Data-driven Process Monitoring Methods in Intelligent Manufacturing Systems

智能製造系統中的數據驅動過程監測方法

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date30 Aug 2024

Abstract

The rapid development of advanced engineering technologies such as artificial intelligence (AI), big data and additive manufacturing (AM) has revolutionized the manufacturing industry with enhanced productivity, product quality and flexibility. However, to effectively monitor processes remains challenging in intelligent manufacturing systems due to the increasing process complexity and process dynamics with growing data volume, velocity and variety. In particular, latent dynamics, network sparsity, lack of process knowledge, and dynamic process settings pose significant challenges to efficacious process monitoring.

In this thesis, we include four main works completed during the degree of Doctor of Philosophy, each addresses different research challenge in the data-driven process monitoring of intelligent manufacturing systems. Through considering complex process dynamics, network characteristics, self-supervised image features and process information, we propose novel data-driven methodologies for dynamic soft sensor modeling, multi-relational network monitoring, image based product quality inspection, and acoustic signal based product geometry prediction.

The first study proposes supervised bidirectional long short-term memory (SBiLSTM) network address the challenge of complex nonlinear process dynamics in data-driven soft sensor modeling. The proposed SBiLSTM network incorporates extended quality information and utilizes a bidirectional architecture to accurately estimate quality variables in industrial processes. With this novel structure, the SBiLSTM can extract and utilize nonlinear dynamic latent information from both process variables and quality variables, which further improve the prediction performance significantly.

In the second study, an ordinal network model is developed to consider more accurate and efficient network representation. Based on the ordinal network model, a Likelihood-based Ordinal Network Monitoring (LONM) scheme is proposed. The proposed LONM considers dyadic interactions with ordinal levels and incorporates ordinal information into the model. The LONM scheme is highly sensitive to network anomalies and is robust to network sparsity. In addition, the matrix representation of the ordinal network model provides efficient computation and parameter estimation for the LONM.

The third study proposes a novel image feature-based self-supervised learning (IFSSL) model for effective and automatic quality inspection in additive manufacturing given absence of label. Through self-supervised learning, the advantages of supervised learning and unsupervised learning are leveraged by requiring no sample label while retaining defect-relevant information. The IFSSL model integrates feature-based image fusion and self-supervised learning to focus on defect-relevant regions in product images without requiring any label.

In the fourth study, we address the challenge of acoustic signal based geometry prediction in wire arc additive manufacturing (WAAM) process by proposing a Residual Self-Attention Bidirectional Long Short-Term Memory (Res-SA BiLSTM) network. The network utilizes multi-head self-attention, BiLSTM, and residual learning to effectively model nonlinear parameter interactions and learn dynamic latent features. To predict geometry of WAAM product with acoustic signal, Mel-frequency cepstral coefficients (MFCC) are extracted from the acoustic signal and combined with the dynamic process information for geometric parameter calculation. With the designed structure, feature extraction and fusion, the proposed Res-SA BiLSTM network is able to predict WAAM geometry accurately.

The studies presented in this thesis contribute to the field of intelligent manufacturing by proposing new data-driven methods to address process monitoring challenges. The developed methods provide methodological novelties in dynamic soft sensor modeling, network monitoring, image-based quality inspection and geometry prediction in additive manufacturing processes, as well as insights for predictive monitoring, quality control and decision making in intelligent manufacturing systems.

The effectiveness of the studies presented in this dissertation has been validated with real case studies of industrial processes, communication network and additive manufacturing processes.