Method for sorting and quickly identifying waste mixed aluminum

Abstract

The application discloses a waste miscellaneous aluminum classification and rapid identification method and relates to the field of waste miscellaneous aluminum classification and rapid identification. S1, grading pretreatment: through a conveying belt, the waste miscellaneous aluminum to be sorted is conveyed into a three-layer vibrating screen machine to be graded and screened, and waste miscellaneous aluminum of different sizes is sorted out; S2, separation of impurities: the graded waste miscellaneous aluminum is conveyed into a drum-type magnetic separator respectively to remove magnetic impurities, then the waste miscellaneous aluminum after magnetic separation is conveyed into an air flow sorting machine to separate light non-metallic impurities, and then a conveying belt conveys the waste miscellaneous aluminum onto a loading platform, full-process automatic processing reduces manual dependence, and the work load of a single production line of sorting workers can be reduced; the classification cost is reduced, compared with manual sorting, the waste product rate of recycled aluminum alloy can be effectively reduced due to accurate classification, and economic benefits are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of classification and rapid identification of scrap aluminum, and particularly to a method for classification and rapid identification of scrap aluminum. Background Technology

[0002] The efficient classification of scrap aluminum is a key bottleneck restricting its recycling. Scrap aluminum of different grades and from different sources has significant differences in composition. If mixed and smelted, it will lead to unstable performance of recycled aluminum alloys and even generate a large amount of waste.

[0003] Currently, the sorting of scrap aluminum mainly relies on manual sorting, which has inherent drawbacks such as low efficiency, high cost, and large errors. Manual sorting speed is typically only 20-30 pieces per minute, and the labor intensity is high; long-term work can easily lead to an error rate of 10-15%. Some companies use simple physical sorting methods, such as preliminary classification based on appearance, color, and shape, but these cannot distinguish between different grades of aluminum alloys with similar compositions.

[0004] Therefore, it is necessary to propose a method for classifying and rapidly identifying scrap aluminum to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for classifying and rapidly identifying scrap aluminum, in order to solve the problem that the current method of classifying scrap aluminum mainly relies on manual sorting, which has inherent defects such as low efficiency, high cost and large error.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for classifying and rapidly identifying scrap aluminum, comprising the following steps:

[0007] S1. Grading and pretreatment: The waste aluminum to be sorted is conveyed to the three-layer vibrating screen by the conveyor belt for grading and screening to separate waste aluminum of different sizes.

[0008] S2. Separation of impurities: The graded waste aluminum is fed into a drum magnetic separator to remove magnetic impurities. Then, the magnetically separated waste aluminum is fed into an air classifier to separate light non-metallic impurities. Finally, the waste aluminum is conveyed to the loading platform by a conveyor belt.

[0009] S3. Data Acquisition and Analysis: Surface images of scrap aluminum on the stage are acquired using a multispectral imaging system, and reflectance characteristic data are extracted. Subsequently, a portable X-ray fluorescence spectrometer is used to perform elemental analysis on the scrap aluminum after image acquisition to determine the content data of the main elements in the scrap aluminum.

[0010] S4. Input the data obtained in the data collection and analysis steps into the trained random forest classification model. The random forest classification model contains 200-300 decision trees with a maximum depth of 15-20 layers. The Gini coefficient is used as the splitting criterion. The training dataset covers 12 types of waste aluminum samples. The model outputs the classification results of waste aluminum through 12 classification nodes.

[0011] S5. Classification result verification and model optimization: 5% of the output classification results are randomly selected for manual review. When the misclassification rate of a certain type of waste aluminum is >2%, the misclassified samples of that type are added to the training dataset and the random forest classification model is retrained.

[0012] Preferably, the reflectivity feature data includes the reflectivity features extracted from the 450nm, 550nm, 650nm, and 850nm wavelength bands, and the main element content data of the waste aluminum is the content of five key elements: Al, Si, Cu, Mg, and Fe.

