Build DFMEA using Generative AI based on Product's Requirements and Field Returns for Handheld Devices

Failure Mode and Effect Analysis (FMEA) is a very powerful Design for Reliability (DfR) tool. It is an iterative process that influences the design by identifying potential failures, assessing their probability of occurrence and their effect on the product, isolating their root causes, and how to prevent them. FMEA should be initiated early in the program during the design phase. It can be applied to hardware and software design, or a process. As stated, it is an iterative and time-consuming process that is usually performed by a team of five to seven people throughout multiple sessions.The proposed work focuses on the integration of Generative Artificial Intelligence (GenAI) and FMEA development. The current FMEA building process faces significant challenges, including a labor-intensive process that consumes excessive time and resources, an inconsistent approach across teams leading to variability in outcomes, incomplete analysis that fails to comprehensively address potential risks, and a limited ability to accurately predict failure modes, ultimately compromising the effectiveness of failure mitigation strategies and product reliability.The goal of this work is to leverage GenAI to streamline the FMEA practice, and integrate the massive database of lessons learned and field returns into the process. This will ensure comprehensive risk analysis and enhance product reliability. The main objectives of this work include reducing the time required for FMEA development by 80% through automation and optimized workflows. Standardize FMEA methodology across all teams and improve failure mode prediction accuracy using AI by integrating historical field failures. Hence, uncover and prevent potential field failures from escaping to the field which lead to customers' satisfaction.We are employing an agentic generative AI approach to this process. Initially, the model inputs, comprising product requirements, usage environments, lessons learned, and field failure data, are defined and preprocessed. This is facilitated by the extraction agent, which identifies pertinent data points, and the chunking and vectorstore agents, which organize this data for Retrieval-Augmented Generation (RAG) and contextual analysis. Subsequently, a primary subsystem is selected to construct and validate the concept. Following this, the model is scaled to encompass all key subsystems that constitute the product. Our results indicate that GenAI can effectively generate a system hierarchy based on the system block diagram and subsequently draft a DFMEA, leveraging a substantial dataset of field returns, in a highly efficient manner. The writer agents are then employed to formalize the DFMEA process documentation. In this paper we shall showcase results obtained for multiple sub systems as part of a handheld product's development process.