JP5263776B2 - Non-magnetic metal identification method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To automatically identify the quality of material of crushed metal flakes of relatively large shapes of copper, aluminum, magnesium, etc. and identify them based on their results, and to automatically identify aluminum alloy crushed metal flakes into ones derived from a flattened material and ones derived from a cast material, and to attain an automatic identification method that identifies metal, is simple, and of high performance based on the results. <P>SOLUTION: The identification method for non-magnetic metals is such that a non-magnetic metal identifying device having a feeder 2 that feeds a crushed metal 1, belt conveyors 3, 5 that convey the fed crushed metal flakes 1, a weigher 4 that measures weights of the crushed metal flakes 1, a photosensor 6, a 3-dimensional laser measuring instrument 7, and a separation/collection mechanism is utilized, integral control is performed on these movements by a controller 19, and the metal flakes are identified for every quality of materials and collected. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

  In the present invention, a relatively large copper, aluminum, magnesium, etc. of about 20 to 200 mm contained in a mixture of non-magnetic metal fragments (mixed metal) collected in a shredder processing facility such as a waste home appliance and a scrap car is used as a heavy liquid. The present invention relates to an identification method suitable for identifying and collecting each material, regardless of the sorting method using the method or the manual sorting method.

  Furthermore, this invention relates to the identification method which isolate | separates into the thing derived from a wrought material and the thing derived from a cast material with respect to the identified aluminum crushing piece.

  In shredder processing facilities such as waste home appliances and scrap cars, after crushing, removing iron by magnetic sorting, removing non-metals such as resin with wind sorters and air tables, etc., using copper and aluminum with eddy current sorters In general, non-magnetic metals are collected together as a mixed metal.

  In order to recycle this mixed metal, it is necessary to further sort by material. Usually, specific gravity sorting using heavy liquid or manual sorting is often used as the means, but the following special means are known because of problems in terms of processing cost and efficiency.

  For example, a method of selecting copper from other metals by a color discriminator is known (see Patent Document 1). Further, as a technique for selecting and collecting metal fragments similar to the present invention, a method by detecting inductance of a metal block is known (see Patent Document 2). Furthermore, a method is known in which the volume of a crushed piece is measured using the fact that the gas volume varies depending on the presence of the crushed piece, and the specific gravity is calculated by separately measuring the weight value to identify the material (Patent Document). 3).

  In addition, aluminum scrap collected at intermediate processing facilities such as scrap home appliances and scrap cars contains aluminum alloys of various components, especially when alloy for wrought material and alloy for casting material are mixed. However, there is a problem that the use of the secondary alloy is limited to castings and die castings due to the mixing of impurity elements, and the economic efficiency of recycling is reduced.

  In order to avoid such a situation, there is a demand for a technology for further sorting aluminum scrap derived from wrought material and cast material. At present, there is no choice but to directly measure the elemental composition of the alloy by laser-induced plasma spectroscopy.

  Furthermore, as an example of applying the object's three-dimensional shape measurement technology to waste identification, there is a known case where the three-dimensional shape of an object is derived by calculation processing of imaging signals acquired by cameras or line sensors arranged in two directions. Yes (see Patent Documents 4 and 5).

  In addition, as described above, with regard to the means for automatically selecting the aluminum alloy fragment from the wrought material and the cast material, there are an identification method by fluorescent X-ray analysis and an identification method by laser-induced plasma spectroscopy. Yes (see Non-Patent Document 1). In addition, a multistage sorting method combining X-ray transmittance measurement and eddy current detection has been developed (see Non-Patent Document 2).

JP-A-6-106091 Japanese Patent Laid-Open No. 9-24344 JP-A-8-71528 JP 11-83461 A JP-A-10-192794

W.L.Dalmijn, Recycling International, (April), 48 (2003) M.B.Mesina, et.al., Int.J. Miner.Process, 82, 222 (2007)

  When non-magnetic metals such as copper, aluminum, and magnesium that are generated in the process of processing waste containing non-metals such as various metals and resins are further separated and recovered for each material, the eddy current sorter has sufficient accuracy. Can not be separated.

  The wind sorting and the air table can be applied to a relatively small fragment having a size of approximately 10 mm or less, but it is difficult to separate a fragment having a larger weight and a larger weight.