[0013] Preferably, the 12 types of scrap aluminum samples include 1050 pure aluminum, 2024 aluminum alloy, 3003 aluminum alloy, 4043 aluminum alloy, 5052 aluminum alloy, 6061 aluminum alloy, 7075 aluminum alloy, aluminum foil scrap, aluminum profile scrap, aluminum casting scrap, aluminum wire scrap, and composite aluminum scrap, with ≥5000 samples in each type and covering different surface states such as oxidation, painting, and rust.

[0014] Preferably, the airflow generator of the airflow separator is a centrifugal fan with a fan pressure of 1500-2000Pa, and a guide plate is installed in the separation chamber. The angle of the guide plate is adjusted within the range of 0-10° according to the density of the waste aluminum, thereby further improving the separation efficiency of light non-metallic impurities.

[0015] Preferably, the three-layer vibrating screen has a first layer with a screen hole diameter of 30mm for separating large pieces of scrap aluminum weighing ≥500g, a second layer with a screen hole diameter of 20mm for separating medium-sized pieces of scrap aluminum weighing 100-500g, and a third layer with a screen hole diameter of 10mm for separating small pieces of scrap aluminum weighing ≤100g. The vibration frequency of the vibrating screen is set to 30-40Hz, the amplitude is set to 5-8mm, and the processing capacity is controlled at 200-300kg / h.

[0016] Preferably, the multispectral imaging system includes a multispectral camera with a resolution of 5 megapixels and a spectral range of 400-1000nm, and an LED ring light source with a color temperature of 5500-6500K. The camera acquisition distance is set to 30-50cm and the exposure time is set to 10-30ms. After acquisition, the image is subjected to grayscale correction, noise reduction processing and edge extraction.

[0017] Preferably, the stage of the multispectral imaging system adopts a conveyor belt structure, and the speed of the conveyor belt is matched with the processing capacity of the vibrating screen in the three-layer vibrating screen machine to realize continuous image acquisition of waste aluminum. The surface of the conveyor belt is made of black matte material to reduce the interference of ambient light on image acquisition.

[0018] Preferably, the training process of the random forest classification model adopts the 5-fold cross-validation method. By adjusting the number of decision trees, the maximum depth, and the splitting criterion parameters, the manual verification process adopts the standard sample comparison method. That is, for the extracted samples, the category is first preliminarily judged by manually observing the appearance and measuring the density, and then compared with the multispectral features and elemental features of the standard samples.

[0019] Preferably, the distance between the detection probe of the portable X-ray fluorescence spectrometer and the detection surface of the waste aluminum is fixed at 5-8 mm, and a protective shield is provided around the probe, with the lead equivalent of the shield being ≥0.5 mmPb.

[0020] Preferably, the sorting process is automated through a PLC control system. The PLC system is connected to the three-layer vibrating screen, magnetic separator, air classifier, multispectral camera, X-ray fluorescence spectrometer and conveyor belt equipment, respectively, to collect the operating parameters of each device in real time and display the operating status on the touch screen.

[0021] The technical effects and advantages of this invention are as follows:

[0022] 1. This invention integrates multispectral surface features and X-ray fluorescence elemental features, combined with an optimized random forest model, achieving a classification accuracy of ≥98%, which is an improvement over traditional single methods. For waste aluminum with complex surface conditions such as oxidation and painting, the misclassification rate is controlled within 2%, solving the problem that traditional methods are greatly affected by surface conditions.

[0023] 2. The three-stage physical pretreatment removes ferromagnetic and non-metallic impurities, improving the purity of waste aluminum; the surface cleaning process reduces the impact of dust and unevenness on testing, lowers elemental analysis errors, and provides a reliable data basis for accurate classification.

[0024] 3. The entire system processes materials faster than traditional manual methods, meeting the assembly line operation needs of large-scale recycled aluminum enterprises with a large daily processing capacity. High automation reduces costs: fully automated processing reduces reliance on manual labor, and a single production line can reduce the workload of sorting workers; classification costs are reduced, and due to more precise classification compared to manual sorting, the scrap rate of recycled aluminum alloys can be effectively reduced, significantly improving economic efficiency.

