Fault Tolerant KPI Evaluation
ANOMALY and EVENT DETECTION
without FALSE-Positives
Everyone loves Watch-Dog systems,
but only when they bark on time
Everyone loves Watch-Dog systems,
but only when they bark on time
Fault-tolerAnt Kpi Evaluation (FAKE) is a generic detection system for anomaly detection and troubleshooting designed to minimize the number of false-positive alerts.
Long time ago, Aesop (Esopo) wrote the well-known fable called "The Boy Who Cried Wolf".
The story learns that if you make false claims, then they won't believe even if you tell the truth.
Most companies define Key Performance Indicators (KPIs) for the analysis of processes. For example, a KPI might be the average number of requests to a service per user/hour.
KPIs are used to represent big-data in a compact and user-friendly manner, but most KPIs are still very similar to the original data and inherits its problems in terms of mathematical function behavior.
The FAKE project started from the background experience of the EVoKE project, an anomaly detection tool designed for wide 2G, 3G, 4G, 5G TLC networks.
FAKE extends the application scenario to different fields such as Computer/IoT networks, environmental monitoring (museums, smart-building, structural health), farming. Moreover, the false-positive reduction techniques have been improved.
FAKE is a general purpose system that can be customized and calibrated for specific application scenarios.
KPIs and business rules are defined by Domain experts.
FAKE natively integrates Learn-By-Example techniques for simplifying the training (setup) phase.
All components of FAKE system are designed to be (quasi) real-time.
Not only tabular data! FAKE-Imaging plugin allow to use images (e.g. from video cameras) as data-source.
FAKE detection system is designed like an onion. User might customize or disable each detection layer.
FAKE exploits user-feedback (e.g. correct/wrong detection) for improving further analyses.
Users might tune parameters of each detector, but also add new components written in Java, Scala or Python.
The FAKE system can run on an isolated node or in a clustered/cloud environment both.
It is very simple to connect FAKE to real-world data-source and HW processes to be monitored.
FAKE detection system is designed like an onion.
Each layer identifies, filters and characterizes events by applying many different algorithms at the same time.
Exploit robust indicators for outlier detection
Exploit Support Vector Regression and Gaussian Processes
Exploit mathematical techniques to select and minimize training set without loosing information
Train multiple-input multiple-output prediction models (e.g., Kriging)
Exploit RNN for time-series regression and characterization
Apply third-generation NN for detecting events in time and space
Apply CS techniques for lossless representation and compression of data streams
Keep learning and improving accuracy from live data and operators feedback