外文翻譯--制造分析進程數(shù)據(jù)使用快速標記技術

1、<p>　　附錄1 翻譯原文及譯文</p><p>　　Doc No: P0193-GP-01-1</p><p>　　Doc Name: Analysis of Manufacturing</p><p>　　Process Data Using</p><p>　　QUICK TechnologyTM</p&

2、gt;<p>　　Issue：1</p><p>　　Data：20 April ，2006</p><p>　　Table of Contents</p><p>　　1Executive Summary4</p><p>　　1.1Introdution4</p><p>　　1

3、.2Techniques Employed4</p><p>　　1.3Summary of Results4</p><p>　　1.4Observations4</p><p>　　2Introdution6</p><p>　　2.1Oxford BioSignals Limited6</p><

4、p>　　3External References7</p><p>　　4Glossary7</p><p>　　5Data Description8</p><p>　　5.1Data types8</p><p>　　5.2Prior Experiment Knowledge8</p><p&

5、gt;　　5.3Test Description8</p><p>　　6Pre-processing10</p><p>　　6.1Removal of Start/Stop Transients10</p><p>　　6.2Removal of Power Supply Signal10</p><p>　　6.3Fr

6、equency Transformation10</p><p>　　7Analysis I-Visualisation13</p><p>　　7.1Visualisation of High-Dimensional Data13</p><p>　　7.2Visualising 5-D Manufacturing Process Data13<

7、;/p><p>　　7.3Automatic Novelty Detection15</p><p>　　7.4Conclusion of Analysis I-Visualisation16</p><p>　　8Analysis II-Signature Analysis17</p><p>　　8.1Constructing

8、Signatures17</p><p>　　8.2Visualising Signatures19</p><p>　　8.3Conclusion of Analysis II-Signature Analysis23</p><p>　　9Analysis III-Template Analysis24</p><p>　　

9、9.1Constructing a Template of Normality24</p><p>　　9.2Results of Novelty Detection Using Template Analysis25</p><p>　　9.3Conclusion of Analysis III-Template Analysis26</p><p>

10、　　10Analysis IV-None-linear Prediction27</p><p>　　10.1Neural Networks for On-Line Prediction27</p><p>　　10.2Results of Novelty Detection using Non-linear Prediction27</p><p>　

11、　10.3Conclusion of Analysis IV-Non-linear Prediction28</p><p>　　11Overall Conclusion29</p><p>　　11.1Methodology29</p><p>　　11.2Summary of Tesults29</p><p>　　11.

12、3Future Work29</p><p>　　12Appendix A-NeuroScale Visualisations31</p><p>　　Table of Figures </p><p>　　Test 90. From top to bottom: Ax, Ay, Az, AE, SP against time t(s)</p>

13、<p>　　Power spectra for Test 19 after removal of 50Hz power supply contribution. The top plot shows a 3-D “l(fā)andspace” plot of each spectrum. The bottom plot shows a “contour” plot of the same information, with incr

14、easing signal power shown as increasing colour from black to red </p><p>　　Power spectra for Test 19 after removal of all spectral components beneath power threshold</p><p>　　Az against time (in

15、 seconds) for Test 19,before removal of low-power frequency components </p><p>　　Az against time (in seconds) for Test 19, after removal of low-power frequency components</p><p>　　SP for an ex

16、ample test, showing three automatically-detecrmined states:S1-drilling in (shown in green); S2-drill-bit break-through and removal (shown in red); S3-retraction (shown in blue)</p><p>　　Example signature of

17、variable plotted against operating-point</p><p>　　Power spectra for test 51, frequency (Hz) on the x-axis between [0 fs/2]</p><p>　　Average significant frequency </p><p>　　Visualis

18、ation of AE signatures for all tests</p><p>　　Visualisation of Ax broadband signatures for all tests</p><p>　　Visualisation of Ax average-frequency signatures for all tests</p><p>　

19、　Novelty detection using a template signature</p><p>　　Executive Summary</p><p>　　Introduction</p><p>　　The purpose of this investigation conducted by Oxford BioSignals was to exami