  Color sorters are not suitable for sorting metals with similar colors, such as aluminum and magnesium, or metals colored by painting. Further, in the specific gravity sorting using heavy liquid, the cost required for the wastewater treatment facility becomes a problem.

  For this reason, the selection of relatively large non-magnetic metal fragments is mainly performed manually, but there are problems in the working environment such as noise and dirt and an increase in cost. For these reasons, the inventions described in Patent Documents 2 and 3 have been made, but the present situation is that they are hardly used in the field of waste treatment.

  In addition, as described above, with regard to the means for automatically selecting the crushed aluminum alloy from the wrought material and the cast material, an identification method by fluorescent X-ray analysis or an identification method by laser-induced plasma spectroscopy ( However, since these are analysis methods for obtaining information in the vicinity of the object surface, when a waste is targeted, accurate analysis is often impossible due to the influence of dirt, painting, and the like. In addition, a multi-stage sorting method that combines X-ray transmittance measurement and eddy current detection has been developed (see Non-Patent Document 2), but the equipment itself is very expensive. Have difficulty.

  The present invention automatically identifies the material of a relatively large shaped piece of copper, aluminum, magnesium, or the like, which has been considered difficult, and identifies the material based on the result. On the other hand, an object of the present invention is to provide a simple and high-performance non-magnetic metal identification method capable of automatically identifying a material derived from a wrought material and a cast material and identifying based on the result.

In order to solve the above-mentioned problems, the present invention is based on the result of an arithmetic processing step using the weight of a non-magnetic metal fragment and the three-dimensional shape information of the fragment obtained by measurement with one laser three-dimensional measuring instrument. In the method of automatically identifying the material and shape of non-magnetic metal fragments,
The arithmetic processing step is performed by comparing the output value of the neural network that inputs the variables calculated from the measured values of the apparent density, volume, area, portrait, landscape, maximum height, and centroid height of the fragment, The neural network structure provides a method for identifying non-magnetic metals, characterized in that a database in which measurement data relating to the above-mentioned variables of the fragments are apparently grouped according to apparent density is determined for each group. To do.

  In the calculation processing step, the value of the discriminant function is calculated using variables calculated from the measured values of the apparent density, volume, area, portrait, landscape, maximum height, and barycentric height of the crushed pieces. And a threshold value set in advance, and for the discriminant function and the threshold value, a database in which the measurement data related to the above-described variables of the fragment is apparently grouped according to the size of the apparent density is created. Comparing the method of determining by performing random analysis with the output value of a neural network that inputs variables calculated from the measured values of apparent density, volume, area, length, width, maximum height, and centroid height of the fragment The neural network structure is divided into groups according to the apparent density of the measured data related to the above-mentioned variables of the fragments. And create the database, or a method of performing a combination of the identification method of the non-magnetic metals and determining for each the group.

By using the nonmagnetic metal identification method of the present invention, the following remarkable effects are produced.
(1) It is possible to identify metal fragments such as copper, aluminum, magnesium, etc. having a relatively large shape of 20 to 200 mm, which have been difficult to efficiently identify with the prior art, for each material at low cost.

(2) With respect to the aluminum alloy, it is possible to distinguish at low cost the one derived from the wrought material and the one derived from the cast material due to the difference in the shape of the crushed pieces. According to the present invention, a crushed piece can be identified without being affected by dirt or coating of the crushed piece.

(3) As an example of applying the three-dimensional shape measurement technique of an object to the identification of waste, the imaging signals acquired by cameras or line sensors arranged in two directions as shown in Patent Documents 4 and 5 are calculated. There is a case where the solid shape of an object is derived by processing. Compared to such a known technique, in the present invention, since the three-dimensional shape of an object is directly measured as numerical data by a single laser three-dimensional measuring instrument, complicated data processing for identification can be performed relatively easily. There are advantages.

It is a figure which shows the whole structure of the apparatus which enforces the identification method of the nonmagnetic metal which concerns on this invention. It is a processing flow figure of metal identification collection by the present invention. It is an enlarged view of the metal crushing metal piece irradiated with the laser beam. It is a figure which shows the example of the algorithm for discriminating the material of a metal. It is a figure which shows the example of the neural network model for discriminating the material of a metal. It is a figure which shows the apparent density distribution of the metal crushing metal piece measured with the apparatus shown in FIG. It is explanatory drawing about the space | gap which cannot be detected with the apparatus shown in FIG. It is a figure which shows the relationship between three types of variables regarding the three-dimensional shape of the crushing metal piece measured with the apparatus shown in FIG. 1, and an apparent density.

