【正文】 metal disk, power is concentrated at higher and higher frequencies, usually reaching a highest frequency(here, approximately KHz with a spike at KHz) when the drillbit passes through the disk. Figuer 3 shows the same test are removal of all ponents with power Pf. This retains the significant peaks in the power spectral, whilst removing ponents assumed to be insignificant due to their low power. Figure 4 and Figure 5 show the corresponding timeseries data for Ax in test 19. After removal of lowpower frequency ponents, the timeseries retains only the episodes in which significantpower vibrations were observed, which are used as the basis for detection of system abnormality by several of the analysis methods used within this investigation. Figure 2Power spectra for Test 19 after removal of 50 Hz power supply contribution. The top plot shows a 3D “l(fā)andscape” plot of each spectrum. The bottom plot shows a “contour” plot of the same information, with increasing signal power shown as increasing colour from black to red. Figure 3Power spectral for Test 19 after removal of all spectral ponents beneath power threshold . Figure 4Az against time(in seconds) for Test 19, before removal of lowpower frequency ponents Figure 5Az against time(in seconds) for Test 19, after removal of lowpower frequency ponents. Analysis IVisualisation This section describes the first of four analysis techniques applied to the manufacturing process dataset. Visualisation of HighDimensional Data Constructing a 2D Visualization The use of large numbers of measured variables introduces problems in the visualization of the resulting data. A collection of temperatures, pressures, etc. forms a highdimensional representation of the state of a system, but this is not readily interpreted by an operator. Neuroscale allows the visualization of systems that have highdimensionality by mapping data to lower numbers of dimensions(typically two,for visual inspection). It attempts to preserve the interpattern distances in the highdimensional data. Data which are close together in highdimensional space are typically kept close together in 2D space, and data that are originally far apart remain well separated after projection. The projection is performed using a nonlinear function from the data’s k dimensional space down to 2D for visualization purposes. In this investigation, k is 5:[Ax, Ay, Az, AE, SP] are the highdimensional sample vectors. The creation of a nonlinear mapping from 5D space to 2D requires sample data from across the range of tests. In order to reduce the large number of available sample data to a quantity suitable for constructing the mapping, a summary of the dataset is required. Each test was summarized by a number of prototype 5D vectors using the kmeans clustering algorithm(in which a large number of data are represented by a smaller number of prototype vectors). The nonlinear mapping was trained using the prototype 5D vectors from all tests. Automatic Test Segmentation To allow the examination of the 5D data using visualization, it is convenient to divide the drilling process in to three stages, corresponding to the typical behaviour of the process described in Section . A heuristic algorithm was produced to perform automatic segmentation into three episodes using the SP channel, as illustrated in Figure 6(which shows a lowpass filtered version of SP superimposed on the original signal as a red line). The three states identified correspond to : State S1: the approximately constantpower (or slightly decreasing) initial period of drilling。 12 Appendix ANeuroScale Visualisations........................................ 錯誤 !未定義書簽。 10 Analysis IVNonelinear Prediction ............................................. 錯誤 !未定義書簽。 8 Analysis IISignature Analysis ....................................................... 錯誤 !未定義書簽。 Constructing Signatures ........................................................................... 錯誤 !未定義書簽。 Neural Networks for OnLine Prediction .............................................. 錯誤 !未定義書簽。 Table of Figures Figure 1 Test 90. From top to bottom: Ax, Ay, Az, AE, SP against time t(s) Figure 2 Power spectra for Test 19 after removal of 50Hz power supply contribution. The top plot shows a 3D “l(fā)andspace” plot of each spectrum. The bottom plot shows a “contour” plot of the same information, with increasing signal power shown as increasing colour from black to red Figure 3 Power spectra for Test 19 after removal of all spectral ponents beneath power threshold Figure 4 Az against time (in seconds) for Test 19,before removal of lowpower frequency ponents Figure 5 Az against time (in seconds) for Test 19, after removal of lowpower frequency ponents Figure 6 SP for an example test, showing three automaticallydetecrmined states:S1drilling in (shown in green)。 State S2: the peakpower period where the drillbit passes through the disk and is removed State S3: the approximately constantpower period of retraction. Note that this segmentation is only the identification of the times of onset and offset of each of the three described states, for the purposes of graphical display as described in the next subsection. 公司機密 牛津信號分析機構(gòu) 文件號： P0193GP01=1 文件名：制造分析進程數(shù)據(jù)使用快速標記技術(shù) 論點： 1 日期： 姓名 簽 名 作者 審核 目錄 1 執(zhí)行概要（文章綜述） 引言 引用的技術(shù) 結(jié)論摘要 觀察資料、報告 2 引言 牛津信號分析機構(gòu) 3 引用國外的參考文獻 4 術(shù)語表 5 數(shù)據(jù)描述 數(shù)據(jù)類型 試驗狀況簡介 測試描述 6 預處理 移除開始、終止