# FMEA based prescriptive model for equipment repair guidance

## TL;DR

This paper presents a machine learning model using Failure Mode and Effects Analysis to guide equipment repair decisions, aiming to reduce downtime in manufacturing.

## Contribution

The novel contribution is a prescriptive model combining FMEA and machine learning for optimal repair action prediction.

## Key findings

- Two modeling approaches were evaluated for predicting repair actions from time-series machine data.
- Classification accuracy was influenced by time-series processing and model architecture choices.
- Effective strategies for prescriptive maintenance were identified to enhance production efficiency.

## Abstract

Accurate prediction of steps required to address machine faults is critical for minimizing downtime and enhancing production efficiency in modern manufacturing. This study utilizes machine failure data and Failure Mode and Effects Analysis to demonstrate how machine learning supports maintenance teams in selecting optimal repair methods.

The research adopts the Design Science Research paradigm, which emphasizes the creation of artifacts to address practical challenges. For the practical component, quality assurance and control frameworks in data science projects were implemented by integrating two widely used methodologies: CRISP-DM and PDCA, to ensure rigorous quality assurance and control in data science initiatives.

Repair actions serve as the target variables, while the input comprises ten multivariate time-series machine parameters. The prediction task is formulated as a classification problem. Two modeling approaches are evaluated. The first approach merges multiple time series into a single sequence, facilitating the application of Multi-Layer Perceptron, Convolutional Neural Networks, and Fully Convolutional Networks. The second approach preserves the time series as three-dimensional arrays, enabling advanced applications of MLP, CNN, Multi-Head CNN, and FCN models.

The models are assessed based on their capacity to predict repair actions, with particular emphasis on the impact of time-series processing and model architecture on classification accuracy. The findings highlight effective strategies for predicting machine repairs and advancing prescriptive maintenance in manufacturing environments.