  Embodiments and examples of a method for identifying a nonmagnetic metal according to the present invention will be described below with reference to the drawings.

  FIG. 1 shows an embodiment of the overall configuration of a nonmagnetic metal identification and recovery apparatus that implements a nonmagnetic metal identification method according to the present invention. Supply device 2 for supplying the crushed metal piece 1, belt conveyors 3, 5 for conveying the crushed metal piece 1 supplied from the supply device 2, weighing scale 4 for measuring the weight of the crushed metal piece 1, photosensor 6, laser 3 The dimension measuring instrument 7 and the separation / collection mechanism are provided, and these operations are configured to be comprehensively controlled by a control device 19 (actually using a computer). When it is necessary to further increase the processing amount per unit time, a plurality of lines shown in FIG.

  The crushed metal pieces 1 are individually supplied from the supply device 2 to the belt conveyor 3. When the crushed metal pieces 1 reach the weigh scale 4, the weight is first measured, and the result is sent to the control device 16. Subsequently, when the crushed metal piece 1 is detected by the photosensor 6, the three-dimensional shape of the crushed metal piece 1 is measured by receiving the detection signal and operating the laser three-dimensional measuring instrument 7 suspended upward.

  In the control device 19, the crushed metal piece 1 is identified by a calculation process using information on the weight and the three-dimensional shape as variables. The algorithm at this time will be described later. The crushed metal piece 1 that has been identified is sorted by a sorting mechanism (actuator) including a compressor 18, electromagnetic valves 15, 16, and 17 and actuators 12, 13, and 14 based on the result. 10 and 11.

  FIG. 2 shows a processing flow of metal identification recovery according to the present invention.

  FIG. 3 shows an enlarged view of a crushed metal piece irradiated with laser light on the belt conveyor of the apparatus shown in FIG. When this laser three-dimensional measuring instrument detects an object moving in a certain direction with a photo sensor, it instantaneously irradiates a slit-shaped laser beam with a width of about 25 cm from diagonally forward, and changes the position of the reflected light line on the object surface. It is detected by the CCD and the three-dimensional shape is recorded as digital data.

  The measured values used for identification in the apparatus shown in FIG. 1 are weight, volume, vertically projected area (hereinafter referred to as area), vertically long (longitudinal length), horizontally long (horizontal length), maximum height, vertically upward The height of the barycentric point of the projection plane (hereinafter referred to as “the barycentric point height”).

  In identifying a crushed metal piece, first, a plurality of representative samples are extracted and repeatedly measured for the crushed metal piece to be identified, and Table 1 (variables used in discriminant analysis and / or neural network is shown. Create a database for the 14 variable values shown in the table. At this time, a combination of 14 variable values obtained when a certain crushed metal piece is measured is registered in the database as one case.

  These variable values depend not only on the three-dimensional shape of the crushed metal piece, but also on the orientation of the crushed metal piece with respect to the transport direction (specifically, the direction of the crushed metal piece placed on the conveyor). Different values are taken when the orientations of the pieces are different. For this reason, for the crushed metal piece having a complicated shape, measurement data for as many orientations as possible are collected and registered in the database as separate cases.

The data here is a combination of numerical values (X _{1 to} X ₁₄ ) obtained when a certain crushed metal piece is measured in a certain orientation. If this is one case and the orientation of the same crushed metal piece is different, it is registered in the database as another case.

  Further, “identification” according to the present invention is an operation of “determining which material a measured case corresponds to in comparison with a large amount of case data registered in advance”. Registration of one case is sufficient for a completely symmetrical shape, but in order to increase the accuracy of identification, a crushed metal piece having a complicated shape should have a larger number of cases to be registered.

Then, sort all cases registered in the database in the order of magnitude of the variable value X _1, grouped into a plurality of case group by setting the appropriate boundary values for X _1. Here, the variable value X _1, a value obtained by dividing the volume value measured weight value in the apparatus shown in FIG. 1, hereinafter referred to as apparent density.

