Presence Of Noisy Signals Research Articles

A Study of High-Frequency Noise for Microplastics Classification Using Raman Spectroscopy and Machine Learning.

Given the growing urge for plastic management and regulation in the world, recent studies have investigated the problem of plastic material identification for correct classification and disposal. Recent works have shown the potential of machine learning techniques for successful microplastics classification using Raman signals. Classification techniques from the machine learning area allow the identification of the type of microplastic from optical signals based on Raman spectroscopy. In this paper, we investigate the impact of high-frequency noise on the performance of related classification tasks. It is well-known that classification based on Raman is highly dependent on peak visibility, but it is also known that signal smoothing is a common step in the pre-processing of the measured signals. This raises a potential trade-off between high-frequency noise and peak preservation that depends on user-defined parameters. The results obtained in this work suggest that a linear discriminant analysis model cannot generalize properly in the presence of noisy signals, whereas an error-correcting output codes model is better suited to account for inherent noise. Moreover, principal components analysis (PCA) can become a must-do step for robust classification models, given its simplicity and natural smoothing capabilities. Our study on the high-frequency noise, the possible trade-off between pre-processing the high-frequency noise and the peak visibility, and the use of PCA as a noise reduction technique in addition to its dimensionality reduction functionality are the fundamental aspects of this work.

Deep learning method based on autoencoder neural network applied to faults detection and diagnosis of photovoltaic system

The paper presents an application of deep learning for fault detection in PV systems located in Algiers –Algeria which has a nominal power of 9.54 kW. Each PV array comprises 30 PV modules each string contains 15 PV modules in series. Fault detection and diagnosis of Photovoltaic Systems (PVS) is an important task for successful solar power generation. Several faults have been occurred such as short-circuit cases and open-circuit string cases in PV generator can leads to fire and risk. The design of efficient defect detection methods for PV systems presents a challenge especially for small scale PV farms. Faults detection methods are failed under the presence of noisy signals. This paper presents a deep learning (DL) based method for detection, diagnosis and classification of the aforementioned defects. The proposed procedure consists of four fundamental steps, first a heuristic optimization approach based on Coyote Optimization Algorithm (COA) the five-unknown electrical parameters of the One Diode Model (ODM) and their insertion into a PSIM-based simulation that aims to mimic the operating PV system. Second, database construction that includes current, voltage and power at the Maximum Power Point (MPP), module temperature and solar irradiance for the PV system is supposed under healthy and faulty operations condition at optimal operational conditions. Then, this is followed by the extraction of new features of the old database using unsupervised learning characteristics of the Auto-Encoder (AE). and last, supervised learning using the new database based on Artificial Neural Network (ANN) construction for PV fault detection and classification. The proposed technique has been validated using monitored data from a real operating PV system situated in Algeria. The obtained results have shown the effectiveness of the proposed technique detection and classification of various PV fault types.

Automated classification of A-DInSAR-based ground deformation by using random forest

ABSTRACT Wide-area ground motion monitoring is nowadays achievable via advanced Differential Interferometry SAR (A-DInSAR) techniques which benefit from the availability of large sets of Copernicus Sentinel-1 images. However, it is of primary importance to implement automated solutions aimed at performing integrated analysis of large amounts of interferometric data. To effectively detect high-displacement areas and classify ground motion sources, here we explore the feasibility of a machine learning-based approach. This is achieved by applying the random forest (RF) technique to large-scale deformation maps spanning 2015–2018. Focusing on the northern part of Italy, we train the model to identify landslide, subsidence, and mining-related ground motion with which to construct a balanced training dataset. The presence of noisy signals and other sources of deformation is also tackled within the model construction. The proposed approach relies on the use of explanatory variables extracted from the A-DInSAR datasets and from freely accessible informative layers such as Digital Elevation Model (DEM), land cover maps, and geohazard inventories. In general, the model performance is very promising as we achieved an overall accuracy of 0.97, a true positive rate of 0.94 and an F1-Score of 0.93. The obtained outcomes demonstrate that such transferable and automated approach may constitute an asset for stakeholders in the framework of geohazards risk management.