[0025] 4. The model is iterative and highly adaptable. Through the sample verification mechanism and dynamic training set update, the model can continuously learn the features of new samples. When new types of waste aluminum are introduced, only supplementary samples are needed to achieve accurate identification, which solves the problem of poor generalization ability of traditional models. The equipment can adjust parameters according to the source of waste aluminum to adapt to the characteristics of waste aluminum in different fields such as automobiles, construction, and electronics. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the process for classifying and rapidly identifying waste aluminum according to the present invention. Detailed Implementation

[0027] This invention provides, for example Figure 1 The method for classifying and rapidly identifying scrap aluminum shown includes the following steps:

[0028] S1. Grading and pretreatment: The waste aluminum to be sorted is transported to the three-layer vibrating screen through a conveyor belt equipped with a wind-powered dust removal mechanism for grading and screening, and is used to remove dust from the waste aluminum to be sorted.

[0029] Specifically, a three-layer vibrating screen is used, with screen aperture diameters of 10mm, 20mm, and 30mm, to classify scrap aluminum by weight. The vibration frequency is set to 30-40Hz, the amplitude to 5-8mm, and the processing capacity to 200-300kg / h. The first layer, with 30mm apertures, separates large pieces of scrap aluminum weighing ≥500g; the second layer, with 20mm apertures, separates medium-sized pieces weighing 100-500g; and the third layer, with 10mm apertures, separates small pieces weighing ≤100g. This classification ensures that scrap aluminum of different sizes enters the corresponding detection channels, avoiding interference from size differences in subsequent testing.

[0030] S2. Impurity Separation: The graded waste aluminum is conveyed to a drum magnetic separator to remove magnetic impurities. The magnetically separated waste aluminum is then conveyed to an air classifier to separate light non-metallic impurities. Subsequently, a conveyor belt equipped with a wind-driven dust removal mechanism transports the waste aluminum to a loading platform. The drum magnetic separator operates at a speed of 30-40 r / min, with the magnetic field strength increasing from 1000 Gs to 1200 Gs along the drum axis. This ensures a removal rate of ≥99% for iron impurities (including micro-iron powder) of varying magnetic strengths, minimizing the impact of iron impurities on subsequent spectral analysis. The airflow angle of the air classifier is 30-45° to the horizontal, with a wind speed controlled at 5-8 m / s. This allows for precise separation of light non-metallic impurities such as plastics and rubber with a density ≤1 g / cm³, achieving a separation efficiency ≥95%. The purity of the pre-treated waste aluminum is increased to over 98%, laying the foundation for subsequent testing.

[0031] S3. Data Acquisition and Analysis: Surface images of scrap aluminum on the stage are acquired using a multispectral camera, and reflectance characteristic data are extracted. Subsequently, a portable X-ray fluorescence spectrometer is used to perform elemental analysis on the scrap aluminum after image acquisition to determine the content data of the main elements in the scrap aluminum.

[0032] For different surface conditions of scrap aluminum (oxidized, painted, smooth), a multispectral camera with a spectral range of 400-1000nm and a resolution of 5 megapixels was used to acquire images of the scrap aluminum surface. The camera acquisition distance was 30-50cm, the exposure time was 10-30ms, and an LED ring light source with a color temperature of 5500-6500K was used to ensure image stability.

[0033] The reflectance characteristics of four key wavelength bands are extracted: 450nm (reflecting surface coating), 550nm (metallic color), 650nm (oxide layer characteristics), and 850nm (near-infrared reflectance). Specifically, the reflectance of the 450nm band is (R1), the reflectance of the 550nm band is (R2), and the difference between the reflectance of the 650nm and 850nm bands is (R3-R4). The R2 / R1 ratio (to distinguish between painted and unpainted surfaces) and the R3-R4 difference (to determine the degree of oxidation) are calculated.

[0034] The acquired images are preprocessed, including grayscale correction using a standard whiteboard, noise reduction using median filtering with a 3×3 pixel window, and edge extraction using the Canny operator with a threshold of 100-200. Finally, a 128-dimensional feature vector is formed to achieve a quantitative description of surface features.