  For each case group grouped according to the apparent density, one or both of discriminant analysis method and neural network, which is a kind of multivariate analysis method, are applied, and the crushed metal measured as a case An algorithm for identifying the material of the piece is determined in advance.

(Application of multivariate analysis method)
The details of the multivariate analysis method and the discriminant analysis method are shown in, for example, “Applied Statistics Handbook p318 to p328, Yokendo (1986)”, and the outline thereof is as follows. The multivariate analysis method refers to the relationship between items of a material (corresponding to “variables” in the present invention) composed of a plurality of items (corresponding to “variables” in the present invention) obtained by simultaneously investigating. This is a statistical analysis method that examines and understands and analyzes the material as a whole. Discriminant analysis is a method of multivariate analysis that searches for a reasonably discriminating criterion (discriminant function in the text) when this material is divided into two groups according to a desired purpose.

Here, the discriminant function Z (X) described by Equation 1 is calculated using the material of the crushed metal piece as an objective variable and the 14 variables shown in Table 1 regarding the three-dimensional shape as explanatory variables. That is, the discriminant function Z (X) has a, b, c. . . Are calculated. Then, these 15 numerical values are converted into a, b, c. . . The discriminant function value (discriminant score for each case) calculated by substituting the variable values of X _{1 to} X ₁₄ for each case into the formula of the discriminant function Z (X) put into the Associated value.

  Here, the material is a type of metal such as copper, aluminum, and magnesium, and this is expressed by different integer values such as 0 for copper (for example, when distinguishing between copper and aluminum), 1 for aluminum, and the like. This value is added to each case data and discriminant analysis is performed using all 15 variables. Such a variable representing a “correct answer” in discriminant analysis is called an objective variable.

  Discriminant analysis can only be separated into two groups, so for the group of cases where three types of metals are mixed, for example, after separating magnesium with “copper and aluminum 0, magnesium 1”, the next discriminant analysis Therefore, a process for separating copper and aluminum is required with “copper as 0 and aluminum as 1”. When discriminating between an aluminum cast material and an aluminum wrought material, which will be described later, the discriminant analysis is performed by quantifying the cast material to 0, the wrought material to 1.

  If the discrimination scores for all cases subject to discriminant analysis are calculated and sorted in the order of their sizes, the discriminant score ranks according to the difference in material, so the discriminant score value with the highest discrimination accuracy is the threshold value. And a case with a discrimination score larger than the threshold is identified as material A, and a case with a discrimination score smaller than the threshold is identified as material B.

  In general discriminant analysis, the discriminant score is divided into two groups (always set the threshold value to 0) according to the sign of the discriminant score. However, when the discrimination score is actually calculated for the case data of the fragmented pieces, the situation shown in Table 6 below often occurs. At this time, if the threshold value is set to 0, the target hit rate is 2/3 for copper and 2/2 aluminum, whereas if the threshold value is set to a value of -0.5 to -2.6, the hit rate is It becomes copper 3/3 and aluminum 2/2, and the accuracy of discrimination is improved. The “discrimination score value with the highest discrimination accuracy” is an arbitrary value of −0.5 to −2.6, and a threshold is set to −1.5, for example.

(Formula 1)
_{Z n (X) = a n} + b n X 1 + c n X 2 + ...... + n n X 13 + o n X 14 (n = 1,2,3 ......)

(Application of neural network)
In the nonmagnetic metal identification method of the present invention, the apparent density, volume, and area of the crushed pieces are used as an arithmetic processing step that uses the solid shape information of the crushed pieces to automatically identify the material and shape of the nonmagnetic crushed pieces. This method may be performed by comparing the output values of a neural network to which variables calculated using measured values of vertical, horizontal, maximum height, and barycentric point height are input. Will be described below as a second embodiment.

  Details of the neural network itself are shown, for example, in “Neurocomputer Engineering p31 to p75, Industrial Research Committee (1992)”, and the outline thereof is as follows. A neural network is an information processing technique that imitates the structure of the human brain, and can perform analysis similar to discriminant analysis in multivariate analysis. When using a neural network, prepare data for finding the relationship between variables (in the present invention, it corresponds to “case group”), learn the neural network using this data, and then input the variable data. Then take the procedure of interpreting the results from the output of the neural network.

