Search

CN-121973355-A - Plastic sorting system based on short-wave infrared multispectral camera imaging

CN121973355ACN 121973355 ACN121973355 ACN 121973355ACN-121973355-A

Abstract

The invention provides a plastic sorting system based on short-wave infrared multispectral camera imaging, which comprises a spectrum database building module, a key wave band screening module, a multispectral sorting model building module and an industrial grade sorting device building module, wherein the spectrum database building module is used for providing data support for subsequent wave band screening and model training, the key wave band screening module is used for executing a data preprocessing and fusion type feature extraction method, the key wave band with the highest distinction degree is screened out from original wave bands, the multispectral sorting model building module is used for building a multispectral sorting model based on the key wave band with the highest distinction degree, so that accurate identification of plastic types is realized, and the industrial grade sorting device building module is used for building a short-wave infrared multispectral imaging plastic sorting device for simulating an industrial production line, so that real-time dynamic sorting is realized. The invention reasonably and effectively realizes dimension reduction of the spectrum band, thereby replacing a high hyperspectral camera and reducing the hardware cost and complexity of the system.

Inventors

  • CHEN LINSEN
  • LI YUANHAO
  • REN YIPENG
  • Cui shaoyang
  • CUI XIANGYU
  • CAO XUN

Assignees

  • 南京大学

Dates

Publication Date
20260505
Application Date
20260122

Claims (10)

