bg

Transform Your Supply Chain Planning and Marketing Strategies with Google Cloud and SAP Integration

Cloud Marketplace

Predictive Maintenance with Machine Learning: A Lixil Case Study

Introduction

Lixil is a renowned market leader in the plumbing fixtures industry, generating annual sales of approximately US$5.0 billion. With well-known brands such as INAX, Grohe, and American Standard, Lixil caters to customers worldwide. The company operates several manufacturing facilities globally, with porcelain being a crucial raw material in its production process. However, porcelain is a delicate material that requires careful handling as it can only withstand a specific temperature before cracking. As Lixil expands its operations, the management realized the need to address manufacturing waste and explore data-driven production methods to minimize waste. To find a solution, Lixil Mexico partnered with Pluto7.

Business Challenge

Lixil faced a manufacturing problem where 8-10% of their products resulted in defects, costing them millions of dollars in losses. The company hypothesized that the high temperature in the kiln was causing the defects, but they were unsure how to prevent it. The company tried several approaches to resolve the issue but was unsuccessful.

Pluto7’s Data Science Approach 

Pluto7 analyzed sample data for 12 months and discovered that temperature alone was not the driving factor behind the defects. Instead, the raw materials (Slip Preparation – resultant of a mixture of several raw materials) were the critical factor affecting the porcelain’s quality. Using machine learning, Pluto7 helped Lixil Mexico identify the ideal temperature bands for each of the twenty-two zones within the kiln and provided a framework for building applications that could provide dynamic recommendations for setting temperature bands and avoiding defects.

Creating a Centralized Data Foundation for Lixil Mexico’s Manufacturing Process

At Pluto7, we recognized the need to establish a centralized data foundation to gather and store sample data for process, temperature, and moisture. With the help of our data science team, we established a data foundation that would allow us to collect and analyze data over a period of 12 months. By gathering this data, we were able to gain insights into Lixil’s manufacturing process and identify areas that needed improvement. 

Manufacturing Process

We followed a comprehensive set of steps to ensure accurate analysis and optimal results. 

  1. Exploratory Data Analysis: helps understand the data pattern and identify any trends or anomalies in the data.
  2. Pre-processing: involves removing duplicates, null values and cleansing the data to prepare it for analysis.
  3. Data Transformation: involves joining multiple tables and restructuring the data to create a unified view for analysis.
  4. Feature Engineering: involves creating new features and preparing the features for modeling by scaling, normalizing, or encoding them appropriately.
  5. Modeling: involves experimenting with different algorithms and selecting the best one that provides the highest accuracy and best performance.
  6. Hyperparameter Tuning: involves optimizing the model performance by fine-tuning the hyperparameters and selecting the best values for them.
  7. Validation: involves validating the outcomes of the model and ensuring that the model provides accurate and reliable results.

Exploring the Relationship between Temperature and Defects

Experiment 1: Defects per day for a low-defect month

We conducted exploratory data analysis to uncover the relationship between temperature and defects. In one of our experiments, we focused on identifying defects per day for a low-defect month. Our findings revealed that temperature fluctuations do not necessarily correspond with defect spikes. 

The area within the red line is the Tolerance band. We can observe that Temperature fluctuations do not directly affect the defect spikes on the upper graph 

low-defect

Experiment 2: Defects per day for a high-defect month

We analyzed a high-defect month with a focus on a single defect category, namely “crack on body.” Our observations revealed that although there were higher fluctuations in temperature compared to the low-defect month, there was no clear correlation between these fluctuations and demand spikes.

ML Models Experiments 

Experiment 1: K-Means Clustering 

We experimented with K-Means clustering to identify different clusters of temperature readings in the observed kiln zones, specifically focusing on low and high defect clusters. We found that the cluster combinations for high and low defects were identifiable and that the minimum and maximum temperatures for each cluster were close to the actual values. However, we also found that outliers had a significant influence on the results. As a result, we decided to try other machine learning algorithms to see if we could obtain more accurate results in predicting the correlation between temperature and defects.

Experiment 2: Gaussian Mixture Clustering 

Through experimentation with Gaussian models, we discovered that Gaussian mixture clustering was effective in identifying the minimum and maximum temperature ranges that would result in fewer defects in our manufacturing process. 

