⚡ electrical engineering

A Data-Driven Decision-Support Framework for Predictive Quality Management in FMCG Manufacturing

This study proposes and validates a data-driven decision-support framework that integrates Six Sigma DMAIC with machine learning, time-series forecasting, FMEA, and SPC to enable predictive quality management and reduce defects in high-volume FMCG biscuit manufacturing.

Original authors: Silviana Hakim, Galuh Zuhria Kautzar, Andy Hardianto, Arie Restu Wardhani, Agung Dwi Cahyo

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Silviana Hakim, Galuh Zuhria Kautzar, Andy Hardianto, Arie Restu Wardhani, Agung Dwi Cahyo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive bakery making millions of biscuits every day. In this high-speed world, even a tiny mistake—like a cookie being slightly burnt or a package being sealed poorly—can ruin thousands of products, costing the company a fortune and making customers unhappy.

This paper is about a team of researchers who helped one of these big biscuit factories (called "PT XYZ") stop guessing why their biscuits were going wrong and start predicting it before it happened. They built a "crystal ball" for quality control, but instead of magic, they used math and computers.

Here is how they did it, broken down into simple steps:

1. The Old Way: Looking in the Rearview Mirror

Traditionally, factories used a method called Six Sigma. Think of this like a detective who only arrives after the crime has been committed. They would count how many broken biscuits they found at the end of the day, figure out what went wrong, and try to fix it for tomorrow.

The Problem: By the time the detective figures it out, thousands of bad biscuits have already been made. It's reactive, not proactive.

2. The New Way: The "Smart Co-Pilot"

The researchers created a new system that acts like a smart co-pilot for the factory. This system combines the old detective work (Six Sigma) with modern Machine Learning (computer programs that learn from patterns).

They followed a five-step roadmap (DMAIC) to build this co-pilot:

Step 1: Define (What are we looking for?)
They asked, "What makes a perfect biscuit?" They identified the "Critical-to-Quality" things: the shape, the color, the packaging, and the machine stability. It's like a chef deciding that the crunch and the crust are the most important parts of the cookie.
Step 2: Measure (How bad is it right now?)
They looked at the factory's data from 2024. They found that the factory was making a lot of mistakes. Using a "Pareto Chart" (which is like a "voting system" for problems), they discovered that 90% of all the bad biscuits came from just two sources:
1. Packaging issues (53%): The wrappers were failing.
2. Process issues (38%): The baking or mixing wasn't quite right.
  Analogy: It's like realizing that 9 out of 10 car accidents in a city happen at just two specific intersections. You don't need to fix the whole city; you just need to fix those two spots.
Step 3: Analyze & Predict (The "Crystal Ball")
This is where the magic happened. They fed the computer data about the oven temperature, the speed of the conveyor belt, and the humidity.
- The Test: They tried different math models to see which one could best guess how many bad biscuits would be made next.
- The Winner: A model called Support Vector Regression (SVR) won.
- Why it matters: The relationship between the oven temperature and a burnt cookie isn't a straight line; it's a curve. Simple math can't see the curve, but SVR can. It's like the difference between a child guessing "it's getting hotter" and a meteorologist predicting exactly when a storm will hit based on complex wind patterns.
Step 4: Forecast (Looking ahead)
They also wanted to know: "Will the number of bad biscuits go up or down next week?"
- They tested three "time-travel" models. The winner was Prophet (a forecasting tool).
- The Result: Prophet was the best at predicting short-term trends, even when the factory was chaotic. It's like a weather app that tells you, "It's going to rain heavily in 2 hours," so you can grab an umbrella before you get wet.
Step 5: Control (The Alarm System)
Finally, they combined their predictions with a Risk Priority System (FMEA).
- They ranked the problems by danger. The biggest danger? Conveyor belts being misaligned and sealing machines failing.
- They set up an alarm system (Statistical Process Control) that doesn't just wait for a broken biscuit to appear. Instead, it watches the "vital signs" (like temperature) and screams, "Hey, the temperature is drifting! Fix it now before we make bad cookies!"

The Big Takeaway

The paper proves that you don't have to choose between "old-school quality control" and "new-school AI." You can mix them.

By using this Data-Driven Decision-Support Framework, the factory can:

Stop guessing: They know exactly which machines are causing the most trouble.
Act early: They can fix a machine before it starts making bad products, rather than cleaning up the mess afterward.
Save money: Less waste means more profit.

In short: The researchers turned a factory that was constantly playing "catch-up" with broken biscuits into one that can see the future and fix problems before they even happen. They didn't just find the broken cookies; they built a system that prevents the cookies from breaking in the first place.

