Extraction and selection of statistical harmonics features for electrical appliances identification using k-NN classifier combined with voting rules method

Abstract

In this paper, we propose a novel framework for electrical appliances identification using statistical harmonic features of current signals and the use of the k-NN classifier combined with a voting rule strategy. Harmonic coefficients are computed over time using short-time Fourier series of the current signals. From these sequences of coefficients, the mean, standard deviation, skewness, and kurtosis are computed, which provide the statistical harmonic features. This framework has three novelties: (i) selecting the best combination of statistical measures in the sense of classification rate (CR); (ii) combining the k-NN classifier with the voting rule method in order to search for the best number of voting vectors; and (iii) selecting relevant features for the task of appliances identification by using one of the relevant feature selection algorithms based on mutual information. Results evaluated on the Plaid dataset clearly show that the mean and standard deviation statistics combination gives the best CR of 92% with 500 features and gives the minimal computing time compared to the system based on HMM models. Moreover, combining the k-NN classifier with the voting rule using the above features increases the CR up to 95%. Using this combination, the results also show that an increase of the training dataset size further improves identification performance results in terms of precision, sensitivity, and F-score. A feature selection procedure based on joint mutual information strategy shows that using a selected subset of five features is sufficient, giving similar CR results to those obtained using the total number of features, whatever the training dataset size.

A novel pulse plethysmograph signal analysis method for identification of myocardial infarction, dilated cardiomyopathy, and hypertension

“…T is the threshold parameter defined by equation 2 In this work, we carried out extensive experimentation via a pool of features from time, frequency, and textural domains. But time and statistical features such as Shannon energy (SE) [43], zero-crossing rate (ZCR) [44], mean (M) [45], log energy (LE) [46], standard deviation (SD) [45] and root mean square (RMS) [47] yielded better performance as compared to frequency and textural domain features. Therefore, these features were computed to extract the discriminant information of PuPG signals related to normal and abnormal classes.…”