[0035] The elemental content of waste aluminum was determined using a portable X-ray fluorescence spectrometer. The spectrometer tube voltage was set to 30-50 kV, the tube current to 50-100 μA, and the detection spot diameter to 3-5 mm. The content analysis of five key elements—Al, Si, Cu, Mg, and Fe—could be completed within 2-3 seconds, with a detection error ≤ ±0.2%. Specifically, the Al content was ≥80%, and the contents of Si, Cu, Mg, and Fe were each 0.1%–10%.

[0036] Before testing, the surface of the scrap aluminum is purged with high-pressure air at a pressure of 0.4-0.6 MPa to remove dust and impurities with a particle size ≤50μm. For scrap aluminum with uneven surfaces, mechanical adjustment is used to ensure the flatness of the testing surface is ≤0.5mm, thus ensuring testing accuracy. The elemental detection data forms a 5-dimensional feature vector, reflecting the intrinsic compositional characteristics of the scrap aluminum.

[0037] S4. Input the data obtained in the data collection and analysis steps into the trained random forest classification model;

[0038] The random forest classification model contains 200-300 decision trees with a maximum depth of 15-20 layers, using the Gini coefficient as the splitting criterion. The model input is a fusion vector of multispectral features (128 dimensions) and elemental features (5 dimensions), and the output is the classification result through 12 classification nodes (corresponding to 12 types of waste aluminum).

[0039] The training dataset covers 12 categories of scrap aluminum samples, with ≥5000 samples in each category, including scrap aluminum in different surface states (oxidized, painted, rusted). Five-fold cross-validation was used to optimize the model parameters, ensuring a classification accuracy of ≥98%.

[0040] A classification result verification step is set up, where 5% of randomly selected samples are manually reviewed. When the misclassification rate of a single class exceeds 2%, the misclassified samples are added to the training set to retrain the model until the classification accuracy requirement is met. Simultaneously, the feature database for each type of scrap aluminum is supplemented with ≥1000 samples monthly to achieve dynamic model optimization and adapt to the identification needs of new types of scrap aluminum.

[0041] The 12 categories of scrap aluminum include: 1050 pure aluminum, 2024 aluminum alloy, 3003 aluminum alloy, 4043 aluminum alloy, 5052 aluminum alloy, 6061 aluminum alloy, 7075 aluminum alloy, aluminum foil scrap, aluminum profile scrap, aluminum casting scrap, aluminum wire scrap, and composite aluminum scrap. The database for each category is updated in real time (≥1000 samples are added monthly).

[0042] S5. Classification result verification and model optimization: 5% of the output classification results are randomly selected for manual review. When the misclassification rate of a certain type of waste aluminum is >2%, the misclassified samples of that type are added to the training dataset and the random forest classification model is retrained.

[0043] This invention integrates multispectral surface features and X-ray fluorescence elemental features, combined with an optimized random forest model, achieving a classification accuracy of ≥98%, which is 13-28 percentage points higher than the traditional single method (accuracy of 70%-85%). For waste aluminum with complex states such as surface oxidation and painting, the misclassification rate is controlled within 2%, solving the problem that traditional methods are greatly affected by surface state.

[0044] The entire system has a processing speed of 150-200 pieces / minute, which is 5-10 times faster than manual sorting (20-30 pieces / minute); the X-ray fluorescence detection time is shortened to 2-3 seconds / piece, which is 1-3 times faster than traditional equipment (5-10 seconds / piece), and can meet the production line operation needs of large-scale recycled aluminum enterprises, with a daily processing capacity of 200-300 tons.

[0045] The three-stage physical pretreatment process achieves a ferromagnetic impurity removal rate of ≥99%, a non-metallic impurity removal rate of ≥95%, and increases the purity of waste aluminum to over 98%. The surface cleaning process reduces the impact of dust and unevenness on testing, ensuring that the elemental analysis error is ≤±0.2%, providing a reliable data foundation for accurate classification.

