Anyone who runs a production facility for months collects data. Process parameters, test results, reject rates, machine running times, tool life, shift comparisons. This data is stored somewhere - in the MES, in the QMS, in the ERP, perhaps in Excel. It is there. But is it being used?
In most companies, the honest answer is: no. Or: rarely. Or: only when an acute problem arises. The data is collected, but not systematically analyzed. The result is a classic paradox: the answers to the most important improvement questions are available - but the questions are not asked.
This article shows you how to systematically evaluate historical process data: with specific analysis methods, a data quality check before the analysis, four documented improvement cases from practice and a clear five-step analysis path that also works without a data science team.
THE MOST IMPORTANT POINTS IN BRIEF
|
BRIEFLY SUMMARIZED
|
If you ask production managers and quality managers why historical data is not used systematically, the answer is rarely 'because we don't have any data'. The actual reasons are of a structural nature.
|
83 % manufacturing companies collect process data but do not evaluate it systematically Fraunhofer IPA Survey 2024 |
Ø 2,3 Systems per company in which relevant historical data is stored CSP project data 2024/25 |
< 5 % Manufacturing companies with defined analysis cycles for process history CSP benchmark data |
Ø 34 % Cpk improvement potential recognizable in data - but unused CSP analysis projects |
The four most common causes of unused historical data
Before an analysis method is selected, the data situation must be assessed. The following maturity matrix helps with this - and shows which analyses are possible at which level.
|
LEVEL 1 Minimum |
LEVEL 2 Structured |
LEVEL 3 Integrated |
LEVEL 4 Data-driven |
|---|---|---|---|
|
DATA SITUATION Paper documentation or Excel islands. Timestamp incomplete. Batches not clear. POSSIBLE ANALYSES Manual case analysis. No trend, no correlation. |
DATA SITUATION Digital test protocols. Time stamp available. Batches partially linked. Systems separated. POSSIBLE ANALYSES Trend evaluation per characteristic. Shift comparison. Cpk time series. |
DATA SITUATION MES, QMS and ERP linked via common key. Process parameters available for each serial number. POSSIBLE ANALYSES Correlation analysis. Root cause analysis. Predictive quality approaches. Multi-variate analysis. |
DATA SITUATION Complete, quality-assured process history. Real-time + history linked. Outliers eliminated. POSSIBLE ANALYSES ML-based pattern recognition. Process optimization. Predictive maintenance. Autonomous control. |
|
Next step Digitize key fields: Time stamp, serial number, batch as mandatory fields. |
Next step System link: Link test result with batch and process parameters. |
Next step Introduce automated evaluation routines and analysis cycles. |
Next step Introduce ML modelling and AI-supported process optimization. |
Historical data is only as good as its quality. Garbage in, garbage out applies in particular to statistical analyses: a false outlier in a time series can completely distort a trend analysis. The following check must be carried out before every analysis.
|
DQ dimension |
Check question before analysis |
✓ Good |
✗ Problem |
Immediate measure |
|---|---|---|---|---|
|
Completeness |
Are measured values missing in the time series? What percentage of fields are ZERO? |
< 5 % missing values |
> 10 % missing values |
Mark missing values, do not interpolate. Restrict analysis period. |
|
Time stamp |
Are all timestamps saved with time zone? Are there any gaps or duplicates? |
ISO 8601, no jumps |
Shift change gaps, duplicates |
Check NTP synchronization. Standardize timestamp format. |
|
Consistency |
Are the same batch or characteristic designations written in the same way everywhere? |
Consistent keys everywhere |
'CH031' vs. 'Charge-031' vs. '031' |
Define standardization rule. Clean up legacy data. |
|
Outliers |
Are there measured values that are physically impossible or clearly indicate incorrect entries? |
All values in the plausible range |
Values × 10 or 0.0 as default |
Limit value filter: Mark values outside ±5σ as candidates. Check manually. |
|
Referenceability |
Can measured values be clearly assigned to a serial number, shift or machine? |
Serial number as primary key |
Date and shift only, no SN |
Retroactive link where possible. Define key date for complete reference. |
|
Sample size |
Are there enough data points for statistically valid statements? (Minimum: 30 per group) |
≥ 50 data points per analysis characteristic |
< 20 data points per characteristic |
Extend analysis period or combine groups. Mark results as trend. |
Different questions require different analysis methods. The following five methods cover the majority of relevant issues in manufacturing process history - from simple to advanced.
