High Performance Machine Learning for High Pressure Engineering
A world-leading engineered surfaces company operates two manufacturing centers that supported 16 warehouses that stock their products in different markets across the globe. They stocked a large variety of product offerings catering to a very broad customer base. The large number of service products with varying demand patterns were competing within limited resource availabilities and strains its supply chain. Warehouse production requirements were not driven by any historical order patterns, rather based on ad-hoc orders.
- Large number of service products with varying demand patterns competing within limited resource availability
- Warehouse production requirements were not driven by any historical order patterns but based on ad-hoc orders.
- Wilsonart business information, sales & product data is not resilient & integrated to provide on-demand analysis that could drive informed business decisions.
- The large variety of product offerings coupled with their high service fulfilment mandate added to the complexity of optimizing their day-to-day manufacturing obligations.
- Without a look forward strategy, they had to simply react to current conditions resulting in non-data driven decisions that strained resource allocations and reduced ROI.
Amazon Web Services (AWS)
We implemented a demand forecasting model to use for strategic and planning purposes. The model was trained on a collection of historical demand data, aggregated at the SKU level to a weekly granularity, across all of their warehouses and covering February 2018 to April 2020. The below solution architecture was developed in AWS to allow access to model predictions on fresh data.
- Built an AWS architecture that would consume historical data using S3 for storage, AWS Glue for ETL and Lambda services for event driven processes.
- Created a Machine Learning model using Amazon Forecast & SageMaker and the DeepAR algorithm to predict demand values in the range of 6 weeks.
- Constructed an automated data pipeline with an event driven architecture to spin up Machine Learning environments using AWS Code Commit, Code Build and CloudFormation for batch processing any subset of the focused products.
- Created a consumable dashboard that would visualize results and inspire action.
- Aggregating historical sales data & business information from various sources on a common accessible AWS platform allowed them to be able to select key products that were fast moving and high value.
PREDICTif established an AWS Machine Learning infrastructure that the customer uses to enhance their operational excellence processes in order to drive maximum business value.
- A machine learning model built on historical data and refined by the DeepAR algorithm can now be used to predict demand allowing them look forward and be better prepared.
- The results can now be used to optimize their production even further based on other product features, labor constraints & economic value.
- The machine learning platform that was delivered with this project can be leveraged to perform additional data science projects in the future.