Work / Pipeline inspection classifier
Fewer false alarmsin pipeline inspection.
Algorithm work on ultrasonic inspection data for VDT Pipeline Integrity Solutions, during a two-month internship.
+50%detection accuracy
ClientVDT Pipeline Integrity Solutions
My roleSoftware Development Intern
WhenFeb to Mar 2025
Built withPython, NumPy, Pandas
StatusCompleted
Screenshot: Signal before and after filtering
The problem
Inspection tools travel through pipelines and record ultrasonic readings of the wall. The raw signal is noisy, and the old classifier flagged so many false defects that engineers spent most of their time dismissing them.
The engineering version
High-frequency sensor streams with poor signal-to-noise ratio. The existing threshold classifier fired on noise. The work was a multi-stage digital filtering pipeline followed by feature thresholding, producing tabular output engineers could review directly.
Pipeline inspection from sensor noise to a short review list
What was hard, and what I did about it
Separating noise from wall loss
Real metal loss and sensor noise look alike in a single reading.
What I did. Filtered the signal in stages and looked at features across neighbouring readings rather than single points.
Implementation notes
Multi-stage digital filters in NumPy, feature extraction over sliding windows, thresholds tuned against labelled runs in Pandas.
Screenshot: Signal before and after filtering
+50%detection accuracy
3xfaster manual review
Ask me about
- How the thresholds were validated without overfitting to one pipeline
- What I would do differently with more labelled data
Built with Python, NumPy, Pandas, Signal processing.
Next case study
Open PVR ticket checkout