基于深度学习的闪烁探测器信号故障识别研究

传统核探测器故障信号诊断研究都需要提前提取信号特征,然后用机器学习、支持向量机、统计方法等对特征进行分类。为了实现对探测器输出信号进行实时识别和故障诊断,本文基于Matlab平台构建了一个用于对图像进行分类的卷积神经网络模型,对核探测器故障信号进行分类诊断。从分类准确率和算法运行时间两个方面对Adam、Sgdm、Rmsprop三种优化算法进行了比较。结果表明Rmsprop算法运行时间最少,但准确度和损失的训练迭代曲线不平稳;Sgdm模型对十组非正常信号图像分类的准确率最高为93.10%,准确度和损失的训练迭代曲线平稳。虽然,本文方法诊断准确率略低于文献报道值,但是不需要对信号进行预处理和特征预提取,使用更为简便。

Research on Fault Identification of Signal for Scintillation Detector Based on Deep Learning

Traditional research on fault signal diagnosis of nuclear detector needs to extract signal features in advance, and then use machine learning, support vector machine, statistical methods to classify the features. In order to realize real-time identification and fault diagnosis of the output signals of the nuclear detector, this paper constructs a convolution neural network model for image classification based on Matlab platform to classify and diagnose nuclear detector fault signals. Three optimization algorithms Adam, Sgdm, and Rmsprop are compared in terms of classification accuracy and algorithm running time. The results show that the running time of Rmsprop algorithm is the least, but the training iteration curve of accuracy and loss is not stable; The highest accuracy of Sgdm model in classifying ten groups of abnormal signal images is 93.10%, and the training iteration curve of accuracy and loss is stable. Although the diagnostic accuracy of this method is slightly lower than the value reported in the literature, it does not require signal preprocessing and feature extraction, so it is easier to use.