  1. 1. The plastic sorting system based on the short-wave infrared multispectral camera imaging is characterized by comprising a spectrum database building module, a key wave band screening module, a multispectral classification model building module and an industrial level sorting device building module; the spectrum database building module is used for executing a method for rapidly collecting short-wave infrared multispectral spectrum data of different types of plastics; the key wave band screening module is used for executing a data preprocessing and fusion type feature extraction method, and 4 key wave bands with highest discrimination degree are screened out of 224 original wave bands; The multispectral classification model construction module constructs a multispectral classification model based on 4 key wave bands with highest discrimination, so as to realize the identification of plastic types; The industrial grade sorting device building module is used for building a short wave infrared multispectral imaging plastic sorting device for simulating an industrial production line, and real-time dynamic sorting is realized.
  2. 2. The system of claim 1, wherein the method for rapidly acquiring short-wave infrared multispectral spectrum data of different types of plastics comprises the following steps: Step a1, a short-wave infrared hyperspectral camera is installed as core acquisition equipment, halogen lamps are installed on two sides of the short-wave infrared hyperspectral camera, the halogen lamps provide illumination light rays in a wave band of 900-1700 nm, and then a white calibration plate providing diffuse reflection reference and a black calibration plate providing noise reference are placed on the head and tail sides of an electric push broom platform; step a2, placing a plurality of plastic samples of the same kind on an electric push-broom platform for shooting each time, and obtaining spectrum data of 224 wave bands of each batch of samples within a range of 900-1700 nm, namely an original brightness value DN obtained by an internal sensor of a short-wave infrared hyperspectral camera; and a3, repeating the step a2, and sequentially obtaining spectrum data sets of different types of plastics.
  3. 3. The system of claim 2, wherein the critical band screening module is configured to perform the steps of: Step b1, preprocessing a spectral data set to obtain a spectral reflectance data set; Step b2, performing primary dimension reduction on the spectral reflectance dataset to generate a pseudo-multispectral dataset; and b3, performing secondary dimension reduction on the pseudo multispectral data set, and finally screening out 4 key wave bands with highest discrimination.
  4. 4. The system of claim 3, wherein step b1 comprises marking the outline of the plastic sample and preserving the effective pixels inside, sorting to obtain a net data set for each plastic sample, and removing the abnormal value by using a quarter-bit-distance IQR method, namely removing DN value in the following steps The above sum The following data, wherein 、 The first and third quartiles of the data set, ; Combining the spectrum data in the areas of the white calibration plate and the black calibration plate, and carrying out normalization operation on all obtained effective data points, wherein the formula is as follows Wherein D is new normalized data, 、 、 And obtaining a spectral reflectance data set according to D, wherein the data are effective data points, diffuse reflection reference data in a white calibration plate area and noise reference data in a black calibration plate area respectively.
  5. 5. The system of claim 4, wherein step b2 comprises performing an integral operation on the data according to different band intervals according to the band and transmittance response curves of the alternative bandpass filter product, and converting the original data into 42-dimensional pseudo-multispectral data according to the formula: , Wherein, the At a wavelength of The transmission coefficient of the corresponding optical filter is, At a wavelength of N-th dimension data corresponding to the wave band, Is wavelength of Is a very small increment of (1), A wavelength lower bound and a wavelength upper bound corresponding to a t new segment band interval respectively, Is the value of the t-th maintenance new spectral reflectivity after one-time dimension reduction according to A primary dimension-reduction spectral reflectance dataset is obtained, Z representing an integer set.
  6. 6. The system of claim 5, wherein step b3 comprises: Step b3-1, randomly selecting partial wave bands as initial feature subsets on the basis of a one-time dimensionality reduction spectral reflectance data set, carrying out feature classification on samples by using a K-means clustering method on the initial feature subsets, calculating an analysis of variance ANOVA-F value of each attempt to verify the significance of the selected wave bands, Step b3-2, repeating the step b3-1, solving the optimal wave band combination by adopting a violence recursion method, and selecting 6 alternative wave bands; And b3-3, analyzing the corresponding molecular structure and characteristic functional groups of various plastics in all alternative wave bands by combining chemical properties, and simultaneously carrying out differential comparison verification, and finally screening out 4 key wave bands with highest discrimination, namely 1140nm, 1200nm, 1245nm and 1310nm.
  7. 7. The system of claim 6, wherein the multispectral classification model construction module performs the steps of: Step c1, selecting a short-wave infrared multispectral camera taking an InGaAs detector as a core component of the machine core, and respectively fixing band-pass filters of 4 selected key wave bands with highest discrimination in front of the short-wave infrared multispectral camera, so that the camera, the filters and the plastic sample are on the same horizontal plane, and simultaneously, a white calibration plate and a black calibration plate are placed on the left side and the right side of the sample; step c2, using a camera to shoot a plastic sample horizontally, and simultaneously adjusting the height of the halogen lamp and the angle of the lampshade baffle plate to enable the incidence plane of the light to form an angle with the horizontal plane Take the value at The exposure time of the camera is adjusted according to the distance, different types of plastic samples are sequentially shot into 1 group of data, then the bandpass filters with different center wavelengths are replaced, shooting is repeated, and finally 4 groups of data are obtained; Step c3, preprocessing the multispectral data set of all kinds of plastic samples under 4 different bandpass filters obtained by shooting in the step c2 in the step b1 to obtain a normalized spectral reflectivity data set; Step c4, taking the normalized spectral reflectance data set as a basic feature, and calculating the following spectral enhancement features: original spectral feature vector Wherein The spectral reflectance value in the characteristic band i; Spectral interdomain ratio Wherein 、 The spectral reflectance values in the characteristic band i and the characteristic band j respectively, Obtaining 6 groups of characteristic data for the inter-domain ratio of the characteristic wave bands i and j; Spectral similarity features ; 6 Groups of characteristic data are obtained for the similarity characteristics of the characteristic wave bands i and j; Finally, 17-dimensional spectrum enhancement characteristic vector data set is obtained by arrangement Wherein Respectively is All data sets, The collection of all of the data, Numbering the belonging plastic types; and c5, constructing a chemical property driven hierarchical decision tree HDT classification model based on the 17-dimensional spectrum enhancement feature vector data set to replace the traditional single-stage decision tree DT.
  8. 8. The system of claim 7, wherein in step c5, the chemical-driven hierarchical decision tree HDT classification model specifically comprises: Separating polypropylene PP and polyethylene PE from other polymers to obtain a polypropylene PP subset and a polyethylene PE subset, and classifying the polypropylene PP subset and the polyethylene PE subset by utilizing the characteristic that the polypropylene PP and the polyethylene PE are nonpolar hydrocarbon polymers and the characteristic absorption generated by C-H skeleton vibration in a near infrared spectrum and the polymer containing polar groups or heterogeneous skeletons have obvious spectrum distinguishing property; in the second layer, in the non-polypropylene PP and polyethylene PE samples, the polyvinyl chloride PVC is identified independently to obtain a subset of the polyvinyl chloride PVC, and the distinction is realized by utilizing the characteristic absorption difference caused by the chlorine-containing chemical structure of the polyvinyl chloride PVC; Dividing the residual sample into an aromatic subset and a non-aromatic subset according to aromatic and non-aromatic polymers, and further forming an aromatic and non-aromatic internal sub-subset according to the combined vibration and overtone characteristics generated by aromatic rings, ester groups, amide groups conjugation or polar groups; A random forest two classifier is adopted at each hierarchical stage, a training set and a testing set are randomly divided for the 17-dimensional spectrum enhancement features, and standardization processing is carried out to eliminate scale differences; After the first and second layers finish coarse classification, fine classification is performed in each chemical subset through a random forest second classifier of a third layer, wherein a polypropylene PP subset and a polyethylene PE subset are internally distinguished by using a classifier, a polyvinyl chloride PVC subset directly outputs a polyvinyl chloride PVC class, an aromatic subset internal classifier further distinguishes polystyrene PS, polymethyl methacrylate PMMA and polyamide PA, and a non-aromatic subset internal classifier distinguishes polyethylene terephthalate PET and polycarbonate PC; n-fold cross validation is carried out on all data in the training set and the test set to obtain the average accuracy of the classification model Wherein To take the nth data as the accuracy of the model when validating the set.
  9. 9. The system of claim 8, wherein the industrial-scale sorting device building module is used for building a short-wave infrared multispectral imaging plastic sorting device simulating an industrial production line by using an InGaAs short-wave infrared detector, an imaging lens, an electric filter wheel, a miniature RGB camera and a mechanical conveyor belt; The electric optical filter rotating wheel is provided with 6 circular grooves, 4 bandpass optical filters of 4 key wave bands with highest discrimination are arranged in the 4 grooves, and the other 2 grooves are not provided with optical filters, and the electric optical filter rotating wheel is used for realizing interframe positioning through full spectrum; The miniature RGB camera monitors the conveyor belt in real time, when unclassified plastics are identified, the system starts a sorting process, then the electric filter rotating wheel rotates at a preset speed, the short-wave infrared multispectral camera shoots at a corresponding speed, the band-pass filters on the electric filter rotating wheel respectively collect spectral data of corresponding wave bands to obtain spectral data of 4 key wave bands with highest discrimination, and inter-frame positioning is achieved through the no-filter grooves.
  10. 10. The system of claim 9, wherein the computer performs the preprocessing operation of step b1 on the spectral data of the 4 key bands with the highest degree of distinction, further calculates to obtain a 17-dimensional spectral enhancement feature vector data set, and then invokes a trained hierarchical decision tree HDT classification model to perform real-time reasoning, and outputs a classification result image through hierarchical staged decision.

