Method and apparatus for assessing valuable metals in electronic scrap

Abstract

The present disclosure relates to the assessment of precious metal content in batches of materials such as electronic scrap, particularly printed circuit boards, an essential initial step if recycling is envisioned. A method is presented that includes the steps of imaging at least a statistically representative number of boards in the batch, processing the images to detect printed circuit boards, and for each detected printed circuit board, extracting a board-related feature vector using image processing techniques, and providing a model that takes at least the board-related feature vector as input and calculates the precious metal content of the batch as output, where the model is calibrated against batch-level precious metal assays. This application deals directly with batches of printed circuit boards, enabling an approach in which the regression model is calibrated based on total batch assays. Printed circuit board component assays are not required.

Description

[Technical Field]

[0001] The assessment of metal values ​​in batches of material such as e-scrap is an essential preliminary step if recycling is envisaged.

[0002] In the present context, the electronic scrap of interest consists primarily of printed circuit boards (PCBs). Such electronic scrap is obtained in batches containing hundreds of miscellaneous boards.

[0003] Highly accurate assays of each batch are required for commercial transactions between material providers and recyclers. Assays are also essential for closely monitoring metallurgical processes, for example, with regard to tracking any metal losses and optimizing process performance.

[0004] The value of e-scrap is most often determined by its precious metal content, i.e., the amount of PGMs (Ru, Rh, Pd, Os, Ir, Pt), Au, and Ag present in the material. Cu can also contribute significantly to this value and is considered a precious metal in this context.

[0005] Because most of these recycled materials are complex and heterogeneous, an accurate and reliable determination of their composition can only be obtained by a 100% sampling scheme, a fairly complex multi-step method that involves smelting operations performed on representative samples to collect the precious metals in the metallic phase, which are then resampled for chemical analysis. These operations are time-consuming, leading to delays in making available important information about the exact content and, consequently, to delays in commercial transactions between donors and recyclers.

[0006] Therefore, early assessment or prediction of precious metal content is advantageous, even if it is somewhat less accurate than chemical analysis.

[0007] A system was developed to address this goal.

[0008] A publication addressing a similar problem is "Material value estimation for recycling of waste printed circuit boards by deep-learning-assisted approach on X-ray images," M. Firsching et al., SBSC 2022, 9th Sensor-Based Sorting & Control, Shaker Verlag, Aachen 2022, pp. 161-178. This publication teaches a metal value estimation system for printed circuit boards. The estimation is based on preliminary chemical assays of component categories such as ICs, particularly pin grid array and ball grid array ICs, tantalum capacitors, and thickness measurements of gold coatings on connectors. The boards are subjected to automated component recognition and counting. The metal value of each board is derived from the number of components in each category. Component assays are required.

[0009] In contrast, the present application deals directly with batches of printed circuit boards, allowing for a different approach in which the regression model is calibrated based on total batch assays; no component assays are required.

[0010] As explained above, complete batch assays are always required and are always performed regardless of other estimates. By using these assays as calibration data, the delay in making accurate batch contents available is greatly reduced. Furthermore, the fact that no extra effort is required to obtain calibration data makes ongoing calibration much more affordable.

[0011] A first embodiment of the present invention is a method for assessing the precious metal content of a batch of printed circuit boards, comprising the steps of: imaging at least a statistically representative number of printed circuit boards in the batch; processing the image to detect the printed circuit board; extracting, for each detected printed circuit board, a board-related feature vector using image processing techniques; providing a model that takes as input at least the substrate-related feature vector and calculates as output the precious metal content of the batch, wherein the model is calibrated against batch-level precious metal assays.

[0012] A batch generally corresponds to a collection of items obtained from a supplier within the framework of a single purchase contract. In the context of the present invention, a batch may include any number of printed circuit boards, some of which may be broken. It may also include partially or completely separated individual components.

[0013] In the context of the present invention, the term "substrate" should be understood as "printed circuit board" or "PCB" and includes all electronic or other components mounted or attached to the substrate.

[0014] Detecting the substrate means that at least a corresponding bounding box is defined, which is preferably supplemented with a mask that more precisely identifies the contours, which makes it possible to carry out a step of extracting substrate-related features.

[0015] An "imaging system" may operate in the visible electromagnetic spectrum, but may also use IR, UV, or X-ray sources and cameras.

[0016] Board-related features are, for example, the type of board, such as motherboard or sound card, and the types of detected components on that board, such as ICs or connectors. Features are mathematically represented by feature vectors.