  An excellent feature of neural networks is that it is not necessary to assume the functional relationship between the objective variable and the explanatory variable in advance as in multivariate analysis, so it is possible to model any functional relationship, and two or more in a single calculation. The point that can be identified in this group. In addition, even in a neural network that has once been learned, if learning data is added and relearned, the identification performance can be improved.

There are several models in a neural network, and any model can be used for identification. Here, an example using the most general hierarchical model will be described. FIG. 4 is an example of a neural network for discriminating crushed metal pieces into four types of copper (Cu), cast aluminum (Alc), expanded aluminum (Alw), and magnesium (Mg). In this case, 14 units (neuron models) corresponding to the variable values X _{1 to} X ₁₄ are set in the input layer, and four corresponding to copper, aluminum casting material, aluminum expanded material, and magnesium are set in the output layer. Set the unit. A plurality of units are set in the intermediate layer according to the characteristics of the case group to be identified. These units are connected through a network having a weight.

Here, the variable values X _{1 to} X ₁₄ normalized between 0 and 1 using (Expression 2) are input to the input layer. In each unit of the intermediate layer, the input value from each unit of the input layer is weighted and summed, and a value of 0 to 1 is output according to a threshold value based on an S-shaped nonlinear function (sigmoid function) Output to each unit in the layer. Each unit of the output layer outputs a value of 0 to 1 by the same calculation.

In the learning of the neural network, each case data obtained by normalizing the variable values X _{1 to} X ₁₄ using the following (Equation 2), in the case of copper (1, 0, 0, 0), 4 integer values (teacher signal) (0, 1, 0, 0) in case of aluminum, (0, 0, 1, 0) in case of aluminum expanded material, and (0, 0, 0, 1) in case of magnesium ) Is added to all 18 variables, the weight of the coupling between the units is corrected by the error back propagation method, and the neural network is optimized.

  Details of the back propagation method are described in “Neurocomputer Engineering p46-p49, Industrial Research Committee (1992)”. Setting the conditions such as the number of learning, the allowable error, the number of supplementary training, etc. in the error back-propagation method is a problem, but these are determined according to the properties of the target case group.

When calculation is performed by inputting variable data X _{1 to} X ₁₄ normalized using the following (formula 2) to the learned neural network, (y ₁ , y ₂ , y ₃ , y ₄ ) are output to the output layer. It is assumed that the following four values are obtained. In this case by comparing the y ₁ ~y _4, if y ₁ is the maximum copper, aluminum cast material when y ₂ is the maximum, the aluminum wrought If y ₃ is maximum, if y ₄ is maximum Is distinguished from magnesium. Thus, when a neural network is used, the material is immediately determined from the four numerical values output to the output layer.

(Formula 2)
Value after normalization = (value before normalization-minimum value) / (maximum value-minimum value)
(Here, the maximum and minimum values in Equation 2 indicate the maximum and minimum values of each variable registered in the database in advance)

(Application of multivariate analysis method and neural network)
When the number of materials to be identified is limited to two, and the number of case data is very small, the material of the crushed metal piece can be identified by a single discriminant function. However, when three or more kinds of materials need to be identified and a large amount of case data exists, it is impossible to accurately identify the material of the crushed metal piece with one discriminant function. Although it is possible to increase the accuracy by setting a plurality of discriminant functions and performing multistage calculation processing, in this case, the algorithm for identification becomes complicated and the setting operation becomes very complicated.

  When a neural network is used, much more cases can be handled than discriminant analysis, but the accuracy of a single-stage calculation process is still limited. Further, as described above, since the order between the case data cannot be determined from the output value of the neural network as in the discriminant analysis, it is not used for further grouping using the result, but used in the final stage of the algorithm. Is suitable.

  Therefore, in the present invention, when a large amount of data is handled, a method of setting an algorithm combining discriminant analysis and a neural network may be used. This method will be further described below as a third embodiment of the present invention.

FIG. 5 is an example of an algorithm that is supposed to identify four types of crushed metal pieces: aluminum wrought material, aluminum cast material, copper, and magnesium. Here, all cases registered in the database are grouped into 13 groups according to the difference in apparent density. In this example, when the apparent density is measured to be 5.0 g / cm ³ or more, it is immediately identified as copper, and when it is measured to be 0.6 g / cm ³ or less, it is immediately identified as aluminum stretched material. To do.

