Neural network system and methods for analysis of organic materials and structures using spectral data

• US 5,218,529 A
• Filed: 07/30/1990
• Issued: 06/08/1993
• Est. Priority Date: 07/30/1990
• Status: Expired due to Fees

• Current Assignee: The University of Georgia Research Foundation, Inc.
• Original Assignee: The University of Georgia Research Foundation, Inc.
• Inventors: Thomsen, Jan U., Meyer, Bernd J., Sellers, Jeffrey P.
• Primary Examiner(s): Hayes, Gail O.
• Application Number:
  US07/559,649
• Time in Patent Office: 1,044 Days
• Field of Search: 364/413.01, 364/497, 364/498
• US Class Current: 702/28
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: Y10S 706/924 Medical

Abstract

Apparatus and processes for recognizing and identifying materials. Characteristic spectra are obtained for the materials via spectroscopy techniques including nuclear magnetic resonance spectroscopy, infrared absorption analysis, x-ray analysis, mass spectroscopy and gas chromatography. Desired portions of the spectra may be selected and then placed in proper form and format for presentation to a number of input layer neurons in an offline neural network. The network is first trained according to a predetermined training process; it may then be employed to identify particular materials. Such apparatus and processes are particularly useful for recognizing and identifying organic compounds such as complex carbohydrates, whose spectra conventionally require a high level of training and many hours of hard work to identify, and are frequently indistinguishable from one another by human interpretation.

Claims

1. Apparatus for analyzing and identifying the structure of a particular organic material by recognizing patterns of information that are characteristic of such materials, the apparatus comprising:

(a) analytical means for applying energy to the organic material under analysis, sensing transformations in the energy imparted by the material, and producing therefrom spectral information corresponding to the energy transformations and the structure of the material;

(b) means for digitizing a plurality of incremental portions of the spectral information into digital data; and

(c) off-line neural network means for utilizing the digital data to identify the structure of the organic material under analysis, comprising;

(i) an input layer comprising a plurality of input nodes, each of which nodes receives the digital data;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer and each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(iv) each of the nodes featuring a threshold value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis.

2. Apparatus according to claim 1 in which each node of the output layer corresponds to an organic material.

3. Apparatus according to claim 1 comprising at least one hidden layer hierarchically intermediate the input and output layers and comprising a plurality of hidden nodes.

4. Apparatus according to claim 1 in which the neural network means is in the form of at least one computer program.

5. Apparatus according to claim 1 in which the nodes are implemented in computer hardware.

6. Apparatus according to claim 1 in which each of the output nodes corresponds to a carbohydrate molecule.

7. Apparatus according to claim 1 in which the off-line neural network means is a forward-feed neural network having a single hidden layer of hidden nodes.

8. Apparatus for analyzing and identifying the structure of a particular organic material by recognizing patterns of information that are characteristic of such materials, the apparatus comprising:

(a) analytical means for applying energy to the organic material under analysis, sensing transformations in the energy imparted by the material, and producing therefrom spectral information corresponding to the energy transformations and the structure of the material;

(b) means for digitizing a plurality of incremental portions of the spectral information into digital data; and

(c) off-line neural network means for utilizing the digital data to identify the structure of the organic material under analysis, comprising;

(i) an input layer comprising a plurality of input nodes, each of which nodes receives the digital data;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes;

(iv) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer and each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(v) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis.

9. Apparatus according to claim 8 in which each node of the output layer corresponds to an organic material.

10. Apparatus according to claim 8 comprising a plurality of hidden layers of hidden nodes.

11. Apparatus according to claim 8 in which the off-line neural network means is in the form of at least one computer program.

12. Apparatus according to claim 8 in which the nodes are implemented in computer hardware.

13. Apparatus according to claim 8 in which each of the output nodes corresponds to a carbohydrate molecule.

14. Apparatus according to claim 8 in which the off-line neural network means is a forward-feed neural network having a single hidden layer of hidden nodes.

15. Apparatus for analyzing and identifying the structure of a particular organic material by recognizing patterns in spectra that are characteristic of such materials, the apparatus comprising:

(a) spectroscopy analytical means for applying radiation energy to the organic material under analysis, sensing transformations in the energy imparted by the material, and producing therefrom spectral information corresponding to the energy transformations and the structure of the material;

(b) means for digitizing a plurality of incremental portions of the spectral information into digital data; and

(c) off-line neural network means for utilizing the spectral information to identify the structure of the organic material under analysis, comprising;

(i) an input layer comprising a plurality of input nodes, each of which nodes receives digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes;

(iv) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer and each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(v) each of the nodes feature a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis.

16. Apparatus according to claim 15 in which each node of the output layer corresponds to an organic material.

17. Apparatus according to claim 15 in which the material under analysis is a carbohydrate molecule and each output node corresponds to a carbohydrate molecule.

18. Apparatus according to claim 15 in which the spectroscopy analytical means is a nuclear magnetic resonance spectroscopy device.

