Fault Diagnosis with Time-Domain Vibration Data

This repository aims to diagnose types of faults based on time-domain vibration data using neural networks. It covers experiments conducted with Temporal Convolution Network-LSTM models.

Experimental Setup

• 4 DC Motors
  • Normal | Misalignment | Unbalance | Damaged Bearing
• 26.5kHz Sampling Rate
• Data Collection: 102400 samples every 5 seconds

Dependencies

Version of libraries may not make a problem except tensorflow.

• tensorflow == 2.9.0
• scikit-learn == 1.0.2
• numpy == 1.21.6
• scipy == 1.9.1
• matplotlib == 3.5.2

Temporal Convolution Network-LSTM

The TCN implementation is based on @philipperemy's work (https://github.com/philipperemy/keras-tcn).

The TCN-LSTM architecture is created by myself.

TCN Framework

• input shape: (, 800)
• Train data: 13639
• Validation data: 3410
• Test data: 1895

Experimental results

Even if the accuracy for sampling points is not 100%, with a total of over 100,000 data points available for real-time diagnosis, the final fault diagnosis is determined based on the majority of predictions obtained during real-time diagnosis. In the experimental environment, as mentioned earlier, the test results confirmed successful fault type diagnosis.

TCN-LSTM

• Test datasets = 1895
• ROC AUC = 0.995
• F1 Score = 0.9857519788918205
• Accuracy = 98.575 %

TCN Confusion matrix