Detection and Analysis of Inter-Area Oscillations Through a Dynamic-Order DMD Approach

The paper deals with a novel method for detecting and analyzing inter-area oscillations on electric power transmission networks. The method starts from the modal analysis performed by the Dynamic Mode Decomposition algorithm, which is able to exploit the synchronized acquisitions of various measurement instruments to detect the mode of a dynamic system. Compared with the classical algorithm, the proposed method presents a fundamental improvement, which ensures its reliability even without having prior information on the type of input signal. In particular, the order of the DMD, i.e. the number of modes characterizing the acquired signal, is dynamically updated according to its energy content. The method has been tested with simulated signals, considering both single-oscillation signals and two-oscillation signals, varying the amplitude, frequency and damping of the oscillatory components. In this way, the improvement with respect to the classical DMD was highlighted and the performance in terms of deviation between the estimated and nominal parameters was evaluated. Furthermore, the assessment on real life acquired signals has been performed; the results confirmed the reliability and accuracy of the measurement method, even in the presence of noisy signals and ambient data.

Fuzzy based relaying scheme for high impedance faults in DG integrated distribution system

Most of the high impedance faults (HIF) remain un-detected by the conventional relays due to non-linear nature of the fault and low magnitude of current. In this work, a combination of discrete wavelet transforms (DWT) and fuzzy inference system (FIS) has been proposed for HIF detection and classification. Modified IEEE 13 node test feeder system has been to validate the proposed scheme. The proposed method uses current signals from one end that are pre-processed using discrete wavelet transform to obtain appropriate input features. The wavelet processed features are given to the FIS for fault detection and classification. Proposed method has been validated using both Mamdani and Sugeno type FIS. Different operating and fault conditions are tested to validate the proposed method such as varying DG parameters, noisy signals, HIFs, evolving faults, fault inception angle, fault resistance, fault location, and non-fault events (e.g. motor load switching, capacitor switching, DG switching, transformer energization, non-linear load switching). The accuracy in detecting and classifying the faults is 100% of all the tested fault cases. Results shows that the overall detection time required to detect the HIFs is minimum 4.25 cycles in most of the cases and maximum 8 cycles in few cases whereas for shunt faults is within 4.25 to 6 cycles only. Advantage of the proposed method is that it can detect adverse situation faults like evolving faults, in presence of noisy signals and remains intact against any switching events. The results of the proposed method are promising and the method is robust against various operating conditions.

Emotion Recognition Using Hybrid Gaussian Mixture Model and Deep Neural Network

This paper aims at recognizing emotions for a text-independent and speaker-independent emotion recognition system based on a novel classifier, which is a hybrid of a cascaded Gaussian mixture model and deep neural network (GMM-DNN). This hybrid classifier has been assessed for emotion recognition on “Emirati speech database (Arabic United Arab Emirates Database)” with six different emotions. The sequential GMM-DNN classifier has been contrasted with support vector machines (SVMs) and multilayer perceptron (MLP) classifiers, and its performance accuracy is indexed at 83.97%, while the other two perform at 80.33% and 69.78% using SVMs and MLP, respectively. These results demonstrate that the hybrid classifier significantly gives higher emotion recognition accuracy than SVMs and MLP classifiers. Our GMM-DNN model yields the results similar to those obtained by human judges in a subjective assessment context. Also, the performance of the classifier has been tested using two distinct emotional databases and in normal and noisy talking conditions. The dominant signal mask provided by the hybrid classifier offers better system performance in the presence of noisy signals.

Time domain damage localization and quantification in seismically excited structures using a limited number of sensors

This paper presents a new algorithm to determine the occurrence, location, and severity of damage in structures subjected to earthquakes. The algorithm is based on the analysis of the time series associated with displacement or acceleration, and provided by a limited number of sensors. The algorithm is formulated in terms of an optimization problem. An objective function is defined based on the moment generating function for a segment of the time histories and an evolutionary optimization strategy, based on the competitive optimization algorithm, is employed to detect damage. The efficiency of the proposed method is numerically validated by studying the response of some structures subjected to the 1940 El-Centro earthquake and the 1994 Northridge earthquake. In order to simulate real conditions, different levels of noise are added to the response’s signals, and then the discrete wavelet transform is used to de-noise the signals. Moreover, the robustness of the method is evaluated by considering an error in the model of the structures. Overall, we find that the proposed algorithm detects and localizes damage even in presence of noisy signals and errors in the model.

