AI-DRIVEN PREDICTIVE MAINTENANCE FOR HIGH-VOLTAGE X-RAY CT TUBES: A MANUFACTURING PERSPECTIVE

Abstract

High-voltage X-ray computed tomography (CT) tubes are critical components in advanced medical imaging, industrial inspection, and non-destructive evaluation systems. These vacuum-based devices operate under extreme electrical, thermal, and mechanical stress, making them highly susceptible to gradual degradation and sudden failure. Unplanned downtime of CT tubes can result in significant operational disruptions, financial loss, and safety risks. Traditional maintenance strategies—such as reactive or preventive maintenance—often fall short in anticipating complex failure mechanisms, especially in high-throughput environments. This study addresses this gap by proposing and validating a predictive maintenance framework powered by artificial intelligence (AI) and designed specifically for high-voltage CT tube systems within industrial manufacturing contexts. The research adopted a hybrid experimental-computational methodology, combining simulated sensor data with real-world failure records to develop and evaluate machine learning and deep learning models. A dataset comprising 18,000 multivariate sensor sequences—including filament current, cathode temperature, vacuum pressure, and rotor vibration—was used to train five predictive models: random forest, support vector machine (SVM), convolutional neural network (CNN), long short-term memory (LSTM), and autoencoder-based anomaly detection. Feature extraction was performed using signal processing techniques, and model performance was assessed using accuracy, F1-score, remaining useful life (RUL) prediction error, and inference latency under real-time constraints. Additionally, a Raspberry Pi-based edge computing prototype was developed to validate real-time deployment feasibility, and a centralized monitoring dashboard was created to visualize health status and facilitate technician interaction. Results showed that LSTM models outperformed other algorithms in temporal degradation forecasting, achieving a ±5% error in RUL estimation and offering a 24–48 hour predictive lead time. Multisensor data fusion significantly improved detection accuracy and model stability across diverse operating scenarios. The autoencoder demonstrated exceptional performance in detecting novel and rare fault patterns without prior labeling, with a 96% detection rate and low false positive incidence. Edge deployment tests confirmed low-latency model inference suitable for real-time applications, while dashboard integration improved decision-making efficiency and technician trust in AI outputs. Overall, the proposed framework enabled proactive intervention, reduced maintenance overhead, and extended CT tube operational uptime. These findings highlight the strategic value of AI-enhanced predictive maintenance in optimizing industrial reliability, aligning with the broader goals of Industry 4.0 and smart manufacturing ecosystems.