Case study · Nov 2023 — Apr 2024
Remaining Useful Life Predictor
CNN-LSTM model estimating remaining useful life of cutting tools from telemetry, at 93% accuracy.
- Python
- CNN-LSTM
- SVM
- Regression
- Express
- React
The problem
A cutting tool replaced too early wastes money; one replaced too late damages the workpiece and the machine. Fixed-interval maintenance schedules guess at a number that actually depends on how the tool has been used.
Approach
- 01Started from 100,000+ points of raw tool telemetry — cleaning, resampling, and extracting the features that actually carry degradation signal.
- 02A CNN-LSTM does the work: convolutional layers pull local patterns out of the sensor windows, the LSTM tracks how those patterns drift across the tool’s life.
- 03Benchmarked against SVM and classical regression baselines to confirm the sequence model was earning its complexity.
- 04Hyperparameter tuning over window size, layer depth, and learning rate; the result is served through an Express API behind a React dashboard so the prediction is readable by someone who is not running notebooks.
Outcome
93% accuracy on remaining-useful-life estimation, turning a fixed maintenance interval into a per-tool prediction.