JP7262345B2 - Material inspection device, material inspection method and material inspection program - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act (1) Date of posting on the website March 13, 2019 (2) Website address https://spie. org/ESD/conferencedetails/artificial-intelligence-and-machine-learning-in-defense-applications? SSO = 1 (3) Publisher Takayuki Higo Shuzo Eto

The present invention relates to a material inspection device, a material inspection method, and a material inspection program.

In recent years, by applying light such as a laser to an object and analyzing the characteristics of the absorbed light and scattered light of the object, it is possible to identify the components of the object or detect that the object contains specific molecules. The use of technology is flourishing. Identifying a component of an object is called identification, and determining the inclusion of a particular molecule in the object is called sensing. Here, they are collectively called identification/detection.

Absorbed light and scattered light of an object have characteristics such that light of a specific wavelength is easily absorbed or the generated scattered light has a strong specific wavelength, depending on the substance that constitutes the object. Therefore, by analyzing the intensity of each wavelength of absorbed light and scattered light, it is possible to identify and detect the object. Here, the intensity of each wavelength of absorbed light or scattered light is called a "spectrum".

Spectral analysis for identification/detection includes the following conventional techniques. In the following, a case where there is a database in which the spectra of absorbed light and scattered light when irradiated with laser exist for each of various substances will be described. The data of each spectrum in the database can be obtained, for example, by irradiating various material samples with light such as laser light in a laboratory environment and measuring the spectrum of absorbed light and scattered light, or by computer simulation.

A substance inspection device that identifies and detects a substance irradiates a substance to be inspected with a laser beam and obtains scattered light from the substance to be inspected. Then, the substance inspection device acquires the spectrum of the substance to be inspected using the acquired scattered light. Thereafter, the substance inspection apparatus obtains a weight that minimizes the difference between the spectrum of the substance to be inspected and the weighted linear sum of the spectra of various substances stored in the database. When there are P types of substances stored in the database and N types of wavelengths forming the spectrum, the calculation of minimization of L is performed by multiple regression on N data with P explanatory variables. is the same as the calculation of Such analytical methods are called classical least squares (CLS).

In addition, as a method of identification and detection, spectral data is repeatedly collected from the scattered light and fluorescence from the object for a certain period of time, and the collected spectral data is analyzed using multivariate analysis to extract Raman scattered light. There is prior art. In addition, there is a prior art that repeatedly decomposes and refines component spectra acquired from an object for identification. Algorithms that incorporate unmodeled components or other spectral information into the multivariate model, maintain the results of calibration with the multivariate model, and iteratively identify the components by estimating them from residual errors. is proposed. Also, a data set of the spectrum of the object is obtained, a data set is obtained for the pure components of the spectrum, the obtained data set is used to determine the concentration coefficients, the spectral shift is performed, the residual error is calculated, and the residual error is calculated. Algorithms have been proposed that repeat the shift until the difference is within a predetermined value.

JP 2017-129389 A Japanese Patent No. 5059106 U.S. Pat. No. 6,711,503 U.S. Pat. No. 6,418,383

However, conventional techniques such as general CLS can obtain highly accurate results when the noise contained in the spectrum of the object or the spectrum of each substance is small, but when the noise is large, erroneous results are obtained. may give. A situation in which this noise is large may be, for example, a case of identifying/detecting a substance from a long distance. In such situations, noise robust spectral analysis techniques are required.

The disclosed technology has been made in view of the above, and aims to provide a material inspection device, a material inspection method, and a material inspection program that perform highly accurate identification and detection even in an environment where noise is assumed. and

In one aspect of the material inspection device, material inspection method, and material inspection program disclosed in the present application, the light emitting unit irradiates the object with light. The measurement unit measures one or both of scattered light and absorbed light from the object irradiated with the light, and obtains a spectrum representing the intensity of the measured light at a plurality of wavelengths. The spectral analysis unit calculates a weight that minimizes the difference between a linear sum of spectra of a plurality of known reference substances weighted for each reference substance and a predetermined spectrum under the condition that the weight is non-negative. perform a spectral decomposition that The spectrum wavelength selection unit selects one or more working wavelengths from among the plurality of wavelengths in the spectrum based on the analysis error of the spectrum decomposition, sets the spectrum of the object as the predetermined spectrum, and selects the The spectral analysis unit is caused to perform spectral decomposition using a limited spectrum representing the light intensity of the object for each wavelength used. The identification detection unit identifies or detects the object and the reference substance based on the weight obtained by spectral decomposition of the spectrum of the object using the restricted spectrum by the spectrum analysis unit.

In one aspect, the present invention can perform highly accurate identification and detection even in an environment where noise is expected.

FIG. 1 is a block diagram of a material inspection device according to a first embodiment. FIG. 2 is a diagram showing errors associated with measurement. FIG. 3 is a diagram for explaining an outline of selection of spectrum wavelengths. FIG. 4 is a diagram showing an example of selection results of spectrum wavelengths. FIG. 5 is a diagram showing a linear sum of spectra of reference substances with weighting factors added. FIG. 6 is a diagram for explaining the linear sum of the spectra of each reference substance when using the selected spectral wavelengths. FIG. 7 is a flowchart of identification/detection processing by the substance inspection device according to the first embodiment. FIG. 8 is a flowchart of spectrum wavelength selection processing by the spectrum wavelength selection unit and the spectrum analysis unit. FIG. 9 is a block diagram of a material inspection device according to the second embodiment. FIG. 10 is a diagram showing the relationship between the laser wavelength and the spectrum of a specific substance. FIG. 11 is a diagram showing an example of analysis error for each laser wavelength. FIG. 12 is a diagram showing an example of identification results. FIG. 13 is a flowchart of identification/detection processing by the substance inspection device according to the second embodiment.

Embodiments of the material inspection device, the material inspection method, and the material inspection program disclosed in the present application will be described in detail below with reference to the drawings. The material inspection device, material inspection method, and material inspection program disclosed in the present application are not limited to the following embodiments. Further, in the following description, identification and detection of substances using scattered light will be mainly described, but identification and detection of substances using absorbed light are also the same.

FIG. 1 is a block diagram of a material inspection device according to a first embodiment. The material inspection device 1 according to this embodiment has a light emitting unit 11, a measuring unit 12, a notification unit 13, an identification detection unit 14, a spectrum analysis unit 15, a spectrum wavelength selection unit 16 and a database 17.