20、ne and determine the suitability of its techniques in analyzing data from an example manufacturing process. This report has been submitted to Rolls-Royce for the expressed of assessing Oxford BioSignals’ techniques with

21、respect to monitoring the example process. </p><p>　　The analysis conducted by Oxford BioSignals (OBS) was limited to a fixed timescale, a fixed set of challenge data for a single process (as provided by Rol

22、ls-Royce and Aachen university of Technology), with no prior domain knowledge, nor information of system failure .</p><p>　　Techniques Employed</p><p>　　OBS used a number of analysis techniques

23、given the limited timescales:</p><p>　　I-Visualisation, and Cluster Analysis </p><p>　　This powerful method allowed the evolution of the system state (fusing all available data types) to be visu

24、alised throughout the series of tests. This showed several distinct modes of operation during the series, highlighting major events observed within the data, later correlated with actual changes to the system’s operation

25、 by domain experts.</p><p>　　Cluster analysis automatically detects which of these events may be considered to be “abnormal”, with respect to previously observed system behavior .</p><p>　　II-Si

26、gnature represents each test as a single point on a plot, allowing changes between tests to be easily identified. Abnormal tests are shown as outlying points, with normal tests forming a cluster.</p><p>　　Mo

27、deling the normal behavior of several features selected from the provided data, this method showed that advance warning of system failure could be automatically detected using these features, as well as highlighting sign

28、ificant events within the life of the system.</p><p>　　III-Template Analysis </p><p>　　This method allows instantaneous sample-by –sample novelty detection, suitable for on-line implementation.&

29、lt;/p><p>　　Using a complementary approach to Signature Analysis, this method also models normal system behavior. Results confirmed the observation made using previous methods.</p><p>　　IV-Neural n

30、etwork Predictor </p><p>　　Similarly useful for on-line analysis, this method uses an automated predictor of system behaviour(a neural network predictor), in which previously identified events were confirmed

31、, and further significant episodes were detected.</p><p>　　Summary of Results</p><p>　　Early warning of system failure was independently identified by the various analysis methods employed. <

32、/p><p>　　Several significant events during the life of the process were correlated with actual known events later revealed by system experts.</p><p>　　Changes in sensor configurations are identifie

33、d, and periods of system stability (in which tests are similar to one another) are highlighted.</p><p>　　This report shall be used as the basis for further correlation of detected events against actual occur

34、rences within the life of the system, to be performed by Aachen University of Technology.</p><p>　　Observations</p><p>　　Based on this limited study, OBS are confident that their techniques are

35、applicable to condition monitoring of the example manufacturing process as follows:</p><p>　　Evidence shows that automated detection of system novelty is possible, compared to its “normal” operation.</p&g

36、t;<p>　　Early warning of system distress may be provided, giving adequate time to take preventative maintenance actions such that system failure may be avoided.</p><p>　　Provision “fleet-wide” analysi

37、s is possible using the techniques considered within this investigation.</p><p>　　The involvement of domain knowledge from system experts alongside OBS engineers will be crucial in developing future implemen

38、tations. While this “blind” analysis showed that OBS modelling techniques are appropriate for process monitoring, it is the coupling of domain knowledge with OBS modelling techniques that may provide optimal diagnostic a

39、nd prognostic analysis.</p><p>　　Introduction</p><p>　　Oxford BioSignals Limited</p><p>　　This document reports on the initial analysis conducted by Oxford BioSignals of manufactur

40、ing process challenge data provided by Rolls-Royce, in conjunction with Aachen University of Technology(AUT).</p><p>　　Oxford BioSignals Limited(OBS) is a world-class provider of Acquisition, Data Fusion, Ne

41、ural Networks and other Advanced Signal Processing techniques and solutions branded under the collective name QUICK Technology. This technology not only provides for health and quality assurance monitoring of the operati

42、onal performance of equipment and plant.</p><p>　　QUICK Technology has been extensively proven in the field of gas turbine monitoring with both on-line and off-line implementations at multiple levels: as a r

43、esearch tool, a test bed system, a ground support tool, an on-board monitoring system, an off-line analysis tool and a “fleet” manager.</p><p>　　Many of the techniques employed by OBS may be described as nov