[0046] Full-process automation reduces reliance on manual labor, with a single production line reducing sorting workers by 80%; sorting costs are reduced to 15-20 yuan / ton, a 75%-85% reduction compared to manual sorting (80-100 yuan / ton); at the same time, due to precise sorting, the scrap rate of recycled aluminum alloys is reduced from 15%-20% to 3%-5%, significantly improving economic efficiency.

[0047] Through a 5% sample verification mechanism and dynamic training set updates, the model can continuously learn the features of new samples. When new types of scrap aluminum are introduced, only 500-1000 additional samples are needed to achieve accurate identification, solving the problem of poor generalization ability of traditional models. The equipment can adjust parameters according to the source of scrap aluminum to adapt to the characteristics of scrap aluminum in different fields such as automobiles, construction, and electronics. Example

[0048] This embodiment uses the above-described technical solution to identify 1050 pure aluminum (Al≥99.5%, Fe≤0.4%).

[0049] Physical pretreatment: Vibrating screen frequency 30Hz, amplitude 5mm, 30mm screen hole to separate large pieces (≥500g) of 1050 pure aluminum; magnetic separator speed 30r / min, magnetic field strength 1000Gs; airflow separation wind speed 5m / s, angle 30°.

[0050] Multispectral imaging: acquisition distance 50cm, exposure time 30ms, light source color temperature 5500K, extracted R1=450nm reflectance 0.35, R2=550nm reflectance 0.42, R2 / R1=1.2, R3-R4=0.05.

[0051] X-ray fluorescence analysis: tube voltage 30kV, tube current 50μA, detection time 3 seconds, detection results were Al=99.6%, Fe=0.32%, Si=0.05%.

[0052] Machine learning classification: Multispectral and elemental features are input into a random forest model with 200 decision trees and a depth of 15 layers. The output classification result is 1050 pure aluminum. The identification is accurate when compared with standard samples. Processing speed is 150 pieces / minute, and the single-class misclassification rate is 0.8%.

[0053] Example 2: Identifying 2024 Aluminum Alloy

[0054] Identify medium-sized (100-500g) 2024 aluminum alloy (Cu=3.8%~4.9%, Mg=1.2%~1.8%).

[0055] Physical pretreatment: Vibrating screen frequency 35Hz, amplitude 7mm, 20mm screen hole; magnetic separator speed 35r / min, magnetic field strength 1100Gs; airflow separation wind speed 7m / s, angle 40°.

[0056] Multispectral imaging: acquisition distance 40cm, exposure time 20ms, light source color temperature 6000K, extraction R1=0.29, R2=0.34, R2 / R1=1.17, R3-R4=0.09.

[0057] X-ray fluorescence analysis: tube voltage 40kV, tube current 75μA, detection time 2.5 seconds, detection results were Cu=4.3%, Mg=1.5%, Al=93.8%.

[0058] Machine learning classification: The input consists of a random forest model with 250 decision trees and a depth of 17 layers. The model classifies correctly. The speed is 170 items / minute, and the misclassification rate is 0.9%.

[0059] Example 3: Identifying 3003 Aluminum Alloy

[0060] Identify small pieces (≤100g) of 3003 aluminum alloy (Mn=1.0%~1.5%, Fe≤0.7%).

[0061] Physical pretreatment: Vibrating screen frequency 40Hz, amplitude 8mm, 10mm screen aperture; magnetic separator speed 40r / min, magnetic field strength 1200Gs; airflow separation wind speed 8m / s, angle 45°.

[0062] Multispectral imaging: acquisition distance 30cm, exposure time 10ms, light source color temperature 6500K, extraction R1=0.31, R2=0.37, R2 / R1=1.19, R3-R4=0.05.

[0063] X-ray fluorescence analysis: tube voltage 50kV, tube current 100μA, detection time 2 seconds, detection results were Mn=1.5%, Fe=0.7%, Al=97.6%.

[0064] Machine learning classification: The input consists of a random forest model with 300 decision trees and a depth of 20 layers. The model classifies the data correctly. The speed is 200 items / minute, and the misclassification rate is 0.7%.