| Method | Analysis method | Key question | Data basis | Tool / Method | Typical result | Analysis effort |
|---|---|---|---|---|---|---|
| 01 | Trend analysis (time series) | Does a process characteristic change systematically over time? | Time series of a quality or process parameter over at least 30 measuring points. Time stamp as X-axis. | Line graph + moving average (7 periods). SPC control chart with trend rules. | Cpk decreases from 1.5 to 1.1 over 8 weeks → Tool wear or material drift identifiable. | Low: 30-60 min. with existing data and Excel/SPC software |
| 02 | Shift/period comparison (ANOVA basis) | Do shifts, days of the week or teams differ significantly in their quality performance? | Test result or reject rate per shift, day of the week and period. At least 20 data points per group. | Box plot per group (shift A/B/C). Mean value comparison. F-test or non-parametric Kruskal-Wallis. | Early shift 0.8 % rejects, late shift 2.4 % - difference statistically significant. Cause: Enrollment gap. | Low-mean: 1-3 hours. Box plots in Excel or QMS software possible. |
| 03 | Correlation analysis | Does a quality characteristic depend on a process parameter - and to what extent? | Pairs of process parameters, e.g. temperature or torque, and quality characteristic, e.g. tensile strength or Cpk. At least 50 data points. | Scatter diagram + Pearson correlation coefficient r. For non-linear correlations: Spearman rank. | Correlation r = -0.74 between coolant temperature and scrap rate → Temperature control improves scrap by approx. 30 %. | Average: 2-4 hours. Requires linked database. |
| 04 | Pareto analysis of the causes of faults | Which 20 % of the causes of defects cause 80 % of the rejects? | Defect/reject categories with frequency from inspection log or 8D database. At least 50 defect cases. | Pareto diagram: bars sorted by frequency + cumulative curve. Draw 80% line. | 3 out of 12 defect types cause 76% of rejects → Focus of improvement measures clear. | Low: 1-2 hours. Standard in every QM software and Excel. |
| 05 | Multi-variate analysis (factor analysis) | Which combination of process parameters determines quality performance - and how do they interact? | Complete data sets with several process parameters and quality characteristics per unit. At least 100-200 data points. | Multiple linear regression. ANOVA with interaction terms. Principal component analysis (PCA) for dimension reduction. | Three parameters (pressure × temperature × speed) explain 84 % of the reject variance. Optimum window identified. | High: 1-3 days. Requires complete, linked database and statistical software such as Minitab, R or Python. |
The following improvement cases show how specific historical process data has led to measurable results - in different manufacturing sectors and with different analysis methods.
| Case | Problem & method | Data insight | Measure & effect |
|---|---|---|---|
| 01 | Screwdriving error on Monday morningShift/period comparisonof NOK rate by day of the week and shift. | Monday morning shift: 1.8 % NOK, rest of the week: 0.2 %. This was caused by the screwdriver calibration after the weekend downtime. | Self-test with reference screwing before the start of the shiftResult: 1.8 % → 0.25 % NOK after 6 weeks. |
| 02 | Cpk drift over tool lifeTrend analysisof Cpk value and number of tool shots. | Cpk decreased linearly from 1.6 to 0.9 within 12,000 shots. Critical drop from 10,000 shots. | Tool change interval reduced to 9,500 shots, SPC warning at Cpk < 1.2Result: Cpk Ø 1.15 → 1.51 after 12 weeks. |
| 03 | Weld seam scrap due to batch changeCorrelation analysisof scrap rate and material batch. | New batch with 15 % higher carbon content required adjusted welding parameters. | Supplier batch as mandatory field, parameter set per batch.result: 4.2 % → 0.6 % scrap after 4 weeks. |
| 04 | Top 3 defects drive 79 % scrapPareto analysisof the 12-month defect data. | Three error codes caused 79 % of total rejects. | Plant management system introduced for the top 3defectsResult: 1.8% → 0.41% total scrap after 16 weeks. |
Historical data is not documentation of the past - it is the mirror in which you really see your process for the first time.
-Amadeus Lederle CTE, CSP Intelligence GmbH
A structured analysis workflow makes the difference between 'we looked at some data' and 'we have a systematic improvement program'. The five steps below are designed for a monthly analysis cycle per production line - total effort: 4-8 hours.
| Step | Focus | What is happening? | Output |
|---|---|---|---|
| 1 | Define analysis question 30-60 min. | Check conspicuous KPIs such as scrap, Cpk, set-up time or downtime. Formulate a specific question, e.g. "Why is the NOK rate increasing on line 3?" Define analysis period, characteristics and owner. | Analysis question, data description, owner and schedule |
| 2 | Check data quality 30-60 min. | Export data set and check for completeness, timestamps, consistency, outliers, references and sample size. Mark missing values and outliers. | Cleaned, annotated data set with quality assessment |
| 3 | Perform analysis 1-3 hours | Select suitable method: Trend, comparison, correlation, Pareto or multi-variate. First visualize, then calculate key figures such as mean value, standard deviation, correlation or p-value. | Analysis graph, key figures and identified anomalies |
| 4 | Verify hypotheses 30-90 min. | Translate anomalies into hypotheses and check with further data or store floor observation. Exclude alternative causes and evaluate causality. | Verified or refuted hypotheses with reasons |
| 5 | Derive measure 30-60 min. | Derive a concrete, measurable measure from the hypothesis. Define KPI, person responsible and schedule. Check effectiveness in the next analysis cycle. | CAPA entry with measure, KPI, owner and date |
In practice, even with good data and clear questions, the same errors occur again and again in the analysis.
