Target Material Identification with High Pressure Water-jet Based on Wavelet Packet Decomposition and PSO-SVM

基於小波包分解和PSO-SVM的高壓水射流靶物材質識別

楊洪濤(Hong-Tao Yang) ； 張東速(Dong-Su Zhang) ； 吳天鳳(Tian-Feng Wu) ； 張偉(Wei Zhang)

《中國機械工程學刊》 39卷1期 (2018/02) Pp. 11-20

摘要

In order to classify the target's material by using the reflection sound signal generated while the target was impacted by the high pressure water-jet, the reflection sound signal was pre-processed and decomposed by wavelet packet in this paper, and the optimum frequency bands of the reflection sound signal was selected through comparative experiments. The relative energy distribution of the optimally selected frequency bands sound signal was calculated as the eigenvalue for the SVM classification model. The standard particle swarm optimization algorithm (PSO) was done in this paper, and the optimized PSO was used to optimize the training parameters (penalty coefficient C and kernel function parameter σ) of the built SVM classification model. As a result, the classification accuracy of the PSO-SVM classification model can be improved, and the time of parameter optimization was reduced. The experimental results show that the classification accuracy (97.78%) was reached by using PSO-SVM classification model, and the modelling time is only 0.92sec. The overall classification accuracy of PSO-SVM classification model was apparently higher than that of BPN, PNN and SVM (K-CV, LOOCV and Grid Search) classification model.

關鍵字

high pressure water-jet ； material identification ； wavelet packet decomposition ； support vector machine ； particle swarm optimization algorithm

並列摘要

為了利用高壓水射流衝擊反射聲信號進行靶 物材質識別，本文對採集的反射聲信號進行預處理和小波包分解，提取各頻率段的相對能量分佈作為特徵向量值輸入支援向量機（SVM）分類模型。為了提高模型識別準確率和減少參數優化時間，本文利用改進的粒子群優化演算法（PSO）來優化SVM分類模型的訓練參數C和σ。詳細介紹了PSO-SVM和短時能量法基本原理和演算法，設計了相關試驗裝置。選擇塑膠地雷、石塊、磚塊作為試驗物件進行試驗，利用上述方法對試驗結果進行處理和驗證。試驗結果證明，以小波包分解後的高12層信號能量分佈作為特徵值能夠得到較好的靶物材質識別效果。應用PSO-SVM靶物分類模型識別靶物材質，正確率可以達到97.78%，而且演算法執行時間僅為0.92 sec，總體識別效果明顯高於BPN、PNN、SVM(K-CV,LOOCV and grid search)分類效果，完全可以用於高壓水射流靶物材質識別。