Innovative and Cognitive Production Technology and Systems, 16 - 18 July 2025
HomeMachine Tools & Special MachinesA Hybrid, Distributed Condition Monitoring System using MEMS Microphones, Artificial Neural Networks, and Cloud Computing
A Hybrid, Distributed Condition Monitoring System using MEMS Microphones, Artificial Neural Networks, and Cloud Computing
Nice work! I would like to ask a couple of questions:
1. I like the advantages proposed by the hybrid method over the traditional CM systems. Have you quantified these advantages in terms of cost?
2. You have used Keras for ANN implementation. Could you elaborate on parameters of this neural network? Also, information on amount of data, training period etc would be interesting.
Thank you.
Thank you very much for your comment!
1. No, the advantages have not been quantified in terms of cost in the scope of this work. However, I fully agree that such quantification would be valuable.
2. The condition detection (classification) model was compiled with an Adam optimizer (0.001 learning rate) and a sparse categorical cross-entropy loss and trained for 10 epochs.
The time-based prediction (regression) model was also compiled with an Adam optimizer (0.001 learning rate) but with a mean squared error (MSE) loss. The regression ANN was trained for 50 epochs. Please find the hyperparameters regarding the topologies of the respective ANNs in the tables below.
Regarding the amount of data used: Both ANNs were trained on 9489 features (spectra corresponding to 3 different conditions) remaining after preprocessing, excluding 20% of the preprocessed dataset reserved for validation.
Nice work! I would like to ask a couple of questions:
1. I like the advantages proposed by the hybrid method over the traditional CM systems. Have you quantified these advantages in terms of cost?
2. You have used Keras for ANN implementation. Could you elaborate on parameters of this neural network? Also, information on amount of data, training period etc would be interesting.
Thank you.
Thank you very much for your comment!
1. No, the advantages have not been quantified in terms of cost in the scope of this work. However, I fully agree that such quantification would be valuable.
2. The condition detection (classification) model was compiled with an Adam optimizer (0.001 learning rate) and a sparse categorical cross-entropy loss and trained for 10 epochs.
The time-based prediction (regression) model was also compiled with an Adam optimizer (0.001 learning rate) but with a mean squared error (MSE) loss. The regression ANN was trained for 50 epochs. Please find the hyperparameters regarding the topologies of the respective ANNs in the tables below.
| Condition Detection (Classification) ANN |
|———————————|———————|————–|
| Layer Type | Output Shape | Activation |
| Input | (None, 2205) | Linear |
| Dense | (None, 11025) | ReLu |
| Dense | (None, 220) | ReLu |
| Dense | (None, 128) | ReLu |
| Dense | (None, 128) | ReLu |
| Dense | (None, 3) | Linear |
| Time-Based Prediction (Regression) ANN |
|———————————|———————|————–|
| Layer Type | Output Shape | Activation |
| Input | (None, 2205) | Linear |
| Dense | (None, 11025) | ReLu |
| Dense | (None, 220) | ReLu |
| Dense | (None, 128) | ReLu |
| Dense | (None, 128) | ReLu |
| Dense | (None, 1) | Linear |
Regarding the amount of data used: Both ANNs were trained on 9489 features (spectra corresponding to 3 different conditions) remaining after preprocessing, excluding 20% of the preprocessed dataset reserved for validation.
Thank you for the detailed and clear explanation.