|
Error |
Why it occurs |
Consequence |
Solution |
|---|---|---|---|
|
Interpret correlation as causality |
Statistically significant correlation looks like a cause |
Measure addresses symptom, not cause |
Always ask: Is there a mechanical/physical explanation? |
|
Remove outliers instead of examining them |
Outliers 'disturb' the analysis picture |
The most interesting information is deleted |
Annotate and analyze outliers separately - they are often the root of problems |
|
Analysis period too short |
Last 2 weeks instead of 3 months - due to data availability |
Seasonal or periodic patterns not recognizable |
At least 3 months, better 6-12 months for stable trend statements |
|
Analyze many characteristics simultaneously |
'Look at everything' without a clear question |
No clear result, no prioritized measures |
One analysis question per cycle. Use Pareto: first top 3, then the rest |
|
No baseline before the measure |
Measure is introduced, no reference value documented beforehand |
Effectiveness of measure not measurable |
Before each measure: document current KPI value as baseline with date |
|
Analysis results not communicated |
Evaluation ends up in folder, not in team |
Improvement potential remains unused |
Monthly 15-minute slot: share analysis results in shopfloor meeting |
PRACTICAL TIP
CSP IPM saves all process parameters, test results and traceability data in an integrated database - with a serial number as a continuous key. This means that all analysis methods from this article can be used directly without having to merge data from different systems.
How many data points do I need as a minimum for a valid trend analysis?
For an initial trend statement: at least 20-30 data points in the time series. For statistically robust statements (e.g. trend significance test): 50-100 points. For Cpk trend analyses: at least 25 subgroups of 5 measurements each. With less data, the trend may be visually recognizable, but not statistically significant - you should communicate this explicitly: 'Trend decreasing, but not yet statistically significant'.
What is the difference between correlation and causality - and why is this important in the production context?
Correlation means: two variables develop together. Causality means: one variable causes the other. Practical example: scrap rate and temperature in the machine room correlate - but not because the temperature causes the scrap, but because high temperature is an indicator of summer, in which the coolant viscosity also decreases, which is the actual driver. Lowering the temperature in the room does not solve the problem. Monitoring the coolant viscosity solves it. Therefore, always ask for the mechanical explanation.
Can we use historical data from Excel for process analyses?
Yes - with restrictions. Excel data can be used if it: has complete and consistent time stamps, key fields (serial number, batch, characteristic) are consistently filled, does not contain any manual additions (or these are clearly marked) and sufficient data points are available. The typical problems with Excel data: inconsistent batch designations, missing time stamps for manual entry and editability (data may have been changed subsequently). Excel is sufficient for initial analyses, but a structured database makes more sense for a continuous analysis program.
How often should we systematically analyze historical data?
As a minimum, CSP recommends: monthly short analysis (1-2 hours, trend check per line), quarterly in-depth analysis (4-8 hours, correlations and period comparisons), annual overall evaluation (1-2 days, multi-variate, long-term trends, benchmark between lines). The monthly cycle is the most important - it ensures that drift trends are recognized before they become problems.
What is the best way to start if we have hardly analyzed any process history so far?
Step 1: Pareto analysis of rejects from the last 3-6 months. This analysis does not require integrated systems, just a list of scrap entries with defect code and date. The result - which 3 types of defects cause 80% of scrap - is immediately actionable and shows the team the concrete value of data analysis. This creates internal momentum for more structured analysis programs.
How do we deal with missing data in the time series?
Missing data points should not simply be interpolated - this creates false continuity. The sensible options: Mark missing values as 'NA' and exclude them from the analysis. If the proportion is < 10 %: Perform analysis anyway with reference to missing data. If > 10 %: Restrict period to the part with complete data. Analyze missing periods yourself: Why is data missing? Shift handover problem? System failure? This is often important information in itself.
Can I also use historical process data for predictive maintenance?
Yes - but only from level 3 of the data maturity model (integrated database, process parameters available for each serial number). Predictive maintenance requires: a complete time series of the relevant sensor values (vibration, temperature, current), documented maintenance and failure times as labels and at least 6-12 months of historical data per system. With these prerequisites, machine learning models can be trained to warn 2-4 weeks before a failure.