[0017] According to a further embodiment, the step of calculating the batch precious metal content is performed based on a combination of substrate-related feature vectors and batch-level features derived from sources other than printed circuit board imaging.

[0018] The substrate-related features are optionally combined with other batch-related features from sources unrelated to substrate imaging, such as supplier identification or the geographic origin of the batch. Historical data from batches originating from the same supplier may also be used. Further examples of batch-related features are the results of alternative rapid analytical methods performed on the batch or a statistically representative sample thereof. Prompt gamma neutron activation analysis is a technique that, when used alone, may not provide the required accuracy.

[0019] According to a further embodiment, the step of calculating the precious metal content of the batch comprises options (1) and (2): (1) summing the substrate-related feature vectors to thereby obtain a batch-related feature vector; and calculating the batch precious metal content based on the batch-related feature vector; or (2) For each printed circuit board, the step is performed by one of the following steps: evaluating the printed circuit board precious metal content based on the board-related feature vector; and summing the printed circuit board precious metal contents, thereby obtaining a batch precious metal content.

[0020] The first option may have advantages when a nonlinear regression model is used, as it better captures the complex interactions between batch and substrate. However, fitting a nonlinear model requires more calibration data. The second option provides an estimate of the precious metal content of individual substrates, allowing for classification and sorting. While prior art techniques can achieve similar results, they rely on preliminary chemical assays of many different components to calibrate the model, whereas in accordance with the present invention, a total batch assay is used instead.

[0021] When using a linear model, both described options calculate the same batch content.

[0022] According to a further embodiment, the method for assessing the precious metal content is carried out prior to metallurgical processing of the batch for precious metal recovery.

[0023] Knowledge of the precious metal content, even approximately, allows for the selection of an optimized metallurgical process: batches rich in precious metals can, for example, be processed with high priority using processes that ensure the maximum yield of precious metals.

[0024] According to a further embodiment, the method for assessing precious metal content is used to determine an advance payment to a supplier of a batch of printed circuit boards.

[0025] The advance payment corresponds to a preliminary assessment of the precious metal content using the methods described above, and is normalized once a more accurate content is determined.

[0026] According to a further embodiment, when selecting the precious metal content evaluation method according to option (2) above, the printed circuit boards are classified into at least two classes based on the evaluated printed circuit board precious metal content.

[0027] According to a further embodiment, each printed circuit board class is individually metallurgically processed for precious metal recovery.

[0028] A batch may contain printed circuit boards with widely varying precious metal contents. Separating them into at least two classes and distinguishing boards with a higher expected precious metal content from those with a lower expected precious metal content allows for the optimization of subsequent metallurgical processes for precious metal recovery and purification. However, the metal content per board can only be assessed if the calculation protocol according to option (2) above is applied.

[0029] According to a further embodiment, a batch level precious metal assay is obtained by smelting a statistically representative sample of the batch to be analyzed using the alloy as a collector.

[0030] A representative sample can be obtained by sampling the batch, followed by grinding the sample, and then resampling the ground product. This scheme of sampling, grinding, and resampling can be carried out repeatedly until a reduced sample volume is obtained that is compatible with the needs of further steps leading to a chemical assay.

[0031] An advantageous further step is smelting the sample, whereby the noble metals are separated from the less noble metals and concentrated in an alloy. Such an alloy is generally homogeneous, which is desirable for accurate chemical assays. The assays can be performed using any accurate analytical method; for example, inductively coupled plasma (ICP) is a suitable technique.

[0032] According to a further embodiment, the total amount of combustible compounds in the batch is assessed.

[0033] Combustible compounds in e-scrap may include plastics and certain metals, especially metallic aluminum. These compounds contribute significantly to the enthalpy when pyrometallurgical refining processes are applied. Therefore, it is useful to estimate the approximate carbon and aluminum content early, i.e., before the refining process actually begins.

[0034] This evaluation can be performed by the same scheme as that applied for precious metal evaluation. The model is then further calibrated against batch level assays for carbon and aluminum.

[0035] A further embodiment relates to an apparatus for assessing the precious metal content of a batch of printed circuit boards, configured to carry out the method described above.

[0036] A device "configured to execute" means that at least a portion of the software necessary to execute the model is present on or made available to the device.

[0037] In more detail, a conveyor belt suitable for exposing the printed circuit board to an imaging system; an imaging system including at least one camera viewing the conveyor belt; one or more computing units coupled to the imaging system; (a) processing an image from an imaging system to detect a printed circuit board; (b) extracting a board-related feature vector for each detected printed circuit board; (c) executing a model based on the vector as input and calculating the precious metal content of the batch as output, wherein the model is calibrated against batch-level precious metal assays.