44、elty detection methods. This approach has a significant advantage over many traditional classification techniques in that it is not necessary to provide fault data to the system during development. Instead, providing a s

45、ufficiently comprehensive model of the condition can be identified automatically. As information is discovered regarding the causes of these deviations it is then possible to move from novelty detection to diagnosis, b&l

46、t;/p><p>　　External References</p><p>　　Accompanying documentation providing further information on the data sets is available in unnumbered documents.</p><p><b>　　Glossary</b

47、></p><p>　　AUT- Aachen University of Technology </p><p>　　GMM- Gaussian Mixture Model </p><p>　　MLP- Multi-Layer Perception</p><p>　　OBS- Oxford BioSigna

48、ls Ltd.</p><p>　　5 Data Description</p><p>　　The following sections give a brief overview of the data set obtained by visual inspection of the data. </p><p>　　Data types</p>

49、<p>　　The data provided were recorded over a number of tests. Each test consisted of a similar procedure, in which an automated drill unit moved towards a static metallic disk at a fixed velocity (“feed”), a hole wa

50、s drilled in the disk at that same feed-rate.</p><p>　　The following data streams were recorded during each test, each sampled at a rate of 20 KHz:</p><p>　　Ax – acceleration of the disk-mountin

51、g unit in the x-plane1 , </p><p>　　Ay- acceleration of the disk-mounting unit in the y-plane1 ,</p><p>　　Az- acceleration of the disk-mounting unit in the z-plane1 ,</p><p>　　AE-RMS

52、 acoustic emission, 50-400 KHz2,</p><p>　　SP-power delivered to the drill spindle3.</p><p>　　Tests considered in this investigation used three drill-prices (of identical product specification) a

53、s shown in Table 1.</p><p>　　Table 1-Experiment Parameters by Test</p><p>　　Note that tests 16,54,128,129 were not provided, thus a series of 190 tests are analysed in this investigation. These

54、190 tests are labeled as shown in Table 2.</p><p>　　Table 2 –Test indices used in this report against actual test numbers</p><p>　　Prior Experiment Knowledge</p><p>　　Normal Tests&l

55、t;/p><p>　　AUT indicated that tests [10110] could be considered “normal processes”.</p><p>　　AE Sensor Placement</p><p>　　AUT noted that the position of the acoustic emission sensor wa

56、s altered prior to test 77, and was adjusted prior to subsequent tests. From inspection of AE data, it appears that AE measurements are consistent after test 84, and so:</p><p>　　·AE is assumed to be un

57、usable for tests [176] –the sensor records only white noise;</p><p>　　·AE is assumed to be usable, but possibly abnormal, for tests [7783] –the sensor position is being adjusted, resulting in extreme va

58、riation in measurements;</p><p>　　·AE is assumed to be usable for tests [94190] –the sensor position is held constant during these tests.</p><p>　　Thus, the range of tests assumed to be nor

59、mal [10110] should be reduced to [84110] when AE is considered.</p><p>　　Test Description</p><p>　　Data recorded for during a typical test are shown in Figure 1. The duration of this test is app

60、roximately t=51 seconds. This section uses this test to illustrate a typical process, as described by AUT.</p><p>　　Drill power-on and power-off events may be seen at the start and end of the test as transie

61、nt spikes in SP.</p><p>　　The drill unit is then moved towards the static disk at the constant feed rata specified in Table 1, between t=12 and 27 seconds. This corresponds to approximately constant values o

62、f SP during that period, approximately zero AE, and very lowamplitude acceleration in x-,y-,and z- planes.</p><p>　　At t=27 seconds, the drill makes contact with the static disk and begins to drill into the

63、metal. This corresponds to a step-change in SP to a higher lever, staying approximately constant until t=38 seconds. During this time, AE increases significantly to a largely constant but non-zero value. The values Ax an

64、d Az increase throughout this drilling operation, while the value of Ay remains approximately zero (as it does throughout the test).</p><p>　　At t=38 seconds, the tip of the drill-bit passes through the rear

