Jiong Tang’s research while affiliated with University of Connecticut and other places

In manufacturing, visual inspection of parts’ surface is traditionally an important examination before the parts can proceed to the next manufacturing step. For example, the timely detection of minor surface defects, such as dents and scratches, in small-sized airfoils of aircraft engine, is typically the final stage of quality assurance before acceptance for assemblage. While such a process is critically important, current practices rely heavily on human operator’s judgment, which is subjective and labor-intensive. In this study, we establish an automated, image-based inspection system that utilizes robotic automation to acquire high-resolution images of the parts under inspection and employs a specifically tailored machine learning technique to facilitate decision-making of inspection. Leveraging deep learning as the underlying methodology, we address a key challenge in the flexible automation of surface inspection, i.e., the scarcity of labeled data during the initial training process. In other words, we tackle the challenge of limited samples with known defects. Specifically, we synthesize an adaptive semi-supervised learning framework, building upon the residual neural network (ResNet) and the deep convolutional generative adversarial network (DCGAN) to extract features from both ground truth and synthetic data. This approach can overcome the shortcomings of the current approaches, leading to more objective and accurate defect detection right from the beginning of implementation with a small labeled dataset. Our results show that the overall classification accuracy on this challenging dataset reaches 92.30%, a 27.79% improvement over the baseline model achieved through optimal use of synthetic and ground truth data. The system also investigates the impact of synthetic data, providing guidelines for integrating it effectively into iterative training. This approach offers a robust solution for surface inspection and quality assurance in diverse manufacturing applications.

This study investigates the mechanical behavior of Ti64 alloy at high temperature (750°C) using digital twinning technology based on finite element analysis in combination with Gleeble testing. In the realm of high-temperature alloy materials, empirical constitutive models, such as the Johnson-Cook constitutive relation, have conventionally been utilized to delineate the mechanical behavior of metallic substances under conditions of elevated temperatures and strain rates. This research applies a modified Johnson-Cook model tailored specifically for characterizing the unique strength hardening mechanisms inherent to Ti64 alloy. Subsequently, computational simulation is conducted to implement the modified Johnson-Cook model within finite element analyses. Furthermore, to address the computational challenge inherent in finite element simulation based inverse analysis through optimization, a multi-response Gaussian process model is constructed as a surrogate model to replicate thermomechanical analyses, thereby facilitating efficient simulation. Finally, the parametric identification is cast into an optimization problem, that minimizes the difference between Gleeble testing results and the model prediction in the parametric space, to determine the model parameters of the modified Johnson-Cook model. This framework is successfully applied to the model updating of Ti64 alloy. The algorithms, including the custom user subroutines and the parameter identification process, provide valuable tools for exploring the constitutive relations of various alloys.

Milling operations are critical in the manufacturing sector, pivotal for achieving desired product specifications and surface quality. However, post-milling deformations, primarily induced by residual stress, can significantly affect the final flatness and surface integrity of workpieces, potentially deviating from manufacturing tolerances and standards. This issue is pronounced in the machining of A2024 aluminum, a material favored for its strength and lightweight properties but susceptible to such stress-related deformations. Finite element Method (FEM) can offer valuable insights but at the cost of extensive computational resources, making them less viable for real-time applications and rapid iteration across large parameter spaces. Addressing this challenge, our study introduces a spatially-aware predictive framework utilizing the physics-based Graph Neural Network (GNN). This framework leverages the spatial relationships between nodes on the milling surface and incorporates physics-guided insights from Jonson-Cook model to forecast post-milling surface flatness with crucial process inputs. A meticulously crafted FEM model drives our automated data generation pipeline, ensuring the creation of high-quality data. The fusion of FEM-derived data with specially tailored GNN modeling represents a paradigm shift in predictive machining results, enabling rapid and reliable flatness prediction to support adaptive manufacturing strategies. This framework is generic, and has the potential of being extended to other manufacturing processes.

Direct inverse analysis of faults in machinery systems such as gears using first principle is intrinsically difficult, owing to the multiple time- and length-scales involved in vibration modeling. As such, data-driven approaches have been the mainstream whereas supervised trainings are deemed effective. Nevertheless, existing techniques often fall short in their ability to generalize from discrete data labels to the continuous spectrum of possible faults which is further compounded by various uncertainties. This research proposes an interpretability-enhanced deep learning framework that incorporates Bayesian principles, effectively transforming convolutional neural networks into dynamic predictive models and significantly amplifying their generalizability with more accessible insights of the model's reasoning processes. Our approach is distinguished by a novel implementation of Bayesian inference, enabling the navigation of the probabilistic nuances of gear fault severities. By integrating variational inference into the deep learning architecture, we present a methodology that excels in leveraging limited data labels to reveal insights into both observed and unobserved fault conditions. This approach improves the model's capacity for uncertainty estimation and probabilistic generalization. Experimental validation on a lab-scale gear setup demonstrated the framework's superior performance, achieving nearly 100% accuracy in classifying known fault conditions, even in the presence of significant noise, and maintaining 96.15% accuracy when dealing with unseen fault severities. These results underscore the method's capability in discovering implicit relations between known and unseen faults, facilitating extended fault diagnosis, and effectively managing large degrees of measurement uncertainties.

