Introduction:

Neural network analysis is an artificial intelligence (AI) approach to mathematical modeling.  Neural networks are systems that are loosely patterned on the human brain, that can learn and discern patterns in real-world conditions where data is incomplete or where the number of variables is vast.  Neural networks can model dynamic, non-linear phenomena that are too complex to be described by analytical methods or empirical rules.  Neural networks can be implemented for advanced control, data and sensor validation, pattern recognition, diagnosis and prediction, fault classification, and multivariable quality control applications.

Benefits:

• Reduced maintenance costs and minimized chances of catastrophic failures through early detection and trend analysis
• Significant reduction in data analysis tasks/time
• Solves difficult process problems that cannot be solved quickly or accurately with conventional methods
• Robust, accurate and operate in real time

Experiences/Applications:

Neural Network Development: Neural network models were developed for the analysis of solenoid valve current and voltage traces to determine the health of a solenoid.  As a solenoid energizes or de-energizes, the manner in which the current changes i s consistent and can serve as a signature or "fingerprint" of a healthy solenoid.  Figure 1 shows a sample of a healthy solenoid energize trace. Trained neural networks were incorporated into an Automated Diagnostic Propulsion System developed at NAS A's Marshall Space Flight Center to analyze the energize and de-energize current and voltage traces of a valve on a propulsion system testbed.  The networks were able to detect anomalous conditions during system operation.  The ability to determ ine the health of a valve provides information useful for early failure detection and diagnostic decision-making. 

Figure 1.  Solenoid Current Trace

Other Applications:

The applications areas and capabilities of neural networks are countless; a few examples are listed below.  Neural networks can be used to:

• Perform computer-chip manufacturing quality control
• Predict product quality
• Maintain product quality specifications
• Perform defect classifications
• Perform electrical load forecasting
• Optimize a set of ingredients and processing attributes

       Figure 2.  Neural Network Topology

References:

1. "The Use of Clustering Analysis and Feature Extraction for the Reduction of Very Large Data Sets, Analyzed Via a RBF Neural Network", Ruby D. Lathon, Proceedings of the AIAA Defense & Space Programs Conference, Huntsville, AL, October 1998.
    
2. "An Application of G2 and NeurOn-Line to an Automated Propulsion Diagnostic System", Ruby D. Lathon, Jonathan Patterson, Proceedings of the 1997 Gensym Users Society Worldwide Conference, Paris, France, April 1997.