Estimating chemical process output from single-object feedstock hyperspectral images

Abstract

The aim is to improve the quality of the processing of objects. An intermediate dataset may be generated based on images of a set of objects, each of which includes plastic. The intermediate dataset may be input into a machine learning model to generate predicted chemometric properties, which may be predicted chemometric properties of a physical output pyrolysis oil produced by performing a pyrolysis process on the set of objects using a pyrolysis reactor. Results associated with the set of objects may be generated, the results being based on or including the predicted chemometric properties.

Description

[Technical Field]

[0001] (CROSS-REFERENCE TO RELATED APPLICATIONS) This application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 427,535, entitled "Estimation of Chemical Process Outputs By Single Object Feedstock Hyperspectral Imaging," filed November 23, 2022, which is incorporated herein by reference in its entirety.

[0002] This specification generally relates to facilitating high-quality downstream processing (e.g., pyrolysis processing) of a set of objects (e.g., automatically sorted objects) by analyzing one or more hyperspectral images using one or more machine learning models to predict chemical process outputs (e.g., chemical kinetic properties of pyrolysis oil produced from the set of objects). This specification further relates to identifying specific features of the images of one or more objects that are predictive of one or more quality metrics of the downstream chemical process. [Background technology]

[0003] Currently, 5 billion metric tons of plastic waste sits in landfills worldwide, and by 2040, an additional 7.7 billion metric tons of plastic will be improperly disposed of (landfilled, incinerated, or leaked) and not recycled. These waste streams represent $5 trillion in commercial value if there were an economical way to reinject the materials into the economy, with the added benefit of reducing gigatonnes of CO2 emissions and creating an impactful and economically viable circular economy.

[0004] In a recycling workflow, a set of objects can be fed into one or more mechanical or chemical processing lines to generate physical output. Mechanical recycling is the dominant strategy for recycling plastics and involves crushing, melting, and re-extrusion of plastic waste. Recycling facilities are often designed to process highly purified and sorted material streams to preserve high levels of material performance in recycled products. However, impurities in raw materials, including complex formulations with additives, as well as physical degradation of materials, reduce the effectiveness of recycling even after several cycles of mechanical recycling. For example, in plastic materials, polylactic acid (PLA) is a common waste plastic that often goes undetected in polyethylene terephthalate (PET) sorting and mechanical recycling operations. As another example, chlorinated compounds such as polyvinyl chloride (PVC) are not tolerated in both mechanical and chemical recycling operations because corrosive compounds are produced during the recycling process, limiting the value of the hydrocarbon output.

[0005] Mechanical recycling is limited in its applicability to mixed, complex, and contaminated waste streams, in part due to its sensitivity to chemical contaminants and the use of mechanical separation and reforming processes that may not be able to alter the chemical structure of the waste materials.

[0006] Chemical recycling can address the limitations of mechanical recycling by breaking down the chemical bonds of waste materials into smaller molecules. For example, in the case of polymeric materials, chemical recycling can provide a way to recover oligomers or even basic molecules (such as monomers) from plastic waste feedstocks. In the case of polymers, chemical recycling processes can include operations to depolymerize and dissociate the chemical makeup of complex plastic products so that by-products can be upcycled into feedstocks for new materials. Chemical recycling elements can allow materials to be repeatedly dissociated into primary raw materials. In this way, rather than being limited to a limited number of plastic types by their chemical structure and material integrity, as is the case with mechanical recycling, chemical recycling can be integrated into an "end-to-end" platform to facilitate the reuse of molecular components of recyclable materials. For example, chemical recycling products can include basic monomers (e.g., ethylene, acrylic acid, butyric acid, vinyl), feed gases (e.g., carbon monoxide, methane, ethane), or elemental materials (e.g., sulfur, carbon). Instead of being limited to a single group of recycled products, products can be identified based on the molecular structure of the input waste material that can be synthesized from intermediate chemicals that can be generated from the waste through chemical reactions. In doing so, the end-to-end platform may manage waste streams by generating chemical reaction schemes that convert waste materials into one or more target products. For example, the end-to-end platform may route different portions of waste feedstock to different chemical recycling facilities for chemical conversion into different target products.

