基于混合生成网络的软件系统异常状态评估

针对现有软件系统异常状态评估方法过度依赖数据标注、对时序数据的时间依赖性关注较低和系统异常状态难以量化等问题，提出一种基于混合生成网络的软件系统异常状态评估方法.首先，通过对长短期记忆网络（long short-term memory network， LSTM）与变分自动编码器（variational auto-encoder， VAE）的融合，设计一种LSTM-VAE混合生成网络，并以该网络为基础构建基于LSTM-VAE混合生成网络的系统异常状态检测模型，由LSTM对系统数据的时序特征进行提取并由VAE对系统数据的分布进行建模.然后，由LSTM-VAE异常状态检测模型处理系统关键特征参数，获取系统关键特征参数的异常度量值.最后，利用耦合度方法对传统的线性加权和方法进行优化，通过加权耦合度优化方法计算得到软件系统异常状态的量化值，从而实现对软件系统的异常状态评估.实验结果表明，本文模型对软件系统的异常时序数据具有较好的检测能力，其对系统异常状态的评估结果更为合理、有效.

Evaluation of Software System Abnormal Status Based on Hybrid Generative

To solve the problems that the existing software system abnormal status evaluation methods over depend on data labeling and pay less attention to the time dependence of time-series data, and then it is difficult to quantify the software system abnormal status. Thus, a software system abnormal status evaluation method based on the hybrid generative network is proposed. Firstly, by combining the long short-term memory network (LSTM) and the variational auto-encoder (VAE), an anomaly detection model based on LSTM-VAE hybrid generative network is designed. The features of the system time-series data are extracted by LSTM and its distribution is modeled by VAE. Then, the LSTM-VAE anomaly detection model detects the software system key feature parameters and obtains the anomaly metric value of system key feature parameters. Finally, the coupling degree method is used to optimize the linear weighted sum method. According to the weighted coupling degree method which is optimized, the software system abnormal status quantitative value is calculated, and the software system abnormal status is evaluated. The experimental results show that the proposed model has a better detection ability for the abnormal time-series data of the software system, and its system abnormal status evaluation result is more feasible and effective.