65、 face of the disk. The value of SP increases until t=44 seconds. During this period, AE reaches correspondingly high values, while Ax and Az decrease in amplitude.</p><p>　　At t=44 seconds, the direction of

66、the drill unit is reversed, and the drill is retracted from the metal disk. Until t=46 seconds, the value of SP and AE decrease rapidly. A transient is observed in Ax and Az at t =44 seconds, with vibration amplitude dec

67、reasing until t=46 seconds.</p><p>　　At t=46 seconds, the drill-bit has been completely retracted from the metal disk, and the unit continues to be withdrawn at the feed rate until the end of the test. The v

68、alue of SP decreases during this period(noting the power-off transient at the very end of the test), while the values of all three acceleration channels and AE are approximately zero.</p><p>　　.Pre-processin

69、g</p><p>　　Removal of Start/Stop Transients</p><p>　　Assuming that normal and abnormal system behaviour will be evident from data acquired during the drilling process, prior to analysis, each te

70、st was shortened by retaining only data between the start and stop events, shown as transients in SP. For example, for the test shown in Figure 1, this corresponds to retaining the period [1350] seconds. </p><

71、p>　　Removal of Power Supply Signal</p><p>　　The 50 Hz power supply appears with in each channel, and was removed prior to analysis by application of a band-stop filter with stop-band [4951] Hz.</p>

72、<p>　　Frequency Transformation</p><p>　　Data for each test were divided into windows of 4096 points. A 4096-point FFT for was performed using data within each window, for Ax,Ay and Az channels. This co

73、rresponds to approximately 5 FFTs per second of data，similar to the QUICK system used in aerospace analysis， shown to provide sufficient resolution for identifying frequency-based events indicative of system abnormality.

74、</p><p>　　For the analyses performed in this investigation, all spectral components of Ax, Ay, and Ay occurring at frequency f with power Pf below some threshold Pf<h were discarded. Time-domain signals w

75、ere reconstructed by performing an inverse FFT operation on each spectral window of 4096 points.</p><p>　　Figure 2 shows the spectral power content of Az for Test 19 after removal of the 50 Hz power supply s

76、ignal, from [021] seconds, with each FFT shown between [0 fs/2] Hz. Frequency content throughout this test is typical for all tests: the majority of significant spectral peaks are concentrated during the drilling operati

77、on(between 14 and 21 seconds, in this test). As a hole is drilled in the metal disk, power is concentrated at higher and higher frequencies, usually reaching a highest frequency(h</p><p>　　Figuer 3 shows the

78、 same test are removal of all components with power Pf<0.1. This retains the significant peaks in the power spectral, whilst removing components assumed to be insignificant due to their low power. </p><p>

79、;　　Figure 4 and Figure 5 show the corresponding time-series data for Ax in test 19. After removal of low-power frequency components, the time-series retains only the episodes in which significant-power vibrations were ob

80、served, which are used as the basis for detection of system abnormality by several of the analysis methods used within this investigation.</p><p>　　Figure 2-Power spectra for Test 19 after removal of 50 Hz p

81、ower supply contribution. The top plot shows a 3-D “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

82、 red.</p><p>　　Figure 3-Power spectral for Test 19 after removal of all spectral components beneath power threshold .</p><p>　　Figure 4-Az against time(in seconds) for Test 19, before removal of

83、 low-power frequency components</p><p>　　Figure 5-Az against time(in seconds) for Test 19, after removal of low-power frequency components.</p><p>　　Analysis I-Visualisation</p><p>

84、;　　This section describes the first of four analysis techniques applied to the manufacturing process data-set.</p><p>　　Visualisation of High-Dimensional Data</p><p>　　Constructing a 2-D Visuali

85、zation</p><p>　　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 high-dimensional represent

86、ation of the state of a system, but this is not readily interpreted by an operator. </p><p>　　Neuroscale allows the visualization of systems that have high-dimensionality by mapping data to lower numbers of

87、dimensions(typically two,for visual inspection). It attempts to preserve the inter-pattern distances in the high-dimensional data. Data which are close together in high-dimensional space are typically kept close together

