Manufacturing Quality Anomaly Detection Dashboard

Overview

Built and deployed an interactive manufacturing-quality dashboard covering a 90-day analysis window with 3,240 source quality rows, 19,635 capped defect-event records, three corrective-action records, and 90 daily aggregates across product, stage, line, equipment, station, lot, and shift dimensions.

The workflow combines an interpretable baseline-derived yield reference, trailing trend rules, defect-code and equipment concentration analysis, elevated retest-pass review, and unsupervised Isolation Forest ranking. Statistical signals explain why a day or entity deserves attention, while the model prioritizes multi-signal daily patterns for engineering review.

The analysis includes overlapping yield degradation, D042 concentration, and EQP_07 concentration with elevated retest-pass behavior, followed by a recovery phase. The workflow evaluates whether these signals are surfaced, but it does not determine root cause, establish real-world model accuracy, or prove causal corrective-action effects.

My contribution covered Python and NumPy data preparation, Polars transformation and feature engineering, statistical yield monitoring, defect and equipment concentration analysis, Isolation Forest ranking, corrective-action pre/post comparisons, ECharts dashboard engineering, automated tests, and Cloudflare Pages deployment.

Problem

Manufacturing-quality reviewers need to connect yield movement with defect concentration, equipment concentration, retest behavior, and action timing without relying on a single opaque anomaly score. Summary-only reporting can hide whether an excursion is broad, localized, transient, or accompanied by a possible screening-related signal.

Data Used

• Daily manufacturing-quality records covering 90 consecutive production days
• 3,240 source quality rows and 90 daily aggregates across product, process stage, line, equipment, station, lot, and shift dimensions
• 19,635 capped defect-event rows for defect-code and equipment concentration analysis
• Three corrective-action records across two action dates
• Overlapping yield, defect-concentration, equipment-concentration, and recovery periods
• Daily tested, failed, yield, retest-pass, defect-share, equipment-share, and rolling-trend signals

Review Scope

• The baseline and comparison windows are fixed so yield, concentration, anomaly-ranking, and corrective-action views use consistent review periods.
• The baseline-derived 3σ lower yield reference is an SPC-style monitoring aid, not a sample-size-adjusted p-chart.
• Isolation Forest is fitted on all 90 daily aggregates and used for prioritization; there is no held-out accuracy evaluation or production-data benchmark.
• Several features are correlated, including yield and fail rate, so the anomaly score is not independent evidence or a calibrated probability.
• Elevated retest-pass behavior is treated as a suspected false-fail signal for engineering review, not proof of false failure.
• Pre/post comparisons evaluate whether subsequent trends are consistent with improvement; they do not establish causal action effects.

Approach

Python and NumPy assemble the 90-day analysis dataset with 3,240 source quality rows, three overlapping abnormal-signal periods, and a subsequent recovery phase.

Polars transformations aggregate tested, failed, yield, retest-pass, defect-share, and equipment-share metrics into 90 daily rows and the comparison datasets used by the dashboard.

Interpretable statistical monitoring uses the first 45 daily aggregates, January 1 through February 14, as the yield baseline. The 98.7992% baseline mean and approximately 0.03829-percentage-point sample standard deviation produce a fixed 98.6843% lower reference at mean minus three standard deviations. No upper reference is calculated, and the limit is not adjusted for daily sample size, so it is an SPC-style baseline-derived 3σ lower yield reference rather than a formal binomial p-chart. A trailing seven-day mean requires all seven periods; the workflow flags daily yield below the lower reference or a trailing rolling-yield delta below -1.5 percentage points.

Concentration analysis compares capped defect-event rows and equipment failure-event rows between the January 1–February 15 baseline and February 16–March 7 comparison windows. It ranks D042 defect-event-share movement and EQP_07 failure-event-share movement, then adds a suspected false-fail review signal when retest-pass rate exceeds 50%. This evidence identifies patterns for investigation, not proven causes or false failures.

An unsupervised Isolation Forest with 200 trees, contamination 0.15, random_state 42, max_samples="auto", max_features=1.0, and bootstrap disabled is fitted across all 90 daily aggregates. Features, in model order, are yield rate, fail rate, retest-pass rate, trailing rolling-yield delta, top daily defect-code share, and top daily equipment failure share. The implementation applies no feature scaling or categorical encoding. The dashboard ranks the top 15 days by the negative score_samples value, where a higher displayed value means more anomalous; it does not present a validated anomaly classification. Yield and fail rate are exact complements, and several model features are correlated, so the ranking combines useful but non-independent signals.

Corrective-action review compares a 15-day pre-action window ending one day before each action with a 20-day post-action window beginning one day afterward. Three corrective-action records over two dates remain visible as context while yield, D042 share, and EQP_07 share are compared. The workflow does not attribute recovery causally to an individual action.

Investigation Focus

• The analysis dataset spans 90 days, 3,240 source quality rows, 19,635 capped defect-event rows, and 90 daily aggregates.
• The yield rules flagged all 20 days from February 16 through March 7, 2025.
• February 27 had the lowest daily yield at 95.62%, compared with the 98.68% baseline-derived lower reference.
• D042 defect-event share increased from 8.20% to 34.92%, a +26.72-percentage-point movement and the largest defect-code share increase.
• EQP_07 failure-event share increased from 12.92% to 39.25%, a +26.32-percentage-point suspicious concentration signal, not causal proof.
• The strongest daily suspected false-fail signal was a 69.30% retest-pass rate on February 27.
• Isolation Forest ranked February 27 first at 0.6968; all top 15 days fell inside the abnormal-pattern or recovery window.
• Three corrective-action records across two dates met the dashboard's pre/post improvement criteria; this does not establish causal effectiveness.

Key Investigation Choices

Methods & Tools

• Python 3.12
• Polars
• NumPy
• scikit-learn
• Isolation Forest
• JavaScript
• ECharts 5
• HTML/CSS
• pytest
• Ruff
• Cloudflare Pages

Result & Impact

• 90 days
  Quality data scale
• 20 days
  Yield-signal window
• +26.72 pp / +26.32 pp
  Concentration movement

The workflow flagged all 20 dates in the yield-signal window from February 16 through March 7, 2025. The lowest-yield date, February 27, reached 95.62% against a 98.68% baseline-derived lower reference and was also the highest-ranked Isolation Forest day with a score of 0.6968. D042 increased from 8.20% to 34.92% of capped defect-event records, a +26.72-percentage-point movement. EQP_07 increased from 12.92% to 39.25% of failure-event records, a +26.32-percentage-point movement. These results surface concentration signals for review; they do not establish root cause. All top 15 ranked days fell within the abnormal-pattern or recovery window. Three corrective-action records met the dashboard's improvement criteria across 15-day pre-action and 20-day post-action windows. Because the actions overlap and there is no control group, the comparison does not establish causal corrective-action effectiveness.

Notes

• Interpretable yield and concentration signals make an unsupervised ranking easier for engineers to review.
• Fitting across the full 90-day window produces an unsupervised prioritization workflow, not a baseline-trained novelty detector or validated classifier.
• Correlated features such as yield and fail rate can reinforce the same condition, so anomaly scores should not be treated as independent evidence or probability.
• Elevated retest-pass behavior can support suspected false-fail review but does not prove that the original failures were false.
• Pre/post recovery patterns support workflow review but do not establish causal corrective-action effectiveness.
• A reproducible data-preparation pipeline keeps calculations and ranking behavior stable across builds.