基于小波包变换和量子神经网络的触电故障类型识别模型

针对农村低压电网中广泛应用的剩余电流保护装置,只能检测到剩余电流有效值的大小作为唯一动作判据,不能自动识别剩余电流与触电故障类型之间所具有的非线性映射规律的难题,提出了一种基于小波包变换和量子神经网络的触电故障类型识别模型。首先应用小波包变换明确了生物体触电故障时,剩余电流中312.475 Hz以下低频带的能量谱波动明显,其中39.062 5~78.125 Hz和119.2~156.25 Hz两频带的波动幅度达9.05和9.00,提取了剩余电流的小波包能量谱8维度特征向量,同时应用特征频带能量占有比之差的平均变化率,实现了生物体发生触电故障的准确检测。然后以小波包能量特征向量为有效信息源,利用量子计算的态叠加思想和神经网络计算的自适应性结合,建立了一种量子神经网络作为触电故障类型识别模型,该网络采用多个量子能级的量子神经元,在学习1 437次时误差精度达到0.000 998 92,快速高效地实现了触电故障类型的识别,其仿真试验准确率达100%。该研究对于研发新一代基于生物体触电电流分量动作的自适应型剩余电流保护装置具有重要的参考价值。

Classification recognition model of electric shock fault based on wavelet packet transformation and quantum neural network

Abstract: Residual current operated protective devices (RCDs) have a wide range of application in low-voltage power grids. RCDs play an important role in preventing electric shock hazard and avoiding fire disaster caused by ground fault. In general, the root mean square (RMS) value of residual current detected is considered as the unique criterion to determine whether the protector acts or not. Theoretical analysis and operation experiences indicate that such a criterion is unavailable in identifying the shocking current signals from animals and human beings. Consequently, it makes human beings unsafe due to the electric shocking, and the disadvantages of the operation principle inherently exist. It results in the malfunction and tripping phenomenon, and greatly decreases the reliability and the rate of proper commissioning for RCDs. Many scholars at home and abroad have made a great deal of breakthrough research on the hardware structure and leakage current detection of residual current protection devices, which improved technical performance of residual current protection device. At present, related research about residual current protection technology focused on the application of a variety of signal processing methods and current amplitude calculation method of biological electrocution branch, which were not related to identification method of electric shock fault of neutral grounded power network. Thus, mathematical expression mechanism between the type of electric shock fault of organism and residual current needs to be studied to identify the type of electric shock fault in time and accurately in the technology of residual current protection. Therefore, time frequency digital feature of residual current signal, and correct identification of the type of electric shock fault are the important preconditions to discover and govern the problems of residual current protection device. In this paper, identification model of electric shock fault based on wavelet packet transform and quantum neural network was proposed to identify the law of nonlinear mapping between residual current and electric shock fault. First, wavelet packet transform was used to analyze energy spectrum fluctuation, and the fluctuation of energy spectrum below 312.475 Hz in residual current was obvious, and reached 9.05 and 9.00 respectively in 119.2-156.25 and 39.062 5-78.125 Hz. Moreover, 8-dimensional eigenvector of wavelet packet energy spectrum of residual current was extracted, and effective threshold for mutation amount of wavelet packet energy was set using average change rate for the difference of characteristic band energy possession ratio, which achieved accurate detection of electric shock fault for organisms. Finally, based on the combination of superposition state of quantum computing and adaptability of neural network computing, a quantum neural network was established as a decision-making system for the type of electric shock failure using energy eigenvectors of wavelet packet as valid sources of information. This network that adopted quantum neurons with multiple quantum levels overcame the problem of local minimum existing in traditional BP (back propagation) algorithm and speeded up the neural network training. The experimental results indicated that the accuracy of the network reached 0.000 998 92 when the number of iterations achieved 1 437, simulation time was 0.146 ms and the accuracy was 100% with the root mean square error (RSME) of 0.108 3, which was superior to EMD-FNN(empirical mode decomposition- fuzzy neural network) algorithm with training times of 125 8, simulation time of 0.398 ms and RSME of 0.193 8. Comparing to EMD-FNN, the time of decomposition of WPT-QNN (wavelet packet transformation- quantum neural network) saved 0.920 s, which could meet the need of actual requirement for quick and accurate action in residual current protection. The method proposed in the paper achieved the recognition of the type of electric shock quickly and efficiently, helpful to develop a new generation of adaptive residual current protection device based on current component action of electric shock for organisms.