Description

Plastic sorting system based on short-wave infrared multispectral camera imaging Technical Field The invention relates to the technical field of industrial sorting and spectral imaging, in particular to a plastic sorting system based on short-wave infrared multispectral camera imaging. Background With the wide application of plastics in daily life and industrial production, the problems of environmental pollution and resource waste caused by waste plastics are increasingly prominent. The key premise of improving the plastic recycling rate is to realize efficient and accurate plastic sorting, and the traditional plastic sorting mode mainly adopts manual sorting, so that the defects of low efficiency, high labor cost, harm to the health of workers and the like are overcome. In order to solve the pain point of manual sorting, various automatic sorting methods including mechanical sorting, electromagnetic sorting, optical sorting and the like appear in the prior art, wherein the optical sorting is a mainstream development direction due to the advantages of high automation degree, high sorting speed, green environmental protection and the like. The Raman spectroscopy, the laser-induced fluorescence technology, the laser-induced breakdown spectroscopy and the like in the optical sorting technology have obvious limitations that the Raman spectroscopy is easy to be interfered by sample fluorescence and has low accuracy in detecting thick samples, the laser-induced fluorescence technology can only detect substances with fluorescence characteristics, is sensitive to environment light and needs to be performed in a controlled environment, and the laser-induced breakdown spectroscopy has high equipment and maintenance cost and strict requirements on the surface evenness of the samples. The near infrared hyperspectral imaging technology is used as an emerging optical sorting technology, can acquire a three-dimensional data cube (comprising image information and spectrum information) of an object, has the characteristics of nondestructive detection and high scanning efficiency, and has certain potential in the field of plastic sorting. However, the plastic sorting scheme related to the existing near infrared hyperspectral imaging technology still has a plurality of defects that firstly, hyperspectral cameras are high in cost and high in system complexity and are difficult to popularize and apply in a large scale in industrial scenes, secondly, the method lacks versatility for extracting spectral features and developing algorithms of specific plastic types and is difficult to adapt to actual scenes with various plastic materials, complex colors and different shapes in industrial production lines, thirdly, the traditional sorting system is mainly static detection, lacks real-time continuous imaging and dynamic sorting capability and is difficult to meet the batch processing requirements of the industrial production lines, and thirdly, hyperspectral data is high in dimensionality and high in redundancy, and is difficult to cause feature loss due to direct dimension reduction, so that the sorting robustness is insufficient, and the dimension reduction method purely relying on data driving lacks material chemical knowledge support and is difficult to ensure effective extraction of key features. In addition, in the existing industrial sorting system, the traditional visual algorithm is easily influenced by appearance factors such as plastic color, shape and the like, the recognition accuracy is low, and the problems of strong band selection blindness, single feature extraction, poor generalization capability of a classification model and the like exist in the partial multispectral sorting scheme, and a dynamic sorting device matched with an industrial production line is lacking, so that real-time and continuous sorting operation cannot be realized. Therefore, the plastic sorting system with low cost, high accuracy and adaptation to the industrial production line is developed, and has important practical significance and application value. Disclosure of Invention Aiming at the problems of high cost, poor universality, low sorting efficiency, insufficient sorting accuracy and the like in the existing plastic sorting technology, the invention provides a plastic sorting system based on imaging of a short-wave infrared multispectral camera, which aims to realize the following purposes: (1) Through dimension reduction screening of spectrum bands, a short-wave infrared multispectral imaging system with low cost and high cost performance is established, and the problem that a hyperspectral camera is difficult to apply in large-scale industry is solved; (2) Providing a key wave band selection method integrating data driving and material chemistry knowledge, realizing accurate distinguishing of plastic types, and avoiding characteristic loss and classification robustness reduction; (3) Constructing a multidimensional feature extraction