Indeed, in the field, industrial use cases are very varied and many of them do not follow cycles or regular cyclical processes, and therefore require treatment with differently optimized algorithms and feature generators.
Mention may be made, for example, of vacuum pumps which only operate when the vacuum must be made and this over a period which cannot be determined in advance, or the flight of an airplane which can vary in time and above all vary in its operational contexts.
Improve the performance of our feature generators and algorithms
Since DiagFit‘s technology is based on feature generators from CNRS research, we have optimized them for the management of acyclic time series.
We have improved their performance as well as that of our unsupervised machine learning algorithms by taking into account all the constraints related to the dynamic management of industrial equipment, cyclical or not.
As a reminder, DiagFit being a blind mode processing software, the optimization of the characteristic generators has enabled it to be functional in all the industrial contexts of our customers.
DiagFit then automatically selects the best generator/algorithm combination and creates the associated pipeline.
Adjust the false/true positive tradeoff of a predictive model in advanced mode
To meet the needs of our most seasoned users, it is now possible to vary the false/true positive compromise of the model.
This balance between True Positive Rate (TPR) and False Positive Rate (FPR) is calculated automatically by DiagFit to theoretically be at the most interesting balance with respect to the data received. Nevertheless, a user who would like to modify this threshold to evaluate several predictive strategies (Risk vs. Gain) can now do so by following these few steps:
- Start training a model on a training dataset
- View the curves resulting from learning and click on “edit”
- Use the slider that appears to modify the sensitivity of the model (link to Makia’s article) and find the compromise that seems most suitable. This modification can be visualized by modifying the threshold on the prediction curve or by varying the TPR and FPR compromise on the curve
- Validate the new threshold/tradeoff for DiagFit to recalculate the model metrics
- Challenge the model with a new dataset using the “Playback” feature or deploy the model in “RUN”