[0065] Example 4: Identifying 5052 Aluminum Alloy

[0066] Identify 5052 aluminum alloy (Mg=2.2%~2.8%, Cr=0.15%~0.35%).

[0067] Physical pretreatment: Vibrating screen frequency 32Hz, amplitude 6mm, 20mm screen hole; magnetic separator speed 32r / min, magnetic field strength 1050Gs; airflow separation wind speed 6m / s, angle 35°.

[0068] Multispectral imaging: acquisition distance 45cm, exposure time 25ms, light source color temperature 5800K, extraction R1=0.33, R2=0.39, R2 / R1=1.18, R3-R4=0.07.

[0069] X-ray fluorescence analysis: tube voltage 35kV, tube current 60μA, detection time 2.8 seconds, detection results were Mg=2.2%, Cr=0.15%, Al=97.4%.

[0070] Machine learning classification: The input consists of a random forest model with 220 decision trees and a depth of 16 layers. The model classifies correctly. The speed is 160 pieces / minute, and the misclassification rate is 0.9%.

[0071] Example 5: Identifying 6061 Aluminum Alloy

[0072] Identify 6061 aluminum alloy (Si=0.4%~0.8%, Mg=0.8%~1.2%).

[0073] Physical pretreatment: Vibrating screen frequency 35Hz, amplitude 7mm, 20mm screen hole; magnetic separator speed 35r / min, magnetic field strength 1100Gs; airflow separation wind speed 7m / s, angle 40°.

[0074] Multispectral imaging: acquisition distance 40cm, exposure time 20ms, light source color temperature 6000K, extraction R1=0.34, R2=0.40, R2 / R1=1.18, R3-R4=0.09.

[0075] X-ray fluorescence analysis: tube voltage 40kV, tube current 75μA, detection time 2.5 seconds, detection results were Si=0.6%, Mg=1.0%, Al=98.1%.

[0076] Machine learning classification: The input consists of a random forest model with 250 decision trees and a depth of 17 layers. The model classifies correctly. The speed is 170 items / minute, and the misclassification rate is 0.6%.

[0077] Example 6: Identifying 7075 Aluminum Alloy

[0078] Identify 7075 aluminum alloy (Zn=5.1%~6.1%, Mg=2.1%~2.9%).

[0079] Physical pretreatment: Vibrating screen frequency 40Hz, amplitude 8mm, 30mm screen hole; magnetic separator speed 40r / min, magnetic field strength 1200Gs; airflow separation wind speed 8m / s, angle 45°.

[0080] Multispectral imaging: acquisition distance 30cm, exposure time 10ms, light source color temperature 6500K, extraction R1=0.28, R2=0.33, R2 / R1=1.18, R3-R4=0.10.

[0081] X-ray fluorescence analysis: tube voltage 50kV, tube current 100μA, detection time 2 seconds, detection results were Zn=6.1%, Mg=2.9%, Al=90.5%.

[0082] Machine learning classification: The input consists of a random forest model with 300 decision trees and a depth of 20 layers. The model classifies correctly. The speed is 200 items / minute, and the misclassification rate is 1.0%.

[0083] Example 7: Identifying aluminum foil waste

[0084] Identify aluminum foil waste (thickness ≤ 0.2 mm, Al ≥ 99%).

[0085] Physical pretreatment: Vibrating screen frequency 30Hz, amplitude 5mm, 10mm screen aperture; magnetic separator speed 30r / min, magnetic field strength 1000Gs; airflow separation wind speed 5m / s, angle 30°.

[0086] Multispectral imaging: acquisition distance 50cm, exposure time 30ms, light source color temperature 5500K, extraction R1=0.40, R2=0.45, R2 / R1=1.13, R3-R4=0.04.

[0087] X-ray fluorescence analysis: tube voltage 30kV, tube current 50μA, detection time 3 seconds, detection results were Al=99.2%, Fe=0.2%.

