外文翻譯--對(duì)醫(yī)療器械行業(yè)注射模系統(tǒng)質(zhì)量檢測(cè)的多元分析（英文）

1、MULTIVARIATE ANALYSIS (MVA) FOR QUALITY DETECTION IN INJECTION MOLDING SYSTEMS IN THE MEDICAL DEVICE COMMUNITY Chris Ambrozic, Umetrics, Inc., Kinnelon, NJ Lee Hutson, Baxter Healthcare Corp., Mountain Home, AR Abstract

2、 We describe a new method of point-of-origin quality detection for injection molding systems. The method encompasses data acquisition, Multivariate modeling, reject control and data reporting, provides in-line quality

3、 detection of injection molded parts, and real-time reports on fault contributors. We discuss real-world production applications in which MVA is applied using real-time molding parameters to predict quality, with a g

4、oal of Parametric Release. Introduction The current quality control techniques for injection molding that are in place within the Medical Device Community (i.e. AQL-based lot sample inspection, Statistical Process Co

5、ntrol (SPC), etc) do not adequately prevent defective parts from continuing on through the manufacturing process. SPC methodologies for example, typically monitor control charts for 3σ excursions as a measure of out-

6、of-control processes. Invariably, waiting for the detection of a 3σ excursion results in a loss of product. Univariate analysis-based fault detection methodologies are also plagued with high false alarm rates. Ad

7、ditionally, univariate SPC techniques do not take into account parameter interactions and correlations. The impact of the failure of conventional QC methods for real-time defect capture and rejection is significant in

8、 terms of downstream production time, cost and recovery, as well as regulatory impact. The overall cost to production is magnified many times relative to the cost of defect capture at or near the point of origin for

9、injection molding systems. There exist Statistical Process Control (SPC) systems for injection molding tools that contain the option to integrate mold cavity sensors and such can help to reduce the numbers of defectiv

10、e units released to downstream processing and these provide documented benefits. These systems do not, however, explain the source of the process variation, and only give an indication of what is actually occurring i

11、n the mold cavity. In order to obtain true parametric release, the source of the variation must be understood and controlled. The addition of Multivariate Analysis (MVA) processing to these solutions is needed to rea

12、lize the greatest processing advantages and true Parametric Release in parts and processes. MVA analysis of an injection molding process can provide critical improvements in injection molding processing including: a)

13、 More accurate and precise fault detection than can be achieved using conventional SPC b) The ability to clearly monitor all variables simultaneously c) Clear views of process drift d) Real-time identification of thos

14、e variables most strongly contributing to a fault e) Real-time, in-line rejection capability f) Fewer false rejects than are observed when using SPC alone g) Increased productivity in terms of troubleshooting an

15、d problem diagnosis MVA technology [1-3] is the science of separating the signal from the noise in data with many variables and presenting this data in a simple graphical format. A key advantage of this technology is

16、the ability to take large, unwieldy data sets and reduce them to simple model representations that can be readily understood and employed for quality control purposes. In MVA technology real-time process data is used

17、 to create a “current” process model which is numerically contrasted with a previously established “known good” process model. The results of this numerical comparison are two relatively simple decision statistics, D

18、ModX and Hotelling’s T2; and these define the nature and extent of observed deviations in the current process from the established “good” process model. The greater the numeric value of Hotelling’s T2, the more likel

19、y it is that the current data deviates significantly from the model and that the product is “out of spec”. The larger the value of DModX, the greater the likelihood that the current data is influenced by factors or i

20、n patterns not present when the original model was formulated. In this report, we describe the incorporation of MVA methodologies into Quality Control for Fault Detection Analysis (FDC) in the injection molding of me

21、dical device components. We describe the use of this methodology for the achievement of improved process understanding, for real-time identification and rejection of A key characteristic of the MVA system is its abili

22、ty to “drill down” through the data and identify fault contributors once the fault has been recognized. Figure 2 shows the onset of each particular fault type to correlate with a particular cycle (recall that in the

23、DModX plot, the points on the X-axis each correspond to one cycle in the injection molding process). Each point along this axis has an associated contribution plot containing values for each of the key variables in

24、the process. Figure 3 shows the contribution plot associated with the cycle in which short shots were intentionally introduced into the process. The plot in Figure 3 shows that, of the variables monitored in this pro

25、cess, two showed significant deviation from the normally accepted model values. The injection pack pressure is seen to be significantly below its expected value during the short shot excursion while the shot cushion

26、shows a small but significant positive deviation from its expected value. Figure 4 shows the contribution plot associated with the cycle in which flashing was introduced to the process. Again, injection pack pressure

27、 and shot cushion are seen to be the primary deviations from the reference model, but in this case the former shows positive deviation from the expected value while the latter is negative. Figure 5 shows the contribu

28、tion plot associated with the cycle in which the simulated double shot occurred. In this plot a different pattern of deviations from expected values is observed than for either short shots or flashing. The DModX appr

29、oach thus shows not only alarms, but also provides information on the kind of alarm. The combined use of the DModX plot with the Contribution Plots DModX provides information on the correlation structure breaks in th

30、e alarm. The breakdown in the pattern of variables is readily apparent in the correlation plots and the pattern of the breakdown is indicative of the characteristics of the fault that has occurred in the process. Thi

31、s identification of fault characteristics is a key concept of multivariate analysis. The methodology monitors not only the value of a variable but also how it relates to other variables in the process. When the corre

32、lation structure / relationship breaks down, MVA detects the break. Such relational breaks are not detected using univariate analysis (UVA), since UVA assumes that all variables are independent. This assumption is no

33、t valid in systems where process variables are correlated (most systems). Results The use of MVA analysis provides a key advantage in process control. With MVA in place, it is possible to automatically identify defect

34、s such as short shots, double shots and excess flash. Once these defects have been identified by the system, it is feasible to automatically divert suspect product and avoid further product loss through downstream pr

35、oduct defects that are directly correlated with the original defect. This improves productivity, reduces potential financial penalties and decreases the need for inspection resources. This study shows that it is possi

36、ble to identify those process parameters that contribute to suspect product and to reduce problem identification and correction time/prevention time through the use of MVA methods in the injection molding of med

37、ical parts. Stability and control of this process has been significantly improved through the increase in process knowledge, resulting in increases in Cp and Cpk values of, typically, 9%. In the long term, it should

38、 be possible to build libraries of defect classifications to assist with a broad range of issue classifications and resolutions. References 1. “Improving Pulp and Paper Process Diagnostics and Knowledge by Means

39、 Multivariate Analytical Techniques (MVA)”, G. Wold, N. Kettaneh-Wold, Pulp and Paper Canada, 104(5) T121-T123 (2003). 2. “Principal Component Analysis”, S. Wold, K. Esbensen and P. Geladi, Chemometrics and Intelli

40、gent Systems, 2, 37 (1987). 3. “Monitoring of a wastewater-treatment plant with a multivariate model”, by N. Bendwell Pulp and Paper Canada 103(7):, T195-198 (2002). 4. “What is Scientific Injection Molding”, J. Bozz

41、elli, “Scientific Molding, Part 1: Filling”, J. Sloan, Injection Molding Magazine, Oct., 1997. 5. “Scientific Molding, Part 2: Cooling”, J. Sloan, Injection Molding Magazine, Nov., 1997. 6. “What Is Scientific Injectio