The database 17 holds information in which the spectra of absorption and scattered light when irradiated with laser are arranged for each of various types of substances. A substance whose information is stored in the database 17 is hereinafter referred to as a “reference substance”. For example, reference substances include _CO2 , _O2 , _N2 , _H2 , _CH4 , _NH3 and _SO2 . In this embodiment, information on spectra of P reference substances #p (p=1 to P) is stored in the database 17 .

The spectrum of each reference substance #p includes the intensity of scattered light for each wavelength when irradiated with a laser of a predetermined type of wavelength. In this way, the spectrum of a specific substance can be obtained as a separate spectrum for each wavelength. Therefore, wavelengths corresponding to individual spectra are sometimes called "spectral wavelengths". In the following, below the spectrum, the wavelength of the irradiated laser may be referred to as "laser wavelength" in order to clearly indicate that it is the wavelength possessed by the laser.

If there are N types of spectral wavelengths, the information on the spectrum of each reference substance #p includes N individual spectra corresponding to each spectral wavelength as elements. That is, the spectrum is represented by an N-dimensional vector when there are N spectral wavelengths. xp (p=1 to P), which is information on the spectrum of the reference substance, is represented by an N-dimensional column vector having N elements x1p, x2p, . . . , xNp. Each element of the column vector representing the spectrum corresponds to an individual spectrum for each spectral wavelength. For example, there are 1000 spectral wavelengths.

The light emitting unit 11 receives from the identification detecting unit 14 an instruction to irradiate a laser beam along with the designation of the laser wavelength. The light emitting unit 11 receives an instruction from the identification detection unit 14 and irradiates the inspection object Z with a laser having a specified wavelength. The object to be inspected Z emits scattered light upon being irradiated with the laser.

The measurement unit 12 receives scattered light emitted from the inspection object Z. As shown in FIG. Then, the measurement unit 12 measures the individual spectrum for each wavelength of the inspection object Z obtained from the acquired scattered light, and outputs information on the spectrum of the inspection object Z obtained from the measurement result to the spectrum analysis unit 15 . In this embodiment, the spectrum of scattered light will be described as an example, but the spectrum of absorbed light may be used. In this case, the measurement unit 12 measures the absorbed light absorbed by the inspection object Z from the scattered light emitted from the inspection object Z, and converts the spectral information of the absorption light of the inspection object Z, which is the measurement result, into a spectrum Output to the analysis unit 15 .

Next, the spectral wavelength selector 16 will be described. The spectrum of a certain inspection target substance Z is information on the intensity of scattered light emitted from a certain inspection target substance Z for each wavelength and the collection thereof. The spectra of absorbed light and scattered light tend to show characteristics at specific spectral wavelengths for each reference substance #p. In addition, individual spectra corresponding to spectral wavelengths other than the spectral wavelength at which the characteristic of each reference substance #p appears may become noise during identification and detection. Therefore, it is preferable to perform spectral decomposition using an individual spectrum having a spectral wavelength in which the characteristic of each reference substance #p appears, in order to improve the accuracy of identification/detection of the inspection target substance Z. The “spectral decomposition” here is a process of decomposing the spectrum of the inspection object Z into the spectrum of the reference material #p registered in the database 17 . By spectrum decomposition, the weight when the inspection target substance Z is represented by a linear sum of weighted reference substances #p is obtained.

The spectrum wavelength selection unit 16 selects a spectrum wavelength used when identifying/detecting the substance Z to be inspected. Specifically, the spectral wavelength selection unit 16 acquires information on the spectrum of each reference material #p from the database 17, and generates a column vector representing each reference material #p using individual spectrum values. Next, the spectral wavelength selector 16 randomly generates weights for each reference material #p as test weights.

For example, if the reference substances #p are CO ₂ , O ₂ , N ₂ , H ₂ , CH ₄ , NH ₃ and SO ₂ , the spectral wavelength selection unit 16 selects 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, 0.18, and 0.18, respectively. Test weights such as 07, 0.06, 0.14, 0.19, 0.11 and 0.25 are generated based on random numbers.

Next, the spectral wavelength selector 16 adds the test weight to the column vector representing each reference material #p, and obtains the linear sum of the column vectors representing each reference material #p to which the test weights have been added.

Next, the spectral wavelength selector 16 randomly generates noise components. For example, the spectrum wavelength selection unit 16 uses the standard deviation of the intensity of each spectrum wavelength for each spectrum wavelength included in the spectrum obtained by linearly summing the weighted spectrum of the reference substance #p to obtain a normal random number generates noise.

FIG. 2 is a diagram showing errors associated with measurement. In FIG. 2, the vertical axis represents spectral intensity, and the horizontal axis represents spectral wavelength. FIG. 2 shows the relationship between the spectrum wavelength and the spectrum intensity when SO ₂ is the inspection object Z and is irradiated with a laser having a wavelength of 202.13 nm. For example, a case where the measurement result of the relationship between the spectrum wavelength and the spectrum intensity measured in advance is obtained as shown in FIG. 2 will be described. Graphs 101 to 103 in FIG. 2 represent the maximum value of the measurement results, the average value of the measurement results, and the minimum value of the measurement results, respectively. Assuming that each measurement result is represented by a normal distribution, the standard deviation is obtained so that the difference between the graphs 101 and 102 and the difference between the graphs 102 and 103 are twice the standard deviation. Since the standard deviation obtained from FIG. 2 is 14%, in this embodiment, the spectrum wavelength selection unit 16 sets the standard deviation to 14% of the intensity of each spectrum wavelength, and generates noise using normal random numbers. .

Then, the spectrum wavelength selection unit 16 adds the generated noise to the linear sum of the test-weighted spectrum of each reference material #p to synthesize a test spectrum. The spectrum wavelength selection unit 16 then transmits the synthesized test spectrum to the spectrum analysis unit 15 and requests spectrum decomposition. Details of spectrum decomposition by the spectrum analysis unit 15 will be described later.