Note that the boundary value of the apparent density when the case groups are grouped and the number of groups after the division may be arbitrarily set according to the size and nature of the population to be identified. Hereinafter, a case group having an apparent density of 2.6 to 2.3 g / cm ³ will be exemplified.

First, the first-stage discriminant analysis 1 is performed for all cases in the group. The discriminant function Z ₁ (X) is calculated by the above method, and the threshold value α ₁ for the discriminant score by the function Z ₁ (X) is determined. Then, divide the all cases into two groups by comparing the discriminant score for each case threshold alpha ₁ and. At this stage, each group contains a mixture of aluminum wrought material, aluminum cast material, copper, and magnesium, and the material cannot be specified due to the difference in discrimination score.

Next, for the two divided groups, the second-stage discriminant analysis 2 and 3 is performed for the case groups belonging to the respective groups, and the discriminant functions Z ₂ (X) and Z ₃ (X) are determined. At the same time, threshold values α ₂ and α ₃ for the discrimination score Z ₂ value and Z ₃ value are determined. Further, by comparing the discrimination score and the threshold value, both groups are divided into two small groups. The number of groups (threshold number) at this time can also be arbitrarily set according to the size and nature of the population to be handled.

  In this way, the classification accuracy is improved by dividing into small groups so that the number of cases targeted in the third-stage neural network is as small as possible. In determining the threshold value in the second-stage discriminant analysis, it is important that the accuracy of discrimination is not a problem and that the number of cases included in the group after division is approximately 700 cases or less.

Note that when the discriminant analysis 2 in the second stage is performed to determine the discriminant function Z ₂ (X) (or Z ₃ (X)), the discriminant functions Z ₁ (X) in the _first and second stages are The discriminant function Z ₂ (X) (or Z ₃ (X)) is not the same. For example, when separating the A, B, C, and D case groups into the A and B case groups and the C and D case groups in the first-stage discriminant analysis, four cases of A, B, C, and D are used. The discriminant function is calculated on the basis of the above information. However, when the A and B case groups are separated into A and B in the second discriminant analysis, only the information on the two cases A and B is used. A discriminant function is calculated.

  As described above, since the “members” of cases included in the case group to be subjected to discriminant analysis are different between the first and second stages, the discriminant functions for separating them are naturally different.

As shown in FIG. 5, a calculation process is performed by the neural network 1 for the case group in which the Z ₂ value has shifted to the group larger than the threshold α ₂ in the second-stage discriminant analysis 2. By setting the hierarchical model and learning the target case group by the above method, the optimal neural network structure (number of intermediate layer units, weight of connection between units and threshold) is determined.

For even migrated case group to the second-stage discriminant analysis a small group Z ₂ value is above the threshold value alpha ₂ in 2, in the neural network 2, performs the same computation as this. For the same reason as in the case of discriminant analysis, the structures of the neural network 1 and the neural network 2 are different.

  The same procedure is applied to the case group that has shifted to discriminant analysis 3 according to the result of discriminant analysis in the first stage. In addition, even for a group of cases belonging to a different apparent density interval, the discriminant function, threshold value, and optimal neural network structure are set in exactly the same procedure, and the algorithm for identifying the material is determined.

For all discriminant functions determined in this way, a, b, c. . . 15 combinations and discrimination threshold values α ₁ , α ₂ , α ₃ are stored in the control device. All the neural network structures are stored in the control device.

  After the above preparation is completed, a sample whose material is unknown is conveyed and identified. Using the measured values obtained by the weigh scale and the three-dimensional measuring instrument, the value of the above discriminant function is calculated, and the algorithm is branched by comparing this value with a preset discriminant threshold. The material of the crushed metal piece is identified by the step discriminant analysis and the one-step neural network.

The identification process will be described more specifically with reference to Tables 2 to 4 for the crushed metal pieces whose material is unknown. Table 2 shows the coefficient values and threshold values of the discriminant function in the apparent density section 2.6 to 2.3 g / cm ³ determined by the above-described procedure.

Assuming that the seven measured values shown in Table 3 are obtained for a certain crushed metal piece, the variable values X _{1 to} X ₁₄ shown in Table 3 are immediately calculated based on the definition shown in Table 1. Is done. In this case, the apparent density value of the crushed metal piece is 2.30 g / cm ³ and can be identified by the procedure shown in FIG.

