Browsing by Author "Sun, Chengyao"

Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    Explainable artificial intelligence for time series modelling and causal inference
    (Cranfield University, 2022-12) Sun, Chengyao; Guo, Weisi; Xing, Yang
    Due to the concern over the trust issue of black-box artificial intelligence (AI) in daily human life, “right to an explanation” requirement is hence proposed for any algorithm in Europe general data protection regulation (GDPR). However, the framework of explain- able AI (XAI) is still in its infancy. Thus, this thesis aims to propose a framework of explainable AI for time series (X-AI4TS) in neural modelling and causal inference. In order to achieve the trustworthy usage of algorithm, at the first stage, this thesis embed the priori knowledge in time series and break it down into neuron-based dynamic system with domain expertise for explainable modelling. Then secondly, this thesis ad- dress the explainabilty in causal inference steps of the neuron-based model — (i) Identifying the “hard to learn” features for neurons; (ii) Revealing and visualizing the inner operations within neurons; (iii) Inferring the original posteriori multimodal distribution via neurons. In addition, the aforementioned methodologies are formulated into an X-AI4TS frame- work and applied to a real business case in semiconductor fabrication for an integrated implementation. A data-driven neuron-based time-series forecasting digital twin is generated, under my proposed framework, for a dynamic manufacturing system. The results offer the diverse explainability to different users (i.e., end-users, data scientist and AI expert) and show the landing potential in dynamic engineering time series application.
  • Thumbnail Image
    ItemOpen Access
    Sampling and inference of networked dynamics using Log-Koopman nonlinear graph fourier transform
    (IEEE, 2020-10-21) Wei, Zhuangkun; Li, Bin; Sun, Chengyao; Guo, Weisi
    Monitoring the networked dynamics via the subset of nodes is essential for a variety of scientific and operational purposes. When there is a lack of an explicit model and networked signal space, traditional observability analysis and non-convex methods are insufficient. Current data-driven Koopman linearization, although derives a linear evolution model for selected vector-valued observable of original state-space, may result in a large sampling set due to: (i) the large size of polynomial based observables (O(N2) , N number of nodes in network), and (ii) not factoring in the nonlinear dependency betweenobservables. In this work, to achieve linear scaling (O(N) ) and a small set of sampling nodes, wepropose to combine a novel Log-Koopman operator and nonlinear Graph Fourier Transform (NL-GFT) scheme. First, the Log-Koopman operator is able to reduce the size of observables by transforming multiplicative poly-observable to logarithm summation. Second, anonlinear GFT concept and sampling theory are provided to exploit the nonlinear dependence of observables for observability analysis using Koopman evolution model. The results demonstrate that the proposed Log-Koopman NL-GFT scheme can (i) linearize unknownnonlinear dynamics using O(N) observables, and (ii) achieve lower number of sampling nodes, compared with the state-of-the art polynomial Koopman based observability analysis.