After that, the spectrum wavelength selection unit 16 acquires the results of spectral decomposition of the test spectrum from the spectrum analysis unit 15 . In this case, spectral wavelength selector 16 obtains an estimated weight for each reference material #p as a result of spectral decomposition. Then, the spectral wavelength selection unit 16 selects the spectral wavelength so as to minimize the analysis error calculated from the difference between the estimated weight and the test weight.

For example, if α _p is the test weight for the P reference substances #p and β _p is the estimated weight for the P reference substances, the spectral wavelength selection unit 16 determines the relationship between the estimated weight and the test weight. An analytical error is obtained by the following formula (1).

E represents the analytical error between the estimated weights and the test weights. |α _p −β _p | represents the absolute value of α _p −β _p .

Expression (1) is for the case of using one test weight, but it is also possible to use a plurality of test weights to obtain the minimum value of the analysis error. For example, when using K test weights, the analytical error is expressed by the following equation (2).

The k subscripted α _p represents the kth test weight. Also, the k subscripted β _p represents the estimated weights when the spectral decomposition is performed using the kth test weights.

Here, the minimization of the analysis error between the estimated weights and the test weights by the spectral wavelength selector 16 according to this embodiment will be described in detail. When the spectral information of the reference substance #p stored in the database 17 is represented by the intensity of scattered light corresponding to N kinds of spectral wavelengths, there are 2 N spectral wavelength combinations. Therefore, the spectrum wavelength selection unit 16 can calculate the analysis error for all combinations of 2 N spectrum wavelengths and find a method of selecting the spectrum wavelength that minimizes the analysis error.

However, when obtaining all combinations of spectral wavelengths, an exponential calculation time is required for N representing the type of reference material #p. Therefore, the method of finding the analysis error for all combinations of spectral wavelengths is not practical when N is large, and it is difficult to select the spectral wavelength that minimizes the analytical error. For example, if it takes 1 second to calculate one analysis error, and N=1000, the calculation time is about 10 to the 300th power.

In order to shorten this calculation time, the spectral wavelength selection unit 16 according to the present embodiment selects the spectral wavelength that minimizes the analysis error using the greedy method. When the greedy method is used, the spectral wavelength selection unit 16 adds spectral wavelengths used for spectral decomposition one by one, and terminates the process when analysis error stops decreasing. The spectral wavelength obtained by the analysis is selected as the spectral wavelength at which the analysis error is minimized. As the spectrum wavelength to be added, the spectrum wavelength selection unit 16 selects a spectrum wavelength that maximizes the decrease in analysis error when adding a new spectrum wavelength to the already added spectrum wavelengths.

More specifically, the spectrum wavelength selection unit 16 according to the present embodiment executes the following calculation as spectrum wavelength selection processing. Here, the N kinds of spectral wavelengths are represented by λ1, λ2, . . . , λN. Also, let t be the number of repetitions of the computation for obtaining the analysis error. Also, let St be a set of spectrum wavelengths already selected in the t-th computation. Also, let Ft be the provisional analysis error obtained in the previous calculation used in the t-th calculation. The spectrum wavelength selection unit 16 sets S1, which is a set of already selected spectrum wavelengths used in the first calculation, to an empty set, and sets F1, which is an analysis error used in the first calculation, to infinity.

Then, the spectral wavelength selection unit 16 obtains spectral wavelengths λi remaining after removing the already selected spectral wavelengths from all spectral wavelengths. For example, if λ1, λ100 have already been selected, λi=(λ2, λ3, . . . , λ99, λ101, .

Next, the spectrum wavelength selection unit 16 requests the spectrum analysis unit 15 to perform spectrum decomposition by adding the newly selected spectrum wavelength to the already selected spectrum wavelengths. After that, the spectrum wavelength selection unit 16 acquires from the spectrum analysis unit 15 each estimated weight for each spectrum wavelength included in the set of spectrum wavelengths when each λi is newly added. Then, the spectral wavelength selector 16 calculates E(St∪λi). Here, E(St ∪ λi) is the estimated weights and test weights when spectral decomposition is performed using the spectral wavelengths obtained by adding λi to the set St. represents the analysis error due to the t-th calculation obtained by

If Ft≦E(St∪λi*), the spectral wavelength selection unit 16 terminates the process and selects spectral wavelengths other than those included in λi as spectral wavelengths that minimize the analysis error. Here, λi* is any one spectral wavelength among λi that minimizes E(St∪λi).

On the other hand, if Ft>E(St∪λi*), the spectral wavelength selector 16 sets St+1=St∪λi. Then, with t=t+1, the calculation of E(St∪λi) and the comparison with Ft are repeated.

Now, with reference to FIG. 3, a general flow of spectrum wavelength selection processing by the spectrum wavelength selection unit 16 will be described. FIG. 3 is a diagram for explaining an outline of selection of spectrum wavelengths. Here, the case where 1000 types of spectral wavelengths Sp1 to Sp1000 are available will be described. In practice, the spectrum wavelength selection unit 16 adds a spectrum wavelength for each calculation to obtain an analysis error. However, in the following description, calculation of estimated weights by the spectrum analysis unit 15 will be omitted.

In the first calculation, the spectrum wavelength selection unit 16 obtains analysis errors Ae1-1 to Ae1000-1 when all the spectrum wavelengths Sp1 to Sp1000 are used one by one. Then, the spectrum wavelength selection unit 16 determines that the analysis error Ae100-1 when using the spectrum wavelength Sp100 among the analysis errors Ae1-1 to Ae1000-1 is the smallest, and selects the spectrum wavelength Sp100 as the spectrum wavelength. Decide to add.

Next, the spectrum wavelength selection unit 16 selects one by one from the spectrum wavelengths Sp1 to Sp1000 excluding the spectrum wavelength Sp100, adds them to the spectrum wavelength Sp100, performs the second calculation, and calculates each analysis error. demand. Here, the spectrum wavelength selection unit 16 obtains the analysis errors Ae1-2 to 99-2 and Ae101-2 to Ae1000-1 when the spectrum wavelengths Sp1 to Sp9 and Sp101 to Sp1000 are added to the spectrum wavelength Sp100. Then, the spectrum wavelength selection unit 16 determines that the analysis error Ae1-1 when using the spectrum wavelength Sp1 among the analysis errors Ae1-2 to Ae99-2 and Ae101-Ae2 to Ae1000-1 is the smallest, and the spectrum It is decided to add the wavelength Sp1 to the spectral wavelengths of choice. Here, the case of F1>E (S1∪Sp1) will be described.

