P/NA.3 Process Modelling and Optimization

 

The Progressive Neural Algorithm P/NA software was developed by N. A. Tech. specifically for continuous value models. Unlike Multi-Layer Perceptron (MLP) models produced using the backpropagation method, P/NA models produce very accurate results for smooth function outputs. Tests conducted by independent reviewers have shown the P/NA models to be significantly more accurate than other neural network techniques, or statistical or non-linear regression models. Version 3.00 (P/NA.3) software is the result of over a decade of research and development by Dr. Jones and his colleagues at N. A. Tech. It represents the state-of-the-art in neural network modeling and offers many advantages not available with other modeling methods.

 

Graphical User Interface (GUI) for a typical T-Fillet weld neural network model

 

P/NA.3 Neural Network models have easy to use graphics that operate on desktop PC computers. A weld cross-section diagram instantly responds to changes in process parameter values, controlled by the slide bars on the screen. Engineering quality measurements change with the weld graphic and are displayed with easy to visualize slide bars and in numerical boxes. Response surface graphics illustrate how variations in control parameters will affect the weld quality. You can optimize the weld quality by pre-planning the weld procedure and testing it on the computer before taking it to the shop floor.

 

  • ROBUST PROCESS - Find process conditions most insensitive to process variations
  • REDUCE COSTS - Use model instead of expensive trial and error to optimize process
  • IMPROVES QUALITY - Ensures consistent quality and conformance to specifications
  • SENSITIVITY ANALYSIS - Ability to pre-determine critical process control parameters
  • SPC ANALYSIS - Identify process setting levels to maximize process stability
  • TROUBLESHOOTING - Use model to find failure conditions and train operators
  • SIMPLE TO USE - On screen graphics for weld performance

Three dimensional dynamic response surface using the N. A. Tech. Dyna-Graph (TM) software for a Flare-Bevel weld neural network model.

The response surface is part of the standard neural network software GUI. The Dyna-Graph is generated dynamically each time a change is made in any of the model input parameters. The graphic has virtually instantaneous response even though the neural network must be run 625 times to produce this surface. The operations "dot" is moved by the mouse to any position on the surface and the weld shape that will be produced using that set of parameter values is automatically displayed to the right of the surface.

The P/NA.3 software has been ported to run on the Texas Instruments TMS320C6xxx DSP (Digital Signal Processor). A model operating in real-time with sensors such as voltage, current, wire feed rate, and travel speed can produce a full three dimensional replica of the weld as the weld in being produced. N. A. Tech. produces the ArcSentry line of weld monitoring equipment, which can have a P/NA.3 neural network embedded.

 

3-D replica of weld, produced real-time by embedded P/NA.3 neural network using N. A. Tech.ArcSentry weld monitor. Colors indicate the weld status (green = acceptable, yellow = marginal, red = reject). Weld status is based on the specification limits for the weld (e.g. weld size, depth of penetration, effective throat, etc.)

The weld base metal can be made transparent so that the weld penetration into the base metal can be viewed. Notice, in this example, when contamination on the surface of the weld caused arc instabilities, the weld penetration was reduced. First, the weld was "flagged" as marginal, and then when the penetration dropped below acceptable levels, the weld status changed to "reject" as indicated by the red color.

To contact us directly about this product, please see our Contact Page.




Home  | Process Improvement  | Flexible Manufacturing  | Services  | Technical Support  | Company Information  | Contact Information

Copyright 2002 Native American Technologies. All rights reserved.