Windowing converts your sequential signal data into vectors for feature extraction. Set the Window size to define the number of samples used to process the training dataset. The platform applies a static window approach.
The window size must be the same for all events in the training dataset, even if the events have different durations. Before loading data into the platform, make sure that the training and validation datasets use the same window size. If needed, apply upsampling or downsampling to align the data.
The platform groups the training dataset by the selected number of samples and calculates all variables from these groups.
The minimum window size is 10 samples, which is required for correct feature calculation. The maximum window size is 1,000 samples, as the platform is designed for TinyML use cases.
Note
If you select a Session ID column during dataset setup, sessions smaller than the selected window size will be removed from the dataset.
You can specify the window size in the following ways:
- Time interval — Specify the window size in milliseconds (ms) and the frequency in Hertz (Hz). If you are using sensors with different frequencies, bring all signals to a single frequency through downsampling or upsampling first.
- Number of Rows (Samples) — Set the number of rows based on your domain knowledge. For example, if your data is sampled at 100 Hz and the longest activity in your dataset takes no longer than one second, 100 rows is a viable option.
- Auto Determination — The platform automatically selects the optimal window size using a built-in algorithm. You can optionally specify minimum and maximum window size boundaries and a sliding shift percentage (for classification and regression).
For frequency domain features, the window size must be a power of 2 and between 128 and 2048 samples. The platform prompts you if your settings do not match these requirements.