The spectral wavelength selection unit 16 selects spectral wavelengths one by one from the spectral wavelengths Sp1 to Sp1000 excluding the spectral wavelengths Sp1 and Sp100, adds them to the spectral wavelengths Sp1 and Sp100, and performs the third calculation. Find the analytical error of Here, the spectrum wavelength selection unit 16 obtains the analysis errors Ae2-3 to Ae99-3 and Ae101-3 to Ae1000-3 when the spectrum wavelengths Sp2 to Sp99 and Sp101 to Sp1000 are added to the spectrum wavelengths Sp1 and Sp100. . Then, the spectrum wavelength selection unit 16 specifies the analysis error with the smallest value from among the analysis errors Ae2-3 to Ae99-3 and Ae101-3 to Ae1000-3. In this manner, the spectrum wavelength selection unit 16 repeats the selection of spectrum wavelengths. Then, when the minimum value of the analysis error becomes larger than the provisional analysis error, the spectrum wavelength selection unit 16 ends the addition of spectrum wavelengths and uses the spectrum wavelengths selected so far for spectrum decomposition. is determined as the spectral wavelength at which The spectrum wavelength used for spectrum decomposition determined by the spectrum wavelength selection unit 16 corresponds to an example of the "use wavelength".

FIG. 4 is a diagram showing an example of selection results of spectrum wavelengths. In FIG. 4, the vertical axis represents the laser wavelength, and the horizontal axis represents the wave number corresponding to the spectral wavelength. Wavenumbers in FIG. 4 are represented by Raman shifts. In FIG. 4 , spectral wavelengths selected by the spectral wavelength selection unit 16 by the greedy method when the laser of each wavelength is irradiated are shown for each laser wavelength on the vertical axis. In this case, 7 to 33 spectral wavelengths are selected for each laser wavelength.

Returning to FIG. 1, the description continues. The spectrum analysis unit 15 receives input of information on the inspection object Z spectrum from the measurement unit 12 . Further, the spectral analysis unit 15 receives from the spectral wavelength selection unit 16 input of information on the spectral wavelength used for identification/detection of the inspection object Z selected according to the wavelength of the irradiated laser. Then, the spectral analysis unit 15 performs spectral decomposition to obtain the weight β _p that minimizes the difference between the spectrum of the inspection object Z and the linear sum of the weighted spectrum of the reference material. Details of spectrum decomposition by the spectrum analysis unit 15 will be described below.

If N spectral wavelengths are used, the spectrum is the intensity for the N spectral wavelengths and is represented by an N-dimensional vector. Each element of this vector corresponds to a separate spectrum. Here, the spectra of the P types of reference substances #p stored in the database 17 are respectively represented by column vectors xp (p=1 to P), and the spectrum of the inspection object Z is represented by the vector y. y is also represented by an N-dimensional vector.

The process of searching for a spectrum similar to y from the column vector xp (p = 1 to P) is realized by searching for the weight so that the difference between the weighted linear sum of the column vector xp and y is minimized. It is possible. This search is represented by the following formula (3).

where β _p (p=1 to P) is the weight for each reference material #p. Σβ _p ×xp is the weighted linear sum of the spectra of the reference materials stored in the database. Σβ _p ×xp is hereinafter referred to as the linear sum of xp. L represents the difference between y and the linear sum of xp, here called the "residual". Also, minL represents a function that searches for a weight β _p that minimizes this residual. β is a column matrix whose elements are weights β _p (p=1 to P), and is hereinafter referred to as a weight matrix. This operation can be regarded as decomposing the spectrum of the inspection object Z into the spectrum of the reference substance, and this operation corresponds to spectral decomposition.

L(β) in Equation (3) above represents the L2 norm of the difference between the linear sum of xp and y as a residual. The L2 norm is the square root of the sum of squares of each element of the vector, and is represented by the following formula (4). In this embodiment, the spectral analysis unit 15 uses the L2 norm when obtaining residuals. Here, d is d=(d ₍₁₎ , d ₍₂₎ , . . . , d _(P) ).

Also, as the residual L, the L1 norm can be used instead of the L2 norm. The L1 norm is represented by the following formula (5).

For example, the weights that minimize L when y = x1 can be expected to be β1 = 1 and βq = (q = 2 to P). interpretable. In general, 0<βp indicates that the test substance Z contains the reference substance #p, and 0=βp indicates that the test substance Z does not contain the reference substance #p. Also, the magnitude of the coefficient can be interpreted as the concentration (ratio) of the reference substance #p. It should be noted that the magnitude of the coefficient and the concentration can be approximated by a linear relationship, and the slope at that time can be determined from the data obtained in the preliminary experiment.

A calculation method for minimizing the residual when the L2 norm is used as the residual will be described. FIG. 5 is a diagram showing a linear sum of spectra of reference substances with weighting factors added. Both vector y and column vector xp are column vectors. Assuming that a matrix in which P column vectors xp are arranged is X, X=(x1 x2 . . . xp).

Here, Xβ is represented by the determinant shown in FIG. Spectra of each reference substance #1 to #P are represented by column vectors x1 to xp shown in FIG. X is a matrix in which column vectors x1 to xp are arranged. In the matrix X, columns are arranged for each substance type, and rows are arranged for each spectral wavelength. That is, each row of matrix X contains the respective spectrum of each reference material #1-#P at the same spectral wavelength. Each element β ₁ to β _P of the weight matrix β is a weight for each reference substance #p, and is a column vector having the same number of elements as P, which is the number of reference substances #p.

At this time, the residual L is represented by the following formula (6).

Here, the spectrum analysis unit 15 according to this embodiment obtains the residual L using the spectrum corresponding to the spectrum wavelength selected by the spectrum wavelength selection unit 16 . FIG. 6 is a diagram for explaining the linear sum of the spectra of each reference substance when using the selected spectral wavelengths. A spectrum wavelength 301 on an arrow representing a spectrum wavelength in FIG. 6 represents a spectrum wavelength selected by the spectrum wavelength selection unit 16 . As shown in FIG. 5, the spectral analysis unit 15 performs spectral decomposition using the individual spectra, which are the elements of the row 302 corresponding to the selected spectral wavelength 301 of each spectrum of the reference substance #p. For example, if 20 spectral wavelengths are selected, the spectral analysis unit 15 weights the spectrum of each reference material #p determined using the 20 rows 302 corresponding to the selected spectral wavelengths 301 in the matrix X. Perform spectral decomposition using linear summation with

For example, the spectrum analysis unit 15 can obtain the weighting matrix β that minimizes the residual L as β=(X'X) ⁻¹ X'y. Here, the subscript dash represents transposition, and the subscript -1 represents the inverse matrix.

