Neural network computing system for pattern recognition of thermoluminescence signature spectra and chemical defense

• US 5,631,469 A
• Filed: 04/15/1996
• Issued: 05/20/1997
• Est. Priority Date: 04/15/1996
• Status: Expired due to Fees

• Current Assignee: United States Department of the Army (Government of the United States of America)
• Original Assignee: The United States Of America As Represented By The Secretary Of The Army
• Inventors: Carrieri, Arthur H., Lim, Pascal I.
• Primary Examiner(s): Porta, David P.
• Assistant Examiner(s): Bruce, David Vernon
• Application Number:
  US08/636,994
• Time in Patent Office: 400 Days
• Field of Search: 395/21, 395/23, 250/341.1, 250/341.5, 250/341.6
• US Class Current: 250/341.5
• CPC Class Codes: G01N 21/3504 for analysing gases, e.g. multi-gas analysis
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A four-layer neural network is trained with data of midinfrared absorption by nerve and blister agent compounds (and simulants of this chemical group) in a standoff detection application. Known infrared absorption spectra by these analyte compounds and their computed first derivative are scaled and then transformed into binary or decimal arrays for network training by a backward-error-propagation (BEP) algorithm with gradient descent paradigm. The neural network transfer function gain and learning rate are adjusted on occasion per training session so that a global minimum in final epoch convergence is attained. Three successful neural network filters have been built around an architecture design containing: (1) an input layer of 350 neurons, one neuron per absorption intensity spanning 700≦ν≦1400 wavenumbers with resolution Δν=2; (2) two hidden layers in 256- and 128-neuron groups, respectively, providing good training convergence and adaptable for downloading to a configured group of neural IC chips; and (3) an output layer of one neuron per analyte--each analyte defined by a singular vector in the training data set. Such a neural network is preferably implemented with a network of known microprocessor chips.

Claims

1. A system for recognizing compositions of matter in accordance with thermoluminescence exhibited by the compositions of matter, the system comprising:

sensor means for detecting the thermoluminescence;

spectral analysis means for analyzing the thermoluminscence detected by the sensor means to produce spectral data for a plurality of wave numbers; and

artificial neural network means for receiving the spectral data and for determining, in accordance with the spectral data, whether the thermoluminescence detected by the sensor means indicates the presence of the compositions of matter.

2. A system as in claim 1, wherein the spectral data are binary data.

3. A system as in claim 1, wherein the spectral data are decimal data.

4. A system as in claim 1, wherein:

the spectral data comprise a plurality of discrete spectral data; and

the artificial neural network means comprises one input neuron for each of the plurality of discrete spectral data.

5. A system as in claim 4, wherein the artificial neural network means further comprises one output neuron for each of the plurality of compositions of matter to be recognized.

6. A system as in claim 5, wherein the artificial neural network means further comprises first and second hidden layers.

7. A system as in claim 6, wherein:

the input layer has 350 neurons;

the first hidden layer has 256 neurons; and

the second hidden layer has 128 neurons.

8. A system as in claim 7, wherein the output layer has nine neurons.

9. A system as in claim 8, wherein the artificial neural network means comprises:

a plurality of artificial neural network chips, each comprising a plurality of neurons; and

motherboard-bus means for interconnecting the plurality of artificial neural network chips in a serial-parallel configuration to provide the input layer, the first and second hidden layers and the output layer.

10. A system as in claim 1, wherein the sensor means comprises:

a Michelsen interferometer for producing raw data; and

fast Fourier transform means for converting the raw data into data indicating the thermoluminescence.

11. A method of training an artificial neural network to recognize compositions of matter in accordance with thermoluminescence exhibited by the compositions of matter, the method comprising:

(a) providing known absorption spectra for the compositions of matter;

(b) differentiating the known absorption spectra with respect to absorption time to derive differentiated absorption spectra; and

(c) training the artificial neural network by backward error propagation with the known absorption spectra and the differentiated absorption spectra to obtain a plurality of neural network weights for the artificial neural network.

12. A method as in claim 11, further comprising:

(d) providing a second artificial neural network having a same configuration of neurons as the artificial neural network; and

(e) exporting the plurality of neural network weights from the artificial neural network to the second artificial neural network;

whereby the second artificial neural network becomes capable of recognizing the compositions of matter.

13. A method as in claim 11, wherein step (c) comprises converting the known absorption spectra into binary values and supplying the binary values to the artificial neural network.

14. A method as in claim 11, wherein:

step (c) comprises converting the spectral data into a plurality of discrete spectral data; and

the artificial neural network comprises one input neuron for each of the plurality of;

discrete spectral data.

15. A method as in claim 14, wherein the artificial neural network further comprises one output neuron for each of the compositions of matter to be recognized.

16. A method as in claim 15, wherein the artificial neural network means further comprises first and second hidden layers.

17. A method as in claim 16, wherein:

the input layer has 350 neurons;

the first hidden layer has 256 neurons; and

the second hidden layer has 128 neurons.

18. A method as in claim 17, wherein the output layer has nine neurons.

19. A method as in claim 11, wherein step (c) comprises converting the known absorption spectra into decimal values and supplying the decimal values to the artificial neural network.