RECOGNITION OF RADIO EXCHANGE VOICE MESSAGES IN AVIATION BASED ON CORRELATION ANALYSIS

The paper considers the problem of speech messages recognition in phraseological radio exchange for tasks of civil aviation. The introduction substantiates the relevance of this problem. The following are research methods based on correlation analysis. Finally, a description of the experiment and the results of the recognition algorithms based on correlation analysis are given. Various variants were recorded for five speech messages and spectral representations of such signals were constructed. Spectral transform can be obtained either using specialized software or based on the Fourier transform of the signal in the time domain. To obtain a more universal reference signal and eliminate the influence of interference, the spectral components of the same speech message recorded several times were averaged. In fact, three spectra of the same speech message were used for averaging. This spectrum averaging over three training components provided a reference sample of phrases or patterns for each phrase, and reduced the influence of additive white Gaussian noise in the reference. Later, on the basis of correlation analysis, the connections between test phrases and all patterns were calculated. On the basis of these connections, a correlation matrix of reference phrases is built. Research has shown that phrases spoken by one person were highly correlated. The analysis showed that the choice of the class (the content of the speech message) when solving the recognition problem corresponding to the value of the correlation coefficient closest to one provides over 90% of correct recognitions on a test sample containing a total of 100 phrases, 20 for each phrase. It should be noted that, when recording test messages, an additive white Gaussian noise was additionally present as a background, reproduced by another audio device. In the case of information analysis without artificially generated noise, the probability of correct recognition for a test sample of 100 phrases, 20 for each phrase, is 100% when using correlation analysis.

Sparse Epistatic Regularization of Deep Neural Networks for Inferring Fitness Functions

Despite recent advances in high-throughput combinatorial mutagenesis assays, the number of labeled sequences available to predict molecular functions has remained small for the vastness of the sequence space combined with the ruggedness of many fitness functions. Expressive models in machine learning (ML), such as deep neural networks (DNNs), can model the nonlinearities in rugged fitness functions, which manifest as high-order epistatic interactions among the mutational sites. However, in the absence of an inductive bias, DNNs overfit to the small number of labeled sequences available for training. Herein, we exploit the recent biological evidence that epistatic interactions in many fitness functions are sparse; this knowledge can be used as an inductive bias to regularize DNNs. We have developed a method for sparse epistatic regularization of DNNs, called the epistatic net (EN), which constrains the number of non-zero coefficients in the spectral representation of DNNs. For larger sequences, where finding the spectral transform becomes computationally intractable, we have developed a scalable extension of EN, which subsamples the combinatorial sequence space uniformly inducing a sparse-graph-code structure, and regularizes DNNs using the resulting greedy optimization method. Results on several biological landscapes, from bacterial to protein fitness functions, showed that EN consistently improves the prediction accuracy of DNNs and enables them to outperform competing models which assume other forms of inductive biases. EN estimates all the higher-order epistatic interactions of DNNs trained on massive sequence spaces—a computational problem that takes years to solve without leveraging the epistatic sparsity in the fitness functions.Significance StatementPredicting the properties of small molecules (such as proteins) from their sequence is an important problem in computational biology. The main challenge is in developing a model that can capture the non-linearities in the function mapping the sequence to the property of interest (e.g., fluorescence) using the limited number of available labeled sequences from biological assays. In this paper, we identify a biologically-plausible sparsity prior and develop a method to infuse this prior into the structure of deep neural networks (DNNs) by regularizing their spectral representation. We demonstrate that our method significantly improves the prediction accuracy of DNNs and enables an interpretable explanation of DNNs—a task that is computationally intractable without leveraging the hidden structure in biological functions.

Theoretical Consideration on the Application of Continuous Wavelet Transform to Time-Series Processing for Magnetotelluric Data