Innovative model-based control approach of a proton exchange membrane fuel cell system

This work highlights the gains of an innovative model-based control approach applied to a proton exchange membrane fuel cell (PEMFC) system, included in a stand-alone hybrid generator. This approach proposes a multivariable setpoint tracking of the PEMFC output power and temperature. The freshness of this approach is based on the combination of a nonlinear model-based predictive control strategy (NMPC) and a global linearizing control (GLC) algorithm.The performance of the proposed control strategy is confirmed thanks to simulations of varying control scenarios. Results show good performance in setpoint tracking, disturbances rejection and robustness against plant/model mismatch in presence of noisy signals. Moreover, for similar setpoint point tracking accuracy, the proposed control strategy appears to be four times faster than a classical multivariable NMPC strategy. According to real-time application objectives, this control strategy appears as a promising option to be implanted in the overall control scheme of the stand-alone hybrid power generator.

On the improved use of Chebyshev expansion for mechanical system identification

Differential equations, and particularly those governing mechanical system behavior, can be transformed into algebraic equations by using the well-known properties of polynomial functions. In this paper, by using Chebyshev polynomials, improvements of classical methods are proposed for the identification and inverse formulation of mechanical systems. The drawbacks of these polynomials compel us to propose alternative formulations by dissociating signal expansions and parameter estimation. Before any identification, excitation and response signals have to be correctly expanded in the polynomial basis. With Chebyshev polynomials, this requires the signals to be in the same narrow frequency bandwidth. In the presence of noisy signals, correct parameter identification can be performed with a new formulation by taking only central points within the time interval. This methodology can be generalized by defining any other point selection criterion with any other polynomials. This new formulation is first tested on a linear 3-DOF mechanical system and then extended to the parameter identification of a nonlinear mechanical component by using the product property of Chebyshev polynomials. The identification carried out on the hysteretic behavior measured and exhibited by a tape-spring actuator gives relevant parameter estimation. The proposed methodologies and formulations can be easily extended to other orthogonal functions in association with differential or integral transformations of differential equations.

Damage detection in bridges using neural networks for pattern recognition of vibration signatures

Damage detection by measurement of vibration signatures is highly attractive for monitoring bridges because it provides the possibility of electronic recording combined with digital processing and report generation. However, changes in frequency and mode shape caused by damage are usually small and methods that have demonstrated successful detection have generally been confined to small-scale laboratory models. In this investigation a damage detection procedure, using pattern recognition of the vibration signature, was assessed using a finite element model of a real structure — a suspension bridge more than 100 years old. Realistic damage scenarios were simulated and the response under moving traffic was evaluated. Feature vectors generated from the response spectra were presented to two unsupervised neural networks for examination. It is shown that the sensitivity of the neural networks may be adjusted so that a satisfactory rate of damage detection may be achieved even in the presence of noisy signals.

Neural networks for kinetic parameters determination, signal filtering and deconvolution in thermal analysis

Feedforward neural networks have been used for kinetic parameters determination and signal filtering in differential scanning calorimetry. The proper learning function was chosen and the network topology was optimized, using an empiric procedure. The learning process was achieved using simulated thermoanalytical curves. The resilient-propagation algorithm have led to the best minimization of the error computed over all the patterns. Relative errors on the thermodynamic and kinetic parameters were evaluated and compared to those obtained with the usual thermal analysis methods (single scan methods). The errors are much lower, especially in presence of noisy signals. Then, our program was adapted to simulate thermal effects with known thermodynamic and kinetic parameters, generated electrically, using a PC computer and an electronic interface on the serial port. These thermal effects have been generated by using an inconel thread.

Computational neural networks for mapping calorimetric data: Application of feed-forward neural networks to kinetic parameters determination and signals filtering

Feedforward neural networks have been used for kinetic parameters determination and signal filtering in differential scanning calorimetry. The proper learning function was chosen and the network topology was optimised using an empiric procedure. The learning process was achieved using various simulated thermoanalytical curves computed for several thermodynamic and kinetic parameters. Various amounts of simulated noise were added on the power signals. The resilient-propagation algorithm led to the best minimisation of the error computed over all the patterns. Relative errors on the thermodynamic and kinetic parameters were evaluated and compared to those obtained with the usual thermal analysis methods. The results obtained are very promising, and the errors are much lower than with usual methods, especially in the presence of noisy signals. This study shows that simulated thermoanalytical curves produced by Joule effect may be used for the deconvolution of the response of the apparatus, by using artificial neural networks.