Method for identifying a type of material in a target material

Abstract

The invention relates to a method for identifying a material type in a target material, characterized in that it comprises the following steps: acquisition of measurement data by a portable measuring device (25) comprising a spectral reflectance measurement of the target material, the reflectance being obtained by emitting light rays; measurement filtering to form a successive sequence comprising sub-bands of measured values ​​that represent characteristic transmission signatures of the article; processing by comparing the signatures with models contained in a database, the processing being configured to determine the material type of the target material. Figure for the abstract: Fig. 6

Description

Title of the invention: Method for identifying a type of material in a target material Technical field of the invention

[0001] The present invention relates to a method for identifying a type of material in a target material.

[0002] It applies, in particular, to the detection of material and sub-families of material in the plastic, textile, metal, organic and mineral fields, of different additives within a material, of multi-materials, and for the combined detection of multi-materials and additives. Prior art

[0003] Many industrial fields including, for example, pharmaceutical, chemical, nutritional, plastic products, due to increasingly strict regulations, require the rapid and non-destructive identification of materials. The use of material identification devices then makes it possible to guarantee the quality and safety of the products.

[0004] Manual sorting has a reliability that depends on the experience of the sorter, which makes this reliability random. Added to this are the poor and even dangerous working conditions linked to respiratory risks, possible fires, and high reject rates due to identification difficulties impacting profitability.

[0005] The use of material identification devices then represents a saving in time, performance, reliability, security, and ultimately, profitability.

[0006] Certain prior art documents also propose a spectroscopic detection system and a method for identifying materials. For example, publication document US2015022802 is known.

[0007] The prior art invention provides a method and system for chemical identification using a light source having a short wavelength range of the near infrared from about 600 to about 1100 nm. A silicon detector is used to detect a reflected beam from the sample to generate a detection signal. A silicon detector may be used to measure the short wavelength of the near infrared region. The detection signal from the detector is sent to a computer. The computer is programmed with software modules for data acquisition of the detection signal and analysis of the data to determine the absorption spectrum. The computer may include memory or remote access to one or more databases for storing a library absorption spectra. The absorption spectrum of the material that provides the highest matching value is identified as the material to be inspected. Data acquisition and data analysis of said detection signal therefore allows the sample to be identified.

[0008] The disadvantages of the invention of the prior art are that there is no indication concerning the identification method, no information is accessible to detail the identification method. Therefore, it does not appear reliable to repeat the described method.

[0009] Furthermore, it appears that it is necessary to present the material to the system and not for the system to read the different materials; it seems complex to set up such a system in a sorting chain for different materials, for example.

[0010] For plastic, some prior art devices for identifying this type of material are simple to use, reliable, and allow automatic interpretation, but have the disadvantages of being non-portable, or portable but with poor ergonomics, or have a busy user interface which hinders ease of use, or requires that the interpretation of the results be done by oneself, or even are expensive. Thus, no prior art document refers to all these characteristics for a single device for identifying plastic material. Presentation of the invention

[0011] The present invention aims to remedy these drawbacks with a completely innovative approach.

[0012] The present invention makes it possible to make the waste treatment chain more efficient by relying on optics and data processing in order to improve sorting.

[0013] More specifically, the invention aims to provide a technique allowing the detection of material and sub-family of material with an easy-to-transport apparatus, the calculation of the percentage of distribution of the different additives within a material, the detection of multi-materials by calculating the percentage of distribution between the different materials, and the combined detection of multi-materials and additives by calculating the percentage of distribution between the different materials and additives.

[0014] In particular, an objective of the invention is to provide such a technique making it possible to dispense with any other complex adjustment system.

[0015] Another objective of the invention is to provide such a technique which is inexpensive to implement and which does not require any particular maintenance.

[0016] An objective of the invention is to provide such a technique, which can be easily adapted to existing systems.

[0017] These objectives, as well as others which will appear subsequently, are achieved, according to a first aspect, using a method for identifying a type of material in a target material, remarkable in that it comprises the following steps: - acquisition of measurement data by a portable measuring device comprising a measurement by spectral reflectance of the target material, the reflectance is obtained by an emission of light rays; - measurement filter to form a successive sequence comprising sub-bands of measured values ​​which represent transmission signatures characteristic of the article; - processing by comparing signatures with models contained in a database, the processing being configured to determine the material type of the target material.

