Explore how a multinational consumer goods company leverages machine learning to achieve 4% reduction in soap manufacturing defects.
The Challenge
Our client is a prominent multinational consumer goods company headquartered in London, England, with operations spanning across food and homecare products.
The client operates a detergent bar manufacturing plant with a production capacity of 1 million soap bars per day. However, they faced two primary types of defects in their manufacturing process: soft bars and bar cracks. These defects collectively contributed to an overall defect rate of approximately 5%.
The client's primary objective is to reduce the defect rate in the manufacturing process. By doing so, they aim to enhance the overall product quality, optimise production efficiency, and minimise costs associated with rework and waste.
The Strategy
1. Identify operational parameters for defects
After processing operational data spanning four months, we identified approximately 1,500 batches, revealing 55 valid parameters.
2. Develop machine learning models
We developed machine learning models to implement clustering and identifying similar batches based on the operational parameters. The objective is to identify the top 18 parameters among the 55 that are responsible for causing defects.
3. Identify the parameter ranges for golden batches
Based on data generated from benchmarked batches, ideal range recommendations were provided for all the identified important features.
4. Develop and deploy a simulator application
A web application was developed to assist floor operators to make informed decisions regarding input parameter ranges and obtain predictions on bar softness and cracks. Regression techniques are employed to predict bar softness, while classification techniques are used to predict bar cracks within a batch.
The Execution
1. Understanding the manufacturing process and factor mapping
We initiated the project by conducting internal research on the soap manufacturing process, engaging with industry experts, and visiting the client's manufacturing plant to communicate with their on-site team. We also collaborated with SMEs to identify potential factors that could contribute to defects. Based on this, a factor map was developed, and hypotheses were formulated for testing.
2. Data ingestion
The client uploaded data from legacy systems onto a drive. However, as the data was not initially intended for analytics, it was scattered across numerous unstructured files. To address this, custom algorithms were developed to clean and consolidate the data. Around 1,500 batches were identified, and the analytical dataset (ADS) was constructed.
3. EDA and Golden batch analysis and recommendations
Once the data processing was complete, an exploratory data analysis (EDA) was performed to assess data sufficiency, accuracy, and eliminate unnecessary information. Clustering techniques were employed to identify similar batches within the dataset, which led to the formation of four distinct clusters. From these clusters, two were singled out: one characterised by low bar softness and the other by high bar softness.
Subsequently, we merged these clusters, and machine learning algorithms were employed to determine key factors contributing to bar softness. Upon further analysis, it was found that out of the 55 parameters, 18 were identified as highly influential in determining bar softness. By leveraging advanced statistical techniques, optimal parameter ranges were derived and shared with the plant team.
During EDA, we also discovered that implementing golden batch recommendations to minimise bar softness resulted in a reduction in bar cracks. To validate this hypothesis, a simulation analysis was conducted.
4. Creating a simulator
A web application was created. This application served as a platform for operators to input parameter values conveniently. In the backend, machine learning models were utilised to retrieve the predicted Penetration Value (PV) and the probability of encountering bar cracks. These models analysed the provided values, delivering precise predictions for PV, and evaluating the likelihood of bar cracks.
The Results
- Overall defect rate was reduced from 5% to 1%
- Successfully identified recommended parameter settings for the most significant features
- Simulations were conducted to demonstrate that by manipulating the top 18 important variables, approximately 75% of the undesirable PV values were eliminated