[0007] However, the efficiency of recycling (e.g., chemical recycling) and the reliability of the recycled output further depend on the degree to which the objects received for processing correspond to one or more target materials, have limited contamination, etc. Furthermore, the degree to which the recycled output possesses the target properties depends on the extent to which the workflow can guarantee that the recycled material meets certain quality constraints.

[0008] However, individual objects initially accepted into recycling lines are often not composed purely of one material. For example, a plastic container may contain multiple types of plastic, labels (including paper), and food contaminants. Routing such objects to downstream processes (e.g., downstream pyrolysis) may result in suboptimal results. For example, pyrolysis of contaminated objects may produce a relatively low yield of target liquid, while producing a relatively high amount of char.

[0009] It would therefore be advantageous to implement improved techniques that would improve the output from downstream processing lines even when the initial raw object is diverse, may be composed of multiple materials, and may be contaminated (e.g., food, dirt, etc.). Summary of the Invention

[0010] In some cases, a computer-implemented method is provided that includes generating an intermediate data set based on images or line scans of a set of objects, each comprising plastic; generating predicted chemometric properties of a physical output pyrolysis oil produced by processing the set of objects using a pyrolysis reactor, the predicted chemometric properties being generated by inputting the intermediate data set into a machine learning model; and generating results associated with the set of objects, the results being based on or including the predicted chemometric properties.

[0011] Generating the intermediate data set may include generating a hypercube based on a set of line scans of the set of objects, a first dimension of the hypercube may correspond to a first spatial dimension in real-world space, a second dimension of the hypercube may correspond to a second spatial dimension in real-world space, and a third dimension of the hypercube may correspond to a frequency dimension, and values ​​in the hypercube represent at least one of intensity, power, reflectance, transmittance, absorbance, and transreflectance.

[0012] Generating the intermediate data set may include generating, for each material in the set of materials, a predicted relative or absolute amount of the material in the set of objects.

[0013] Generating the intermediate data set may include generating, for each material in the set of materials, a fraction of the weight or mass of the set of objects that is predicted to be attributable to the material.

[0014] The predicted chemometric property of the pyrolysis oil can be the American Petroleum Institute (API) weight, density, or relative density of the pyrolysis oil.

[0015] A predicted chemometric property of the pyrolysis oil may be the vapor pressure of the crude oil produced using the pyrolysis oil.

[0016] A predicted chemometric property of the pyrolysis oil may be the pour point of the pyrolysis oil.

[0017] The predicted chemometric property of the pyrolysis oil may be or may be based on the amount of one or more halogens in the pyrolysis oil.

[0018] The predicted chemometric properties of the pyrolysis oil may be or may be based on the amounts of inorganic and organic contaminants in the pyrolysis oil. The inorganic contaminants may include at least one of sulfur, chlorine, and phosphorus. The organic contaminants may include at least one of sulfur, polyfluorinated substances (PFAS), caprolactam, organic acids, perfluorinated and fluorinated compounds, halogenated organic compounds, and oxygen as measured by neutron activation.

[0019] The method may further include controlling whether the set of objects is routed to a pyrolysis process pipeline based on the results.

[0020] The results may include the selection or identification of one or more other objects for combination with the set of objects before the pyrolysis process is performed on the set of objects.

[0021] An image of a set of objects can be generated based on a set of line scans acquired at different wavelengths.

[0022] The wavelength may be selected from the group of wavelength ranges consisting of 1000 to 1700 nm, 2200 to 5000 nm, and 400 to 1000 nm.

[0023] The images of the set of objects may be generated by performing one of a line scan, an area scan, and point mapping.

[0024] Generating the intermediate data set may include performing a hydrocarbon analysis on the set of objects for each material in the set of materials, wherein the hydrocarbon analysis may provide a profile of at least one paraffin, isoparaffin, and aromatics present in the set of objects.

