Case Study: Festo cuts ML experiment setup from 8 hours to 20-30 minutes with Weights & Biases
Key results
The challenge
Festo developed machine-learning models that detect pneumatic cylinder leaks from audio signals to predict equipment failures before they cause downtime. Setting up and tuning each new experiment with new data was slow and manual.
The solution
Festo used Weights & Biases experiment tracking to record hyperparameters and metrics automatically, Sweeps for hyperparameter optimization, and Launch for distributed training across GPU machines.
“W&B has saved us so much time and effort by streamlining our workflow and making it easy to compare experiments side-by-side and see the impact of different approaches on our results.”
DSDaniel SpiesMachine Learning Engineer, Festo
The results, in context
Festo reduced the time to set up a new experiment from an average of eight hours to twenty to thirty minutes, including Keras callback and plot logging.