Technical Summary: A Data-Driven Decision-Support Framework for Predictive Quality Management in FMCG Manufacturing

Problem Statement
In the Fast-Moving Consumer Goods (FMCG) sector, particularly in high-volume biscuit manufacturing, maintaining consistent product quality is hindered by complex process interactions, high throughput, and operational disturbances. While Six Sigma's DMAIC (Define–Measure–Analyze–Improve–Control) methodology is widely adopted for reducing defects, its traditional application is predominantly retrospective and reactive. It relies on historical data analysis and post-production evaluation, limiting the ability of managers to anticipate emerging quality issues in real-time. Consequently, there is a gap in integrating predictive analytics holistically across the entire DMAIC cycle to support forward-looking decision-making in dynamic production environments.

Methodology
This study proposes and validates a data-driven decision-support framework that integrates Six Sigma DMAIC with machine learning (ML), time-series forecasting, Failure Mode and Effects Analysis (FMEA), and Statistical Process Control (SPC). The research was conducted as an industrial case study at PT XYZ, a high-volume biscuit manufacturer in Indonesia, focusing on "Line 6," which exhibited recurring quality losses.

The methodology followed five analytical stages:

Define & Measure: Critical-to-Quality (CTQ) parameters (e.g., product structure, packaging integrity) were identified via SIPOC mapping. Baseline process capability was quantified using Defects Per Million Opportunities (DPMO) and Sigma levels. Pareto analysis was used to identify dominant defect categories.
Predictive Modeling (Regression): To model defect occurrence as a function of process variables (e.g., oven temperature, conveyor speed, humidity), three regression models were evaluated: Multiple Linear Regression, Ridge Regression, and Support Vector Regression (SVR). Data preprocessing involved handling missing values and excluding temporal identifiers to prevent leakage. Hyperparameter tuning was performed using GridSearchCV, with performance assessed via Mean Squared Error (MSE) and $R^2$ .
Time-Series Forecasting: To project short-term defect trends, daily production data were resampled and interpolated. Three forecasting models—ARIMA, Prophet, and Exponential Smoothing—were tested. Model selection was based on Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).
Analyze & Improve (Risk Prioritization): Root causes were diagnosed using Fishbone diagrams and ML-derived feature importance. These findings were translated into prioritized actions using FMEA to calculate Risk Priority Numbers (RPNs).
Control: The framework integrated SPC charts (X̄–R) with ML-based predictive outputs to enable early detection of deviations before they exceeded conventional control limits.

Key Results

Defect Distribution: Analysis of 2024 production data (372 observations, ~26.4 million kg output) revealed that packaging defects (53.6%) and process-related defects (37.89%) collectively accounted for over 91% of total quality losses.
Regression Performance: Among the regression models, Support Vector Regression (SVR) demonstrated superior predictive performance compared to linear models. This indicates that the relationship between key process variables and defect occurrence is inherently non-linear.
Forecasting Performance: In time-series forecasting, the Prophet model achieved the lowest error rates (RMSE: 3.12; MAPE: 2.45), outperforming ARIMA and Exponential Smoothing in capturing trend changes and irregular fluctuations within the dynamic production environment.
Risk Prioritization: FMEA identified conveyor misalignment (RPN 280) and packaging seal failures (RPN 270) as the highest-risk failure modes, providing a structured basis for resource allocation.

Significance and Contributions
The paper claims three primary contributions to management science and Quality 4.0 literature:

Methodological Extension: It extends Six Sigma from a reactive improvement methodology to a predictive decision-support system by embedding analytical models (regression, forecasting) directly within the DMAIC structure.
Operationalization of Predictive Analytics: It demonstrates how ML and forecasting can be practically operationalized to support managerial decision-making in high-volume manufacturing, moving beyond isolated defect classification to a unified framework.
Scalable Framework: It provides empirical evidence of a scalable approach that integrates prediction, forecasting, and risk-based prioritization (FMEA) with continuous monitoring (SPC).

Limitations and Future Scope
The authors modestly acknowledge that the study is based on historical, aggregated data from a single facility over one year, which may limit generalizability. The models currently lack real-time sensor integration for autonomous closed-loop control. Future research is suggested to extend the framework to multiple sites, integrate real-time IoT data for continuous model updating, and evaluate long-term economic impacts. The study does not claim to have solved all quality issues but rather offers a structured mechanism for earlier identification of defect risks and more timely intervention.

1. The Old Way: Looking in the Rearview Mirror

2. The New Way: The "Smart Co-Pilot"

The Big Takeaway

Technical Summary: A Data-Driven Decision-Support Framework for Predictive Quality Management in FMCG Manufacturing

More like this