using a sealed box with a rubber-aluminum alloy sealing interface, which validates the effectiveness of this new characterization principle for determining the leakage rate and the leakage location. The results indicate that the feasibility of the leakage characterization method for the sealing interface based on the temperature field and the temperature gradient field obtained from the infrared thermography technique.

Urban water supply network is a hidden underground asset; therefore, it is difficult to make a direct diagnosis towards the faults of pipe network by means of the detecting technology. Based on the historical data of the rupture of water supply pipe network, a spatial cluster analysis will be carried out towards the historical data of spatial fault in water supply pipeline, using density- based cluster analysis and the DBSCAN algorithm. K-neighborhood graph is well-used in optimizing the input parameter radius of neighborhood (Eps) and density (MinPts) and classifying the historical data of the rupture of pipeline. According to the pipe network failure point coordinates (X,Y), adopting density-based cluster analysis, a danger level classification will be researched without knowing the reasons of the rupture of pipe network, which lays the foundation for finding out pipeline rupture in local areas in the future. Meanwhile, a noise analysis will be processed in terms of the historical data of pipeline rupture in water supply network, avoiding the ineffective data caused by recording errors and regular analysis. This research focuses on the reasons that result in the rupture of water supply network, which plays a significant role in the identification of high-density and high-danger zones in water supply networks.

In this paper, combining the robustness of vector norm with that of the approximation components after the discrete wavelet transform (DWT), a blind and adaptive audio watermarking algorithm is proposed. In order to improve the robustness and imperceptibility, a binary image encrypted by Arnold transform as watermark is embedded in the vector norm of the segmented approximation components, the count of which depends on the size of the watermark image, after DWT of the original audio signal through quantization index modulation (QIM) with an adaptive quantization step selection scheme. Moreover, a detailed method has been designed to search the suitable quantization step parameters. Experimental results indicate that even though the capacity of the proposed algorithm is high, up to 102.4bps, this algorithm is still able to maintain good quality of the audio signal and tolerate a wide class of common attacks such as additive white Gaussian noise (AWGN), Gaussian Low-pass filter, Kaiser Low-pass filter, resampling, requantizing, cutting, MP3 compression and echo.

Critically sampled perfect reconstruction lapped transforms (LT), apart from signal representation, are difficult to use for signal analysis and processing. We tackle this problem by establishing a novel connection between the LT and the DFT based on matrix analysis. Specifically, for the modified discrete cosine transform (MDCT) with an arbitrary symmetric window, one of the most widely used LT, a sparse matrix representation for the conversion to and from the DFT is developed, leading to an efficient implementation in the form of low-order frequency domain FIR filtering. We give two example applications, MDCT domain group delay estimation and aliasing reduction for the temporal noise shaping (TNS).