Hybrid feature recognition method for pulse based on CEEMDAN and extreme learning machine
Accurate identification of traditional Chinese medicine pulse patterns is beneficial for diagnosing human diseases.In response to the problem of fuzzy pulse pattern characteristics,a hybrid feature extraction method based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN),fundamental pulse waveform time-domain features,energy entropy,and sample entropy is proposed.First,human pulse signal data are collected and subjected to wavelet decomposition to remove high-frequency noise and baseline drift.Second,the CEEMDAN method is applied to process pulse signals of five different types,resulting in intrinsic mode functions(IMF)at various orders.The selection of 5-7 order IMF components for computing energy entropy and sample entropy is based on their energy proportions and correlation with the pulse signal.Finally,a hybrid feature vector combining time-domain features,energy entropy,and sample entropy is input into a Sparrow algorithm-optimized Extreme Learning Machine(SSA-ELM)for pulse pattern recognition.The results demonstrate an accuracy rate of 98.60%,with an average precision of 98.64%.Moreover,recall rates and F1 scores for pulse pattern recognition are consistently above 97%.Compared to traditional methods,the proposed strategy exhibits superior recognition performance.
万方数据
维普
