signal-to-noise ration, SNR) 时的广义相位排列熵熵值变化情况. 最后, 将广义相位排列熵运用于异常 心电信号分析. PE q,δ , q=1, δ=1 PE q,δ , q=1, δ=1 PE q,δ , q=0.5, δ=1 PE q,δ , q=1, δ=2 PE q,δ , q=0.5, δ=1 PE q,δ , q=1, δ=2 PE q,δ , q=0.8, δ=1.5 PPE q,δ , q=1, δ=1 PPE q,δ , q=0.5, δ=1 PPE q,δ , q=1, δ=2 PPE q,δ , q=0.8, δ=1.5 PE q,δ , q=0.8, δ=1.5 PPE q,δ , q=1, δ=1 PPE q,δ , q=0.5, δ=1 PPE q,δ , q=1, δ=2 PPE q,δ , q=0.8, δ=1.5 Figure 7 (Color online) The PE q,δ and PPE q,δ in different length of data and signal-to-noise ration (SNR) when m = 5, τ = 1. (a) The sequence of Gaussian wave packets; (a1) enlarged view of each wave packet in a Gaussian wave packet sequence; (b) the PE q,δ and PPE q,δ in different length of data when m = 5, τ = 1; (c) the PE q,δ and PPE q,δ in different SNR when m = 5, τ = 1 此, 在之后的分析时可以同时使 q, δ 取较优值, 但是要尽可能的选用 PPE q,δ 去测量, 从而使检测结果 更为显著. 为了探究数据长度和噪声对于广义相位排列熵的影响, 产生一数据长度为 10000 的 Gauss 波包 7 康欢等: 基于两参数熵的广义相位排列熵算法 序列, 如图 7(a) 所示. 图 7(b)如图 10(a) 所示为 cu01 心室纤颤信号, 数据点 1100∼2500 为心室纤颤病发阶段, 其余处为正常心 电信号. 图 10(b) 和 (c) 分别为 q = 1, δ = 1, q = 0.5, δ = 1.5, q = 0.4, δ = 1.6 时心室纤颤信号的 PE q,δ 和 PPE q,δ 熵值图. 相比于图 8 和 9, 无论 q, δ 取何值, PE q,δ 和 PPE q,δ 都能明显地检测出心室纤颤病 Abstract The generalized permutation entropy (PE q,δ ) with appropriate parameters can amplify the minor changes of the system, however, the phase of the signal contains more critical information than the amplitude. This paper introduces the phase information into the generalized permutation entropy and proposes the generalized phase permutation entropy (PPE q,δ ) algorithm. At the same time, we verify the advantages of PPE q,δ in detecting the dynamic changes of the system, analyze the influence of q, δ selection on the dynamic change of the system, and explore the effect of data length and noise for the PPE q,δ . Finally, the PPE q,δ is applied to the analysis of abnormal ECG signals. The results show that when the value of q, δ are the same, the PPE q,δ has a more significant effect on the detection of the same dynamic change. Whether for the logistic map or the abnormal ECG signal dynamic changes detection, when q > 0 and δ > 0, the effect of PE q,δ and PPE q,δ can be improved by decreasing q value, or increasing δ value, or both changing simultaneously. Furthermore, the change of data length has no effect for the PE q,δ and PPE q,δ mainly, and both of them are robust to noise.