[0018] Spectral reflectance of the target material means whether the light is reflected light or passing light.

[0019] The present invention offers a more simplified use since no training is required to learn the sorting standards.

[0020] The present invention is portable, which means that it can be easily adapted to the process already in place.

[0021] The present invention has very high reliability and makes it possible to comply with market standards.

[0022] The present invention makes it possible to promote the safety of agents and working conditions are therefore improved.

[0023] In the field of sorting, the present invention limits sorting waste while guaranteeing better profitability.

[0024] In one example, the database is updated which allows for better reliability with respect to new elements detected as new signatures exist.

[0025] The present invention is rapid (the detection time is of the order of a second, which is suitable for manual sorting)

[0026] The present invention is non-destructive (no alteration of the material, however a version could be released later for black plastics with a jet of sparks which can alter the material)

[0027] The present invention operates on battery power, it is indeed a portable measuring device.

[0028] The present invention makes it possible to provide statistics useful for managing material stocks, for example, the results of scans of the devices to indicate what percentage of which plastic is present at the sorting output.

[0029] The invention is advantageously implemented according to the embodiments and variants set out below, which are to be considered individually or according to any technically effective combination.

[0030] In one embodiment, the processing step comprises the following sub-steps: - constitution of sub-bands of the filter stage, defining sub-bands from the sub-bands of measured values ​​to form transmission signatures characteristic of the article.

[0031] This improves the accuracy of the signatures.

[0032] In one embodiment, the processing step comprises a forest of decision trees or an artificial neural network, said artificial neural network being characterized in that it comprises the following sub-steps: - constitution of said artificial neural network, said artificial neural network comprises a layer of input neurons and a layer of output neurons which are all connected to each other; any artificial input neuron of said neural network is connected to an artificial output neuron via an artificial synapse; said artificial neural network is contained in a microcontroller; - implementation of said neural network by testing the different models contained in the database and selecting the signatures closest to the transmission signatures resulting from the measurement filter step in relation to models contained in the database.

[0033] In one embodiment, the emission of light rays of the data acquisition step comes from a near-infrared radiation source, the spectral band of the near-infrared light rays being between 900 and 1700 nm.

[0034] In one embodiment, the emission of light rays of the data acquisition step comes from an infrared radiation source.

[0035] In one embodiment, the emission of light rays of the data acquisition step comes from a radiation source in the visible spectrum.

[0036] In one embodiment, the treatment step relates to plastic, textile, metal and / or organic materials.

[0037] In one embodiment, said method further comprises a step of displaying the type of material of the article on a screen of the measuring device.

[0038] In one embodiment, in the measurement filter step, a value of the transmission signatures characteristic of the item is between 0 and 1; and wherein in the processing step, there is a correlation threshold that is configured to validate the material type of the target material when the value of the transmission signatures characteristics of the item is equal to or greater than 94% of the value of one of the models contained in the database.

[0039] In one embodiment, said method comprises a prior calibration step by acquiring a measurement with a reference surface, said measurement is stored in the database. Brief description of the figures

[0040] Other advantages, aims and characteristics of the present invention emerge from the following description given, for explanatory and non-limiting purposes, with reference to the appended drawings, in which:

[0041] [Fig.l] represents the original spectral curve acquired by the measuring device;

[0042] [Fig.2] represents the spectral curve obtained after identification of the sub-bands discriminating;

[0043] [Fig.3] represents the spectral curve obtained after grouping the identified discriminating sub-bands into a single band;

[0044] [Fig.4] represents a diagram of the principle of the decision tree;

[0045] [Fig.5] shows a diagram of the measuring device for detecting materials;

[0046] [Fig.6] shows an exploded view of the internal architecture of the material detection apparatus.

[0047] [Fig.7] represents a curve resulting from the acquisition of measurement data from the material detection device.

[0048] [Fig.8] represents another curve resulting from the acquisition of measurement data from the material detection device. Description of the embodiments

[0049] [Fig.l] represents the original spectral curve acquired by the measuring device.

[0050] The abscissa corresponds to the wave number in nm, and the ordinate corresponds to the percentage of transmittance.

