不相干约束项的判别字典表示电能质量扰动分类研究

针对稀疏表示电能质量扰动识别中判别字典学习的冗余性,提出一种具备精简性和不相干约束项的判别字典学习电能质量扰动分类方法。首先,将不同电能质量扰动样本训练获得子字典,公共字典和判别字典。接着,利用判别字典优化方法求解出降维测试信号的稀疏表示。最后,利用稀疏表示重构方法求解测试样本,由冗余残差最小值确定电能质量扰动信号的类型。不相干约束项的判别字典学习方法是在训练字典的过程中直接驱使字典具有判别性,获得更加精简且具有判别性的稀疏字典来提升最终的识别性能。实验结果表明8类电能质量扰动信号在40、30、20 d B信噪比递减时,平均扰动识别率有所降低但平均识别精度仍高达96%以上。仿真实验结果表明该方法能有效的对不同电能质量扰动进行识别并提高识别结果的精确度,并且不相干约束项的判别字典算法更优化于判别字典学习算法的分类识别性能。

Research on power quality disturbances classification based on discriminative dictionary learning with structured incoherence

In order to solve the redundancy of discriminative dictionary learning (DDL) in the sparse representation of the power quality disturbance identification, the dictionary learning with structured incoherence (DLSI) is proposed to make discriminant dictionary more concise.Firstly, the various types of power quality disturbances are trained to obtain sub-dictionaries, public dictionary, and discriminant dictionary.Then, the sparse representation of the reduced dimension test signal is solved by the method of discriminative dictionary optimization.Finally, using sparse representation reconstruction method to solve the test samples, and the type of power quality disturbance signals are determined by the minimum of the residual error.DLSI could directly drive the discriminative dictionary to can discriminate various types of power quality disturbances, and could obtain a more compact and discriminative sparse dictionary to improve the final recognition rates for identification of power quality disturbances.The experimental results demonstrate that the average recognition rate is higher than 96% for identifying eight types of power quality disturbances, while the classification accuracy decreases slightly with the ratio of signal to noise ratio (SNR) varying from 40, 30 to 20 dB.The simulation results show that DLSI can effectively identify different types of power quality disturbance signals and improve the accuracy of the identification results, in the meantime, DLSI algorithm shows better classification and recognition performance than DDL algorithm.