Driver distraction results.

Recently, distributed systems have become the backbone of technological development. It serves as the foundation for new trends technologies such as blockchain, the internet of things and others. A distributed system provides fault tolerance and decentralization, where a fault in any component does not result in a whole system failure. In addition, deep learning model enables processing data to find patterns, which helps in classification, regression, prediction, and clustering. This work employs deep learning to handle faults within distributed systems in three scenarios. Firstly, a faulty processor may not be able to produce the right output. Therefore, deep learning model uses the inputs and outputs of other processors to find patterns and produces the proper output of the faulty processor. Secondly, if a faulty possessor corrupts its inputs as well, then the deep learning model learns from the inputs and the outputs of successful processors and produces the proper output of the faulty processor, even with corrupted inputs. Thirdly, for unrelated data, in which the patterns of the input of the faulty processors differ from the patterns of the inputs of successful ones. In this case, the model is able to discover the new pattern and to be labeled as unknown. In the experiments, we use deep learning models like VGG16, VGG19, AlexNet LSTM and ResNet34, to investigate the performance of the deep learning in the three mentioned scenarios. For unstructured datasets, the accuracy of the models is affected by the size of the faulty data. The accuracy of all models lies between 60% when the size of the faulty data is 90%, and 96%, when the size of the faulty data is 90%. The structured datasets are not significantly affected by the portion of the faulty data and the accuracy reaches 99%.

近年来，分布式系统（distributed system）已成为技术发展的核心支柱，为区块链（blockchain）、物联网（internet of things）等新兴技术筑牢了发展根基。分布式系统具备容错（fault tolerance）与去中心化（decentralization）特性，任一组件发生故障均不会引发整个系统的瘫痪。此外，深度学习模型（deep learning model）可通过数据处理挖掘潜在模式，助力分类（classification）、回归（regression）、预测（prediction）与聚类（clustering）等任务的开展。本研究采用深度学习方法处理分布式系统中的三类故障场景：其一，故障处理器（processor）无法生成正确输出，此时深度学习模型可借助正常处理器的输入与输出数据挖掘模式，进而生成故障处理器的正确输出；其二，若故障处理器同时损毁自身的输入数据，深度学习模型仍可通过正常处理器的输入与输出数据进行学习，即便输入已遭破坏，仍能生成故障处理器的正确输出；其三，针对无关数据场景——即故障处理器的输入模式与正常处理器的输入模式存在显著差异——该模型可识别出新的模式并将其标记为未知类别。实验环节中，我们采用VGG16、VGG19、AlexNet、LSTM以及ResNet34等深度学习模型，对上述三类场景下深度学习方法的性能展开探究。对于非结构化数据集（unstructured datasets），模型的准确率受故障数据规模的影响：当故障数据规模为90%时，所有模型的准确率介于60%至96%之间。而结构化数据集（structured datasets）的准确率受故障数据占比的影响较小，准确率可达99%。