[0051] According to one example, the present invention proposes a preliminary calibration step.

[0052] Calibration is accessible by means of selection made through the menu, a calibration is made by acquiring a spectral curve from a white reference surface. This standard curve is then used to process each spectral curve acquired by the spectrometer on a type of material.

[0053] According to this step, it is possible for the measuring device to be recalibrated to ensure good performance. This prevents it from going out of adjustment.

[0054] The reference surfaces are created from paint, wood and paper on which a layer of protective varnish has been applied.

[0055] The present invention proposes a first acquisition step during which a measuring device is used.

[0056] This measuring device receives as input the light reflected from the surface, or the light passing through the material.

[0057] According to a variant, the acquisition of the spectral curve of a material sample begins with the illumination of the surface to be scanned.

[0058] According to another variant, the acquisition of the spectral curve of a sample of material begins with illumination through the material by lamps located behind the material relative to the measuring device.

[0059] Thus, at the input of the system, the original scan of the spectral band in the infrared and / or near infrared or visible range is obtained.

[0060] The measuring device then returns the data associated with the spectral curve in the form of a list of values.

[0061] Near infrared spectroscopy has a spectral range between 900-1700 nm. This frequency range is chosen because: the vast majority of plastics have absorption bands in this range, this range is sufficient to discriminate one plastic from another, sub-bands have been identified in this band in order to identify the types of plastics, and for design cost reasons.

[0062] According to one embodiment, the acquisition curve for plastic and textile transmitted by the measuring apparatus is defined from 900nm to 1700nm. According to one example, with 80 points received from the curve, each representing the average value over a spectral width of 9.37nm, the whole having an exposure time to capture the light of 1.270 ms.

[0063] [Fig.2] represents the spectral curve obtained after identification of the sub-bands discriminating.

[0064] The present invention proposes a second measurement filtration step forming a sequence comprising sub-bands of measured values ​​which represent transmission signatures characteristic of the sample.

[0065] The filtration begins with the identification of the discriminating sub-bands corresponding to the hatched areas.

[0066] There are two types of model, either by neural networks or by decision tree.

[0067] In the case of neural networks, the method involves a design which is originally schematically inspired by the functioning of biological neurons, and which subsequently moved closer to statistical methods.

[0068] According to an example of neural networks, each input corresponds to a value, processed or not, of the spectral curve, or of a sub-curve. One or the inputs passes through nodes, hence the name neural network and which is a regression model Nonlinear supervised network trained by the database. Each neural network used is different and specific to the recognition dedicated to the node to which it is attached. The database contains each characteristic of each node. It should be noted that the neural network that is used can have from 1 to N sub-layers (N Layers). The number of sub-layers is defined according to the most convincing results per node, by training the neural network with the database.

[0069] In a variant, instead of using neural networks, a decision tree forest is used, this is an algorithm that combines the concepts of random subspaces.

[0070] Concerning the decision tree forest, the principle is as follows: the data are input to several decision trees which each return a result. The final result is that of the majority of the results obtained by each of the trees.

[0071] [Fig.3] represents the spectral curve obtained after grouping the identified discriminating sub-bands into a single band.

[0072] Following the identification of the discriminating sub-bands, and if there is more than one sub-band which is used by the neural network, the sub-bands are grouped into a single band.

[0073] The present invention proposes a processing step by comparing the signatures with models established within a database, the processing makes it possible to determine the type of material of the sample.

[0074] The processing of the curve received by the measuring device in relation to the standard curve includes the calculation of the absorbance according to the following formula:

[0075] For each value i of the points of the spectral curve:

[0076] I = Iogl0(standard_curve(i))-logl0(spectral_curve(i))

[0077] The set of 80 I values ​​obtained defines the curve in a space of 80 dimensions.

[0078] All the acquisitions made for the database are used to initially calculate the PCA model, to reduce the number of dimensions of each curve.

[0079] PCA stands for Principal Component Analysis. This allows data to be projected from an N-dimensional space to a K-dimensional space, with K <N.

[0080] The number of dimensions is different depending on the types of materials to be detected.

[0081] PCA consists of the projection of the spectrum onto some calculated vectors in order to concentrate the information.