The spectral analysis unit 15 performs multiple regression calculation for N data having P explanatory variables as a calculation for obtaining a weighting matrix β that minimizes L, which is a residual. However, the spectrum analysis unit 15 according to the present embodiment performs multiple regression calculation under the condition that each weight βp, which is a coefficient in Equation (4), is non-negative. Since the spectrum analysis unit 15 calculates the residual L using the L2 norm in this embodiment, the weight matrix β that minimizes the residual L is obtained using the quadratic programming method. Here, when obtaining the residual using the L1 norm, the spectral analysis unit 15 can obtain the weight matrix β that minimizes the residual L by using linear programming.

In this way, the weight βp (here, since the weight βp is the coefficient of the formula, it will be referred to as a “coefficient” for explanation) is non-negative and the multiple regression calculation is performed, thereby facilitating the interpretation of the coefficient. and the calculation can be easily performed. Finding the weight matrix β that minimizes the residual L without imposing the condition that the coefficients are non-negative may result in negative regression coefficients. The problem of interpretation arises. On the other hand, if the condition that the coefficients are non-negative is imposed, there is no problem of interpreting negative regression coefficients, and the spectrum analysis unit 15 can easily perform spectrum decomposition.

In addition, the coefficients can be stabilized by performing multiple regression calculations with non-negative coefficients. For example, the stabilization of coefficients will be described by taking as an example a case where reference material #a and reference material #b having similar spectra are present. Let xa and xb be the spectra of reference material #a and reference material #b, respectively. Let βa and βb be the weights of the reference material #a and the reference material #b, respectively. In this case, the linear sum of these spectra is expressed as β _a ×xa+β _b ×xb.

Due to the similarity of spectra xa and xb, the linear sum of the spectra of reference material #a and reference material #b is approximately zero when β _a =−β _b . In this way, when there are similar spectra, if the coefficients are allowed to be negative, the respective coefficients may be determined such that the spectra cancel each other out, leading to destabilization of the coefficients. This destabilization problem is called multicollinearity in multiple regression. On the other hand, in this embodiment, by imposing the condition that the weights are non-negative, the destabilization of the coefficients is alleviated, and the spectrum analysis unit 15 can appropriately calculate the weights.

After calculating the weight β _p as described above, the spectrum analysis unit 15 outputs information on the calculated weight β _p corresponding to each reference substance #p to the identification detection unit 14 .

Also, the spectrum analysis unit 15 receives the test spectrum from the spectrum wavelength selection unit 16 and receives a request for spectrum decomposition. In this case, the spectrum analysis unit 15 treats the acquired test spectrum information in the same manner as the spectrum information of the inspection object Z, performs spectrum decomposition, and calculates the weight for each reference material #p. However, in the spectral wavelength selection process by the greedy method executed by the spectral wavelength selection unit 16, the spectral analysis unit 15 performs spectral decomposition using the spectral wavelengths included in the set specified by the spectral wavelength selection unit 16 at each calculation stage. to run. The spectrum analysis unit 15 transmits the calculated weight for each reference substance #p to the spectrum wavelength selection unit 16 as an estimated weight.

Returning to FIG. 1, the description continues. The identification detection unit 14 receives an execution instruction for identification and detection of the inspection object Z from an operator from an input unit (not shown). Upon receiving an instruction to execute identification/detection of the inspection object Z, the identification detection unit 14 instructs the light emitting unit 11 to emit laser light.

After that, the identification detection unit 14 calculates the weight _βp of each reference substance #p obtained from the spectrum of the inspection object Z obtained from the scattered light when the inspection object Z is irradiated with the laser having each laser wavelength. Obtained from the spectrum analysis unit 15 .

Here, the weight β _p acquired by the identification detection unit 14 from the spectrum analysis unit 15 is a weight that is the result of spectrum decomposition using the selected spectrum wavelength. That is, the identification detection unit 14 detects the element of the inspection target substance Z corresponding to the selected spectral wavelength, and the spectral linearity of each reference substance _#p corresponding to the selected spectral wavelength to which the weight βp is added. A weight β _p obtained by using the sum is acquired from the spectrum analysis unit 15 . This is a weight obtained using the spectral wavelength at which the characteristic of each reference material #p appears, and can be said to be a weight obtained by reducing noise in spectral decomposition. The spectrum of each reference material #p corresponding to the selected spectral wavelength is an example of a "restricted spectrum."

The identification detection unit 102 executes identification/detection using the acquired weight β _p and information on the reference material #p stored in the database 17 . The identification detection unit 102 can perform accurate identification/detection by using the weight _βp calculated by spectrum decomposition with reduced noise. After that, the identification detection unit 14 outputs the identification/detection result to the notification unit 13 .

The notification unit 13 receives an input of identification/detection results from the identification detection unit 14 . Then, the notification unit 13 causes a display device or the like to output the identification/detection result, and notifies the operator of the identification/detection result.

Next, with reference to FIG. 7, the flow of identification/detection processing by the substance inspection device 1 according to the present embodiment will be described. FIG. 7 is a flowchart of identification/detection processing by the substance inspection device according to the first embodiment.

The spectral wavelength selection unit 16 executes spectral wavelength selection processing (step S1). After that, the spectrum wavelength selection unit 16 notifies the spectrum analysis unit 15 of information on the selected spectrum wavelength.

The identification detection unit 102 receives an execution instruction to identify and detect the inspection object Z from the operator, and instructs the light emission unit 11 to emit laser light. The light emitting unit 11 receives the instruction from the identification detection unit 102 and irradiates the inspection object substance Z with laser light (step S2).

The measurement unit 12 receives the scattered light from the inspection target substance Z (step S3). Then, the measurement unit 12 generates information on the spectrum of the substance Z to be inspected and outputs the information to the spectrum analysis unit 15 .