19. Apparatus according to claim 15 in which the spectroscopy analytical means is an infrared absorption spectroscopy device.

20. Apparatus according to claim 15 in which the spectroscopy analytical means is an x-ray analysis device.

21. Apparatus according to claim 15 in which the spectroscopy analytical means is a mass spectrometer.

22. Apparatus according to claim 15 in which the spectroscopy analytical means is a gas chromatograph.

23. Apparatus according to claim 15 in which the spectroscopy analytical means is ultraviolet spectroscopy.

24. Apparatus for analyzing and identifying the structure of a particular carbohydrate material by recognizing patterns in spectra that are characteristic of such materials, the apparatus comprising:

(a) magnetic resonance analytical means for subjecting the carbohydrate material under analysis to a magnetic field and radio-frequency radiation, and producing spectral information corresponding to the absorption of the radiation and the structure of the material;

(b) means for digitizing a plurality of incremental portions of the spectral information into digital data; and

(c) off-line neural network means for utilizing the spectral information to identify the structure of the material under analysis, comprising;

(i) an input layer comprising a plurality of input nodes, each of which nodes receives digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes;

(iv) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer and each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from carbohydrate materials analogous to the material under analysis; and

(v) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from carbohydrate materials analogous to the material under analysis.

25. A method for analyzing and identifying the structure of a particular organic material by recognizing patterns in spectra that are characteristic of such materials, comprising the steps of:

(a) subjecting an organic material under analysis to energy in an analytical means;

(b) sensing transformations in the energy imparted by the material;

(c) producing spectral information corresponding to the energy transformations and the structure of the material;

(d) digitizing a plurality of incremental portions of the spectral information;

(e) supplying at least one off-line neural network which comprises;

(i) an input layer comprising a plurality of input nodes, each of which is capable of receiving digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer; and

each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(iv) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis and

(f) applying the digital data corresponding to incremental portions of the spectral information relating to the material under analysis to the input nodes of the neural network in order to generate at the output nodes information that is useful to indicate and identify the structure of the material under analysis.

26. A method according to claim 25 in which the step of supplying at least one off-line neural network comprises supplying at least one off-line neural network which includes at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes.

27. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using a nuclear magnetic resonance spectroscopy device.

28. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using an infrared absorption spectroscopy device.

29. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using an x-ray analysis device.

30. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using a mass spectrometer.

31. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using a gas chromatography device.

32. A method according to claim 25 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using an ultraviolet spectroscopy device.

33. A method according to claim 25 in which the generated information indicates the material under analysis.

34. A method according to claim 25 in which the generated information indicates portions of the material under analysis.

56. A method according to claim 25 in which the generated information corresponds to close structural relatives to the material under analysis.

35. A method for analyzing and identifying the structure of a particular carbohydrate material by recognizing patterns in spectra that are characteristic of such materials, comprising the steps of:

(a) subjecting a carbohydrate material under analysis to energy in a spectroscopy analytical means;

(b) sensing transformations in the energy imparted by the material;

(c) producing spectral information corresponding to the energy transformations and the structure of the material;

(d) selecting portions of the spectral information which are desired for use in identifying the structure of the material;

(e) digitizing a plurality of incremental portions of the selected spectral information;

(f) supplying at least one off-line neural network which comprises;

(i) an input layer comprising a plurality of input nodes, each of which is capable of receiving digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes;

(iv) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer; and

each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from carbohydrate materials analogous to the carbohydrate material under analysis; and

(v) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from carbohydrate materials analogous to the carbohydrate material under analysis; and

(g) applying the digital data corresponding to incremental portions of the spectral information relating to the material under analysis to the input nodes of the neural network in order to generate at the output nodes information that is useful to indicate and identify the structure of the material under analysis.

36. A method according to claim 35 in which the step of supplying at least one off-line neural network comprises supplying at least one feed forward off-line neural network.

37. A method according to claim 35 in which the steps of subjecting the material under analysis to energy in a spectroscopy analytical means, sensing transformations in the energy imparted by the material and producing spectral information corresponding to the energy transformations and the structure of the material are performed using a nuclear magnetic resonance spectroscopy device.

38. A method according to claim 35 further comprising the step of deselecting undesired spectral information that corresponds to impurities in the material.

39. A method according to claim 35 further comprising the step of normalizing the spectral information with respect to a predetermined point in the spectrum.

40. A method according to claim 35 in which the generated information indicates the material under analysis.

41. A method according to claim 35 in which the generated information indicates portions of the material under analysis.

57. A method according to claim 35 in which the generated information corresponds to close structural relatives to the material under analysis.