Unlike K-Means clustering, the Gaussian mixture results were not affected by temperature outliers, and the algorithm identified temperature ranges that were even closer to the current minimum and maximum.

 In fact, the outcomes fell within the Lixil-approved temperature ranges currently used, demonstrating the algorithm’s effectiveness. This promising approach provides a better-performing and leaner temperature range, making it an attractive option for Lixil to improve its manufacturing process

low risk

The comparison of the above two clusters clearly indicates that the temperature ranges for low defect clusters are narrower than those for high defect clusters. Additionally, the low defect cluster has minimum and maximum temperatures that are closer to the current temperature ranges being used.

Experimentation with Moisture Data 

We conducted an experiment to determine if a combination of temperature and moisture data was causing defects in our manufacturing process. Using the ‘Bowl’ product, we filtered the existing kiln datasets to include only rows with no defect and those with the ‘Crack on body’ defect. 

We then carried out a hypothesis testing to check for correlations between temperature, moisture, and defects for this dataset. Promising results were obtained, indicating a positive correlation between moisture data and defects. 

To further explore this correlation, we employed various defect algorithms, including Logistic Regression, Random Forest, XGBoost, and CatBoost. The Random Forest algorithm provided balanced results, accurately predicting the formation of defects with 88% accuracy. These findings suggest that moisture data may be contributing to defects in our manufacturing process and provide a pathway for further improvements.

table

Random Forest Results showing a positive correlation with Moisture Data 

Advanced Analytics with Machine Learning 

With data centralization, we now had the capability to run advanced analytics and get answers to questions that would have otherwise taken days to compile and generate. Some of the insights generated through our data-driven approach are highlighted below:

What are the most common defect zones? 

Through our analysis, we were able to identify six key zones that were contributing to the defects in our manufacturing process. This helped us narrow the search further, enabling us to take targeted corrective measures.

Which are the most prevalent defects? 

With over 40 prevalent defects identified, we were able to pinpoint two key defects that were most commonly occurring in our manufacturing process. This allowed us to focus our efforts on minimizing the occurrence of these defects.

Which kiln has the optimum production-to-defect ratio? 

By analyzing the data, we identified a particular kiln responsible for 40% of the overall production. However, it was also contributing around 55% to manufacturing defects, making it an area of concern for Lixil. The blue bar indicating the share of product manufacturing for the kiln was overshadowed by other bars indicating a higher defect rate. 

By identifying this kiln, we were able to prioritize corrective measures, ensuring that our production processes were optimized and defects were minimized.

Conclusion 

In conclusion, the exploratory data analysis that we conducted for Lixil Mexico has highlighted the immense benefits of data centralization. By bringing together various datasets and analyzing them, we were able to identify key areas of focus for further investigation. With this centralization in place, we can now conduct rapid experiments that were previously impossible and leverage advanced data science techniques to optimize manufacturing processes. Over the next few months, we will continue to analyze data, synthesize it, and build scalable data models on the Google Cloud Platform. By doing so, we believe that Lixil Mexico will be able to unlock new insights, optimize its manufacturing processes, and drive business growth.

For more information
www.pluto7.com/success-stories

Enable Decision Intelligence Into Every Corner Of Your Product And Operations.

industry image

Industry Manufacturing

Organization Name: Lixil

Challenges

  • Lixil’s manufacturing plants produce finished products with a defect rate of 8-10%, with the top 3 defects being Pre-heating, Cracks on the body, and Glaze Boil.
  • Lixil suspected that the Kiln stage, specifically the heating and cooling process, is a contributing factor to these defects.
  • They aimed to investigate the correlation between temperature, moisture content, and other parameters with the occurrence of defects.

Results

  • Established a centralized Data Foundation to gather and store sample data for Process, Temperature, and Moisture over a period of 12 months.
  • Conducted an Exploratory Data Analysis (EDA) to investigate the correlation between temperature and defects, specifically at the Kiln stage.
  • Developed a Machine Learning (ML) model capable of predicting the likelihood of defects based on a given combination of parameters, specifically temperature, and moisture.

Products Used

  • BigQuery
  • Vertex AI