Frontiers of Engineering Reports on Leading-Edge Engineering from the 2013 Symposium (2014) / Chapter Skim
Currently Skimming:
Distributed Anomaly Detection for Timely Fault Remediation in Modern Manufacturing--Dragan Djurdjanovic
Distributed Anomaly Detection for Timely Fault Remediation in Modern Manufacturing--Dragan Djurdjanovic
Pages 29-44
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 29... ...
First, the condition monitoring procedures have to be trained to recognize a large number of potential faults, some of which cannot be anticipated during the design stage. Second, because modern manufacturing equipment can perform a variety of operations, it displays highly dynamic behavior.
|
From page 30... ...
One of those downtimes was finally resolved after a teleconference with a quantum physics PhD, who led the control design and development of the PECVD tool in question. (I have seen the same type of tool cause similar problems in two other semiconductor manufacturing companies, again because it took a long time to find the root causes and do appropriate repairs.)
|
From page 31... ...
Hence, instead of identifying various faulty behavior regimes, the more useful focus would be on localizing the source of the anomalous behavior, using the paradigm of distributed anomaly detection. Essentially, monitoring can be realized through a set of anomaly detectors (ADs)
|
From page 32... ...
Pressure Exhaust Temp ADcooler EGR TEGR Ambient Temperature Cooler FIGURE 1 Distribution of anomaly detectors (ADs) in a diesel engine exhaust gas r ecirculation (EGR)
|
From page 33... ...
from anomaly detectors (ADs) that proliferate through the exhaust gas recirculation (EGR)
|
From page 34... ...
s Advanced signal processing, statistical analysis, and time-series modeling, which worked so well with rotating machinery and traditional manufacturing, do not help in this case. A new mathematical construct was recently devised for degradation modeling and anomaly detection in inherently unobservable processes in variable operating regimes.
|
From page 35... ...
introduces a genetic algorithm (GA) –based procedure to identify parameters for the degradation-describing HMMs using sensor readings emitted by a system in arbitrarily mixed operating regimes.
|
From page 36... ...
The dashed vertical line labeled "Training" demarcates the data used to identify the parameters of the operating regime–specific HMMs. Two major downtimes caused by severely unacceptable tool behavior were observed during this period and are labeled in Figure 4 as "Big particle event" and "Coulomb crystals." In addition, based on the analysis of particle counts obtained via particle monitoring wafers (special non-production wafers occasionally passed through the system to assess particle contamination in the system)
|
From page 37... ...
. Between the two major downtimes, several minor particle excursions were observed, during which particle monitoring wafers had significantly elevated counts (slightly less severe levels than what was seen during the first minor particle excursion and order of magnitude less severe than what was seen during the first major downtime)
|
From page 38... ...
To automate identification of the root cause using distributed anomaly detection, it is necessary to understand the causal interactions between various sub systems of the monitored system -- what variables describe each subsystem and FRU of the PECVD tool, and through which variables and in what ways these systems interact while the tool operates. In the case of the automotive EGR system, all relevant variables were adequately sensed, and it was possible to see exactly how various subsystems interacted (what were the inputs and outputs of each subsystem and component)
|
From page 39... ...
In addition, research is needed to explore agent distribution policies that optimize tradeoffs between computational resources needed for each AD, their sensitivity to anomaly detection, and the speed of localization of the sources of anomalous behavior. A simple example in Figure 6 illustrates how the use of computational resources can increase fault sensitivity and speed of localization.
|
From page 40... ...
Djurdjanovic Figure 6_R02544.eps
|
From page 41... ...
Humans as Distributed Agents for Removal of Faults ("Antigens") in a Manufacturing System The unprecedented-fault localization process described above resembles to a degree that of a natural immune system, which identifies and labels an antigen by coating it with appropriately generated antibodies: the diagnostic system described here uses ADs to identify and label faulty subsystems and FRUs.
|
From page 42... ...
42 FIGURE 7 Plot (a) illustrates the proposed tree-based representation of a machine and its subsystems and field replacement units (FRUs)
|
From page 43... ...
2011. Dynamic feature monitoring technique applied to thin film deposi tion processes in an industrial PECVD tool.
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.