[0082] The PCA model will subsequently be used to reduce all new acquisitions from 80 to N dimensions.

[0083] Subsequently, the identification of the type of plastic is enabled by artificial intelligence using a neural network.

[0084] These N values ​​are the input values ​​of the artificial intelligence models, first trained by the database, then receiving values ​​from new scans.

[0085] Indeed, the neural network has an identification algorithm that allows the synaptic weights to be modified according to a data set presented as input to the network. This training allows the neural network to identify from the given examples and in particular here from different real materials. If the training is correctly carried out, the network is capable of providing output responses very close to the original values ​​of the training data set.

[0086] The spectral database is designed with different samples of plastics already identified.

[0087] According to one embodiment, said spectral database is enhanced as and when, and in particular by data recovered via devices for identifying materials in circulation.

[0088] [Fig.4] represents a diagram of the principle of the decision tree.

[0089] To obtain the composition of the sample from which the spectrum is obtained in several elements of materials and additives, a decision tree is used in which each node represented by a circle is a neural network.

[0090] Each neural network interprets a set of sub-bands of the spectrum specific to its field of discrimination and allows the sub-families to be identified to be directed to the tree structure.

[0091] The sub-bands for each neural network are predetermined in advance in the system, and are each time sub-bands of the initial spectrum to be analyzed.

[0092] Within the decision tree, each node is connected to N branches of the tree and represents a material family.

[0093] In addition, the branches of a node determine the field of subfamilies of the parent node.

[0094] Thus, each node is specialized in detecting the type of subfamily to which the family of matter detected at the preceding node belongs.

[0095] Each neural network is associated with a set of K sub-bands (K>=1) allowing the discrimination of sub-families specific to a node. The neural network has been trained specifically for a set of specific sub-families. However, the sets of sub-families of two neural networks of two distinct nodes may overlap or intersect.

[0096] Each node returns a degree of confidence on the prediction as a percentage of the subfamily to which the spectrum belongs.

[0097] The decision tree is traversed from node to node, that is to say from sub-family to sub-family, by selecting the best detection score given by a neural network.

[0098] Then, the node that returns the best result is selected to go to the corresponding subfamily and refine the detection.

[0099] If the neural network does not give any prediction score or below a certain threshold, the path stops at the corresponding node, and therefore at the corresponding subfamily as matter prediction.

[0100] Each threshold is specific to a node.

[0101] At the output of the process, the type of material detected is indicated, as well as a confidence value in the detection ranging from 0 to 100% (also called correlation threshold).

[0102] According to one embodiment, in order to detect several materials, the decision tree is traversed while keeping only the first two or three nodes which have the best result in their category.

[0103] According to one embodiment, in order to detect several materials, and from several decision trees according to the type of material (plastic, additives, textile, metals, etc.), several decision trees are traversed according to the desired analysis to compile the results between families of materials.

[0104] Several decision trees are used in order to avoid having too many daughter nodes from the first branch, which would generate a computational overhead.

[0105] According to one embodiment, to detect multi-materials, the spectral characteristics of each material are visible in the spectral curve. The discriminating sub-bands are cut into sub-bands specific to each material. Subsequently, the ratio between each peak is calculated to determine the proportion between each material.

[0106] The peaks are identified by a neural network or a derivative calculation, in order to determine the local extrema of the peaks. Finally, the ratio of each corresponding peak in the database is calculated.

[0107] [Fig.5] shows a diagram of the material detection apparatus.

[0108] The device is composed of a screen (22), a system "standby" button (20), a menu selection button (21), an external USB (acronym for "Universal Serial Bus") socket for updating or retrieving data (24), and an external USB charging socket (23).

[0109] The screen (22) constitutes the display allowing the user to launch the scans and to read the results of the identification algorithm on the screen.

[0110] According to one example, the display mode makes it possible to display only the name of the plastic on the screen (22), and possibly a percentage confidence rate of the recognized plastic.

[0111] According to one example, the menu selection button (21) allows access to the menu in order to carry out the calibration and / or the update.

[0112] The system's "standby" button (20) allows the device to be turned on or off.

[0113] Data concerning the scans are also stored and then retrieved via a remote connection or a USB key using the external USB socket for updating or retrieving data (24).

