Simulation-based test design using Variational Auto-Encoder (VAE)

4347 단어 딥러닝DeepLearning

Introduction



This post is based on the original publication [1].

배경



Design of experiments (DoE) gives designers a guidance to define an adequate set of tests. Defining tests to evaluate a system comprehensively is an important step in a system design. As a complimentary to DoE, we would like to show a way to enumerate similarity of system's responses during these tests and restructure tests with respect to the similarity.

Idea



As a first step, we run a lot of simulations and monitor some selected signals. This becomes training data.

Data is fed to a variational auto-encoder (VAE) to learn both reduced representation of signals and reconstruction of inputs from the reduced representation.

When VAE is trained successfully, we have following tools for our test design:
  • Encoder to convert high dimensional inputs to low dimensional representation
    If you feed two signals to Encoder, they will become two points in the low dimensional space. The distance between these two points gives a similarity of two signals by means of features that VAE learned.
  • Cost function that gives a high value when a give signal is not alike training data set.
    By feeding any new signal to Encoder, the cost function tells if the new signal is alike training signals or not.

  • With these two tools, we can try to search for test cases which are not covered well in the original tests, and will be able to improve test quality by adding these tests and eliminate redundant tests.

    Success of this process does depend on the coverage of the original tests. This is in a way alike our learning process where we depend on our experiences as a clue to understand new concepts.

    Example



    1. Simple Hydraulic arm with three points to be monitored



    Three hydraulic pressure will be monitored during tests.



    2. Test patterns



    Three types of test patterns are defined simply with respect to the arm move.



    Test outputs look like:


    3. VAE network training



    A VAE network was trained on these hydraulic pressure signals.



    4. Encoding test outputs



    After training the VAE, all tests defined at 2 are encoded to four dimensional latent variables. Visual representation of the encoded signals are given below (3 components of 4-D data is shown):



    Purple Green, and Yellow color represent points corresponding encoded outputs from Periodic, Down, and UP test patterns.

    5. Test reduction



    Test cases result in similar latent variables may also be redundant since we would like to find tests stimulate systems to result in wide variety of states. If we set a threshold value of distance and eliminate similar tests, it result in a reduced test set. In the figure below, points are reduced if the distance in the 4-D latent space is less than 0.5.



    References:



    [1] Yasunori Yokojima and Toshihiko Nakazawa, Applying deep learning to test design process in hydraulic systems design 심층 학습을 응용한 유압 시스템 설계의 테스트 설계 프로세스. Journal of the Japan Fluid Power System Society 49, 71-74, 2018.

    좋은 웹페이지 즐겨찾기