[0025] Generating the intermediate data set can include performing a simulated distillation of the set of objects for each material of the set of materials. The simulated distillation can provide a volumetric distillation profile for the set of objects.

[0026] In some embodiments, a system is provided that includes one or more data processors and a non-transitory computer-readable storage medium that includes instructions that, when executed on the one or more data processors, cause the one or more data processors to perform some or all of one or more of the methods disclosed herein.

[0027] In some embodiments, a computer program product is provided that is tangibly embodied in a non-transitory machine-readable storage medium and includes instructions configured to cause one or more data processors to perform some or all of one or more of the methods disclosed herein.

[0028] The details of one or more implementations of the subject matter described herein are set forth in the accompanying drawings and the description below. Other potential features, aspects, and advantages of the subject matter will become apparent from the description, drawings, and claims. [Brief explanation of the drawings]

[0029] The present disclosure is described in conjunction with the accompanying drawings.

[0030] [Figure 1] FIG. 1 is a block diagram of an example system 100 implemented to perform chemical process output estimation. [Figure 2] FIG. 1 illustrates a process flow for processing image data to train a machine learning model to predict properties of a raw material. [Figure 3] FIG. 10 identifies an example specification of 30 batches used to generate training data for training a model. [Figure 4]FIG. 10 identifies an example specification of 16 batches used to generate training data for training a model. [Figure 5] FIG. 1 shows an example of using three hyperspectral images of an object to predict the pyrolysis output of the batch containing the images. DETAILED DESCRIPTION OF THE INVENTION

[0031] According to one innovative aspect of the subject matter described herein, image data of a raw material can be collected. The image data can include an image or one or more line scans. In one implementation, the image can be generated based on a set of line scans of a set of objects acquired at different wavelengths. The wavelengths can be selected from the group of wavelength ranges consisting of 1000-1700 nm, 2200-5000 nm, and 400-1000 nm. In another implementation, the image of the set of objects can be generated by performing one of a line scan, an area scan, and point mapping. In some cases, the image data includes hyperspectral data and / or a hyperspectral cube (e.g., including a first spatial dimension, a second spatial dimension, and a frequency dimension, and indicating intensity, power, reflectance, transmittance, absorbance, and trans-reflectance for each of a set of points in three-dimensional space).

[0032] At least a portion of the image data can then be fed to a machine learning model capable of generating predicted chemometric properties of the physical output of a downstream processing line. The physical output is the output predicted to be produced when the object represented in at least a portion of the image data is transformed using the downstream processing line. For example, the machine learning model may generate predicted chemometric properties of a pyrolysis oil produced by performing a pyrolysis process using the object. Exemplary chemometric properties of pyrolysis oil include American Petroleum Institute (API) weight, density, relative density, pour point, the amount of halogens in the pyrolysis oil, the amount of inorganic contaminants (e.g., sulfur, chlorine, phosphorus, etc.) in the pyrolysis oil, the amount of organic contaminants (e.g., sulfur, polyfluorinated substances (PFAS), caprolactam, organic acids, perfluorinated and fluorinated compounds, halogenated organic compounds, and oxygen as measured by neutron activation), and the like. An overall quality metric or classifier is derived by a collection of individual chemometric properties or other predictive functions. Another exemplary chemometric property of pyrolysis oil is the vapor pressure of crude oil produced using the pyrolysis oil.

[0033] At least a portion of the image data provided to the machine learning model may represent a single object or multiple objects. For example, in some cases, the image data is provided to a segmentation model that predicts which portions of the image data correspond to different types of materials (e.g., PET bottles, paper labels, PP bottle caps). At least a portion of the image data may be defined to include (for example) one or more portions of the image data that correspond to a single object or a particular set of objects.

[0034] In some cases, other processing is performed to identify a single object or a set of specific objects represented in at least a portion of the image data. For example, other processing may include predicting a compositional attribute of each of one or more objects (e.g., the absolute or relative amount of an object of a given material, and / or determining that the absolute or relative amount exceeds a predefined threshold). Other processing may include predicting the weight or mass of a specific type of material (e.g., a specific type of plastic) or a specific material (e.g., a specific plastic) within an object, and / or predicting the total weight or mass of the object. Thus, the absolute or relative weight or mass of a specific type of material can be predicted.

