A Bayesian algorithm is introduced into the training of Dynet, and applied it to the problem of modelling damage formation. This has enabled reliable estimates of the uncertainty of the predictions to be made; an aspect from many neural network approaches. The hyperparameters relating to the noise level and weight decay have been optimised through an evidence framework. Synthetic data for damage evolution has been generated by simulating its dependence on temperature, strain and strain rate. The performance of Dynet in capturing the causal relationship has been evaluated as a function of uncertainty in the measurement of the target output as well as the amount of the target information. Even when only the final damage levels were used in the model it was able to predict not only the final damage state but also its evolution through the deformation trajectories. The model was found to perform well even when the training measurements were uncertain to within 20%. Under such conditions, its predictions of the true values of unseen data were more accurate than if they had actually been measured. The uncertainty in the predictions were found to be a fair measure of its predictive capability and its performance on unseen data only marginally worse than for training data indicative of good generality and the absence of overfitting. The advantage of the algorithm is that the complexity of Dynet can be controlled to a certain degree. With the optimal hyperparameters, the possibility of overfitting is reduced. In comparison with the results of fixed hyperparameters, the algorithm is trying to learn the law governing the mechanics of damage formation rather than measurement noise. In addition, this algorithm can give not only the expected value of the target, but also the uncertainty of the prediction, i.e. the errorbar. This study has demonstrated the capability of Dynet to model evolutionary processes, in this case damage evolution during forging, but the model could find utility across a wide range of processes from the natural to the social sciences. It should be pointed out that the values of the external inputs, temperature, strain and strain rate are set within a specific region. In real damage process, the situation could be more involved, and therefore it might be harder to achieve good performance with limited experimental data. One possible approach might be to modify the structure of Dynet by introducing a physically sensible part. The present work provides a good stalling point towards this target. Finally, the recently developed kernel-based learning technique has achieved good success in the statistical modelling community. The application of this technique to model damage formation and the testing of these methods with experimental data constitute a future task.


    Zugriff

    Zugriff über TIB

    Verfügbarkeit in meiner Bibliothek prüfen

    Bestellung bei Subito €


    Exportieren, teilen und zitieren



    Titel :

    An evaluation of recurrent neural network modelling for the prediction of damage evolution during forming


    Beteiligte:
    Xiong, Y.S. (Autor:in) / Withers, P.J. (Autor:in)

    Erschienen in:

    Erscheinungsdatum :

    2005


    Format / Umfang :

    12 Seiten, 11 Bilder, 1 Tabelle, 19 Quellen




    Medientyp :

    Aufsatz (Zeitschrift)


    Format :

    Print


    Sprache :

    Englisch




    COMMERCIAL VEHICLES ROTOR CRACKING PREDICTION USING RECURRENT NEURAL NETWORK

    SUBRAMANIAN CHIDAMBARAM | Europäisches Patentamt | 2022

    Freier Zugriff

    COMMERCIAL VEHICLES ROTOR CRACKING PREDICTION USING RECURRENT NEURAL NETWORK

    SUBRAMANIAN CHIDAMBARAM | Europäisches Patentamt | 2023

    Freier Zugriff

    COMMERCIAL VEHICLES ROTOR CRACKING PREDICTION USING RECURRENT NEURAL NETWORK

    SUBRAMANIAN CHIDAMBARAM | Europäisches Patentamt | 2023

    Freier Zugriff

    Optimized Graph Convolution Recurrent Neural Network for Traffic Prediction

    Guo, Kan / Hu, Yongli / Qian, Zhen et al. | IEEE | 2021


    Vehicle motion prediction control method based on recurrent neural network

    WU XIALAI / LIN LING / YI JIAN et al. | Europäisches Patentamt | 2023

    Freier Zugriff