چكيده به لاتين
The advent of additive manufacturing (AM) has brought about a paradigm shift in industrial production, enabling the creation of intricate designs and complex geometries. However, the crit- ical challenge of ensuring consistent high-quality product properties has limited AM’s adoption in vital applications. As a response to this challenge, this doctoral thesis embarks on an innovative journey to enhance the effectiveness of AM by employing state-of-the-art methodologies for intel- ligent process monitoring and predictive modeling, with a specific focus on the fused deposition modeling (FDM) technique. The research commences by exploring the potential of long short-term memory (LSTM) net- works as a robust tool for real-time FDM process monitoring. Leveraging sophisticated pre-processing methods encompassing handcrafted feature extraction and image-based representation, the pro- posed hybrid LSTM models attain unparalleled precision and adaptability. Notably, these image- based LSTM models distinguish themselves by achieving a remarkable mean accuracy of 99.85%. Building upon these promising outcomes, the pursuit of excellence extends into the second section, which delves into signal-to-image encoding and deep learning fusion models for multi-sensor data fusion (MSDF) in the realm of process monitoring. Through the implementation of feature-level fu- sion, centered around recurrence plot (RP) anomaly images, exceptional accuracies of up to 99.6% are attained, illustrating remarkable resilience against signal anomalies and surpassing conventional methodologies. This seamless progression from deep learning models to feature-level fusion rein- forces our commitment to advancing the effectiveness of Additive Manufacturing, harnessing the power of image-based representations for precision and adaptability. The thesis further delves into AM process parameter optimization using the Taguchi method. In predictive modeling, artificial neural networks (ANNs) emerge as the frontrunners, boasting corre- lation coefficients of up to 98%, thereby reinforcing AM’s reliability and transformative potential in critical applications. A data-driven predictive model is introduced for FDM, leveraging thermo- graphic and vibration data to achieve an impressive correlation coefficient of approximately 99%, facilitating real-time predictions of specific mechanical properties. The study culminates with an exploration of unsupervised machine learning for enhancing AM by intelligently clustering printing states in the FDM framework. This investigation includes a comparison of manual feature extraction with image-based representation for processing raw sensor data. Utilizing supervised and unsupervised feature selection and reduction techniques, it efficiently clusters printing states. evaluation metrics, including the F1-like score, v-measure, and silhouette, consistently support the superiority of image-based representation. In conclusion, this integration of advanced techniques foresees a future characterized by precise, high-quality products, reaffirming AM’s standing as an innovation powerhouse within the industrial landscape.
