The paper examines the usage of Convolutional Bidirectional Recurrent Neural Network (CBRNN) for a problem of quality measurement in a music content. The key contribution in this approach, compared to the existing research, is that the examined model is evaluated in terms of detecting acoustic anomalies without the requirement to provide a reference (clean) signal. Since real music content may include some modes of instrumental sounds, speech and singing voice or different audio effects, it is more complex to analyze than clean speech or artificial signals, especially without a comparison to the known reference content. The presented results might be treated as a proof of concept, since some specific types of artefacts are covered in this paper (examples of quantization defect, missing sound, distortion of gain characteristics, extra noise sound). However, the described model can be easily expanded to detect other impairments or used as a pre-trained model for other transfer learning processes. To examine the model efficiency several experiments have been performed and reported in the paper. The raw audio samples were transformed into Mel-scaled spectrograms and transferred as input to the model, first independently, then along with additional features (Zero Crossing Rate, Spectral Contrast). According to the obtained results, there is a significant increase in overall accuracy (by 10.1%), if Spectral Contrast information is provided together with Mel-scaled spectrograms. The paper examines also the influence of recursive layers on effectiveness of the artefact classification task.
This paper presents a universal approximation of the unit circle by a polygon that can be used in signal processing algorithms. Optimal choice of the values of three parameters of this approximation allows one to obtain a high accuracy of approximation. The approximation described in the paper has a universal character and can be used in many signal processing algorithms, such as DFT, that use the mathematical form of the unit circle. One of the applications of the described approximation is the DFT linear interpolation method (LIDFT). Applying the results of the presented paper to improve the LIDFT method allows one to significantly decrease the errors in estimating the amplitudes and frequencies of multifrequency signal components. The paper presents the derived formulas, an analysis of the approximation accuracy and the region of best values for the approximation parameters.
This paper presents the general solution of the least-squares approximation of the frequency characteristic of the data window by linear functions combined with zero padding technique. The approximation characteristic can be discontinuous or continuous, what depends on the value of one approximation parameter. The approximation solution has an analytical form and therefore the results have universal character. The paper presents derived formulas, analysis of approximation accuracy, the exemplary characteristics and conclusions, which confirm high accuracy of the approximation. The presented solution is applicable to estimating methods, like the LIDFT method, visualizations, etc.
This paper derives analytical formulas for the systematic errors of the linear interpolated DFT (LIDFT) method when used to estimating multifrequency signal parameters and verifies this analysis using Monte-Carlo simulations. The analysis is performed on the version of the LIDFT method based on optimal approximation of the unit circle by a polygon using a pair of windows. The analytical formulas derived here take the systematic errors in the estimation of amplitude and frequency of component oscillations in the multifrequency signal as the sum of basic errors and the errors caused by each of the component oscillations. Additional formulas are also included to analyze particular quantities such as a signal consisting of two complex oscillations, and the analyses are verified using Monte-Carlo simulations.
Quality of energy produced in renewable energy systems has to be at the high level specified by respective standards and directives. One of the most important factors affecting quality is the estimation accuracy of grid signal parameters. This paper presents a method of a very fast and accurate amplitude and phase grid signal estimation using the Fast Fourier Transform procedure and maximum decay side-lobes windows. The most important features of the method are elimination of the impact associated with the conjugate’s component on the results and its straightforward implementation. Moreover, the measurement time is very short ‒ even far less than one period of the grid signal. The influence of harmonics on the results is reduced by using a bandpass pre-filter. Even using a 40 dB FIR pre-filter for the grid signal with THD ≈ 38%, SNR ≈ 53 dB and a 20‒30% slow decay exponential drift the maximum estimation errors in a real-time DSP system for 512 samples are approximately 1% for the amplitude and approximately 8.5・10‒2 rad for the phase, respectively. The errors are smaller by several orders of magnitude with using more accurate pre-filters.
Fast and accurate grid signal frequency estimation is a very important issue in the control of renewable energy systems. Important factors that influence the estimation accuracy include the A/D converter parameters in the inverter control system. This paper presents the influence of the number of A/D converter bits b, the phase shift of the grid signal relative to the time window, the width of the time window relative to the grid signal period (expressed as a cycle in range (CiR) parameter) and the number of N samples obtained in this window with the A/D converter on the developed estimation method results. An increase in the number b by 8 decreases the estimation error by approximately 256 times. The largest estimation error occurs when the signal module maximum is in the time window center (for small values of CiR) or when the signal value is zero in the time window center (for large values of CiR). In practical applications, the dominant component of the frequency estimation error is the error caused by the quantization noise, and its range is from approximately 8×10-10 to 6×10-4.