Mechanical engineer to data scientist: SPC charts and OEE first

Dear Rohit

You wrote that you have spent six years as a production engineer at an auto-components plant, that you maintain the SPC charts for two machining lines, and that you want to move into data science without throwing those years away. You asked whether the move is realistic and where to start. Here is my honest answer, in the order I would do things.

You have been doing statistics every shift

When your line supervisor asks why the bore diameter on line 2 drifted after lunch, you pull the X-bar and R chart, check whether points fall outside the control limits, look for runs of seven on one side of the centre line, and then walk to the machine to check tool wear. That sequence — look at the distribution, test for a signal against noise, form a hypothesis, check it physically — is the heart of data analysis. Many people who come into data science from software have never had to separate a special cause from common-cause variation with money on the line.

Your other artefacts transfer too:

• OEE sheets: availability, performance and quality as a product of three ratios. That is metric decomposition, which analysts do constantly.
• Downtime logs with reason codes: an event log with categories. Pareto charts on reason codes are exploratory data analysis.
• MSA and gauge R&R studies: measurement error thinking. In data science this is the question "is my label reliable?"
• Maintenance records: time-between-failures data, the starting point for predicting breakdowns.

Rebuild your OEE sheet in code first

Before any machine learning, reproduce something you already trust. Take a synthetic shift log like the one below and compute OEE per shift. If the numbers match what you would get in Excel, you have learned the tools without also learning a new domain.

import pandas as pd

# Synthetic shift log for one machining cell
log = pd.DataFrame({
    "shift":          ["A","B","C","A","B","C"],
    "planned_min":    [480,480,480,480,480,480],
    "downtime_min":   [35, 60, 90, 20, 55, 110],
    "ideal_cycle_s":  [30, 30, 30, 30, 30, 30],
    "total_parts":    [850,780,700,880,790,650],
    "good_parts":     [835,760,665,870,775,610],
})

g = log.groupby("shift").sum()
run_min = g["planned_min"] - g["downtime_min"]
g["availability"] = run_min / g["planned_min"]
g["performance"] = (g["ideal_cycle_s"] / 2 * g["total_parts"]) / (run_min * 60)
g["quality"] = g["good_parts"] / g["total_parts"]
g["oee"] = g["availability"] * g["performance"] * g["quality"]

print(g[["availability","performance","quality","oee"]].round(3))

Note the ideal_cycle_s / 2: grouping summed the cycle time over two days, so dividing by the day count restores the per-part ideal time. That is the sort of silent error that grouping introduces, and catching it is exactly the checking habit you already have. Output:

       availability  performance  quality    oee
shift
A             0.943        0.956    0.986  0.888
B             0.880        0.929    0.978  0.799
C             0.792        0.888    0.944  0.664

Shift C loses on all three factors, but mostly on availability. You would have guessed that from the floor. The code is not telling you anything new yet — and that is the point of this step.

Then ask a question the chart cannot answer

A control chart tells you a process has shifted. It does not tell you whether tomorrow's shift C will lose more than 90 minutes. That is a prediction problem, and it is where you begin learning data science proper:

1. 1Build a table where each row is one shift, with features known before the shift starts: machine age since last overhaul, tool changes due, operator experience band, the previous shift's downtime, material batch.
2. 2The target is next shift's downtime minutes, or a flag for "more than 60 minutes".
3. 3Split by time, never randomly. Train on the earlier weeks and test on the later ones, because in a plant the future must not leak into the past.
4. 4Compare your model against a simple rule, such as "predict the same downtime as the last shift". If you cannot beat that, say so.

The Data Science programme covers Python, SQL, statistics and model validation in that order, which is the order you need.

What you are missing, specifically

I would not pretend the gap is small:

• Python and SQL fluency — pandas, joins, reading messy exports from an MES or ERP.
• Machine learning evaluation — train/test splits, leakage, choosing a metric from costs. Your SPC knowledge helps with the statistics but not with this.
• Working with far more features than a control chart has — dozens of columns, missing values, categories.
• Explaining models, not processes — why the model flagged this shift.

Manufacturing analytics before generalist data science

A pure data scientist title at a consumer internet company would throw away your advantage. Look instead for manufacturing data analyst, process data analyst, quality analytics engineer or data scientist — operations at manufacturers, auto suppliers and industrial software firms. Read each description for words like OEE, SPC, predictive maintenance or MES; those tell you the team values the floor. The free skill check will show whether your CV already surfaces those terms or buries them under job duties.

Leaving the shop floor without disowning it

I've maintained SPC charts and OEE reports for two machining lines for six years, so I'm used to separating real signals from noise and acting on them. What the charts couldn't do was warn me before a bad shift. I rebuilt our OEE calculation in Python to check I could trust the tools, then built a model to predict downtime for the next shift from things we know before it starts, split by time so nothing leaked. It beat a "same as last shift" baseline only modestly, and I can explain why. I want to do this work full-time, ideally close to manufacturing.

Replace the baseline sentence with whatever your own project actually showed; a modest honest result survives follow-up questions, an inflated one does not.

Write back when your OEE numbers match your Excel sheet. That is the first milestone, and it is closer than you think.