The spectrum analysis unit 15 receives notification of the selection result of the spectrum wavelength from the spectrum wavelength selection unit 16 . The spectrum analysis unit 15 also receives input of information on the spectrum of the substance Z to be inspected from the measurement unit 12 . Then, using the selected spectral wavelength, the spectral analysis unit 15 imposes a condition that the weight β _p is non-negative, and the linear sum of the spectrum of the inspection target substance Z and the spectrum of the reference substance to which the weight β _p is added. Spectral decomposition is performed by finding the weight that minimizes the difference (step S4). After that, spectrum analysis section 15 outputs information on calculated weight β _p to identification detection section 14 .

The identification detection unit 14 receives input of information on the calculated weight β _p from the spectrum analysis unit 15 . Then, the identification detection unit 14 uses the spectral information of the reference substance #p stored in the database 17 to identify and detect the inspection target substance Z (step S5). After that, the notification unit 13 notifies the operator of the identification/detection result.

Next, the flow of spectrum wavelength selection processing will be described with reference to FIG. FIG. 8 is a flowchart of spectrum wavelength selection processing by the spectrum wavelength selection unit and the spectrum analysis unit. The process shown in the flowchart of FIG. 8 corresponds to an example of the process performed in step S1 of FIG.

The spectral wavelength selector 16 generates test weights using random numbers (step S101).

Next, the spectrum wavelength selector 16 obtains a linear sum of the test-weighted spectra of the reference material. Furthermore, the spectral wavelength selection unit 16 adds randomly generated noise components to the linear sum of the test-weighted reference substance spectra to synthesize a test spectrum (step S102). After that, spectrum wavelength selection section 16 transmits the test spectrum to spectrum analysis section 15 .

Next, the spectrum wavelength selection unit 16 selects all the N types of spectrum wavelengths as the spectrum wavelengths λi to be selected. That is, λi=(λ1, . . . , λN). Also, the spectrum wavelength selection unit 16 sets S1, which is a set of already selected spectrum wavelengths used in the first calculation, to an empty set, and sets F1, which is an analysis error used in the first calculation, to infinity. Furthermore, the spectrum wavelength selection unit 16 sets the number of calculations t of the analysis error to 1 (step S103).

Next, the spectrum wavelength selection unit 16 selects one unselected spectrum wavelength from among the spectrum wavelengths λi to be selected. Let λs be the selected spectral wavelength (step S104).

Next, the spectrum wavelength selection unit 16 notifies the spectrum analysis unit 15 of information on the set of spectrum wavelengths selected by adding λs to the set St of spectrum wavelengths already selected for use in the t-th calculation, and Instructs the execution of decomposition. The spectral analysis unit 15 then performs spectral decomposition of the test spectrum using the selected spectral wavelength (step S105). After that, the spectral analysis unit 15 transmits the estimated weights β _p obtained by spectral decomposition to the spectral wavelength selection unit 16 .

The spectral wavelength selector 16 receives the estimated weights β _p from the spectral wavelength selector 16 . Then, the spectrum wavelength selection unit 16 calculates the analysis error E(St∪λs) when using the spectrum wavelength selected by adding λs to the already selected spectrum wavelength set St (step S106).

After that, the spectrum wavelength selection unit 16 determines whether or not calculation of the analysis error when each spectrum wavelength is added has been completed for all spectrum wavelengths included in λi (step S107). If there remains a spectral wavelength for which an analysis error has not been calculated for the addition (step S107: No), the spectral wavelength selection process returns to step S104.

On the other hand, when the calculation of the analysis error in the case of addition has been completed for all spectrum wavelengths included in λi (step S107: Yes), the spectrum wavelength selection unit 16 performs the analysis in the case of addition in λi. λi*, which is the spectral wavelength that minimizes the error, is specified (step S108).

Next, the spectrum wavelength selection unit 16 performs analysis when the provisional analysis error Ft calculated in the t-1th calculation is added to the set St of already selected spectrum wavelengths with the selected spectrum wavelength λi*. It is determined whether or not the error is equal to or less than E(St∪λi*) (step S109).

If Ft is greater than E(St∪λi*) (step S109: No), the spectrum wavelength selection unit 16 sets λi as a set obtained by removing the spectrum wavelength λi* from the spectrum wavelengths included in λi (step S110).

Furthermore, the spectrum wavelength selection unit 16 increments t, which is the number of repetitions of the calculation, by one. In addition, the spectrum wavelength selection unit 16 converts E(St∪λi*), which is an analysis error when adding the spectrum wavelength λi* to the already selected set St of spectrum wavelengths, into a provisional is an analytical error Ft. Furthermore, the spectrum wavelength selection unit 16 sets the set obtained by adding the spectrum wavelength λi* to the already selected set St of spectrum wavelengths as the set St of already selected spectrum wavelengths in the next calculation (step S111). Thereafter, the spectral wavelength selection process returns to step S104.

On the other hand, if Ft is equal to or less than E(St∪λi*) (step S109: NO), the spectrum wavelength selection unit 16 selects a spectrum included in the already selected set St of spectrum wavelengths as a spectrum wavelength Determine (step S112).

As described above, the material inspection apparatus according to this embodiment selects spectral wavelengths to be used for spectral decomposition. Then, using the selected spectral wavelength, the substance inspection apparatus imposes a weight that minimizes the linear sum of the weighted reference substance spectrum and the test target substance spectrum, and imposes a condition that the weight is non-negative. ask for By performing spectral decomposition using the selected spectral waveform, noise in obtaining the weight can be reduced, and more accurate weight calculation can be performed. Therefore, highly accurate identification and detection can be performed even in an environment where noise is assumed. Moreover, by imposing the condition that the weights are non-negative, the weights can be easily interpreted and the calculated weights can be stabilized, so that the weights can be calculated accurately. Therefore, more accurate identification and detection can be performed.

FIG. 9 is a block diagram of a material inspection device according to the second embodiment. The substance inspection apparatus 1 according to the present embodiment selects a laser wavelength with a minimum analysis error, irradiates the selected laser wavelength to perform spectral decomposition, and executes identification and detection, which is different from the first embodiment. . The material inspection apparatus 1 according to the present embodiment further includes a laser wavelength selection section 18 in addition to the components of the material inspection apparatus 1 according to the first embodiment. In the following description, the description of the operation of each unit similar to that of the first embodiment will be omitted.

