Robust and Unsupervised KPI Anomaly Detection Based on Conditional Variational Autoencoder
论文：《Robust and Unsupervised KPI Anomaly Detection Based on Conditional Variational Autoencoder》
Thus， in this paper we propose Bagel, a robust and unsupervised anomaly detection algorithm for KPI that can handle time information related anomalies.
主要细节去看Donut的论文比看这篇论文更好，两篇结合查看更有助于理解。
KPI: key performance indicator
- CVAE(conditional variational auto-encoder, 条件变分自编码器）:http://arxiv.org/abs/1511.06988
用于处理时序信息
- dropout layer：to avoid overfitting.
- KPI v = (v1, v2, …, vn)
- the i-th window of the KPI is x(i) = (vi, vi+1, …, vi+W−1)
- z prior：p(z|y)=p(z)=N(0, i)
- z posterior: q ϕ ( z ∣ x , y ) = N ( μ z , d i a g ( σ z 2 ) ) q_{\phi}(z|x,y)=N(\mu_{z},diag(\sigma^2_z)) qϕ(z∣x,y)=N(μz,diag(σz2))
- x posterior: p θ ( x ∣ z , y ) = N ( μ x , d i a g ( σ x 2 ) ) p_{\theta}(x|z,y)=N(\mu_x,diag(\sigma^2_x)) pθ(x∣z,y)=N(μx,diag(σx2))
- μ z = W μ z T f ϕ ( x ) + b μ z \mu_z = W^T_{\mu_z}f_{\phi}(x)+b_{\mu_z} μz=WμzTfϕ(x)+bμz
- μ x = W μ x T f θ ( z ) + b μ x \mu_x=W^T_{\mu_x}f_{\theta}(z)+b_{\mu_x} μx=WμxTfθ(z)+bμx
- σ z = l n ( 1 + e x p ( f ϕ ( x ) ) ) + Δ \sigma_z=ln(1+exp(f_\phi(x)))+\Delta σz=ln(1+exp(fϕ(x)))+Δ
- σ x = l n ( 1 + e x p ( f θ ( z ) ) ) + Δ \sigma_x=ln(1+exp(f_\theta(z)))+\Delta σx=ln(1+exp(fθ(z)))+Δ