本論文提出一個特徵篩選機制結合粒子群演算法(particle swarm optimization, PSO)及螢火蟲演算法(glowworm swarm optimization, GSO)之粒子群螢火蟲演算法(particle glowworm swarm optimization, PGSO),用於調整機率類神經(probabilistic neural network, PNN)之平滑矩陣,並以留一交叉驗證法(leave-one-out cross validation, LOOCV)估測機率神經網路分類準確率作為螢火蟲演算法之適應值,從數個候選特徵中移除影響力較小的特徵,進而建立一組理想的特徵向量。 研究首先,藉由S轉換(S transform, ST)與TT轉換(TT transform, TT)分析13種電力品質(power quality disturbance, PQD)取得6條時間特性曲線及5條頻率特性曲線,接著藉由11條特性曲線計算出分別於無雜訊、SNR=30dB、SNR=25dB、SNR=20dB及SNR=15dB雜訊環境下之時間-頻率關係與時間-時間關係,由上述兩種關係計算出62個候選特徵;其次,經由特徵篩選機制在不降低估測分類準確率條件下,移除影響力較小的特徵,篩選出一個理想的特徵向量。 結果顯示,於上述雜訊環境下,使用粒子群螢火蟲演算法特徵篩選機制所建構之特徵向量,使用倒傳遞類神經(back propagation neural network, BPNN)進行分類準確率估測,其分類準確率及運算時間皆有提升。
This study proposes an optimization feature selection scheme, which combines a glowworm swarm optimization (GSO) algorithm with a particle swarm optimization (PSO), namely particle glowworm swarm optimization (PGSO). The proposed PGSO-based scheme optimizes the smoothing parameters of LOOCV of probabilistic neural network (PNN). The least influenced features, rarely degrading the cross-validation accuracy, are removed from the candidate features by using the optimal smoothing parameters to reconstruct the optimal feature vectors set. This paper illustrates time-frequency and time-time relationships of 13 types of power quality disturbance (PQD) by using S-transform (ST) and TT-transform (TT) in the conditions of no noise, SNR=30dB, SNR=25dB, SNR=20dB and SNR=15dB. By observing the ST and TT contours, 6 types of time characteristic curves and 5 types of frequency characteristic curves are depicted. According to the time-frequency and the time-time relationships, 62 candidate features are calculated. The results show that the classification accuracies and runtimes of classifiers by using to the back propagation neural network (BPNN) is superior to that by using the optimal feature vectors set obtained by the proposed PGSO-based scheme, even in the environment with various noise interference.