Demonstration of anomaly detection on the UC Berkeley milling data set using a disentangled-variational-autoencoder (beta-VAE).

The method is described in the article "Self-supervised learning for tool wear monitoring with a disentangled-variational-autoencoder" in IJHM. Link to the preprint is here. The method is also described in my MASc thesis, Feature Engineering and End-to-End Deep Learning in Tool Wear Monitoring.

I also have detailed blog posts exploring the UC Berkeley milling data set (here), describing how the VAE is constructed (here), and analyzing the results of the anomaly detection model (here). The code is well explained in the blog posts (and associated colab notebooks).

Feel free to cite my research if you use it in any academic research.

@article{hahn2021self,
  title={Self-supervised learning for tool wear monitoring with a disentangled-variational-autoencoder},
  author={Hahn, Tim Von and Mechefske, Chris K},
  journal={International Journal of Hydromechatronics},
  volume={4},
  number={1},
  pages={69--98},
  year={2021},
  publisher={Inderscience Publishers (IEL)}
}

milling-tool-wear-beta-vae.ipynb is the notebook to replicate the results and make the figures. I recommend using google colab. The notebook is optimized for it, it will run in your browser, and no package installation required!

A disentangled-variational-autoencoder, with a temporal convolutional neural network, was used to model and trend tool wear in a self-supervised manner. Anomaly detection was used to make predictions in both the input and latent spaces. The experiment was performed on the UC Berkeley milling data set.

The method achieved a precision-recall area-under-curve (PR-AUC) score of 0.45 across all cutting parameters in the milling data set, and a top score of 0.80 for shallow depth cuts. The study presents the first known use of a disentangled-variational-autoencoder for tool wear monitoring.

Figure 1: The tool wear trend for case 13, generated from the latent space.

Figure 2: The distribution of the test data samples in the latent space. The y-axis shows the true state of the data samples (either healthy, degraded, or failed) and the x-axis shows the anomaly predictions (either normal or abnormal). The dotted line represents a decision threshold. Each small grey dot represents one data sample.

Figure 3: The precision-recall curve (left) and the ROC curve (right) for the best model on the test data (using latent space anomaly detection). The no-skill model (equivalent to a random model) is plotted as a comparison. In the context of this experiment, precision is the proportion of abnormal predictions that are truly abnormal (the tool is in a failed state). Recall is the proportion of truly abnormal samples (failed) that were identified correctly.

Figure 4: The PR-AUC score in the latent space, while looking at one parameter in a pair at a time.

Figure 5: An example of one cut (out of 167) from the milling data set. Six signals are collected during each cut.