AI REGIO Supervised Learning for Multivariant Time Series Forecasting
A Jupyter Notebook demo to demonstrate the partial content of FZI asset "Process integrated feedback management" which uses the historical multivariant time series data from a simulative robot manipulator to predict data in the future time steps based on supervised regression learning.
- Demo task: predict the robot tool center point (TCP) deviation caused by external forces and robot internal control error based on measured joint position, joint velocities and joint accelerations. The data is collected previously in a simulated environment. When the translation error of TCP can be well predicted, it is quite possible to replace some external measurement devices of robotic TCP.
- Preprocessing: provide the method of choosing the related data columns, data quality check, dataset separation, dataset standardisation, and creating the time window slices by using rolling window with configured time length and shift steps.
- AI Models: in this demo, we create 7 different model combinations based on LSTM, TCN, and Convlutional 1D block.
- Parameter Tuning: in this demo, we integrate the hyperparameter tuning tool: Keras Tuner. For each model, we tune model parameter, e.g. kernel size, filter size, unit number, activation function, dropout rate, and training parameter, e.g. batch size, training epochs, learning rate etc. We use Bayesian optimization approach to find the optimal parameter combinations.
- Model Evaluation: provide prediction visualization and metric output.
The resource belongs to the FZI Asset "Process integrated feedback management" and is used in the experiment collaborated with Kautenburger GmbH. The Experiment is a part of the EU Project "AI-REGIO".
The resource has been packaged in a zip folder. To use the notebook, it is recommended to follow the "configuration guide" to configure a virtual environment before running the notebook "Asset_demo_notebook.ipynb". We will maintain this asset and plan iterative updates.