Engineering intelligence for mining data.

01 The engineering problem

Drilling telemetry is valuable, but field data is often messy. Depth traces, rod string length, and measured hole depth can drift out of alignment, creating tedious review work before the data can be trusted.

The drift is systematic rather than random: every time a rod is added to the string, the relationship between logged depth and the hole shifts by a step, and the error accumulates across a shift.

02 Why it matters

Depth is the index every other measurement is filed against. Load, torque, feed rate and RPM are only interpretable if you know where in the hole they were recorded.

Cleaner telemetry creates better downstream decisions. When the data is easier to inspect and trust, teams can spend more energy on analysis, planning, and operational judgment instead of repetitive manual cleanup.

03 The data

Multi-file TDMS shift telemetry from a drill rig. The shift shown spans 2026-05-07 22:08 to 2026-05-08 08:59 — a single night shift, roughly 380,000 samples.

Channels used: load cell, hydraulic torque, feed rate, RPM, head position, rod string length, the connection channel, raw bit depth, and reviewed hole depth.

04 My approach

I built a workflow that compares the available depth references, highlights likely mismatches, and supports algorithm-assisted correction. The goal is not to replace engineering review, but to make the review process faster, clearer, and more reliable.

Candidates are scored by confidence and put in front of an engineer to merge, delete, or approve. Where alignment is weak the tool says so plainly instead of quietly guessing — the reviewer keeps the decision, and the file keeps a record of it.

05 Tools and methods

• Python, reading multi-file TDMS telemetry into one continuous record
• Channel logic over the connection and rod-length signals to propose candidates
• Confidence scoring against expected rod length and target depth
• Matplotlib comparison plots, decimated for interactive review
• Desktop GUI with a review table and a synced graphs tab
• A job queue and worker, so shifts are processed in batches rather than one at a time

06 How the result was validated

Tolerance check. Each candidate's rod string length is compared against its assigned target; anything outside ±1.05 m is flagged for the reviewer rather than applied.

Reconciliation. Corrected hole depth and corrected bit depth are confirmed synced, checked against surveyed depth, and the count of approved intervals is reconciled against the expected count — 8 against 8 in the shift shown.

07 My contribution

I designed and built the tool as the sole developer: the detection logic, the confidence scoring, the review interface, the synced comparison plots, the queue and worker, and the export path.

The drilling domain context and the data came from the research centre.

08 Current status

I do not have adoption figures, measured time savings, or accuracy against a labelled benchmark — so I am not reporting any. Those numbers would need a study I have not yet run.

09 Known limitations

• Confidence scoring is heuristic, not a learned model
• Weak-alignment cases still require an engineer's judgement
• Decimation is for display only — corrections apply to full-resolution data
• Validation depends on surveyed depth being available for the hole
• No labelled ground-truth set yet, so detection precision and recall are unmeasured

10 The next engineering step

Assemble a labelled set of reviewed shifts, then measure precision and recall on rod-add detection against it.

Only with that baseline does it make sense to ask whether a learned classifier beats the current heuristic — otherwise there is nothing to compare against.