18 July 2022
Deployment of machine learning and artificial intelligence solutions in the manufacturing environment
Artifi cial Intelligence (AI) is expected to contribute up to $ 15.7 trillion to the global economy by 2030. Applied to the production context, AI and Machine Learning (ML) as a subdomain have the potential to improve manufacturing through multiple different applications. However, studies and surveys from practice show that a high percentage - more than 95 % - of ML and AI projects fail. An alternative source indicates that 87 % of data science projects never make it into production. One major reason for this is that many solutions are never being deployed. There are multiple last-mile delivery problems occurring during the deployment of AI and ML. Enterprises are discovering that it is easier to build AI than it is to integrate it into existing processes. In case of unsuccessful deployments, companies waste a substantial amount of resources as working hours of data scientists. Nonetheless, in order to make use of a developed ML model it must be deployed to production. The goal of deployment is understood as making a resulting model available in a specific environment in order to make the results usable where they are needed. This very step of deploying ML models into the running production process proves to be enormously difficult due to numerous technological but also organizational challenges. From a technological point of view, a main challenge consists in the constantly changing landscape of software tools for data and AI applications. Organizational barriers for deployment include the coordination between different roles and missing support by business leaders. The goal of the study "Deployment of ML and AI solutions in the manufacturing environment" was to create a guideline for a successful deployment of ML solutions in order to generate added value in production. In this report, a guideline was developed which supports companies during the complex deployment process of ML models. The guideline presents the relevant decisions and steps for bringing an ML model into production. Broken down into five components, the guideline covers all aspects of deployment from collaboration between different roles identified during the study to practical methods and software tools. The relevance and usefulness of the contents of the five components was demonstrated by an exemplary deployment based on an existing ML model which was trained on a publicly available data set from collaboration between different roles identified during the study to the monitoring and retraining of models. The ML model predicting the quality of products in the manufacturing was wrapped into a web application to let the end user interact with the model and retrieve the predictions, monitor its behavior and retrain the model if necessary.