[0038] The computing units may comprise one or more general-purpose computers that contain or have access to specialized hardware and / or software needed to perform tasks. Some tasks may be performed by resources in the cloud.

[0039] A further embodiment relates to a computer program product comprising instructions that, when executed by a computer, cause the computer to implement the model described above.

[0040] A further embodiment relates to a data carrier comprising instructions which, when executed by a computer, cause the computer to execute the model described above.

[0041] The following examples are provided to further illustrate the present invention. Figures 1 to 4 show the following embodiments. [Brief explanation of the drawings]

[0042] [Figure 1]1 shows the imaging and detection of PCBs on a conveyor belt, where reference number 1 corresponds to the back side of a computer motherboard, reference number 2 corresponds to the front side of a sound card, and reference number 3 corresponds to the front side of a RAM memory card. [Figure 2] Extraction of board-related features from the imaged front side of a motherboard is shown. Reference number 1 is the CPU, reference number 2 is the chipset IC, reference number 3 is the parallel port, and reference number 4 is the PCI slot. [Figure 3] We show the mean squared error (MSE) of Au for 10 different training scenarios expressed in ppm on the y-axis against the number of training batches on the x-axis. Three linear models are fitted using general least squares (x), Lasso (+), and Ridge penalty (o). [Figure 4] Prediction results (x) and training results (o) of Lasso regression for Au are shown. The y-axis is the predicted Au concentration (ppm), and the x-axis is the true / simulated Au concentration. One marker represents one batch. The solid line represents the ideal response where the predicted value is equal to the true value.

[0043] To illustrate the invention, a realistic simulation of various printed circuit boards is used, accurately representing those typically encountered in the e-waste recycling industry. A database of components and the boards on which these components are mounted is created. Simulated batches of boards are then generated, each with an associated total precious metal content. Finally, an optical recognition system is simulated, including typical errors that would be considered in a practical method.

[0044] A regression model is fitted or trained to correlate the recognized features with the precious metal content of each batch.

[0045] Component Database Twenty-five components are defined. Each component type is characterized by a randomly selected mean value between 5 and 50 g for component weight and between 5 and 1000 mg for precious metal content for each of Ag, Pd, and Au. A normal distribution for weight and precious metal content is applied to each generated component. The standard deviations are 25% for weight, 30% for Ag, 35% for Pd, and 25% for Au. Thus, each individual component is unique in its type, weight, and composition. An offset with a normal distribution and a relative standard deviation of 10% is applied to each batch for the component metal concentration.

[0046] PCB Database Ten substrates are defined. Each substrate type is characterized by a substrate weight distribution and the maximum allowed number of components of each type. For each generated substrate, the number of components of each type is randomly selected between 0 and the maximum allowed value. The substrate weight per type is characterized by a normal distribution with a randomly selected mean weight between 100g and 1000g. The relative standard deviation is 25%. The mean weight for all types in the batch is biased by multiplying it by a random number that deviates from a normal distribution with a mean of 100% and a standard deviation of 10%. Therefore, each individual substrate is unique in its type, number, and selection of components, composition, and weight. The component and substrate database closely replicates industrial reality.

[0047] A simulated batch of boards 100 batches of substrates are generated, each consisting of a randomly selected subset of an average of 5 substrate types out of the 10 available, with each selected type having a random number of substrates between 0 and 1000. Thus, on average, each batch contains 2500 unique substrates.

[0048] Optical Recognition of Substrates and Components In the simulation, the board type is correctly detected and typed 90% of the time, and also recognizes whether the board is presented upside down. A 5% chance of incorrect typing and a 5% chance of not being detected at all are built into the simulation. There is a 50% chance that the board will be presented upside down, in which case the board type is detectable but the components are not.

[0049] A component has a 95% chance of being detected and typed when presented with a correctly oriented board, regardless of whether the board is correctly typed or not. There is a 2.5% chance of not detecting a component and a 2.5% chance of incorrectly typed a component.

[0050] The above parameters are derived from our experience with camera-based feature detection using Mask R-CNN ("Mask R-CNN," K. He, G. Gkioxari, P. Dollar, and R. Girshick, 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980–2988) on substrates moving on a conveyor belt. Mask R-CNN was trained for substrate and component annotation. Python was used as the programming language, and TensorFlow, Keras, and imgaug were used to create and train the neural network. Pre- and post-processing of images and results was performed using a combination of the following packages: sci-kit image, PILLOW, sci-kit learn, scipy, and numpy. Pre-processing steps consisted of image resizing and color channel normalization. Post-processing steps consisted of image stitching and "non-maximum suppression," which counts substrates and components only once.