[0114] According to one embodiment, the data extracted from the scans recovered via USB key can subsequently feed the recognition database.

[0115] [Fig.6] shows an exploded view of the internal architecture of the material detection device. All internal parts constituting the device are indicated.

[0116] The trigger (26) triggers the acquisition of a spectral curve by the measuring device (25), a curve which is then sent to an electronic card (27) comprising a microcontroller (for example of the Raspberry type) in order to carry out the analysis, and the result of the analysis is then displayed on the screen (22).

[0117] According to one embodiment, the measuring device (25) is a spectrometer having a spectral resolution of 10 nm.

[0118] The external USB charging socket (23) connecting external peripherals is connected to the battery (28), allowing the device to be recharged.

[0119] [Fig.7] represents a curve resulting from the acquisition of measurement data from the material detection device.

[0120] At the measurement filter stage, a value of the characteristic transmission signatures of the article is between 0 and 1, in this case the value is 0.951. In the database, this model corresponds to the model of the PET value (acronym for Polyethylene terephthalate) since it is greater than 94%, it is indeed a measurement of PET.

[0121] [Fig.8] represents another curve resulting from the acquisition of measurement data from the material detection device.

[0122] In the present case, the value of the characteristic transmission signatures of the article is 0.428. In the database, this model corresponds to the model of the PE value (acronym for polyethylene), since it is less than 94%, it is not a PE measurement. List of reference signs

[0123] [Tables 1] References Designations 20 Standby button 21 Selection button 22 Screen 23 External USB charging socket 24 External USB socket for updating or retrieving data 25 Measuring device 26 Trigger 27 Electronic card 28 Battery

Claims

1.

2. Claims Method for identifying a type of material in a target material, characterized in that it comprises the following steps: - acquisition of measurement data by a portable measuring device (25) comprising a measurement by spectral reflectance of the target material, the reflectance is obtained by an emission of light rays; - measurement filter to form a successive sequence comprising sub-bands of measured values ​​which represent transmission signatures characteristic of an item; - processing by comparing the signatures with models contained in a database, the processing being configured to determine the type of material of the target material by detecting material and sub-family of material, and by calculating the percentage of distribution of the different additives within a material, said processing being configured for the detection of multi-materials by calculating the percentage of distribution between the different materials, and the combined detection of multi-materials and additives by calculating the percentage of distribution between the different materials and additives; the processing step further includes the following sub-step: - constitution of other sub-bands of the filter stage, defining other sub-bands from the sub-bands of measured values ​​to form transmission signatures characteristic of the article. Method according to claim 1, in which the processing step uses a model of a decision tree forest or a model of an artificial neural network, said method being characterized in that it comprises the following sub-steps: - constitution of said artificial neural network, said artificial neural network comprises a layer of input neurons and a layer of output neurons which are all connected to each other; any artificial input neuron of said neural network is connected to an artificial output neuron via an artificial synapse; said artificial neural network is contained in a microcontroller; - implementation of said neural network by testing the different models contained in the database and selecting the signatures closest to the transmission signatures resulting from the measurement filter step in relation to models contained in the database.

3. The method of claim 1, wherein the emission of light rays in the data acquisition step comes from a near-infrared radiation source, the spectral band of the near-infrared light rays being between 900 and 1700

4. 11111. The method of claim 1, wherein the emission of light rays in the data acquisition step comes from an infrared radiation source.

5. The method of claim 1, wherein the emission of light rays in the data acquisition step comes from a radiation source in the visible spectrum.

6. A method according to claim 1, wherein the treatment step relates to plastic, textile, metal and / or organic materials.

7. The method of claim 1, wherein said method further comprises a step of displaying the material type of the article on a screen (22) of the measuring apparatus (25).

8. The method of claim 1, wherein in the measuring filter step, a value of the characteristic transmission signatures of the item is between 0 and 1; and wherein in the processing step, there is a correlation threshold that is configured to validate the material type of the target material when the value of the characteristic transmission signatures of the item is equal to or greater than 94% of the value of one of the models contained in the database.

9. Method according to claim 1, wherein said method comprises a prior step of calibration by acquiring a measurement with a reference surface, said measurement is stored in the database.

Images