JP2913106B2 - Infrared spectrum identification method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an infrared spectrum identification method, and more particularly, to an improved infrared spectrum identification method for performing identification analysis of unknown compounds using a composite neural network.

[Conventional technology]

An infrared spectrum, which is an absorption or emission spectrum of a substance appearing in an infrared wavelength region, is a powerful means for identifying and analyzing a substance. In the infrared wavelength region, transitions between vibration energy levels of molecules, that is, vibration spectra mainly appear, and the natural frequency of a molecule differs depending on its chemical structure, so that the infrared spectrum also differs depending on the substance. For this reason, the infrared spectrum, together with the Raman effect, is widely used in a method for identifying and analyzing a compound. As a conventional method to help identify and analyze unknown compounds from the spectral data obtained by the infrared spectrometer, the infrared spectrum data is stored in the mainframe as a database, and the absorption peak of the unknown compound is calculated from the infrared spectrum of the unknown compound. There is a method in which the wave number and transmittance are artificially read, and a database search is performed using these as keys.

[Problems to be solved by the invention]

However, the database-based method not only requires a huge computer system for storing / retrieving data, but also has a disadvantage that a lot of time is spent for retrieval processing. Further, the wave number and transmittance of the absorption peak, which are the keys of the search, are artificially read out from the spectrum chart, which is labor-intensive and further causes human error.

Accordingly, an object of the present invention is to provide an excellent infrared spectrum identification method that can solve the above-mentioned difficulties of the conventional database-based method and can perform substance identification and analysis in a short time without human intervention. It is to be.

[Means for solving the problem]

In order to achieve the above object, the infrared spectrum identification method of the present invention comprises: a first step of inputting spectral data of samples of the number of categories to be classified and learning the different categories into a neural network; The spectral data of the additional sample is input to the neural network trained in the step, and if the output value from the neural network does not reach a certain value or more, the learning of the spectral data is not performed, and the output value is constant. For those that have reached the above, a second step of learning the category for the spectrum data by the neural network, and a step of learning the sample not learned in the second step by the neural network trained in the second step. Input spectrum data and output value exceeds a certain level For those that have not, the learning of the spectrum data is not performed, and for those whose output values have reached a certain value or more, the third step of learning the category for the spectrum data by the neural network is repeated. Using the neural network of the former stage and the neural network of the latter stage which is trained to output the score for each compound using the output from the neural network of the previous stage as learning data, and using each compound as teacher data; Is input to perform identification analysis of unknown compounds.

[Action]

The infrared spectrum identification method of the present invention is broadly divided into a neural network learning phase and a neural network recognition phase. In the former phase, neural network learning is performed using the infrared spectrum data of the standard sample, and the data is classified into multiple categories according to the similarity in multiple stages, and finally individual data can be identified A complex neural network is constructed. In the latter phase, the measured infrared spectrum data of the unknown compound is directly input, not artificially, to the composite neural network constructed as described above, and data recognition is performed by the learned neural network. Then, the output of the recognition result is compared with the output value of the infrared spectrum of each standard sample, and a mark is calculated for each standard sample, and a candidate compound is obtained based on the mark.

〔Example〕

Hereinafter, the present invention will be described in detail based on embodiments with reference to the drawings.

FIG. 1 is a conceptual diagram showing the structure of a neural network (NN) which is the basis of the system used in the present invention, and FIG. 2 is a conceptual diagram showing the structure of the entire complex neural network formed by combining individual neural networks. is there. In the present invention, as shown in FIG. 2, a plurality of neural nets are combined in a multi-stage (hierarchical) manner to constitute the whole. However, as shown in FIG. , A perceptron model having a three-layer structure of an intermediate layer and an output layer. As the scale of the neural network, the units of each layer are appropriately set in order to keep the calculation speed sufficiently fast. In this embodiment, the input layer is about 300 units, the intermediate layer is about 10 units, and the output layer is about 10 units. . The function of each neural network is to input n (10 in this example) infrared spectrum data of the input unknown compound.
Is classified into any of the categories. That is, in the first stage neural network, unknown data is 1 /
n (1/10 in this example), the possibility of 1 / m (1/10 in this example) in the second stage neural network, It is designed to output a score for.

In the present invention, since a neural network is used, the work flow can be divided into a phase of “learning a neural network” and a phase of “recognition using a neural network”. In identifying an unknown compound, only the recognition phase may be used, but the learning phase is indispensable for the construction of the system. Therefore, in the following, it is divided into the learning phase and the recognition phase,
The method of the present invention will be described.

Neural Network Learning Phase Of the neural nets used in the present invention, those other than the last stage are used for pattern classification. FIG. 3 is a flowchart showing the procedure of the method for learning the neural network.