[0051] Figures 1 and 2 provide typical examples of photos recorded by the camera. The boards sometimes overlap and appear upside down half the time. The machine learning model Mask R-CNN was used to detect the type, orientation, and masks of the boards and components. The masks of the boards and components are used to link them to each other. Metadata such as broken or partially covered boards, board color, and area can be added.

[0052] Regression Model The fitting / training method is performed on simulated batches. Feature vectors are derived for each substrate and each component. The feature vectors are aggregated across the entire batch, and this aggregate vector is used in a regression model aimed at predicting the precious metal content of that batch. Linear regression models are fitted using least squares, Lasso, and Ridge penalization. Lasso regression performed slightly better than Ridge regression and outperformed least squares as the number of training batches increased. Least squares is sensitive to cross-correlation present within the dataset.

[0053] result Using this scheme, Figure 3 shows that meaningful predictions of precious metal content can be achieved after training on approximately 30 batches. This is extremely fast, considering the many different substrates and components involved. This efficiency is likely due to the cross-correlations that exist between substrate types and the components mounted on them. Such cross-correlations are implied by the statistics used when creating the synthetic batches. These same cross-correlations are likely to exist in industrial reality.

[0054] The accuracy of the precious metal prediction is shown in Figure 4. The results are more than sufficient for the purpose of providing early feedback on material composition to batch suppliers, so that, for example, payments can be expedited or further processing of the material can be planned without having to wait for accurate chemical assay results.

[0055] When such a system is put into practical use, the training database can be easily completed with each batch of actual data, thus strengthening the correlation between the feature vectors and the chemical assays. This results in a self-learning system without any additional cost or effort beyond periodically running the regression model fitting or training algorithm.

[0056] This achievement is particularly significant as major e-waste recyclers can handle several batches per day.

Claims

1. 1. A method for assessing the precious metal content of a batch of printed circuit boards, comprising: imaging at least a statistically representative number of said printed circuit boards in said batch; processing the image to detect the printed circuit board; extracting, for each detected printed circuit board, a board-related feature vector using image processing techniques; providing a model that takes as input at least the substrate-related feature vector and calculates as output the precious metal content of the batch, the model being calibrated against batch-level precious metal assays; A method comprising:

2. The method of claim 1 , wherein the batch precious metal content is calculated based on a combination of board-related feature vectors and batch-level features derived from sources other than printed circuit board imaging.

3. Calculating the precious metal content of the batch is performed using options (1) and (2): (1) summing the substrate-related feature vectors, thereby obtaining a batch-related feature vector; and calculating the precious metal content of the batch based on the batch-related feature vector; or (2) for each printed circuit board, assessing a printed circuit board precious metal content based on the board-related feature vector, and summing the printed circuit board precious metal contents to obtain a precious metal content of the batch; The method according to claim 1 or 2, wherein the method is carried out by any one of the following methods:

4. 4. The method of any one of claims 1 to 3, wherein the step of assessing the precious metal content is carried out before metallurgically processing the batch for recovery of the precious metals.

5. The method of claim 3, wherein the printed circuit boards are classified into at least two classes based on the evaluated printed circuit board precious metal content.

6. 6. The method of claim 5, wherein each printed circuit board class is individually metallurgically processed for recovery of said precious metals.

7. 7. The method of any one of claims 1 to 6, wherein the batch level precious metal assay is obtained by smelting a statistically representative sample of the batch to be analyzed using an alloy as a collector.

8. The method according to any one of claims 1 to 7, wherein the total amount of combustible compounds of the batch is assessed.

9. An apparatus for assessing the precious metal content of a batch of printed circuit boards, configured to carry out the method according to claims 1 to 8.

10. a conveyor belt suitable for exposing the printed circuit board to an imaging system; an imaging system including at least one camera viewing the conveyor belt; one or more computing units coupled to the imaging system; (a) processing an image from the imaging system to detect the printed circuit board; (b) extracting a board-related feature vector for each detected printed circuit board; (c) running a model based on the vector as input and calculating the precious metal content of the batch as output, the model being calibrated against batch-level precious metal assays; one or more computing units configured to perform the tasks of The apparatus of claim 9, comprising:

11. A computer program product comprising instructions which, when executed by a computer, cause the computer to implement a model according to the method of any one of claims 1 to 3.

12. A data carrier comprising instructions which, when executed by a computer, cause the computer to implement the model according to the method of any one of claims 1 to 3.

Images