Design of a conversational assistant for standardized issue description in manufacturing

Abstract

This thesis describes the design and development of a conversational assistant (CA) for standardized issue descriptions in manufacturing. The CA was specifically targeted at the acquisition of data on standardized issue descriptions and at the same time reduce operator friction with the system. CAs can be used to handle and process repetitive data, such as gathering data at scale. For Diversey specifically, the conversational assistant can gather data on issues and in the future on root causes, solutions and best practices. At Diversey, agile manufacturing is applied to the production process. Due to the complexity of this approach, the production process has many stoppages. Context analysis at Diversey showed that operators primarily resort to their own intuition and experience to resolve stoppages. Evaluations with operators showed that the system for data acquisition on stoppages currently in place, is an inconvenience and a time sink. The data that was captured with the system was often ambiguous, incomplete and non-descriptive. However, such data can be of significant value to the company as it can be used for process improvement of the production line. A Wizard-of-Oz-like experiment was conducted with the operators where the researcher role-played a CA for acquiring issue description, which resulted in 71 issue capturing dialogues. Even though these dialogues were filled with implicit and explicit knowledge, without proper understanding of the context, the dialogues were too chaotic and unstructured for a machine learning (ML) algorithm to understand and process. For a complete issue description, that a ML algorithm can process, several key entities were identified: machine location, machine component, machine component state, product component and product component state. Furthermore, several challenges were found related to the speech patterns of the operators. Two of these were tackled in the prototype: Usage of synonyms and pronouns. With the open-source framework Rasa, a prototype was developed for a CA that would capture product related issues through a form structure into a database. The form re-quires the following entities: machine location, product component and product component state. The form is used to guide the user to capture good, structured data, while al-lowing some flexibility in the input. With the synonym and pronoun handling features, some of this flexibility is realized.
A conversational flow test was conducted to test and improve the CA prototype. Although failing the accuracy requirements set for intents by 7.4% and for entities by 2%, it provided an indication that with some information on the functionality of the CA, the participants were able to correctly capture the set issue. With some training of the operators, it would be feasible to implement such a CA in a manufacturing environment. Through contextual filtering, the model can filter towards the input of the user, allowing for only context-specific issue to be captured. With a qualitative test it was concluded that this feature decreased the duration of issue capturing. The contextual filtering and issue description control provide feedback to the user whether correct information is captured.