42. A method for analyzing and identifying the structure of a particular organic material by recognizing patterns in free induction decay information that are characteristic of such materials, comprising the steps of:

(a) subjecting the organic material under analysis to energy in a nuclear magnetic resonance device;

(b) sensing transformations in the energy imparted by the material;

(c) producing free induction decay information corresponding to the energy transformations and the structure of the material;

(d) selecting portions of the free induction decay information which are desired for use in identifying the material;

(e) digitizing a plurality of incremental portions of the selected free induction decay information;

(f) supplying an off-line neural network which comprises;

(i) an input layer comprising a plurality of input nodes, each of which is capable of receiving digital data corresponding to an incremental portion of the free induction decay information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer; and

each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(iv) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(g) applying the digital data corresponding to incremental portions of the free induction decay information relating to the material to the input nodes of the off-line neural network in order to generate at the output nodes information that is useful to indicate and identify the structure of the material under analysis.

43. A method according to claim 42 in which the step of supplying at least one off-line neural network comprises supplying at least one off-line neural network which includes at least one hidden layer hierarchically intermediate the input and output layers comprising a plurality of hidden nodes.

44. A method for analyzing and identifying the structure of a particular organic material by recognizing patterns in spectral information that are characteristic of such materials, comprising the steps of:

(a) subjecting the organic material under analysis to energy in at least two analytical means;

(b) sensing transformations in the energy imparted by the material;

(c) producing spectral information corresponding to the energy transformations and the structure of the material;

(d) selecting portions of the spectral information which are desired for use in identifying the material;

(e) digitizing a plurality of incremental portions of the selected spectral information;

(f) supplying at least one off-line neural network which comprises;

(i) at least one input layer comprising a plurality of input nodes, each of which is capable of receiving digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis; and

(iii) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer; and

each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(iv) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(g) applying the digital data corresponding to incremental portions of the spectral information relating to the material to the input nodes of the off-line neural network in order generate information at the output nodes that is useful to indicate and identify the structure of the material under analysis.

45. A method according to claim 44 in which the step of supplying at least one off-line neural network comprises supplying an off-line neural network that includes at least one hidden layer hierarchically intermediate an input and the output layer comprising a plurality of hidden nodes.

46. A method according to claim 44 in which the analytical means comprise gas chromatograph and mass spectrograph means.

47. A method according to claim 44 in which the off-line neural network contains two input layers of neurons, one input layer corresponding to mass spectral information and the other input layer corresponding to gas chromatograph information.

48. A method according to claim 45 in which the off-line neural network contains a single hidden layer, and the neurons in the mass spectral information input layer and the neurons in the gas chromatograph information input layer are each connected to each neuron in the hidden layer.

49. A method according to claim 45 in which the off-line neural network contains two hidden layers, a first in which each neuron is connected to every neuron in the mass spectral information input layer and a second in which each neuron is connected to every neuron in the gas chromatograph information input layer, and the neurons in each hidden layer are each connected to each neuron in the output layer.

50. A method for analyzing and identifying the structure of a particular organic material by recognizing patterns in spectral information that are characteristic of such materials, comprising the steps of:

(a) subjecting the organic material under analysis to a first and a second type of energy in at least one spectroscopy device;

(b) sensing transformations in the energy imparted by the material in the device;

(c) producing a first and second set of spectral information corresponding to the transformations in the first and second energy types and to the structure of the material;

(d) selecting portions of the spectral information which are desired for use in identifying the material;

(e) digitizing a plurality of incremental portions of the selected spectral information;

(f) supplying at least one off-line neural network which comprises;

(i) at least one input layer comprising a plurality of input nodes, each of which is capable of receiving digital data corresponding to an incremental portion of the spectral information;

(ii) an output layer hierarchically lower than the input layer comprising a plurality of output nodes, for indicating and identifying the structure of the material under analysis;

(iii) a plurality of synaptic connections, each of which connects a node in a hierarchically higher layer to a plurality of nodes in a hierarchically lower layer; and

each of which features a synaptic strength value which has been generated during a back-propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(iv) each of the nodes featuring a threshold value which has been generated during a back propagation learning process using spectral data from organic materials analogous to the organic material under analysis; and

(g) applying the digital data corresponding to incremental portions of the spectral information relating to the material to the input nodes of the neural network in order to generate information at the output nodes that is useful to indicate and identify the structure of the material under analysis.

51. A method according to claim 50 in which the step of supplying at least one off-line neural network comprises supplying at least one off-line neural network which includes at least one hidden layer hierarchically intermediate an input and the output layer comprising a plurality of hidden nodes.

52. A method according to claim 50 in which the off-line neural network contains two input layers of neurons, one input layer corresponding to the first information set and the other input layer corresponding to the second information set.

53. A method according to claim 51 in which the off-line neural network contains a single hidden layer, and the neurons in the first information set input layer and the neurons in the second information set input layer are each connected to each neuron in the hidden layer.

54. A method according to claim 51 in which the off-line neural network contains two hidden layers, a first in which each neuron is connected to every neuron in the first information set input layer and a second in which each neuron is connected to every neuron in the second information set input layer, and the neurons in each hidden layer are each connected to each neuron in the output layer.

55. A method according to claim 50 in which steps (a) through (c) are performed in separate spectroscopy devices.