[0035] In some cases, other processing is performed to determine a profile of at least one of paraffins, isoparaffins, and aromatics present in the set of objects. The profile can be determined by performing a hydrocarbon analysis (DHA) on the set of objects. The purpose of the detailed DHA is to determine the composition of bulk hydrocarbon group classes such as PONA: paraffins, olefins, naphthenes, and aromatics.

[0036] In some cases, other processing is performed to determine a volumetric distillation profile of the set of objects. The volumetric distillation profile can be determined by performing a simulated distillation of the set of objects. Simulated distillation can be a gas chromatography method and can develop into an essential tool in the petroleum industry to determine the distillation behavior of various petroleum products and ensure fuel quality. In contrast to standard physical distillation, simulated distillation can exhibit various advantages, including comparatively small sample sizes and the ability to be automated, which are of great importance in research and the development of new fuels.

[0037] Other processing may include predicting or determining whether to include an object in a particular process flow (e.g., based on a predicted weight or mass of a particular type of material), whether to complete a particular process flow (at least incorporating the object), how to adjust parameters to determine which other objects to include in a particular process flow, how to define process flow parameters, etc.

[0038] This determination can be made based on other objects currently or tentatively assigned to the same lot or raw material in a processing line associated with the particular material (e.g., to extract the particular material). The determination can include comparing a predicted absolute or relative weight or mass of a particular type of material to a threshold value. The threshold value can be specific to a given processing line and can be determined based on (for example) input from a client system, one or more recent outputs from the processing line, and / or compositional data of one or more other objects already assigned or routed to the processing line. For example, the threshold value can be determined by identifying multiple other objects flagged to proceed to the next raw material or next lot in the processing stream and determining a threshold value for the objects based on predefined criteria for the processing stream, an estimated total weight or mass of the one or more other objects, and / or an estimated weight or mass of a particular type of material or specific material within the one or more other objects. In some cases, the threshold value further varies depending on the estimated total weight or mass of the objects.

[0039] To illustrate, at a given time and on a given processing line, 10 objects may already be routed, assigned, or tentatively assigned to a given processing line. These 10 objects may have a cumulative predicted or actual mass of m1 and a cumulative predicted target mass of t1. The processing line may be configured to receive lots of approximately m2 and may require a target mass of at least a target mass threshold t2 (the target mass corresponding to a target absolute mass or weighted mass of one or more specific materials or material types and / or the total mass corresponding to a total target absolute mass or total weighted mass of one or more specific materials or material types). A given object may have a predicted mass of m2 - m1 (e.g., corresponding to the total mass of the object, the total mass of a given type of material in the object, or the mass of a specific material in the object) and a predicted target mass of t3. If the sum of the object's cumulative predicted target mass t1 and the predicted mass meets or exceeds the target mass threshold t2, it may be determined to add the given object to the lot. In other cases, a given object may be routed to a different processing line or waste bin. In some cases, for each object, the ratio of predicted target mass to predicted total mass is evaluated and compared to a threshold. However, as different objects are added to the lot, the threshold may be changed. For example, if the ratio for the first group of objects added to the lot is very high, a lower ratio may be acceptable for the second group of objects added to the lot so that the entire lot can still meet the target overall ratio.

[0040] 1 is a block diagram of an exemplary system 100 implemented to perform chemical process output estimation. System 100 includes a camera system 110 for taking images (e.g., hyperspectral images) or line scans (e.g., hyperspectral line scans) of objects. Each image may be of part or all of one or more objects being moved by a conveyor belt 112. The objects may include objects initially collected from multiple individual recycling bins.

