首页>
外国专利>
INDUSTRIAL GEARBOX FAILURE DIAGNOSIS APPARATUS AND METHOD USING CONVOLUTIONAL NEURAL NETWORK BASED ON ADAPTIVE TIME-FREQUENCY REPRESENTATION
INDUSTRIAL GEARBOX FAILURE DIAGNOSIS APPARATUS AND METHOD USING CONVOLUTIONAL NEURAL NETWORK BASED ON ADAPTIVE TIME-FREQUENCY REPRESENTATION
展开▼
机译:基于自适应时频表示的卷积神经网络工业齿轮箱故障诊断装置和方法
展开▼
页面导航
摘要
著录项
相似文献
摘要
The present invention relates to an industrial gearbox failure diagnosis apparatus using an adaptive time-frequency representation-based convolutional neural network, wherein a multi-channel multi-scale convolution filter is convolutionally multiplied by time series data having the health state of the gearbox as a label. a data processing unit for learning an information derivation model from which time-frequency information is derived and adaptive time-frequency information that is time-frequency information that is optimized through learning from the plurality of time-frequency information is derived; a characteristic factor extracting unit that receives the adaptive time-frequency information and learns an extraction model from which the characteristic factors are extracted through a convolutional neural network from the adaptive time-frequency information; a calculation unit for classifying the health state from the characteristic factor by a calculation model, calculating a loss function, and correcting the weights of the data processing unit and the characteristic factor extraction unit; And time series data without a health state label is input, and the data processing unit and the characteristic factor extraction unit are re-learned a predetermined number of times based on the corrected weight, and the final information derivation model and the final extraction model and the and a diagnostic unit for diagnosing a health state by a calculation model, wherein the adaptive time-frequency information is provided as an average of a plurality of time-frequency information. According to the present invention, by using deep learning to autonomously extract the characteristic factors required for fault diagnosis from data, there is an effect that can be universally applied to fault diagnosis of various systems by reducing the dependence on diagnosis-related expertise. In addition, according to the present invention, there is an effect of improving the performance of fault diagnosis compared to the existing method by indicating more abundant fault-related information than the existing TFR through an adaptive time-frequency expression using a learnable multi-scale convolution filter as a basis function. have.
展开▼