First, Z ₁ = −0.91 is obtained by substituting the coefficients a to o shown in the column of Z _{1 in} Table 2 and the variable values X _{1 to} X ₁₄ in Table 3 into Equation 1. If this value is compared with a preset threshold value α ₁ = 0.6, Z ₁ > α ₁ , so the algorithm moves to the discriminant analysis 2 side.

Subsequently, Z ₂ = −0.06 is obtained by the same calculation from the coefficients and variable values shown in the column of Z _{2 in} Table 2. If this value is compared with a preset threshold value α ₂ = 0.7, Z ₂ > α ₂ , so the algorithm moves to the neural network 1 side.

Subsequently, output values Y ₁ = 1.0, Y ₂ = 0.0, Y ₃ = 0.0, and Y ₄ = 0.0 are obtained by calculation using the preset neural network 1. Since the largest value among the output values Y _{1 to} Y ₄ is Y ₁ , the crushed metal piece is finally identified as copper.

  Table 4 shows an example of a crushed metal piece that is identified as an aluminum casting by a similar calculation procedure.

  Hereinafter, the reason why the crushed metal piece can be identified by the above-described method will be further described. FIG. 6 shows three kinds of metal crushing metal pieces (300 aluminum, 250 magnesium, and 105 copper (including brass, hereinafter referred to as copper)) obtained from a scrap car processing facility on a conveyor. The frequency distribution of the apparent density at the time of measuring 15 times for each crushed metal piece in a different orientation is shown.

In the figure, for aluminum, a cast material and a wrought material are separately shown. For magnesium, all samples are cast materials, and for copper, cast materials and wrought materials are not distinguished. From this figure, the apparent density measured by the apparatus shown in FIG. 1 is smaller than the true density (Al2.7, Mg1.7, Cu7.9-8.9 g / cm ³ ) regardless of the type of crushed metal pieces. It turns out that it becomes a value.

  The reason is that this 3D measuring instrument is based on information viewed from a certain direction, so when a blind spot area where the laser beam does not reach occurs, the part corresponding to the shadow is included in the volume and counted extra. It is.

  In crushed metal pieces with bent plate-like and thin members often found in wrought materials, there are many such undetectable voids inside the crushed metal pieces and in the gap with the conveyor, and the proportion occupied by these pieces is relatively It is considered that the error in the apparent density was larger than that of the massive and thick crushed metal pieces often found in the cast material (this conceptual diagram is shown in FIG. 7).

As for copper, about half of the cast material is measured to have an apparent density of 3.4 g / cm ³ or more. However, the thin tube material also has a large error, and the distribution extends to a low density range. . Therefore, it is difficult to accurately identify these by simply measuring the weight and volume of the crushed metal pieces and comparing the apparent densities.

FIG. 8 is a plot of average values for each section of apparent density at intervals of 0.2 g / cm ³ for three types of variables related to the three-dimensional shape of the crushed metal pieces for the same sample as FIG. As can be seen in FIG. 8 (a), the crushed metal pieces of magnesium generally have an apparent density value of X ₁ = 0.7 to 1.8 g / cm ^3, but an aluminum wrought material having a similar apparent density. compared to and aluminum cast material, there is a tendency that the area X ₃ than these decreases.

Further, as seen in FIGS. 8B and 8C, when X ₁ = 1.8 g / cm ³ or more where the overlap of the apparent density of the aluminum casting material and copper is a problem, the height of the object is concerned. there is a difference between the two to _{X 8} and _{X 11} value, which is an index. In addition, regarding the aluminum cast material and the aluminum wrought material, there is a significant difference in any variable in the range of X ₁ = 1.2 to 1.6 g / cm ³ where these apparent densities overlap.

  Therefore, if the apparent density is limited to a certain narrow section, there is some difference in the three-dimensional shape of the aluminum cast material, aluminum wrought material, magnesium and copper crushed metal pieces existing in the same section from a statistical point of view. These features can be identified by extracting their features well using multivariate analysis and neural networks.

  Four types of metal crushing metal pieces (20 ~), collected from non-ferrous metal sorting line of scrap car processing facility of scrap car, including copper (including brass), aluminum alloy (stretched material), aluminum alloy (casting material), and magnesium alloy The test which identifies about 250 mm according to material was implemented.

