基于改进生成对抗网络的时序数据异常检测

Anomaly detection of time series data based on improved generative adversarial networks

针对时序数据复杂的时间相关性,以及现有异常检测模型存在准确性低、训练不稳定等问题,提出一种结合BiLSTM和WGAN-GP的无监督时序数据异常检测模型。使用BiLSTM作为生成器和判别器的基础网络来捕获时序数据的时间相关性;为保证训练过程的稳定性,使用Wasserstein距离取代原有的衡量方法,在判别器损失中加入梯度惩罚项;将重构损失与判别损失相结合定义异常函数,采用局部自适应阈值方法判别异常,提高异常检测的准确性。为验证模型性能,在涉及多个领域的5类数据集上进行实验,其结果表明,该模型相比于Arima、LSTM等模型具有最高的平均F1分数。

Aiming at the complex time correlation of time series data and the problems of poor accuracy and instability of training in existing anomaly detection models,an unsupervised time series data anomaly detection model combining BiLSTM and WGAN-GP was proposed.BiLSTM was used as the basic network of generator and critic to capture the time correlation of time series data.To ensure the stability of the training process,the Wasserstein distance was used to replace the original measurement method,and the gradient penalty was added to the critic loss.The reconstruction loss and the discrimination loss were combined to define the anomaly function,and the local adaptive threshold method was used to distinguish the anomaly to improve the accuracy of anomaly detection.To verify the performance of this model,experiments were carried out on five kinds of data sets involving various domains.The results show that this model has the highest averaged F1 score compared with that of Arima,LSTM and other methods.