There are many valuable insights for GRU implementation in industry, thank you for the presentation Ms. Garghetti! In your case studies, what are training dataset sizes and training times?
Thank you for your question! Traning phase is composed by 46 signals, acquired during the milling operations of 46 vacum pumps with tool copy n.1.
Traning acquisition time is limited to few minutes (around 4 minutes).
The training time for the implementation of the GRU is related to the Bayesian optimization, since you can fix a maximum time. In the current case study, Bayesian Optimizator has a maximum optimization time equal to 20 minutes.
Thanks Ms. Garghetti for this interesting presentation. Indeed available commercial and literature TCM systems have many drawbacks as you mentioned in your presentation. One of the main drawbacks is the lack of generalization, which increase the time and effort needed to train the model. I am wondering out of the 46 operations you acquired, how many operations were used to train the GRU model and how many were used for testing and validating? Was the data selected randomly for training?
I have used the RNN before for real-time detection of tool wear “state” using a “biased” learning data and it showed high accuracy within a very wide range of machine power. I am interested in a collaboration opportunity to apply the GRU on the same data and benchmark the results. Please find more details in the publication below:
Hassan, M., A. Sadek, and M. H. Attia. “A generalized multisensor real-time tool condition–monitoring approach using deep recurrent neural network.” (2019): 41-52. https://www.astm.org/ssms20190020.html
Thank you very much! During the training phase, I implemented a Bayesian optimization to find the optimal hyperparameters values. During each iteration, leave one out cross validation is applied:
• Training on N-1 cycles (45 cycles)
• Testing on the remaining N-th cycle.
At the end of the optimization, I select the hyperparameters corresponding to the minimum RMSE. I think you can find the entire procedure in detail as soon as the conference paper will be published.
There are many valuable insights for GRU implementation in industry, thank you for the presentation Ms. Garghetti! In your case studies, what are training dataset sizes and training times?
Dear Gamze,
Thank you for your question! Traning phase is composed by 46 signals, acquired during the milling operations of 46 vacum pumps with tool copy n.1.
Traning acquisition time is limited to few minutes (around 4 minutes).
The training time for the implementation of the GRU is related to the Bayesian optimization, since you can fix a maximum time. In the current case study, Bayesian Optimizator has a maximum optimization time equal to 20 minutes.
I am avaible for any clarifications,
Best regards,
Federica
Thanks Ms. Garghetti for this interesting presentation. Indeed available commercial and literature TCM systems have many drawbacks as you mentioned in your presentation. One of the main drawbacks is the lack of generalization, which increase the time and effort needed to train the model. I am wondering out of the 46 operations you acquired, how many operations were used to train the GRU model and how many were used for testing and validating? Was the data selected randomly for training?
I have used the RNN before for real-time detection of tool wear “state” using a “biased” learning data and it showed high accuracy within a very wide range of machine power. I am interested in a collaboration opportunity to apply the GRU on the same data and benchmark the results. Please find more details in the publication below:
Hassan, M., A. Sadek, and M. H. Attia. “A generalized multisensor real-time tool condition–monitoring approach using deep recurrent neural network.” (2019): 41-52. https://www.astm.org/ssms20190020.html
Dear Mahmuoud,
Thank you very much! During the training phase, I implemented a Bayesian optimization to find the optimal hyperparameters values. During each iteration, leave one out cross validation is applied:
• Training on N-1 cycles (45 cycles)
• Testing on the remaining N-th cycle.
At the end of the optimization, I select the hyperparameters corresponding to the minimum RMSE. I think you can find the entire procedure in detail as soon as the conference paper will be published.
For any other discussions, please send me an email to federica.garghetti@polimi.it
Best regards,
Federica