Research

Intelligent Control

Intelligent control refers to a class of control methods that use artificial intelligence (AI) to compute control signals such as artificial neural network (ANN), machine learning (ML), genetic algorithm, fuzzy logic, and neuro-fuzzy (hybrid). Although existing control methods including PID, LQR, backstepping, H-infinity, sliding mode, and feedback linearization have showed great utility in engineering fields, their performance and implementation are often limited to simple or simplified systems (i.e., linearized and/or approximated). Furthermore, there are critical challenges of applying these methods to fault detection and identification, anomaly mitigation, adaptation, and gain-scheduling due to their significant reliance on existing model information. For this reason, intelligent control has garnered vast interest in robotics community. Nonetheless, simply replacing a controller by training a ML-based controller and predicting the control signal directly causes performance degradation due to poor generalization. To address these limitations, methods that combine ML to commonly used controllers are more desirable. 

Physics-guided System Modeling

Data-driven models, including ANNs have found widespread uses in various engineering applications due to their prominent capability and accuracy of representing complex systems. Compared to data-driven models, the foremost challenges associated with physics-based (PB) models are unknown system dynamics and parameter uncertainty, causing errors that are difficult to quantify or compensate. However, pure data-driven models in general suffer from poor extrapolability, and therefore, they require a tremendous amount of data with sufficient diversity to cover the full range of parameter space, which is often difficult to obtain. To address issues associated with both ANN and PB models, PB models are utilized in ANN training, viz., physics-guided neural network (PGNN). PB models are embedded in PGNN to introduce physical constraints to predict system states and guide ANN models to obey the physical laws, leading to more accurate prediction and extrapolation of actual system behavior. Hence, PGNN mitigates the requirement of training data in both the volume and the diversity compared to pure data-driven ANNs and dramatically enhances generalization ability. 

Data Analytics and Estimations

For many engineering applications, real-time data analytics and situation awareness assessment have become indispensable for system health monitoring and decision-making. Recently, AI has been used extensively for real-time data analytics due to their fast inference time and notable ability to extract key features and learn complex relationships in latent space. For instance, online anomaly detection of manufacturing system was developed that uses a deep autoencoder model to detect cyber and physical attacks. In addition, grease life of bearing was modeled to predicted using a novel deep learning structure that fuses acoustic and vibration sensor data. In order to enhance the prediction accuracy, convolutions filters of different sizes were implemented in parallel to feature-rich frequency bands of individual sensors.  Lastly, ML technique to predict computational fluid dynamics (CFD) solutions have been developed. CFD is a great tool for generating high-fidelity simulation data, however, its demanding computation time and cost precludes it from being used in many real-time and fieldable applications. 

Design Optimization

“Optimization” is perhaps the best word to describe the work of engineers. As a mechanical and aerospace engineer, it is essential to investigate a diversity of optimization problems in various fields including control, design, modeling, and navigation. Due to their intense computation cost, use of optimization in many applications is limited if not prohibitive. Optimization in this group mainly focuses on reducing its time and computation load. Surrogate models trained using different approaches such as Kriging, co-Kriging, polynomial regression, and ANN, can be employed in lieu of high-fidelity models without appreciably sacrificing the model accuracy. In addition, dramatically boosting computing parallelization subject to given computing resources is another important strategy for optimization acceleration. GPU-enabled heuristic optimization allows to run genetic algorithm, particle swarm optimization, and differential evolution while fully utilizing massive GPU threads in parallel. That is, a large number of candidate solutions are evaluated in parallel, therefore expanding the search space for each iteration and decreasing the total number of iterations to find the optimum. 

Research Projects