This research aims to establish a systematic framework for parametric identification of materials undergoing high temperatures and high strain rates. While advanced testing equipment, such as the Gleeble physical simulator, can produce controlled measurements of specimens under various conditions, significant challenges remain in determining the parameters of constitutive relations. Temperature gradients inevitably arise during Gleeble testing, leading to nonuniform strain distribution caused by complex thermal–mechanical coupling. Although finite element analysis of Gleeble testing can be performed, such simulations are computationally expensive, making brute-force optimization to minimize the difference between experimental data and finite element simulation across the parametric space infeasible. Furthermore, since the related constitutive relations are semi-empirical in nature, the ground truth of the constitutive parameters is generally unknown. In this context, a single-objective optimization based on a number of testing conditions may yield biased results or become trapped in local minima. In this research, we employ finite element analysis simulating Gleeble operation as the foundation, leveraging a suite of computational intelligence tools to address these challenges. We first develop a multi-response Gaussian process surrogate model, trained using a relatively small amount of finite element data, to rapidly emulate the forward analysis. We then implement a multi-objective optimization approach using simulated annealing to individually minimize the differences between experimental results and emulations under various testing conditions. AISI 9310 steel and the Johnson–Cook model are adopted for methodological demonstration. The development of the finite element model, Gaussian process surrogate model, and inverse optimization is detailed, and the results obtained are discussed. This framework can be extended to the parametric identification of other materials and heat treatment conditions using Gleeble testing.

Microperforated panels (MPPs) display excellent capacity in noise control applications owing to their high strength, simple design, and efficacy in low-frequency sound absorption. Traditionally, the development of MPPs has relied on a trial-and-error design approach. Although simple optimization-based methods have recently begun to be employed, these designs often overlook practical considerations, such as the increased costs associated with adding more MPP layers, which presents a gap to achieve the practical feasibility of MPP deployment. To address this, the study aims to develop an inverse multi-objective optimization-empowered framework for MPP design to enhance low-frequency noise mitigation while minimizing fabrication costs. Specifically, a finite element (FE) model is established to conduct the acoustic analysis of MPPs, followed by thorough experimental validation. A novel multi-objective particle swarm optimization algorithm (MOPSO) is then developed to cope with mixed-type design variables with interrelations inherent to the MPP architecture. Using the high-fidelity FE model as a cornerstone, the MOPSO guides the inverse optimization analysis to yield multiple non-dominant solutions. These solutions not only avoid the trap of local optima, but also allow for continuous screening to ensure the engineering viability based on empirical judgment. The results clearly demonstrate the effectiveness of the proposed methodology. The MPPs designed in this study show great potential for mitigating indoor noise in buildings, addressing noise issues arising from rapid urbanization and transportation development. Furthermore, the novel optimization strategy proposed in this study holds wide applicability for other sound absorption materials.

I n the evolving landscape of Prognostics and Health Management (PHM) enhanced by the Internet of Things (IoT), diagnosing machinery system faults is critical for ensuring operational efficiency and safety across various industries. This research introduces a novel, interpretable deep learning architecture designed to overcome key limitations in existing fault detection methods, such as the high demand for extensive training data and the lack of transparency in feature extraction. Our model uniquely integrates dual branches: one processing raw time-series data through a spatially transformed convolutional neural network, and another incorporating wavelet transform coefficients. This dual-branch approach not only maximizes the effective use of limited data but also significantly enhances model interpretability, eliminating the need for extensive feature engineering and manual feature selection. The significance of this research lies in its innovative methodology, which bridges the gap between advanced deep learning techniques and practical applicability in industrial settings. By leveraging IoT sensors and real-time data processing, our model exemplifies a practical application of IoT in PHM. The proposed algorithm is rigorously evaluated on experimental gearbox data and further validated on a publicly available bearing dataset, demonstrating its generalizability and scalability. Through comprehensive parametric investigations, we elucidate the impact and robustness of the physics-integrated parallel architecture, showcasing its potential to significantly improve fault diagnosis accuracy in diverse operational conditions. This study not only advances the state-of-the-art in fault diagnosis but also provides a framework for developing more interpretable and efficient deep learning models for industrial applications.

With the growing interest in space exploration, the need for advanced technologies to support future missions has become increasingly crucial. Among these technologies, a highly reliable power system is needed to withstand extreme space environments and external disturbances. A cyber-physical testbed (CPT) emerges as a valuable platform for testing and validating extraterrestrial microgrids, enabling the assessment of their resilience and performance under realistic conditions. This article presents a CPT based on power hardware in the loop (PHIL). The testbed incorporates physical components such as solar arrays, pressure and temperature control systems, and electronic loads that mimic habitat loads, along with cyber components like nuclear generation, energy storage systems, and other space habitat loads. This testbed provides unique capability to thoroughly evaluate the behavior of the space microgrid under various load management scenarios, particularly in response to unexpected disturbances. By emulating realistic conditions, the testbed allows us to analyze the power system’s performance, assess its robustness, and identify areas for improvement in terms of power management, load prioritization, and fault mitigation within a space microgrid.

Efficient and accurate bearing fault diagnosis and prognostics are crucial for predictive maintenance in industrial settings. Deep learning approaches have significantly advanced the field of time series classification, bringing powerful models to the forefront of prognostic and health management. Despite these advancements, current methods, including conventional Transformer models, encounter specific limitations with computational efficiency and the effective processing of long sequences. Addressing these challenges, our research presents an optimized Transformer-based model tailored for surmounting these hurdles, enhancing efficiency and ensuring high performance. By eschewing the traditional decoder in favor of an encoder-only architecture, our model capitalizes on the encoder’s ability to distill contextual information from sensor data. Enhanced with random convolutional kernel transform (ROCKET) for its exceptional feature extraction in time sequences and refined through principal component analysis (PCA) for dimensionality reduction, the proposed approach significantly boosts model efficiency, markedly trimming training duration, while achieving an exemplary accuracy of 99.10% on the benchmark bearing dataset. The proven efficiency and outstanding performance of our model hold significant promise for real-time predictive maintenance, offering a powerful tool for prompt and accurate fault detection.