Robust and Unsupervised KPI Anomaly Detection Based on Conditional Variational A

Paper: “Robust and Unsupervised KPI Anomaly Detection Based on Conditional Variational Autoencoder” in this paper we propose Bagel, a robust and unsupervised anomaly detection algorithm for KPI that can handle time information related anomalies.

It is better to read Donut’s paper than this one for the main details. It is more helpful to understand the two papers in combination.

KPI: key performance indicator

1. CVAE(Conditional Variational auto-encoder) :arxiv.org/abs/1511.06…

Used to process timing information

2. 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)
• | y z the prior: p (z) = p (z) = N (0, I)
• Posterior: z q ϕ (z ∣ x, y) = N (mu z, diag (sigma z2) q_ {\ phi} (z | x, y) = N (\ mu_ {} z, diag (\ sigma _z ^ 2)) q ϕ (z ∣ x, y) = N (mu z, diag (sigma z2))
• X posterior: p theta (x ∣ z, y) = N (mu x, diag (sigma x2)) p_ {\ theta} (x | z, y) = N (\ mu_x, diag (\ sigma _x ^ 2)) p theta (x ∣ z, y) = N (mu x, diag (sigma x2))
• 
  $\mu_z = W^T_{\mu_z}f_{\phi}(x)+b_{\mu_z}$
• 
  $\mu_x=W^T_{\mu_x}f_{\theta}(z)+b_{\mu_x}$
• 
  $\sigma_z=ln(1+exp(f_\phi(x)))+\Delta$
• 
  $\sigma_x=ln(1+exp(f_\theta(z)))+\Delta$