[0041] It should be understood that the images or line scans collected and analyzed by camera system 110 may include non-hyperspectral images, and that the disclosure herein referring to hyperspectral images may be adapted to use non-hyperspectral images. For example, camera system 110 may include a lens with low absorption in the visible, near-infrared, short-wave, or mid-wave infrared ranges. Thus, the captured images may exhibit signals in the visible, near-infrared, short-wave, or mid-wave infrared ranges, respectively. Camera system 110 may include any of a variety of illumination sources, such as light-emitting diodes (which may be synchronized to camera sensor exposure), incandescent light sources, lasers, and / or blackbody illumination sources. The light-emitting diodes included in camera system 110 may have a peak emission that matches the peak resonance of a given target chemical (e.g., a given type of plastic) and / or a bandwidth that matches multiple molecular absorptions. In some cases, multiple LEDs are included in camera system 110 to cover a given spectral range. LED light sources provide coherent light, and control of emission and signal-to-noise may be more feasible than with other light sources. Additionally, they are generally more reliable and consume less power than other light sources.

[0042] In some cases, the camera system 110 is configured so that the optical axis of the camera's lens or image sensor is at an angle between 75 and 105 degrees, 80 and 90 degrees, 85 and 95 degrees, 87.5 and 92.5 degrees, 30 and 60 degrees, 35 and 55 degrees, 40 and 50 degrees, 42.5 and 47.5 degrees, or less than 15 degrees relative to the surface supporting the object being imaged (e.g., a conveyor belt). In some cases, the optical system includes multiple cameras, and the angle between the optical axis of a first camera relative to the surface supporting the object is different from the angle between the optical axis of a second camera relative to the surface. The difference may be (for example) at least 5 degrees, at least 10 degrees, at least 15 degrees, at least 20 degrees, at least 30 degrees, less than 30 degrees, less than 20 degrees, less than 15 degrees, and / or less than 10 degrees. The difference may facilitate detecting signals from objects having different shapes or positioned at different angles (e.g., having different inclinations) relative to the underlying surface. In some cases, the first camera filters a different type of light relative to the second camera. For example, the first camera may be an infrared camera and the second camera may be a visible light camera. In some cases, the camera system 110 includes a light source that is in a specular condition relative to the camera and a second light source that is in a diffuse condition relative to the camera (e.g., to facilitate detection of objects with different specular and diffuse reflectances).

[0043] The camera system 110 may include a light guide (e.g., fiber optic, hollow, solid, or liquid-filled type) for transmitting light from the illumination source to the imaging location, which may reduce heat emitted at the imaging location. The camera system 110 may include a type of light source or optics such that light from the light source is focused into a line or focused to match the projected size of the entrance slit into the spectrometer. The camera system 110 may be configured so that the illumination source and imaging device (camera) are positioned under specular conditions (to produce a bright-field image), non-specular (or diffuse) conditions (to produce a dark-field image), or a mixture of conditions. Most hyperspectral images have image data for each of several or even dozens of wavelength bands, depending on the imaging technique. In many applications, it is desirable to reduce the number of bands in a hyperspectral image to a manageable amount because processing images with a large number of bands is computationally expensive (resulting in delays in obtaining results and high power usage), high-dimensional spaces can prove infeasible to search or have inappropriate distance metrics (the "curse of dimensionality"), or bands are highly correlated (the "correlated regressor problem"). Many different dimensionality reduction techniques have been presented in the past, such as principal component analysis (PCA) and pooling. However, these techniques often remain computationally expensive, require specialized training, and do not always provide the desired accuracy in applications such as image segmentation. Furthermore, many techniques attempt to use most or all bands for segmentation decisions, even though different wavelength bands often have dramatically different information values ​​for segmenting different types of boundaries (e.g., boundaries of different types of regions with different properties such as material, composition, structure, texture, etc.). This has traditionally led to the inefficiency of processing image data for more wavelength bands than are required for the segmentation analysis. Also, accuracy is limited because data in bands with low relevance to the segmentation boundary obscures key signals in data with noise and weakly relevant data.In some embodiments, the techniques disclosed herein can generate trained embeddings using a machine learning model (e.g., an autoencoder).