[0088] Machine learning classification: The input consists of a random forest model with 200 decision trees and a depth of 15 layers. The model classifies correctly. The speed is 150 cases / minute, and the misclassification rate is 1.2%.

[0089] Example 8: Identifying aluminum profile scrap

[0090] Identify architectural aluminum profile waste (6063 series, Si=0.2%~0.6%).

[0091] Physical pretreatment: Vibrating screen frequency 35Hz, amplitude 7mm, 20mm screen hole; magnetic separator speed 35r / min, magnetic field strength 1100Gs; airflow separation wind speed 7m / s, angle 40°.

[0092] Multispectral imaging: acquisition distance 40cm, exposure time 20ms, light source color temperature 6000K, extraction R1=0.36, R2=0.42, R2 / R1=1.17, R3-R4=0.08.

[0093] X-ray fluorescence analysis: tube voltage 40kV, tube current 75μA, detection time 2.5 seconds, detection results were Si=0.4%, Mg=0.7%, Al=98.6%.

[0094] Machine learning classification: The input consists of a random forest model with 250 decision trees and a depth of 17 layers. The model classifies correctly. The speed is 170 cases / minute, and the misclassification rate is 0.7%.

[0095] Example 9: (Identifying aluminum casting scrap)

[0096] Identify automotive aluminum casting scrap (ADC12, Si=9.6%~12%).

[0097] Physical pretreatment: Vibrating screen frequency 38Hz, amplitude 7mm, 30mm screen aperture; magnetic separator speed 38r / min, magnetic field strength 1150Gs; airflow separation wind speed 7.5m / s, angle 42°.

[0098] Multispectral imaging: acquisition distance 35cm, exposure time 15ms, light source color temperature 6300K, extraction R1=0.25, R2=0.30, R2 / R1=1.20, R3-R4=0.12.

[0099] X-ray fluorescence analysis: tube voltage 45kV, tube current 90μA, detection time 2.2 seconds, detection results were Si=10.5%, Cu=2.5%, Al=86.7%.

[0100] Machine learning classification: The input consists of a random forest model with 280 decision trees and a depth of 19 layers. The model classifies correctly. The speed is 190 items / minute, and the misclassification rate is 0.9%.

[0101] Example 10: Composite Aluminum Scrap (High Limit Parameters)

[0102] Identify aluminum-plastic composite waste (aluminum content 60%–70%).

[0103] Physical pretreatment: Vibrating screen frequency 40Hz, amplitude 8mm, 20mm screen hole; magnetic separator speed 40r / min, magnetic field strength 1200Gs; airflow separation wind speed 8m / s, angle 45°.

[0104] Multispectral imaging: acquisition distance 30cm, exposure time 10ms, light source color temperature 6500K, extraction R1=0.22, R2=0.26, R2 / R1=1.18, R3-R4=0.15.

[0105] X-ray fluorescence analysis: tube voltage 50kV, tube current 100μA, detection time 2 seconds, detection result is Al=65%, the rest are non-metallic.

[0106] Machine learning classification: The input random forest model, consisting of 300 decision trees and 20 layers, was correctly classified. The speed was 200 samples / minute, with a misclassification rate of 1.1%. After verification with 5% of the samples and adjustment of the training set, the misclassification rate decreased to 0.8%.

