Home

FEDOT.Industrial framework provides a set of tools for industrial type of tasks, such as time series forecasting, classification, and regression, as well as anomaly detection. The project is based on the FEDOT AutoML framework using its algorithmic kernel in terms of evolutionary optimization and obtaining graph-like pipelines.

It is a versatile tool for the time-series engineer. The software is designed to minimize user involvement in the programming while maximizing the quality of the results obtained for tasks such as: time series classification and forecasting, anomaly detection, and regression tasks.

We developed a unified automated machine learning approach for univariate and multivariate time series that can solve tasks of classification, extrinsic regression, forecasting in a unified way. It is achieved by the implemented kernel-based approach that allows one to design the adaptive feature space (that includes statistical-based, frequency-based, and topology-based representation) for each modeling task.

The interface of the framework is standard for modern AutoML solutions. It accepts dataset, task type and timeout for optimisation. Raw data series (that can include different properties - e.g. high frequency and pulses) are preprocessed to regular form with unified time step using interpolation. AutoML setup (hyperparameters, search space) is configured according to the modeling task. Then, feature space transformation step is conducted (framework generate sets of the features that can be used for trainig of the models). The full modelling pipeline is optimised using evolutionary algorithm. Finally, the best obtained pipeline (or several pipelines in multi-objective case) is using to generate the prediction.