本文關鍵詞：集成參數自適應調整及隱含層降噪的深層RBM算法　出處：《自動化學報》2017年05期 　論文類型：期刊論文

  更多相關文章： 限制玻爾茲曼機 特征提取 降噪 齒輪箱

【摘要】：深度置信網絡是由若干層無監(jiān)督的限制玻爾茲曼機(Restricted Boltzmann machines,RBM)和一層有監(jiān)督的反饋神經網絡組成的深層結構,該結構通過對低層輸入的逐層抽象轉化提取復雜輸入及復雜分類數據的有效信息.然而,深度置信網絡模型存在隱含層數及特征維數難以確定,后向有監(jiān)督過程存在"導數消亡"問題,使得低層結構參數得不到有效的訓練,而且噪聲干擾直接影響識別結果的問題.針對以上問題,提出以下解決方法:每個隱含層位置構建當前層輸出與樣本標簽之間的映射轉換矩陣,根據理論標簽與實際標簽之間的差異,實現隱含層特征維數的自適應調整,緩解"導數消亡"問題,同時在第一隱含層位置進行特征空間降噪,保證計算效率及提高診斷模型的識別效果.復雜工況的齒輪箱故障模擬實驗,驗證所提方法的有效性.
[Abstract]:Deep belief network is composed of several layers of unsupervised restricted Boltzmann machine (Restricted Boltzmann machines, RBM) and a layer of supervised feedback neural network composed of deep structure, the structure of the low level input layer transform to extract information from complex input and complex data classification. However, there are hidden layers and features it is difficult to determine the depth dimension of belief network model, to the supervised process "derivative die", the lower level of structure parameters to the lack of effective training, but the noise directly affects the recognition results. To solve the above problems, put forward the following solutions: building the current position of each hidden layer mapping between the output layer and sample label the conversion matrix, based on the difference between the theoretical and actual label label, to achieve adaptive hidden layer feature dimension, alleviate the problem, the same number of guide die " In the first hidden layer, we denoise the feature space to ensure the computation efficiency and improve the recognition effect of the diagnosis model. The gearbox fault simulation experiment under complex working conditions proves the effectiveness of the proposed method.

【作者單位】： 廈門理工學院機械與汽車工程學院;
【基金】：國家自然科學基金(51605406,51475170,51605405,51405272) 廈門理工學院科研啟動項目(YKJ14042R) 福建省自然科學基金青年基金(2014J05065) 廣東高校青年創(chuàng)新人才項目(2014KQNCX176)資助~~
【分類號】：TH132.41;TP183
【正文快照】： 引用格式張紹輝.集成參數自適應調整及隱含層降噪的深層RBM算法.自動化學報,2017,43(5):855-865齒輪箱是旋轉機械系統(tǒng)的重要組成部件,其運行狀態(tài)的好壞直接影響到相應設備的工作狀況,因此,國內外學者從機理、信號分析等方面對齒輪箱部件的故障診斷方法展開研究.然而,實際齒輪，

本文編號：1419895