scholarly journals On the Use of Concentrated Time–Frequency Representations as Input to a Deep Convolutional Neural Network: Application to Non Intrusive Load Monitoring

Entropy ◽  
2020 ◽  
Vol 22 (9) ◽  
pp. 911 ◽  
Author(s):  
Sarra Houidi ◽  
Dominique Fourer ◽  
François Auger
Keyword(s):  
Pattern Recognition ◽  
Recognition Task ◽  
Time Frequency ◽  
Neural Architecture ◽  
Signal Representations ◽  
Current State ◽  
Signal Features ◽  
Load Monitoring ◽  
Pattern Recognition Task

Since decades past, time–frequency (TF) analysis has demonstrated its capability to efficiently handle non-stationary multi-component signals which are ubiquitous in a large number of applications. TF analysis us allows to estimate physics-related meaningful parameters (e.g., F0, group delay, etc.) and can provide sparse signal representations when a suitable tuning of the method parameters is used. On another hand, deep learning with Convolutional Neural Networks (CNN) is the current state-of-the-art approach for pattern recognition and allows us to automatically extract relevant signal features despite the fact that the trained models can suffer from a lack of interpretability. Hence, this paper proposes to combine together these two approaches to take benefit of their respective advantages and addresses non-intrusive load monitoring (NILM) which consists of identifying a home electrical appliance (HEA) from its measured energy consumption signal as a “toy” problem. This study investigates the role of the TF representation when synchrosqueezed or not, used as the input of a 2D CNN applied to a pattern recognition task. We also propose a solution for interpreting the information conveyed by the trained CNN through different neural architecture by establishing a link with our previously proposed “handcrafted” interpretable features thanks to the layer-wise relevant propagation (LRP) method. Our experiments on the publicly available PLAID dataset show excellent appliance recognition results (accuracy above 97%) using the suitable TF representation and allow an interpretation of the trained model.

The Heartbeat Classification Based on PCA

2012 ◽  
Vol 195-196 ◽  
pp. 402-406
Author(s):  
Xue Qin Chen ◽  
Rui Ping Wang
Keyword(s):  
Principal Component Analysis ◽  
Pattern Recognition ◽  
Recognition Task ◽  
Principal Component ◽  
True Positive Rate ◽  
True Positive ◽  
Heartbeat Classification ◽  
Positive Rate ◽  
Electrocardiogram Ecg ◽  
Pattern Recognition Task

Classify the electrocardiogram (ECG) into different pathophysiological categories is a complex pattern recognition task which has been tried in lots of methods. This paper will discuss a method of principal component analysis (PCA) in exacting the heartbeat features, and a new method of classification that is to calculate the error between the testing heartbeat and reconstructed heartbeat. Training and testing heartbeat is taken from the MIT-BIH Arrhythmia Database, in which 8 types of arrhythmia signals are selected in this paper. The true positive rate (TPR) is 83%.

scholarly journals Electric Source Localization Adds Evidence for Task-Specific CNVs

1998 ◽  
Vol 11 (1) ◽  
pp. 21-28 ◽  
Author(s):  
Ina M. Tarkka ◽  
Luis F. H. Basile
Keyword(s):  
Pattern Recognition ◽  
Source Localization ◽  
Clinical Neurophysiology ◽  
Recognition Task ◽  
Source Analysis ◽  
Visual Spatial ◽  
Spatio Temporal ◽  
Electric Source ◽  
Pattern Recognition Task ◽  
Spatial Recognition

This study was an attempt to replicate recent magnetoencephalographic (MEG) findings on human task-specific CNV sources (Basile et al., Electroencephalography and Clinical Neurophysiology 90, 1994, 157–165) by means of a spatio-temporal electric source localization method (Scherg and von Cramon, Electroencephalography and Clinical Neurophysiology 62, 1985, 32–44; Scherg and von Cramon, Electroencephalography and Clinical Neurophysiology 65, 1986, 344-360; Scherg and Berg, Brain Electric Source Analysis Handbook, Version 2). The previous MEG results showed CNV sources in the prefrontal cortex of the two hemispheres for two tasks used, namely visual pattern recognition and visual spatial recognition tasks. In the right hemisphere, the sources were more anterior and inferior for the spatial recognition task than for the pattern recognition task. In the present study we obtained CNVs in five subjects during two tasks identical to the MEG study. The elicited electric potentials were modeled with four spatio-temporal dipoles for each task, three of which accounted for the visual evoked response and one that accounted for the CNV. For all subjects the dipole explaining the CNV was always localized in the frontal region of the head, however, the dipole obtained during the visual spatial recognition task was more anterior than the one obtained during the pattern recognition task. Thus, task-specific CNV sources were again observed, although the stable model consisted of only one dipole located close to the midline instead of one dipole in each hemisphere. This was a major difference in the CNV sources between the previous MEG and the present electric source analysis results. We discuss the possible basis for the difference between the two methods used to study slow brain activity that is believed to originate from extended cortical patches.

scholarly journals POTENTIAL FUNCTION METHOD APPROACH TO PATTERN RECOGNITION APPLICATIONS