In the present invention, first, data for learning a neural network is created using infrared spectrum data of a standard sample. Ten data are randomly selected from these data (representative data), and the neural network learns (step). Next, the trained network is made to recognize the pattern of the remaining data, and the pattern whose output value has reached a certain value (generally, a value between 0 and 1 and empirically set to 0.5) is extracted. (Step). Then, the pattern extracted in the step is added to the learning data, and learning is performed again. In this process, similar patterns are gathered into one category. For a category in which the collection of patterns does not progress, the representative data is changed, and the same operation is repeated. And finally all compounds are 10
Steps 1 to 5 are repeated until the category can be classified. If a pattern does not belong to any category, it is judged by a person and included in some category. Finally, learning is performed using the infrared spectrum data of all the standard samples to complete the composite network (step).

In the final stage neural network, a general learning method for pattern recognition is used. That is, the third
The data belonging to the same category in the network of the stage is used as training data, and the individual compounds are used as teacher data.
Learning is performed so as to output a score for each compound.

Phase of Recognition by Neuralnet The procedure of a method for identifying infrared spectrum data of an unknown compound using the neural network constructed as described above is shown in the flowchart of FIG.

First, an infrared spectrum pattern of an unknown compound is read into the first-stage neural network. For example, when the input layer of the neural network is set to 300 units, the reading of data of 300 points on the infrared spectrum pattern is performed directly from the infrared spectrometer (without artificially reading) from each of the input layers. This is done by reading it into the corresponding unit. When the reading is completed, the learned neural network performs a calculation based on the learning data and outputs a value to an output layer (step). The score is calculated based on the values obtained here, the next-stage neural net to be used next is selected (step), and the selected neural network recognizes the pattern data of the unknown compound again (step). The above operation is repeated until the final neural network is reached. Finally, a score is calculated based on the value output by the final-stage neural network, and a compound name of an unknown pattern is obtained (step). In this case, the score is naturally 1 if the unknown compound is the same as the standard sample. Note that, depending on the pattern data of an unknown compound, it may not reach a single compound in some cases. In that case, candidate compounds are determined based on the scores. As described above, the identification analysis of an unknown compound is performed at high speed and without requiring a huge computer system.

As described above, the present invention has been described based on one embodiment. However, the present invention is not limited to the above embodiment, and various modifications and changes can be made. For example, the unit of each layer of the neural network can be set to an appropriate number of units as long as the calculation speed is kept fast enough, and the number of categories and the number of stages of the composite neural network should be set to appropriate numbers according to the number of standard samples. Can be.

〔The invention's effect〕

In the present invention, by using a neural network to determine the score by comparing the infrared spectrum data of the unknown compound with the infrared spectrum data of the standard sample,
Compared with the conventional database search method, not only does a huge computer system for storing / retrieving data become unnecessary, but also because of the principle feature of the neural network,
Processing time can be greatly reduced. Further, there is an effect that it is possible to prevent human error and to avoid individual differences in determination without having to artificially read out the wave number and transmittance of the absorption peak, which are the keys of the search, from the spectrum chart.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram showing the structure of each neural network. FIG. 2 is a conceptual diagram showing the structure of a composite neural network (the entire system). FIG. 3 is a flowchart showing a learning method for classifying the infrared spectrum patterns of the standard sample into categories. FIG. 4 is a flowchart showing a method of recognizing a pattern of an unknown compound and searching for a candidate compound.

Continuation of front page (56) References Appl. qot. 28 (15) (1989) p3129-3133. / IEICE Technical Report MBE 87-156 (1986) p. 407-414 Proc IEEE Int. Conf Aconst Speech Speech Sigrol Procedures Vol 1989. No. Vol. 3 (1989) p1775-1778. Mag. Res. 84 (1) (1989) p212-217 (58) Fields investigated (Int. Cl. ⁶ , DB name) G01N 21/00-21/01 G01N 21/17-21/74 G06F 15/18 G01J 3/78 JOIS

Claims (1)

(57) [Claims] 1. A first step of inputting spectral data of samples corresponding to the number of categories to be classified and learning the different categories in a neural network, and an additional step in the neural network learned in the first step. If the spectral data of the sample is input and the output value from the neural network does not reach a certain value or more, the learning of the spectral data is not performed. A second step of learning the category for the neural network in the neural network; and inputting the spectral data of the unlearned sample in the second step to the neural network trained in the second step; For those that have not been reached, the spectral data A neural network of a former stage constructed by repeating a third step of causing a neural network to learn a category for the spectral data for an output value having reached a certain value or more without performing learning; a neural network of a preceding stage; Is used as learning data, the individual compound is used as teacher data, and a neural network at the subsequent stage trained to output a score for each compound is used; spectral data of the unknown compound is input, An infrared spectrum identification method, characterized in that identification analysis is performed.