Claims

1. A method for classifying and rapidly identifying scrap aluminum, characterized in that: Includes the following steps: S1. Grading and pretreatment: The waste aluminum to be sorted is conveyed to the three-layer vibrating screen by the conveyor belt for grading and screening to separate waste aluminum of different sizes. S2. Separation of impurities: The graded waste aluminum is fed into a drum magnetic separator to remove magnetic impurities. Then, the magnetically separated waste aluminum is fed into an air classifier to separate light non-metallic impurities. Finally, the waste aluminum is conveyed to the loading platform again by a conveyor belt. S3. Data Acquisition and Analysis: Surface images of scrap aluminum on the stage are acquired using a multispectral imaging system, and reflectance characteristic data are extracted. Subsequently, a portable X-ray fluorescence spectrometer is used to perform elemental analysis on the scrap aluminum after image acquisition to determine the content data of the main elements in the scrap aluminum. S4. Input the data obtained in the data collection and analysis steps into the trained random forest classification model. The random forest classification model contains 200-300 decision trees with a maximum depth of 15-20 layers. The Gini coefficient is used as the splitting criterion. The training dataset covers 12 types of waste aluminum samples. The model outputs the classification results of waste aluminum through the 12 types of waste aluminum samples. S5. Classification result verification and model optimization: 5% of the output classification results are randomly selected for manual review. When the misclassification rate of a certain type of waste aluminum is >2%, the misclassified samples of that type are added to the training dataset and the random forest classification model is retrained. The reflectivity feature data includes the reflectivity features extracted from the 450nm, 550nm, 650nm, and 850nm wavelength bands, and the main element content data of the waste aluminum is the content of five elements: Al, Si, Cu, Mg, and Fe. Among them, the reflectance of the 450nm band is R1, the reflectance of the 550nm band is R2, and the difference between the reflectance of the 650nm and 850nm bands is R3-R4. The R2 / R1 ratio is calculated to distinguish between painted and unpainted, and the R3-R4 difference is calculated to determine the degree of oxidation. The training process of the random forest classification model adopts the 5-fold cross-validation method to optimize the model parameters. The manual verification process adopts the standard sample comparison method, that is, for the extracted samples, the category is initially judged by manually observing the appearance and measuring the density, and then compared with the multispectral characteristics and elemental characteristics of the standard samples.

2. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The 12 categories of scrap aluminum samples include 1050 pure aluminum, 2024 aluminum alloy, 3003 aluminum alloy, 4043 aluminum alloy, 5052 aluminum alloy, 6061 aluminum alloy, 7075 aluminum alloy, aluminum foil scrap, aluminum profile scrap, aluminum casting scrap, aluminum wire scrap, and composite aluminum scrap. Each category has ≥5000 samples and covers different surface states such as oxidation and painting.

3. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The airflow generator of the airflow separator uses a centrifugal fan with a fan pressure of 1500-2000Pa. A guide plate is installed inside the separation chamber, and the angle of the guide plate is adjustable within the range of 0-10° according to the density of the waste aluminum.

4. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The three-layer vibrating screen has a first layer with a screen hole diameter of 30mm, used to separate large pieces of scrap aluminum weighing ≥500g; a second layer with a screen hole diameter of 20mm, used to separate medium-sized pieces of scrap aluminum weighing 100-500g; and a third layer with a screen hole diameter of 10mm, used to separate small pieces of scrap aluminum weighing ≤100g. The vibration frequency of the vibrating screen is set to 30-40Hz, the amplitude is set to 5-8mm, and the processing capacity is controlled at 200-300kg / h.

5. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The multispectral imaging system includes a multispectral camera with a resolution of 5 megapixels and a spectral range of 400-1000nm, and an LED ring light source with a color temperature of 5500-6500K. The camera acquisition distance is set to 30-50cm and the exposure time is set to 10-30ms. After acquisition, the image is subjected to grayscale correction, noise reduction processing and edge extraction.

6. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The stage of the multispectral imaging system adopts a conveyor belt structure. The speed of the conveyor belt is matched with the processing capacity of the vibrating screen in the three-layer vibrating screen machine to realize continuous image acquisition of waste aluminum. The surface of the conveyor belt is made of black matte material.

7. The method for classifying and rapidly identifying scrap aluminum according to claim 1, characterized in that: The distance between the detection probe of the portable X-ray fluorescence spectrometer and the detection surface of the waste aluminum is fixed at 5-8 mm, and a protective shield is set around the probe, with a lead equivalent of ≥0.5 mmPb.

8. The method for classifying and rapidly identifying scrap aluminum according to claim 5, characterized in that: The PLC control system is connected to the three-layer vibrating screen, drum magnetic separator, air classifier, multispectral camera, portable X-ray fluorescence spectrometer and conveyor belt respectively, to collect the operating parameters of each device in real time and display the operating status on the touch screen.

Images