Learning Based Intelligent Modeling of Distributed Parameter Systems

基於學習的分佈式参數系統的智能建模

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date13 Aug 2019

Abstract

Distributed parameter systems (DPSs) are a common kind of industrial processes where the input and output may vary in both time and space dimension. Despite of the difficulty, modeling such complex systems is essential to industrial simulation, control and optimization. The family of time-space separation based spatiotemporal modeling methods has been verified to be effective in modeling of unknown DPSs, which has attracted wide interests from both academia and industry. The thesis is based on this spatiotemporal modeling framework, and we mainly investigate three new perspectives:

1. A reinforcement learning (RL) perspective for optimal sensor placement. Optimizing the sensor locations within a distributed process is challenging since most distributed processes are intrinsically nonlinear with infinite dimensions. The self-learning property from unknown environments makes RL a promising candidate for the optimization or control of real systems. We develop an integrated RL-based optimal sensor placement method for spatiotemporal modeling of DPSs. The sensor placement configuration is mathematically formulated as a Markov decision process (MDP) with specified elements, and the sensor locations are optimized through learning the optimal policies of the MDP according to the spatial objective function.

2. An incremental learning perspective for online modeling. Traditional spatiotemporal modeling methods are performed in batch-mode, assuming that all the output data is available and accessible at the beginning of the modeling process. Therefore, they are feasible for offline implementations only. We propose an incremental learning method that recursively updates the spatial basis functions and the temporal model. In this way, the time-space synthesis is inherited and updated efficiently as the streaming data increases over time. This incremental properties enables the spatiotemporal modeling methods to be applicable for online settings, while being computationally effective.

3. A multimode perspective for modeling complex distributed processes. For complex DPSs with strong nonlinearities and time-varying dynamics, the conventional spatiotemporal modeling methods become ill-suited since the elementary assumption that the process data follow a unimodal Gaussian distribution usually becomes invalid. We propose a multimode method is for modeling of such systems. The original operating space is partitioned into several subspaces by modified dissimilarity analysis. Each subspace represents the local spatiotemporal characteristics of the original system. An ensemble model is obtained using the soft weighting sum of the local models, where the corresponding weights are calculated by principal component regression. By properly decomposing the original space into several local parts, the ensemble model is capable of handling the strong nonlinearities and time-varying dynamics of the system.

We synthesize the above perspectives into the "learning based intelligent modeling" framework in this thesis. With these state-of-the-art learning techniques, we can provide more intelligent methods for modeling the challenging DPSs, such as a model-free sensor placement method for unknown systems, a feasible incremental method for online settings, or a more precise method for modeling complex systems. Thus, we provide an option for those interested in spatiotemporal modeling, but also who pursue the advantages of state-of-the-art learning techniques to solve various challenging modeling problems in real-world domains.

    Research areas

  • Distributed parameter systems, Machine learning, Intelligent agents (Computer software)