88、 in 2-D space, and data that are originally far apart remain well separated after projection.</p><p>　　The projection is performed using a non-linear function from the data’s k dimensional space down to 2-D

89、for visualization purposes. In this investigation, k is 5:[Ax, Ay, Az, AE, SP] are the high-dimensional sample vectors. </p><p>　　The creation of a non-linear mapping from 5-D space to 2-D requires sample da

90、ta 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 data-set is required. Each test was summarized by a nu

91、mber of prototype 5-D vectors using the k-means clustering algorithm(in which a large number of data are represented by a smaller number of prototype vectors). The non-linear mapping was trained using t</p><p

92、>　　Automatic Test Segmentation</p><p>　　To allow the examination of the 5-D data using visualization, it is convenient to divide the drilling process in to three stages, corresponding to the typical behav

93、iour of the process described in Section 5.3.</p><p>　　A heuristic algorithm was produced to perform automatic segmentation into three episodes using the SP channel, as illustrated in Figure 6(which shows a

94、low-pass filtered version of SP superimposed on the original signal as a red line). The three states identified correspond to :</p><p>　　State S1: the approximately constant-power (or slightly decreasing)

95、initial period of drilling;</p><p>　　State S2: the peak-power period where the drill-bit passes through the disk and is removed</p><p>　　State S3: the approximately constant-power period of r

96、etraction.</p><p>　　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 ne

97、xt sub-section.</p><p><b>　　公司機密</b></p><p><b>　　牛津信號分析機構</b></p><p>　　文件號：P0193-GP-01=1</p><p>　　文件名：制造分析進程數(shù)據(jù)使用快速標記技術</p><p><b

98、>　　論點：1</b></p><p>　　日期：2006.4.20</p><p><b>　　目錄</b></p><p>　　執(zhí)行概要（文章綜述）</p><p><b>　　引言</b></p><p><b>　　引用的技術</b&

99、gt;</p><p><b>　　結論摘要</b></p><p><b>　　觀察資料、報告</b></p><p><b>　　引言</b></p><p><b>　　牛津信號分析機構</b></p><p><b&g

100、t;　　引用國外的參考文獻</b></p><p><b>　　術語表</b></p><p><b>　　數(shù)據(jù)描述</b></p><p><b>　　數(shù)據(jù)類型</b></p><p><b>　　試驗狀況簡介</b></p>

101、<p><b>　　測試描述</b></p><p><b>　　預處理</b></p><p>　　移除開始、終止瞬態(tài)數(shù)據(jù)</p><p><b>　　移除電源干擾信號</b></p><p><b>　　頻率變換</b></p>

102、<p><b>　　分析處理1-可視化</b></p><p><b>　　高維數(shù)據(jù)分析</b></p><p><b>　　5維機械加工數(shù)據(jù)</b></p><p><b>　　自動信號檢測</b></p><p>　　分析方案1-可視化的結

103、論</p><p>　　分析處理2-信號處理分析</p><p><b>　　構建信號系統(tǒng)</b></p><p><b>　　波形分析信號</b></p><p><b>　　分析結論</b></p><p>　　分析處理3-基于模板分析的數(shù)據(jù)分析&l

104、t;/p><p><b>　　構建普通信號模板</b></p><p>　　使用模板分析捕獲信號的結論</p><p><b>　　分析結論</b></p><p>　　分析處理5-非線性預測分析</p><p>　　基于在線預測的神經(jīng)網(wǎng)絡</p><p>

105、;　　基于非線性預測的神經(jīng)網(wǎng)絡得到的結論</p><p><b>　　非線性預測結論</b></p><p><b>　　系統(tǒng)結論</b></p><p><b>　　方法學</b></p><p><b>　　結論概述</b></p>&l

106、t;p><b>　　前景工作</b></p><p>　　12 附錄：神經(jīng)網(wǎng)絡分析</p><p><b>　　關于圖表的列表</b></p><p>　　圖表1-測試90,從上到下分別是:AX,AY,AZ,AE,SP相對于時間的坐標</p><p>　　圖表2-在移除50HZ電源干擾信號后測