MLOps with Snowflake for Scalable and Reproducible Model Training : A Case Study in Laser Welding Quality Control

Sammanfattning

This thesis presents the design and implementation of a cloud-based Machine Learning Operations (MLOps) platform, built on Snowflake, to support automated quality control in the laser welding process at Scania, a well-known manufacturer of heavy-duty industrial vehicles. The underlying Artificial Intelligence (AI) system utilizes a Siamese Deep Neural Network (DNN) to detect defective axles based on ultrasonic scan data. While the model architecture itself is well-established, the focus of this work lies in addressing key challenges in the model training workflow and its integration with the edge inference system. The existing setup, based on manual training and lacking standardization or version control, hinders reproducibility and scalability. To resolve these issues, a custom MLOps solution was developed, emphasizing three main objectives: (1) building a scalable cloud-based infrastructure for training DNNs, (2) ensuring end-to-end versioning and reproducibility of all Machine Learning (ML) assets, and (3) creating a reliable interface for deploying trained models to the factory floor for real-time inference. Initially, the system leveraged Snowflake Stored Procedures to encapsulate and execute training jobs, but was later re-architected using Snowpark Container Services to enable greater flexibility and runtime support. This dual-prototype approach reflects a practical evolution based on development constraints and learnings. The system was evaluated through both qualitative feedback from developers and data scientists, and quantitative metrics, including an 800% reduction in training time compared to the prior setup. While development costs were significant, the platform demonstrated low operational overhead and high usability, justifying its adoption. This thesis contributes not only to the successful deployment of the laser welding use case but also provides a reusable blueprint for integrating highly customized ML models within Snowflake-based MLOps workflows. The solution stands out for its innovative approach to managing complex model training pipelines and reinforces the importance of automation, reproducibility, and system integration in real-world AI deployments.