2017 ◽  
Vol 2 ◽  
pp. 30
Author(s):  
Peter Grabusts
Keyword(s):  
Pattern Recognition ◽  
Potential Function ◽  
Function Method ◽  
Recognition Task ◽  
Potential Functions ◽  
Learning Procedure ◽  
Potential Methods ◽  
Pattern Recognition Task ◽  
Method Approach ◽  
Potential Function Method

Potential function method was originally offered to solve the pattern recognition tasks, then it was generalized to a wider range of tasks, which were associated with the function approximation. Potential function method algorithms are based on the hypothesis of the nature of the function that separates sets according to different classes of patterns. Geometrical interpretation of pattern recognition task includes display of patterns in the form of vector in the space of input signal that allows to perceive the learning as approximation task. The paper describes the essence of potential function method and the learning procedure is shown that is based on practical application of potential methods. Pattern recognition applications with the help of examples of potential functions and company bankruptcy data analysis with the help of potential functions are given.

Channel and feature selection for a surface electromyographic pattern recognition task

2014 ◽  
Vol 41 (11) ◽  
pp. 5190-5200 ◽  
Author(s):  
Iker Mesa ◽  
Angel Rubio ◽  
Imanol Tubia ◽  
Joaquin De No ◽  
Javier Diaz
Keyword(s):  
Pattern Recognition ◽  
Feature Selection ◽  
Recognition Task ◽  
Selection For ◽  
Pattern Recognition Task

scholarly journals Training Load and Its Role in Injury Prevention, Part I: Back to the Future

2020 ◽  
Vol 55 (9) ◽  
pp. 885-892
Author(s):  
Franco M. Impellizzeri ◽  
Paolo Menaspà ◽  
Aaron J. Coutts ◽  
Judd Kalkhoven ◽  
Miranda J Menaspà
Keyword(s):  
Injury Prevention ◽  
Injury Risk ◽  
Training Load ◽  
Training Process ◽  
Current State ◽  
Training Principles ◽  
Load Monitoring ◽  
Quantitative Indication ◽  
Practical Recommendations

The purpose of this 2-part commentary series is† to explain why we believe our ability to control injury risk by manipulating training load (TL) in its current state is an illusion and why the foundations of this illusion are weak and unreliable. In part 1, we introduce the training process framework and contextualize the role of TL monitoring in the injury-prevention paradigm. In part 2, we describe the conceptual and methodologic pitfalls of previous authors who associated TL and injury in ways that limited their suitability for the derivation of practical recommendations. The first important step in the training process is developing the training program: the practitioner develops a strategy based on available evidence, professional knowledge, and experience. For decades, exercise strategies have been based on the fundamental training principles of overload and progression. Training-load monitoring allows the practitioner to determine whether athletes have completed training as planned and how they have coped with the physical stress. Training load and its associated metrics cannot provide a quantitative indication of whether particular load progressions will increase or decrease the injury risk, given the nature of previous studies (descriptive and at best predictive) and their methodologic weaknesses. The overreliance on TL has moved the attention away from the multifactorial nature of injury and the roles of other important contextual factors. We argue that no evidence supports the quantitative use of TL data to manipulate future training with the purpose of preventing injury. Therefore, determining “how much is too much” and how to properly manipulate and progress TL are currently subjective decisions based on generic training principles and our experience of adjusting training according to an individual athlete's response. Our message to practitioners is to stop seeking overly simplistic solutions to complex problems and instead embrace the risks and uncertainty inherent in the training process and injury prevention.

scholarly journals The role of echolocation strategies for niche differentiation in bats

2018 ◽  
Vol 96 (3) ◽  
pp. 171-181 ◽  
Author(s):  
Annette Denzinger ◽  
Marco Tschapka ◽  
Hans-Ulrich Schnitzler
Keyword(s):  
Pattern Recognition ◽  
Open Space ◽  
Recognition Task ◽  
Positional Information ◽  
Niche Differentiation ◽  
Trophic Niche ◽  
Barro Colorado Island ◽  
Narrow Space ◽  
Pulse Echo

Guilds subdivide bat assemblages into basic structural units of species with similar patterns of habitat use and foraging modes, but do not explain mechanisms of niche differentiation. Bats have evolved four different echolocation strategies allowing the access to four different trophic niche spaces differing in niche dimensions. Bats foraging in open and edge spaces use the “aerial hawking or trawling strategy” and detect and localize prey by evaluating pulse–echo trains in which the prey echo is unmasked. The pulse–echo pairs deliver mainly positional information on the prey and only little information on its nature. Signals are highly variable and are adapted for detection and localization in open space and (or) edge space. In narrow space, bats identify prey by solving a pattern recognition task. Bats using the “flutter detecting strategy” evaluate glint pattern in prey echoes; bats using the “active gleaning strategy” evaluate the spectral–temporal pattern of the prey–clutter echo complex; and bats using the “passive gleaning strategy” evaluate the pattern of prey-generated cues to find food and use echolocation only for spatial orientation. The less variable signals of narrow space bats are adapted for pattern recognition. The diverse and species-rich tropical bat assemblage at Barro Colorado Island, Panama, is here used as an exemplar for assigning bats to guilds, and we discuss the role of echolocation and other adaptations for niche differentiation within guilds.

Sign in / Sign up

Export Citation Format

Share Document