Laboratory for 

      Manufacturing $ystem Realization and Synthesis

 
SOVA DIAGNOSIS
MA/RS HOME->MA/RS ROADMAP->SOVA->DIAGNOSIS->FAULT LOCALIZATION ANALYSIS

Title: Multiple Fault Diagnosis in Multi-Station-Assembly Systems

Sponsor: NIST-ATP

Collaborators: R. Kumar, J. Kong, Y. Zhou, S. Gogoneni (Dimensional Control Systems)
Team: P. Tsai, K. Mannar, and D. Ceglarek
Description:

SOVA diagnosis involves three major steps – fault identification, fault localization and fault isolation. In this project it investigates the fault transmission in the rigid assembly process to illustrate the need of sensor localization and develop best combination of methodologies in these three steps. It shows the complexity of MMP (multi-stations manufacturing process) usually beyond the capability of multivariate statistical methods used in isolation phases. The complex MMP system should be break down and reduce dimension to minimize the influence of structural noise before applying pattern mapping or extractionZ methods in the isolation phase. Diagnosability of isolation methods, PCA (Principal Component Analysis) and ICA (Independent Component Analysis) are investigated to illustrate when sensor localization should be utilized.

SOVA Diagnosis Framework

Fault Generation and Propogation

(a-1) Single part

(a-2) Single part on fixture

(a-3) fixture fault pattern on Single part
 
The (a-3) animation shows the rigid part pattern due to fixture faults such as locating pin broken or wearing out.
 

Single part and the locating fixture

(b-1) But joint
 

(b-2) Lap-but joint
 

 (b-3) Lap joint
 

Error propagation due to part-to-part interaction. Faults transmit from one part to another through joint

Figure (b-2) shows an example of fixture faults which is further manifested by sensor variation patterns. Fault propagation depends on part-to-part joint design. Product architecture (parts, subassemblies and the interfaces/ joints between them) significantly influences fault propagation which can be classified as global or local impact. The global or local impact of fault determines if the fault can be identified, localized and/or isolated by end-of-line or distributed sensing systems.
 

Diagnosis strategy
The diagnosis of dimension variation failure in multi-station assembly system with end-of-line or distributed sensing systems can be conducted in three steps.
Step (1) Fault identification: The objective of this step is to select candidate set of sensor (CSS) base on two criteria:
  • variation threshold

  • Correlation threshold
Step (2) Fault localization: The objective of this step is to identify candidate component/ candidate station.

Step (3) Fault isolation: The objective of this step is identifying root cause by mapping symptom to fault pattern base on the previously determined CSS, candidate component/station and is supported by statistical analysis (PCA/ICA).

 
Need for localization

Traditional statistical approaches apply directly to fault isolation on all set of sensor data. This approach works very well for single station system. However, it is not efficient for diagnosis in multi-station assembly system with end-of-line/distributed sensing. In most of the case, fault identification and localization steps are necessary before fault can be isolated. For example, a multi-station assembly system can be represented in the form of state space model as

where represents the active faults we intend to isolate, is the structural noise and represent the error term (Tsai et. al 2006). The diagnosability of dimensional fault using PCA approach can be defined as

which indicates the importance of reducing the error term
To isolate the fault root cause, it is critical to localize the fault and structural noise can be reduced before applying fault isolation methods to pattern extraction or mapping. The fault localization could be done in two steps:
  •  Identification to localize to subsets of sensors
  • localization to decide candidate stations and components
Step (1) Fault Identification
Approach 1: Variation Threshold/ correlation Method (Ceglarek et al. 1994)
Approach 2: FRL method (Tsai et al. 2006)

 

Fault Identification is to select the subset of measurements that detect the fault. This is can be done by only applying data analysis criteria such as standard deviation magnitude and correlation as contrast function to select sensor which represent the largest fault cause by single root cause.
The FRL method uses classification tree approach to obtain critical values sets that could distinguish faulty and normal data. The critical values and the corresponding sensors are then grouped where each subset is associated with one set of faults.

 
Step (2) Fault localization
 Approach 1: Localization based on Process Tree (Ceglarek et al. 1994)

 

Fault localization is the method to use the sensor measurement subset from fault Identification step to localize to the station and component that fault occurs. The process tree localization method map the faulty sensor subset to the process tree to calculate membership in order to reduce the scope of problem and track where the fault come from. The output of fault localization is candidate station(s) and component(s). This is also important to the next isolation step because the designated pattern pool can be reduced after candidates station/component are identified.
 
Step (3) Fault Isolation
Approach 1: PCA based pattern recognition (Ceglarek and Shi 1996 )
Approach 2: ICA/ BSS based pattern extraction and recognition
Approach 3: Orthogonal Diagonalization Analysis (Kong et al. 2006 )

Fault isolation is the method to extract fault pattern from the CSS. When process model is available, patterns can be mapped to the designated pattern pool to identify the root cause. The candidate station/component information from fault localization step could facilitate the mapping process by reducing the pattern pool.
PCA is the most used databased method to identify the major variation direction. In single dominant fault scenario, sensor variation pattern could be matched to fault pattern and root cause can be recognized. In multiple fault scenario, ICA is the databased used to identify independent source of signal, which is fixture fault of our interest. ICA methods use higher order statistic, such as fourth order moment as contrast function to separate mixed signal and identify the most independent directions as fault patterns. With SOVA model, these extracted patterns can be further used in pattern mapping to identify physical root cause.
Orthogonal Diagonalization Analysis (ODA) is the model based method to project sensor variation to the designated patterns provided by the SOVA model. The projection length represents the variation on a specific direction (pattern) to provide analysis for multiple faults .
 
 
Related Papers
 

 
 
 
website stat