The database 17 holds the spectrum of the reference substance for each usable laser wavelength for each reference substance. FIG. 10 is a diagram showing the relationship between the laser wavelength and the spectrum of a specific substance. Different laser wavelengths change the intensity for each spectral wavelength. That is, each time the laser wavelength is different, the individual spectral value for each spectral wavelength, which is each element of the column vector representing the spectrum of a particular substance, changes. For example, if there are K types of laser wavelengths, the column vector representing the spectrum of a specific substance is expressed as spectrums ##1 to ##R of a specific substance that differ for each laser wavelength, as shown in FIG. be. For example, the database 17 holds each element of a column vector representing a spectrum corresponding to each laser wavelength as shown in FIG. 10 for each reference substance.

Here, since the spectrum differs for each wavelength of the laser, the accuracy of spectral resolution depends on the wavelength of the irradiated laser. Therefore, it is preferable to use a laser beam with a wavelength that minimizes the analysis error.

Returning to FIG. 9, the description continues. The laser wavelength selection unit 18 stores in advance the wavelengths of usable lasers. For example, if there are K types of usable laser wavelengths, the laser wavelength selector 18 stores R laser wavelengths.

The laser wavelength selection unit 18 selects one laser wavelength from available laser wavelengths. Next, the laser wavelength selection unit 18 acquires information on the spectrum corresponding to the selected laser wavelength of each reference material #p from the database 17, and generates a column vector representing the spectrum of each reference material #p. Next, the laser wavelength selection unit 18 randomly generates a weight for each reference material #p as a test weight. Next, the laser wavelength selector 18 adds a test weight to the column vector of each reference material #p, and obtains a linear sum of the column vectors representing each reference material #p to which the test weight is added.

Next, the laser wavelength selector 18 randomly generates noise components. Then, the laser wavelength selection unit 18 adds the generated noise component to the linear sum of the test-weighted spectrum of each reference material #p to synthesize a test spectrum for laser wavelength selection. Then, the laser wavelength selection unit 18 transmits the synthesized test spectrum for laser wavelength selection to the spectrum analysis unit 15 and requests spectrum decomposition.

After that, the laser wavelength selection unit 18 obtains from the spectrum analysis unit 15 the estimated weight that is the result of the spectral decomposition of the test spectrum for laser wavelength selection. Next, the spectral wavelength selector 16 uses the estimated weights and test weights to obtain an analysis error. For example, the laser wavelength selection unit 18 uses equation (5) or (6) in the same way as the spectrum wavelength selection unit 16 to find the analysis error.

The laser wavelength selector 18 obtains an analysis error for each usable laser wavelength. And the laser wavelength selection unit 18 is: The laser wavelength that minimizes the analysis error is determined as the irradiation laser wavelength. Then, the laser wavelength selection unit 18 notifies the identification detection unit 14 of information on the irradiation laser wavelength. This laser wavelength selection unit 18 corresponds to an example of the "optical wavelength selection unit".

FIG. 11 is a diagram showing an example of analysis error for each laser wavelength. The vertical axis in FIG. 11 represents the analysis error, and the horizontal axis represents the type of laser wavelength. For example, when the analysis error when using each laser wavelength is obtained as shown in FIG. 11, the minimum analysis error is 0.007 when the laser wavelength shown in the bar graph 400 is 213.51 nm. Therefore, the laser wavelength selection unit 18 determines the irradiation laser wavelength to be 213.51 nm.

The identification detection unit 14 receives notification of information on the irradiation laser wavelength from the laser wavelength selection unit 18 . Further, the identification detection unit 14 receives an instruction from the operator to identify and detect the inspection object Z, and instructs the light emission unit 11 to irradiate the laser having the irradiation laser wavelength notified from the laser wavelength selection unit 18. do. After that, the identification detection unit 14 detects each reference substance obtained from the spectrum of the inspection object Z obtained from the spectrum of the inspection object Z obtained from the scattered light when the laser having the irradiation laser wavelength from the inspection object Z is irradiated from the spectrum analysis unit 15 The weight β _p of #p is received from the spectrum analysis unit 15 . Then, the identification detection unit 14 executes identification/detection using the acquired weight _βp .

For example, when the spectral wavelength selected for each laser wavelength is shown in FIG. Spectral decomposition is performed using the spectral wavelengths represented by points 200 in .

FIG. 12 is a diagram showing an example of identification results. FIG. 12 shows the identification result when the irradiation laser wavelength is determined to be 213.51 nm by the laser wavelength selection unit 18, and the spectrum analysis unit 15 performs spectral decomposition using the spectrum wavelength represented by the point 200 in FIG. be. The database 17 also holds spectra of CO ₂ , O ₂ , N ₂ , H ₂ , CH ₄ , NH ₃ and SO ₂ as reference substances.

In this case, the test weights for _CO2 , _O2 , _N2 , _H2 , _CH4 , _NH3 and _SO2 are respectively 0.18, 0.07, 0.06, 0.14, 0.19, Determine 0.11 and 0.25. Then, the substance inspection apparatus 1 spectrally decomposes the spectrum generated using the test weights, and obtains values of 0.18, 0.07, 0.06, 0.14, and 0 representing weights for the respective reference substances. .18, 0.11 and 0.25 are obtained. In this way, an estimation result having substantially the same value as the estimation result by the material inspection device 1 is obtained for the test weight. Therefore, it can be said that the substance inspection device 1 according to the present embodiment can perform highly accurate identification and detection.

Next, with reference to FIG. 13, the flow of identification/detection processing by the substance inspection apparatus 1 according to the present embodiment will be described. FIG. 13 is a flowchart of identification/detection processing by the substance inspection device according to the second embodiment.

The spectral wavelength selection unit 16 executes spectral wavelength selection processing (step S11). After that, the spectral wavelength selection unit 16 notifies the information of the selected spectral wavelength to the spectrum analysis unit 15 .

