Manufacturing focus: ‘Smart Twin’ of the Factory with factory data and AI models
In this solution, Databricks is the platform we use to build the data pipelines, train the models on existing data, and deploy the models for inference using MLflow experiment tracking.
The data foundation is built on Unity Catalog, which is connected to Databricks. The prediction results are uploaded to a manufacturing execution system (MES).
The goal is to translate product data from the factory floor into concrete actions.
Every step of a product manufacturing cycle generates data and our job is to figure out which steps are significant indicators of what the product quality and performance will be.
We then use the data of those steps to deploy AI models that are built to deliver almost instantaneous predictions and recommendations about the final product.
Up to now, we have deployed several AI models:
- an auto-encoder plus regression model – to predict product characteristics based on data of the product’s specific shape;
- an ensemble model – to leverage existing predictions on the same shape with additional datapoints. This improved existing predictions by as much as 25 percent;
- an image analysis model – to classify defects. It already does so better than the naked eye.
At every significant step in the manufacturing process, these models are trained on the available data up to and including that step. Then a “check gate” is implemented on the prediction result. This allows us to track the product dynamically throughout its manufacturing cycle.
Correct, scrap, or continue
With the uploaded results available in the MES tool, the operators on the factory floor get direct feedback. Based on the results of the model at each gate, we’re able to make recommendations:
- perform one or more corrective actions to ensure the part meets the specified requirements;
- scrap the part altogether if it is expected not to perform at an acceptable level, even with corrective actions;
- continue as planned if the outcome meets expectations.
This approach shifts quality control – from only at the end of the manufacturing process to every step along the way. Each defective part discovered before final testing represents a predicted cost avoidance of $30,000. At the moment, we detect multiple at-risk parts per month, which implies substantial savings on material and labor costs for our customer.
Long-term tracking benefits
Another advantage of our method is the ability to track long-term performance and visualize it using the dashboarding tool Grafana. This makes it possible to monitor if any issues emerge in individual steps of the manufacturing process, and directly respond to it. For example, if a factory machine operates outside its optimal state, the dashboard quickly detects the issue, which prevents costly downtime. Alerts are then sent to the factory operators for further investigation.
After Manufacturing: Cradle-to-Grave and Inline Tuning
The product’s lifecycle doesn’t end when the product leaves the factory. For this reason we focus on two other areas as well. These terrains also profit from developing an AI layer on top of the data foundation.
After Manufacturing: Cradle-to-Grave and Inline Tuning
The product’s lifecycle doesn’t end when the product leaves the factory. For this reason we focus on two other areas as well. These terrains also profit from developing an AI layer on top of the data foundation.
We’re developing AI models to predict the performance and longevity of the product in the field. Our customer also wanted to understand how logistics, such as storing the product for a longer period of time or movement during transport, affect the product’s quality. By analyzing these factors, we help prevent "dead-on-arrival" (DoA) issues and trace failures back to manufacturing conditions.
Processing data that is collected while the product is in use in the field provides a better understanding of the product’s quality and performance. At the same time, it tells us if the product is still operating in optimal conditions. By modelling the product’s behavior, we aim to improve the longevity of the product when it’s in use.
Toward a Digital twin
The ultimate goal is to create a full digital twin of the product lifecycle in collaboration with our customer. To achieve this, we will integrate the AI models and data from the individual steps and elements. The twin will help to control the manufacturing process in (near) real time, catch failures by implementing early warning systems on potential process issues, and tighten product manufacturing specs by gaining insights into the impact of each of the manufacturing steps toward actual field performance.
