Synthetic data-driven predictive maintenance for computed tomography scan machine using neural network
Year: 2025
Type: dissertation
Abstract: The absence of abnormal data in predictive maintenance (PdM) for CT scan machines poses significant challenges, particularly in identifying crucial condition indicators, estimating health indices, and modeling degradation patterns necessary for predicting machine failures and remaining useful life. These limitations hinder the development of robust PdM models capable of effectively detecting anomalies and predicting failures in critical medical equipment. To address these challenges, this study aims to apply the Mahalanobis Distance (MD) method to analyze the real IoT sensor data collected from CT scan machines for detecting anomalies. A critical MD threshold value of 3.5485, corresponding to a 95% confidence level, was used to identify abnormal data points. Following the anomaly detection, synthetic abnormal data is generated using two methods which are Noise Addition (NA) and Tabular Generative Adversarial Networks (TGAN). The study aims to assess the reliability of the generated synthetic data and evaluate the performance of machine learning (ML) models in classifying anomalies for PdM in CT scan machines. An Artificial Neural Network (ANN) was employed to classify the normal and abnormal including real and synthetic data. The results showed that the NA method achieved a classification accuracy of 98.64%, while the TGAN method demonstrated a higher accuracy of 99.53%. Additionally, TGAN-generated data closely resembled real-world anomalies, resulting in superior performance metrics such as 100% for precision, recall, and F1- scores, with fewer misclassifications. In conclusion, TGAN proved to be a more reliable approach for generating synthetic data and improving PdM model performance, particularly in situations where real abnormal data is limited. These findings highlight the importance of advanced data generation techniques like TGAN in enhancing the accuracy, reliability, and robustness of predictive maintenance systems for CT scan machines.
Institution: Universiti Putra Malaysia
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7168113629