The laser wavelength selection unit 18 generates test weights and uses the generated test weights to synthesize a test spectrum for laser wavelength selection. Then, the laser wavelength selection unit 18 obtains an analysis error between the estimated weight obtained by spectrally decomposing the test spectrum for the laser wavelength and the test weight when each laser wavelength is used, and determines that the analysis error is the minimum. is selected as the laser wavelength of the laser for irradiation (step S12). After that, the laser wavelength selection unit 18 notifies the identification detection unit 14 of information on the selected laser wavelength.

The identification detection unit 102 receives notification of information on the selected laser wavelength from the laser wavelength selection unit 18 . After that, upon receiving an execution instruction for identification and detection of the inspection object Z from the operator, the light emitting unit 11 is instructed to irradiate the laser having the selected laser wavelength. Upon receiving the instruction from the identification detection unit 102, the light emission unit 11 irradiates the inspection object substance Z with a laser beam having the selected laser wavelength (step S13).

The measurement unit 12 receives the scattered light from the inspection target substance Z (step S14). Then, the measurement unit 12 outputs information on the spectrum of the substance Z to be inspected to the spectrum analysis unit 15 .

The spectrum analysis unit 15 receives notification of information on the selected spectrum wavelength from the spectrum wavelength selection unit 16 . The spectrum analysis unit 15 also receives input of information on the spectrum of the substance Z to be inspected from the measurement unit 12 . Then, using the selected spectrum wavelength, the spectrum analysis unit 15 imposes a condition that the weight is non-negative so that the difference between the spectrum of the inspection target substance Z and the linear sum of the weighted reference substance spectrum is the minimum. Spectral decomposition is performed by obtaining a weight that is equal to (step S15). After that, the spectrum analysis unit 15 outputs the calculated weight information to the identification detection unit 14 .

The identification detection unit 14 receives input of the calculated weight information from the spectrum analysis unit 15 . Then, the identification detection unit 14 uses the spectral information of the reference substance stored in the database 17 to identify and detect the inspection target substance Z (step S16). After that, the notification unit 13 notifies the operator of the identification/detection result.

As described above, the substance inspection apparatus according to the present embodiment selects a laser with a wavelength that minimizes the analysis error, irradiates the substance to be inspected, and uses the scattered light to perform identification and detection. As a result, it is possible to further suppress errors in spectral resolution and to perform identification and detection with higher accuracy.

1 substance inspection device 11 light emitting unit 12 measurement unit 13 notification unit 14 identification detection unit 15 spectrum analysis unit 16 spectrum wavelength selection unit 17 database 18 laser wavelength selection unit R inspection object

Claims (7)

a light emitting unit that irradiates an object with light;
a measurement unit that measures either one or both of scattered light and absorbed light in the object irradiated with the light and obtains a spectrum representing the intensity of the measured light at a plurality of wavelengths;
Perform spectral decomposition to calculate a weight that minimizes the difference between a linear sum of spectra of a plurality of known reference substances weighted for each reference substance and a predetermined spectrum under the condition that the weight is non-negative a spectrum analysis unit for
Based on the analysis error of the spectrum decomposition, one or more working wavelengths are selected from the plurality of wavelengths in the spectrum, and the spectrum of the object is set as the predetermined spectrum, and each of the working wavelengths of the scattered light a spectrum wavelength selection unit that causes the spectrum analysis unit to perform spectrum decomposition using a limited spectrum that represents the intensity of
an identification detection unit that identifies or detects the object and the reference substance based on the weight obtained by spectral decomposition of the object spectrum using the restricted spectrum by the spectrum analysis unit; A substance inspection device characterized by: The spectrum wavelength selection unit obtains a linear sum of the spectra of the plurality of reference materials to which test weights are added, and causes the spectrum analysis unit to perform the spectral decomposition on the test spectrum obtained by adding a noise component to the linear sum. 2. A material inspection apparatus according to claim 1, wherein the estimated weight is obtained by: and the working wavelength is selected so as to minimize the analytical error between the estimated weight and the test weight. The spectrum wavelength selection unit repeatedly obtains the analysis error while sequentially adding different wavelengths, and determines the use wavelength that minimizes the analysis error between the estimated weight and the test weight. 3. The material inspection device according to claim 2. further comprising a light wavelength selection unit that selects the wavelength of the light emitted by the light emitting unit based on the analysis error for the test spectrum in the spectrum decomposition,
The light emitting unit irradiates the object with light of a wavelength selected by the light wavelength selection unit,
The substance inspection device according to claim 2 or 3, wherein the measurement unit acquires a spectrum of scattered light from the object irradiated with the light of the wavelength selected by the light wavelength selection unit. 5. The optical wavelength selection unit according to claim 4, wherein the wavelength of the light is selected based on each analysis error for the test spectrum when the spectrum decomposition is performed by changing the wavelength of the light. A material inspection device as described. Irradiate the object with light,
measuring either one or both of scattered light and absorbed light from the object irradiated with the light;
obtaining a spectrum representing the intensity of the measured light at multiple wavelengths;
Perform spectral decomposition to calculate a weight that minimizes the difference between a linear sum of spectra of a plurality of known reference substances weighted for each reference substance and a predetermined spectrum under the condition that the weight is non-negative Select one or more working wavelengths from the plurality of wavelengths based on the analysis error in the case of
The inspection weight that minimizes the difference between the linear sum of the weighted for each reference substance of the restricted spectrum representing the intensity of the scattered light of the reference substance for each of the used wavelengths and the restricted spectrum of the object, Calculated under the condition that the test weight is non-negative,
A substance inspection method, wherein the object and the reference substance are identified or detected based on the calculated inspection weight. Irradiate the object with light,
measuring either one or both of scattered light and absorbed light from the object irradiated with the light;
obtaining a spectrum representing the intensity of the measured light at multiple wavelengths;
Perform spectral decomposition to calculate a weight that minimizes the difference between a linear sum of spectra of a plurality of known reference substances weighted for each reference substance and a predetermined spectrum under the condition that the weight is non-negative Select one or more working wavelengths from the plurality of wavelengths based on the analysis error in the case of
The inspection weight that minimizes the difference between the linear sum of the weighted for each reference substance of the restricted spectrum representing the intensity of the scattered light of the reference substance for each of the used wavelengths and the restricted spectrum of the object, Calculated under the condition that the test weight is non-negative,
A material inspection program for causing a computer to execute a process of identifying or detecting the object and the reference material based on the calculated inspection weight.

Images