build-model (AD)
1. Data
2. Configuration
configuration parameter | default value |
---|---|
determined from dataset end | |
first non-timestamp column | |
none | |
automatic | |
automatic | |
True | |
automatic | |
True | |
automatic | |
automatic | |
polynomial, time_offsets, identity, intercept, rest_of_week, piecewise_linear, exponential_moving_average, periodic |
target_column
Unlike forecasting, where the primary focus is on predicting future values, in KPI-driven anomaly detection, the target column refers to the specific column for which normal behavior will be modeled. This column typically represents a sensor or a key performance indicator (KPI) suspected to have a strong correlation with potential failures. If there is no prior knowledge indicating which sensor or KPI might be most relevant, it is advisable to model all available sensors or KPIs to detect anomalies.
"target_column": 'pump_stage_4'
holiday_column
In contrast to forecasting, where the holiday column usually marks actual holidays, in anomaly detection, it may be beneficial to also consider special events that could affect normal behavior. However, in general this setting can be neglected.
"holiday_column": 'maintenance_event'
transformations
Tangent tries to enhance the model building process with artificially created featured derived from the original predictors. Here is some guidance to tweak the transformation settings for anomaly detection:
Default Configuration
Use Case: This configuration is ideal when your data involves human activities and cyclical patterns that may influence the behavior being modeled.
Recommended for: General applications, including those involving consumer apps, financial markets, authentication logs, or any domain where human behavior or calendar events play a significant role.
Sensor Data Configuration
Use Case: This configuration is tailored for scenarios where data is purely mechanical or environmental, such as sensor data from industrial equipment or systems where human influence is minimal.
Recommended for: Manufacturing, IoT sensor data, machine monitoring, or any application where the data is not affected by time-based human activities. Since normal behavior modeling functions similarly to a nowcast (or regression), it’s recommended to set target_offsets to None. This approach focuses on capturing the current relationship between predictors and the target, without relying too much on past values of the target. However, using predictor_offsets can be highly beneficial, as they help detect structural changes that occur over time.
transformation | Default configuration (calendar features) | alternative configuration (no calendar features) |
---|---|---|
✅ | ✅ | |
✅ | ✅ | |
✅ | ❌ | |
❌ | ❌ | |
✅ | ✅ | |
✅ | ✅ | |
✅ | ✅ | |
❌ | ✅ | |
✅ | ✅ | |
❌ | ❌ | |
❌ | ❌ | |
❌ | ❌ | |
❌ | ❌ | |
✅ | ✅ |
"transformations": ['exponential_moving_average', 'rest_of_week', 'periodic',
'intercept', 'piecewise_linear', 'time_offsets', 'polynomial',
'identity', 'simple_moving_average', 'month', 'trend',
'day_of_week', 'fourier', 'public_holidays', 'one_hot_encoding']