This project realizes the anomaly detection and root cause location of multimodal data. The anomaly detection part adopts MTAD-GAT model (metric, trace) and DeepLog model (log), The root cause localization part adopts the Squeeze model.

Experimental setup

• Train:
  • Train data (update model): 2022-03-24 15:20:00 ~ 2022-03-25 08:06:00
  • Valid data (prevent overfitting): 2022-03-25 08:07:00 ~ 2022-03-25 15:19:00
• Test:
  • Valid data (search threshold): 2022-03-26 08:30:00 ~ 2022-03-26 11:29:00
  • Test data (evaluation model): 2022-03-26 11:30:00 ~ 2022-03-26 20:29:00

Result

anomaly detection

		P	R	F1
metric	-	0.5329	0.7945	0.6379
metric	+	0.8873	0.7412	0.8077
trace	-	0.1943	0.3527	0.2506
trace	+	0.2073	0.8706	0.3348
log	-	0.1382	0.4027	0.2058
log	+	0.1759	1.0000	0.2992
metric+trace	-	0.3190	0.6218	0.4217
metric+trace	+	0.7917	0.8941	0.8398
metric+trace+log	-	0.3347	0.6359	0.4386
metric+trace+log	+	0.8085	0.8941	0.8492

root cause localization

		PR@1	PR@2	PR@3	PR@4	PR@5	PR@Avg
RootCause	-	0.2783	0.4001	0.5192	0.5953	0.6217	0.4829
RootCause	+	0.5739	0.7652	0.8522	0.9217	0.9391	0.8104

More

• See Log for training and testing logs.
• See Loss for loss visualization.