In the magnetotelluric (MT) method, the responses of the natural electromagnetic fields are evaluated by transforming time-series data into spectral data and calculating the apparent resistivity and phase. The continuous wavelet transform (CWT) can be an alternative to the short-time Fourier transform, and the applicability of CWT to MT data has been reported. There are, however, few cases of considering the effect of numerical errors derived from spectral transform on MT data processing. In general, it is desirable to adopt a window function narrow in the time domain for higher-frequency components and one in the frequency domain for lower-frequency components. In conducting the short-time Fourier transform, because the size of the window function is fixed unless the time-series data are decimated, there might be difference between the calculated MT responses and the true ones due to the numerical errors. Meanwhile, CWT can strike a balance between the resolution of the time and frequency domains by magnifying or reducing the wavelet, according to the value of frequency. Although the types of wavelet functions and their parameters influence the resolution of time and frequency, those calculation settings of CWT are often determined empirically. In this study, focusing on the frequency band between 0.001 Hz and 10 Hz, we demonstrated the superiority of utilizing CWT in MT data processing and determined its proper calculation settings in terms of restraining the numerical errors caused by the spectral transform of time-series data. The results obtained with the short-time Fourier transform accompanied with gradual decimation of the time-series data, called cascade decimation, were compared with those of CWT. The shape of the wavelet was changed by using different types of wavelet functions or their parameters, and the respective results of data processing were compared. Through these experiments, this study indicates that CWT with the complex Morlet function with its wavelet parameter k set to 6 ≤ k < 10 will be effective in restraining the numerical errors caused by the spectral transform.

Fractional Soliton Dynamics and Spectral Transform of Time-Fractional Nonlinear Systems: A Concrete Example

In this paper, the spectral transform with the reputation of nonlinear Fourier transform is extended for the first time to a local time-fractional Korteweg-de vries (tfKdV) equation. More specifically, a linear spectral problem associated with the KdV equation of integer order is first equipped with local time-fractional derivative. Based on the spectral problem with the equipped local time-fractional derivative, the local tfKdV equation with Lax integrability is then derived and solved by extending the spectral transform. As a result, a formula of exact solution with Mittag-Leffler functions is obtained. Finally, in the case of reflectionless potential the obtained exact solution is reduced to fractional n-soliton solution. In order to gain more insights into the fractional n-soliton dynamics, the dynamical evolutions of the reduced fractional one-, two-, and three-soliton solutions are simulated. It is shown that the velocities of the reduced fractional one-, two-, and three-soliton solutions change with the fractional order.

Deep Spectral Kernel Learning

Recently, spectral kernels have attracted wide attention in complex dynamic environments. These advanced kernels mainly focus on breaking through the crucial limitation on locality, that is, the stationarity and the monotonicity. But actually, owing to the inefficiency of shallow models in computational elements, they are more likely unable to accurately reveal dynamic and potential variations. In this paper, we propose a novel deep spectral kernel network (DSKN) to naturally integrate non-stationary and non-monotonic spectral kernels into elegant deep architectures in an interpretable way, which can be further generalized to cover most kernels. Concretely, we firstly deal with the general form of spectral kernels by the inverse Fourier transform. Secondly, DSKN is constructed by embedding the preeminent spectral kernels into each layer to boost the efficiency in computational elements, which can effectively reveal the dynamic input-dependent characteristics and potential long-range correlations by compactly representing complex advanced concepts. Thirdly, detailed analyses of DSKN are presented. Owing to its universality, we propose a unified spectral transform technique to flexibly extend and reasonably initialize domain-related DSKN. Furthermore, the representer theorem of DSKN is given. Systematical experiments demonstrate the superiority of DSKN compared to state-of-the-art relevant algorithms on varieties of standard real-world tasks.

Multispectral Transforms Using Convolution Neural Networks for Remote Sensing Multispectral Image Compression

A multispectral image is a three-order tensor since it is a three-dimensional matrix, i.e.one spectral dimension and two spatial position dimensions. Multispectral image compression canbe achieved by means of the advantages of tensor decomposition (TD), such as NonnegativeTucker Decomposition (NTD). Unfortunately, the TD suffers from high calculation complexity andcannot be used in the on-board low-complexity case (e.g., multispectral cameras) that the hardwareresources and power are limited. Here, we propose a low-complexity compression approach formultispectral images based on convolution neural networks (CNNs) with NTD. We construct anew spectral transform using CNNs, where the CNNs are able to transform the three-dimensionspectral tensor from large-scale to a small-scale version. The NTD resources only allocate thesmall-scale three-dimension tensor to improve calculation efficiency. We obtain the optimizedsmall-scale spectral tensor by the minimization of original and reconstructed three-dimensionspectral tensor in self-learning CNNs. Then, the NTD is applied to the optimized three-dimensionspectral tensor in the DCT domain to obtain the high compression performance. We experimentallyconfirmed the proposed method on multispectral images. Compared to the case that the newspectral tensor transform with CNNs is not applied to the original three-dimension spectral tensorat the same compression bit-rates, the reconstructed image quality could be improved. Comparedwith the full NTD-based method, the computation efficiency was obviously improved with only asmall sacrifices of PSNR without affecting the quality of images.