After acquiring a dataset, I embarked on a thorough exploration to identify which features were most meaningful, both generally and for predicting asset failure. I employed various methods, including statistical analysis, decision trees, visualization, and Principal Component Analysis (PCA). This multifaceted approach yielded numerous valuable insights.

One major finding was the notable scarcity of specific failure modes and the detailed operating conditions—speeds, temperatures, and times—associated with each failure mode. By focusing on and filtering failures to the forefront, I was able to produce clear and impactful visuals. This process revealed significant patterns that were previously obscured by the non-failure data, providing a much clearer understanding of the factors leading to asset failure.

Once I had a thorough understanding of the dataset, I began building a predictive model. Initially, I encountered concerning results as my model's recall dropped over multiple epochs. This decline was due to the unbalanced nature of the dataset and the equal weighting between failure and non-failure events.

To address this issue, I first attempted to adjust the weights to account for the imbalance, but this approach felt "hacky" and involved manually determining weights through a formula, which I found unsatisfactory. Consequently, I decided to bootstrap the dataset using a tabular Generative Adversarial Network (GAN). By generating synthetic data similar to my existing dataset, I was able to extend the dataset and improve its balance.

Using Keras tuners, I fine-tuned the model and ultimately achieved a recall of approximately 95%. This approach not only enhanced the model's performance but also provided a more robust solution to the dataset imbalance problem.