Loading...
Publication
Modelos preditivos da manutenção de endoscópios
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 8.77 MB | Adobe PDF |
Advisor(s)
Abstract(s)
A endoscopia consiste num procedimento médico efetuado com um instrumento designado
"endoscópio", o qual permite a observação do interior de um órgão oco ou de uma cavidade do
corpo. A ausência de um sistema centralizado e automático de gestão de tarefas pode resultar
em falhas na sua programação, levando à omissão de manutenções e à realização de reparações
tardias. Tal aumenta o risco de falhas nos equipamentos médicos e influencia a segurança dos
pacientes.
A tese/dissertação desenvolveu-se em parceria com a empresa Fujifilm, para o
desenvolvimento de algoritmos de machine learning (ML), para auxiliar na gestão da
manutenção de endoscópios. O objetivo é identificar padrões que permitam determinar o tipo
de intervenção seguinte (inspeção ou reparação), prever a quantidade de intervenções para os
anos futuros, de 2025 a 2030; prever o tempo médio até à intervenção futura e otimizar a
gestão do inventário de peças de reposição com base em dados históricos.
O estudo abrange o conjunto de dados de todos os equipamentos de clientes de Portugal
submetidos a reparação ou inspeção no Centro de Serviços da Fujifilm entre 2020 e 2024. Foram
aplicadas técnicas de pré-processamento de dados, análise estatística e métodos preditivos,
incluindo regressão linear, regressão polinomial de segundo grau, modelos Random Forest e
redes neuronais.
Os resultados demonstraram que o modelo de regressão linear aplicado à previsão do número
de reparações anuais entre 2025 e 2030 obteve um coeficiente de determinação (R²) de 0,9722,
indicando que cerca de 97 % da variação nas reparações pode ser explicada pelo ano. O modelo
de regressão polinomial de segundo grau apresentou um desempenho ligeiramente superior,
com um R² ajustado de 0,976.
Na previsão do tempo entre reparações, os modelos de regressão linear apresentaram
desempenho mais limitado, com R² ajustado de 0,1242 e, com validação cruzada, de
aproximadamente 0,11. O modelo Random Forest obteve um erro médio absoluto (MAE) entre
122 e 123 dias, e erro quadrático médio (RMSE) entre 158 e 159 dias, o que equivale a uma
margem de erro de cerca de cinco meses.
Relativamente à previsão do tipo de manutenção, o modelo Random Forest alcançou uma
exatidão entre 77,37 % e 78,10 %, demonstrando uma boa capacidade de classificação. O
modelo de redes neuronais também apresentou um desempenho positivo, com erros médios
quadráticos (RMSE) de aproximadamente 0,58 tanto nos conjuntos de treino como de teste, o
que indica ausência de sobreajuste (overfitting).
Endoscopy is a medical procedure performed with an instrument called an “endoscope,” which allows observation of the interior of a hollow organ or body cavity. The absence of a centralized and automatic task management system can result in scheduling failures, leading to missed maintenance and delayed repairs. This increases the risk of medical equipment failures and affects patient safety. The thesis/dissertation was developed in partnership with Fujifilm to develop machine learning (ML) algorithms to assist in the maintenance management of endoscopes. The goal is to identify patterns that allow determining the type of next intervention (inspection or repair), predict the number of interventions for future years, from 2025 to 2030; predict the average time until future intervention and optimize spare parts inventory management based on historical data. The study covers the dataset of all equipment submitted for repair or inspection at the Fujifilm Service Center between 2020 and 2024. Data preprocessing techniques, statistical analysis, and predictive methods were applied, including linear regression, second-degree polynomial regression, Random Forest models, and neural networks. The results showed that the linear regression model applied to predicting the number of annual repairs between 2025 and 2030 obtained a coefficient of determination (R²) of 0,9722, indicating that approximately 97 % of the variation in repairs can be explained by the year. The second-degree polynomial regression model performed slightly better, with an adjusted R² of 0,976. In forecasting weather between repairs, linear regression models showed more limited performance, with an adjusted R² of 0.1242 and, with cross-validation, approximately 0.11. The Random Forest model obtained better results, with a mean absolute error (MAE) between 122 and 123 days, and a root mean square error (RMSE) between 158 and 159 days, which is equivalent, on average, to a margin of error of about five months in predicting the time between interventions. Regarding the prediction of the type of maintenance, the Random Forest model achieved an accuracy between 77.37 % and 78.10 %, demonstrating good classification ability. The Random Forest model obtained a mean absolute error (MAE) between 122 and 123 days and a root mean square error (RMSE) between 158 and 159 days, which is equivalent to a margin of error of about five months.
Endoscopy is a medical procedure performed with an instrument called an “endoscope,” which allows observation of the interior of a hollow organ or body cavity. The absence of a centralized and automatic task management system can result in scheduling failures, leading to missed maintenance and delayed repairs. This increases the risk of medical equipment failures and affects patient safety. The thesis/dissertation was developed in partnership with Fujifilm to develop machine learning (ML) algorithms to assist in the maintenance management of endoscopes. The goal is to identify patterns that allow determining the type of next intervention (inspection or repair), predict the number of interventions for future years, from 2025 to 2030; predict the average time until future intervention and optimize spare parts inventory management based on historical data. The study covers the dataset of all equipment submitted for repair or inspection at the Fujifilm Service Center between 2020 and 2024. Data preprocessing techniques, statistical analysis, and predictive methods were applied, including linear regression, second-degree polynomial regression, Random Forest models, and neural networks. The results showed that the linear regression model applied to predicting the number of annual repairs between 2025 and 2030 obtained a coefficient of determination (R²) of 0,9722, indicating that approximately 97 % of the variation in repairs can be explained by the year. The second-degree polynomial regression model performed slightly better, with an adjusted R² of 0,976. In forecasting weather between repairs, linear regression models showed more limited performance, with an adjusted R² of 0.1242 and, with cross-validation, approximately 0.11. The Random Forest model obtained better results, with a mean absolute error (MAE) between 122 and 123 days, and a root mean square error (RMSE) between 158 and 159 days, which is equivalent, on average, to a margin of error of about five months in predicting the time between interventions. Regarding the prediction of the type of maintenance, the Random Forest model achieved an accuracy between 77.37 % and 78.10 %, demonstrating good classification ability. The Random Forest model obtained a mean absolute error (MAE) between 122 and 123 days and a root mean square error (RMSE) between 158 and 159 days, which is equivalent to a margin of error of about five months.
Description
Keywords
Endoscope Maintenance Machine Learning Medical Equipment Management Endoscópio Manutenção Machine Learning Gestão de Equipamentos Médicos