  First, measurements were repeated for randomly extracted pieces of crushed metal pieces while changing the orientation using the apparatus shown in FIG. 1, and a database of 14 variables shown in Table 1 was created. The number of samples registered in the database is copper: 4022 cases (number 105), cast aluminum: 5011 cases (number 104), aluminum expanded material: 7193 cases (number 192), and magnesium: 10544 cases (number 246). .

  The judgment algorithm is the same as that shown in FIG. 5, and the apparent density is divided into 13 sections. In each section, three types of discriminant functions, their threshold values, and four types of neural networks are set. The material is identified by a neural network.

  Thereafter, 200 pieces of crushed metal pieces not registered in the database (50 copper, 50 cast aluminum, 50 aluminum wrought material, 50 magnesium) were measured 5 times each while changing the orientation, The results of examining the median are shown in Table 5. Depending on the difference in orientation, the hit ratio fluctuated between 85% and 98%, but the average of all measurements made it possible to discriminate between copper, aluminum, and magnesium with a hit ratio of 90% or more.

  As described above, in a shredder processing facility such as a scrap car, a nonmagnetic metal sorting device according to the present invention is arranged after an eddy current sorter, thereby comparing about 20 to 200 mm included in a mixture of nonmagnetic metal crushed metal pieces. Therefore, it is possible to identify and collect large copper, aluminum, magnesium, etc. for each material regardless of a sorting method using heavy liquid or a manual sorting method. In addition, since it is possible to learn by adding data sequentially as the characteristic of the neural network, the nonmagnetic metal identification method according to the present invention can improve the identification accuracy according to the accumulation of measurement data.

  As mentioned above, although embodiment and the Example of the identification method of the nonmagnetic metal which concern on this invention were described with reference to drawings, this invention is not limited to such an Example, As described in a claim It goes without saying that there are various embodiments within the technical scope.

  The nonmagnetic metal identification method according to the present invention can be applied to identification other than metal by appropriately changing an object in data collection performed in advance. For these reasons, it is considered that the present invention contributes to improving the efficiency of recycling and economical efficiency of scrapped automobiles and the like.

  Furthermore, the method for identifying a nonmagnetic metal according to the invention can be applied to the identification of other objects if the identification object is changed in advance data collection. For example, it can be applied to fields such as manufacturing process management of various industrial products and identification of foods, agricultural and marine products.

DESCRIPTION OF SYMBOLS 1 Crushed metal piece 2 Supply apparatus 3, 5 Belt conveyor 4 Weigh scale 6 Photo sensor 7 Laser three-dimensional measuring instrument 8, 9, 10, 11 Collection container 12, 13, 14 Actuator 15, 16, 17 Electromagnetic valve 18 Compressor 19 Control apparatus

Claims (2)

Based on the result of the calculation process using the weight of the non-magnetic metal fragment and the three-dimensional shape information of the fragment obtained by measurement with one laser three-dimensional measuring instrument, the material and shape of the non-magnetic metal fragment are determined. In the automatic identification method,
The arithmetic processing step is performed by comparing the output values of the neural network that inputs the variables calculated using the measured values of the apparent density, volume, area, portrait, landscape, maximum height, and centroid height of the fragment. Done
In the structure of the neural network, a database in which measurement data relating to the above-mentioned variable of the fragment is apparently grouped according to apparent density is created and determined for each group. Based on the result of the calculation process using the weight of the non-magnetic metal fragment and the three-dimensional shape information of the fragment obtained by measurement with one laser three-dimensional measuring instrument, the material and shape of the non-magnetic metal fragment are determined. In the automatic identification method,
The arithmetic processing step includes
Calculate the value of the discriminant function using the variables calculated using the measured values of the apparent density, volume, area, portrait, landscape, maximum height, and barycentric height of the fragment, and this value and a preset threshold value For the discriminant function and the threshold value, a database in which the measurement data related to the above-mentioned variables of the fragments is preliminarily grouped according to the size of the apparent density is created, and multivariate analysis is performed for each group. This is done by comparing the output value of the neural network with the variables calculated using the measured values of the apparent density, volume, area, portrait, landscape, maximum height, and centroid height of the fragment. In the structure of the neural network, the measurement data relating to the above-mentioned variables of the fragment is preliminarily grouped according to apparent density. Create an identification method of a non-magnetic metal which is characterized in that a combination and determining for each the group.