Neural networks can be trained to recognize patterns and associations in complex data sets.  Specifically, they can be trained to:

  • Predict or forecast future outcomes of a variable (such as a stock price)
  • Classify data sets into one of many categories.

HD Laboratories can implement neural networks in both software and hardware. Hardware versions can be used with live data or on production lines.  A description of the technology follows:

There are two major types of networks, predictors and classifiers.  These are diagrammed below.

Neural Networks
 Predictor Network
 Classifier Network

Predictor networks generally have multiple inputs (2-100) and one output.  The network is trained to recognize relationships between various input data patterns and the output.  It then can be used to predict future behavior of the output channel, based on the current input data.  Prediction accuracy depends on the quality of the network training.

Classification networks have multiple outputs.  They are trained to recognize patterns of the input data as one of the output classes.   There are two primary classification network types, binary and linear.  Binary networks activate only one of the outputs when an input data pattern is recognized.  Linear networks, however, present a numeric value for each output.  The value  indicates the probability that the input data set is a member of the corresponding output class.

Example:  A Land Mine Classifier

HD Laboratories’ president conducted a program for the Department of Defense to detect  land mines and differentiate them from other battlefield debris using  an electromagnetic search coil.  A 15-output classifier network was successfully trained to identify 6 types of land mines and 9 debris items. 

The classifier network was trained using four channels of input data from a   two-frequency eddy current tester.  The objective was to distinguish     metallic-component land mines from other debris such as gum wrappers, cartridge cases, coins, washers, bottle caps and tin can lids. 

The most complex data was from can lids.  A lid can have as many electromagnetic signatures as it has degrees of rotation relative to the search coil.  When the lid is parallel to the coil, it sustains maximum eddy current losses but has minimal effects on the magnetic circuit of the coil.  However, when the lid’s plane is normal to the coil, its maximum effect is to alter the reluctance of the magnetic flux path, and eddy currents are minimized. These conditions have dramatically different signatures in the complex impedance plane of the search coil.

The network was able to successfully classify can lids at multiple angles as well as all the other objects included in the simulated mine field.

                                                                                                                                                                                                     Back to Top


HD Laboratories’ vice president has a Ph.D in electrical engineering and an extensive background in theoretical and basic research.  We maintain Mathcad and Origin software for mathematical analysis and development.  Typical applications include:

  • Analysis and modeling of ultrasonic and electromagnetic wave propagation
  • Solving complex equation systems
  • Optimization (maximization or minimization of an objective function)
  • Solving differential equations

                                                                                                                                                                                                      Back to Top


Do you have a process or system response that needs an equation to define it?  We can generate highly accurate equations for a variety of unusually-shaped curves.

An example is shown below.  The customer brought a data set he had measured in a processing plant.  An equation was needed for process control.  We used a second order exponential decay function in our curve-fitting software.  This produced a precise fit by selecting constants Yo and To and coefficients A through D of the equation shown.  The result is shown in the chart.

Output = Yo + A{exp[(To-T)/B]} + C^[(To-T)/D]



Back to Top

[Home] [About] [Staff] [Contact]

Copyright 2000-2017 by HD Laboratories, Inc.