CST: Complex Sparse Transformer for Low-SNR Speech Enhancement.

Speech enhancement tasks for audio with a low SNR are challenging. Existing speech enhancement methods are mainly designed for high SNR audio, and they usually use RNNs to model audio sequence features, which causes the model to be unable to learn long-distance dependencies, thus limiting its performance in low-SNR speech enhancement tasks. We design a complex transformer module with sparse attention to overcome this problem. Different from the traditional transformer model, this model is extended to effectively model complex domain sequences, using the sparse attention mask balance model's attention to long-distance and nearby relations, introducing the pre-layer positional embedding module to enhance the model's perception of position information, adding the channel attention module to enable the model to dynamically adjust the weight distribution between channels according to the input audio. The experimental results show that, in the low-SNR speech enhancement tests, our models have noticeable performance improvements in speech quality and intelligibility, respectively.

An Efficient and Scalable Algorithm to Mine Functional Dependencies from Distributed Big Data.

A crucial step in improving data quality is to discover semantic relationships between data. Functional dependencies are rules that describe semantic relationships between data in relational databases and have been applied to improve data quality recently. However, traditional functional discovery algorithms applied to distributed data may lead to errors and the inability to scale to large-scale data. To solve the above problems, we propose a novel distributed functional dependency discovery algorithm based on Apache Spark, which can effectively discover functional dependencies in large-scale data. The basic idea is to use data redistribution to discover functional dependencies in parallel on multiple nodes. In this algorithm, we take a sampling approach to quickly remove invalid functional dependencies and propose a greedy-based task assignment strategy to balance the load. In addition, the prefix tree is used to store intermediate computation results during the validation process to avoid repeated computation of equivalence classes. Experimental results on real and synthetic datasets show that the proposed algorithm in this paper is more efficient than existing methods while ensuring accuracy.

Research on RSS Data Optimization and DFL Localization for Non-Empty Environments.

Device-free localization (DFL) is a new technique which can estimate the target location through analyzing the shadowing effect on surrounding radio frequency (RF) links. In a relatively complex environment, the influences of random disturbance and the multipath effect are more serious. There are kinds of noises and disturbances in the received signal strength (RSS) data of RF links and the data itself can even be distorted, which will seriously affect the DFL accuracy. Most of the common filtering methods adopted in DFL field are not targeted and the filtering effects are unstable. This paper researches the characteristics of RSS data with random disturbances and proposes two-dimensional double correlation (TDDC) distributed wavelet filtering. It can filter out the random disturbances and noise while preserving the RSS fluctuations which are helpful for the DFL, thus improving the quality of RSS data and localization accuracy. Furthermore, RSS variation rules for the links are different in complex environments and hence, it is difficult for the collected training samples to cover all possible patterns. Therefore, a single machine learning model with poor generalization ability finds it difficult to achieve ideal localization results. In this paper, the Adaboost.M2 ensemble learning model based on the Gini decision tree (GDTE) is proposed to improve the generalization ability for unknown patterns. Extensive experiments performed in two different drawing rooms demonstrate that the TDDC distributed wavelet filtering and the GDTE localization model have obvious advantages compared with other methods. The localization accuracy rates of 87% and 95% can be achieved respectively in the two environments.

Synthesis of Ultrahigh Molecular Weight Poly(methyl Methacrylate) via the Polymerization of MMA Initiated by the Combination of Palladium Carboxylates with Thiols.

A novel synthesis of ultrahigh molecular weight poly(methyl methacrylate) (PMMA) using organosulfur compounds combined with a catalytical amount of transition metal carboxylates as an initiator has been developed. The combination of 1-octanethiol with palladium trifluoroacetate (Pd(CF3COO)2) was found to be a very efficient initiator for the polymerization of methyl methacrylate (MMA). An ultrahigh molecular weight PMMA with a number-average molecular weight of 1.68 × 106 Da and a weight-average molecular weight of 5.38 × 106 Da has been synthesized at the optimal formulation of [MMA]:[Pd(CF3COO)2]:[1-octanethiol] = 94,300:8:23 at 70 °C. A kinetic study showed that the reaction orders with respect to Pd(CF3COO)2, 1-octanethiol, and MMA are 0.64, 1.26, and 1.46, respectively. A variety of techniques such as proton nuclear magnetic resonance spectroscopy (1H NMR), electrospray ionization mass spectroscopy (ESI-MS), size exclusion chromatography (SEC), X-ray photoelectron spectroscopy (XPS), transmission electron microscopy (TEM), and electron paramagnetic resonance spectroscopy (EPR) were employed to characterize the produced PMMA and palladium nanoparticles (Pd NPs). The results revealed that Pd(CF3COO)2 was firstly reduced by the excess of 1-octanethiol to form Pd NPs at the early stage of the polymerization, followed by the adsorption of 1-octanethiol on the surface of nanoparticles and subsequent generation of corresponding thiyl radicals to initiate the polymerization of MMA.