Hello,
Anomaly detection projects have a unique feature which is the learning on device (you can retrain the model directly on the board).
What you get from the benchmark is the structure of models that works well for the data used, but not a pre trained model. You need to use "learning signals" to retrain the model.
In the validation step, the knowledge is not available, so you need to use a "learn file" containing nominal data, for the model to retrain itself and then two other files to do the tests.
It is advised to use a learning file containing nominal signals different than the nominal file. You can import the nominal data from the benchmark, but use new files for the "regular file" and "abnormal file".
Later, when compiling the model, you have the choice to include, or not, the knowledge from the benchmark:
- If you use the knowledge, it will use it as learning data and work as in the benchmark.
- If you don't, you will need to learn new nominal signals to retrain the model.
- You can do both if you want. You can complete the knowledge from the benchmark by learning other signals later.
.
A use case example of the learning on device is the following. Let's say that you have multiple machines of the same model. Some are newer than other, so they will have slightly different signals. With the learning on device, you can collect data from few machines, use the benchmark to get the model and deploy the same model on all machines. The trick is that you can now learning nominal signals from each machine and get model specifically trained for each machine!
Best regards,
Julian
In order to give better visibility on the answered topics, please click on 'Accept as Solution' on the reply which solved your issue or answered your question.