[0044] Thus, in some cases, band selection techniques may be implemented, which may include (for example) one or more of the actions disclosed in U.S. Patent Application Publication No. 17 / 811,766, filed July 11, 2022, and incorporated by reference in its entirety for all purposes.

[0045] To illustrate, in some implementations, synthetic or modified bands are generated. The synthetic bands may be generated by processing the image data for one or more of the bands selected in the first iteration. For example, each band in the subset of bands may be subjected to one or more operations (e.g., image processing operations, mathematical operations, etc.), which may include one or more operations that combine data from two or more different bands (e.g., of the bands selected in the first iteration). Various predetermined functions may each be applied to the image data for a different combination of the selected bands (e.g., for each pair of bands or each permutation in the selected subset of bands). This may create a new set of synthetic bands, each representing a different modification to the bands selected in the first iteration or combination of bands. The synthetic bands may also (for example) additionally or alternatively be derived from a convolution or projection applied to the image data. A convolution or projection is a function that maps a group of pixels and a single pixel to a single number, respectively. A single convolution, multiple convolutions, a single projection, and / or multiple projections may be applied across the entire image to create new synthetic bands.

[0046] One or more original bands and / or one or more composite bands can be evaluated and / or scored to determine the level at which they are predicted to provide information about a target of interest. For example, a band may be evaluated and / or scored based on how predictive the intensity within the band is of whether and / or to what extent a particular type of material is included in a corresponding depicted object or group of corresponding objects. Selected bands can then be used to analyze other images (e.g., to predict compositional characteristics).

[0047] In some cases, an image has two dimensions representing spatial dimensions (e.g., corresponding to the width and length axes) and another dimension representing different wavelength (or frequency) bands. In some cases, an image is generated based on a set of line scans (e.g., each line scan may generate an output identifying an intensity for each position along one spatial dimension and for each of multiple wavelength bands). For example, a line scan may be generated by scanning across the width dimension of the conveyor belt 112 (e.g., for each of multiple wavelength bands). The conveyor belt may move so that the next line scan scans a different material. Multiple line scans may be combined to generate an image corresponding to two different dimensions (e.g., multiple frequency bands).

[0048] Thus, for each collected image, image data 115 may be generated that identifies intensity, power, reflectance, transmittance, absorbance, and trans-reflectance for each of a plurality of locations and for each of a plurality of wavelength bands. The image data 115 is transmitted via network 120 to computing system 130, which may process the image data 115 (e.g., potentially to perform segmentation, identify sorting instructions for individual objects or groups of objects, predict the output of processing lines if used to process a particular group of objects, etc.). In the example of FIG. 1 , camera system 110 includes or is associated with a computer or other device that can communicate via network 120 with computing system 130 that processes the hyperspectral image data and returns segmented images or other data derived from the segmented images. In other implementations, the functions of computing system 130 (e.g., generating a profile, processing hyperspectral image data, performing segmentation, etc.) may be performed locally at the location of camera system 110. For example, system 100 may be implemented as a stand-alone unit housing camera system 110 and computing system 130 .

[0049] Network 120 may include a local area network (LAN), a wide area network (WAN), the Internet, or a combination thereof. Network 120 may also include any type of wired and / or wireless network, a satellite network, a cable network, a Wi-Fi network, a mobile communication network (e.g., 3G, 4G, etc.), or any combination thereof. Network 120 may utilize communication protocols including packet-based and / or datagram-based protocols such as Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), or other types of protocols. Network 120 may further include several devices that facilitate network communication and / or form the hardware infrastructure for the network, such as switches, routers, gateways, access points, firewalls, base stations, repeaters, or a combination thereof.

[0050] The computing system 130 may be configured to process the image data 115 using one or more techniques and / or one or more models (e.g., one or more machine learning models). The one or more models may include (for example) neural networks, convolutional neural networks, deep neural networks, clustering algorithms, etc.

[0051] In some instances, the computing system 130 may first process the image data using a segmentation machine learning model to predict segmentation data 135. The segmentation data 135 may include predictions regarding which data points in the image data depict or represent points or voxels within each of multiple individual objects (or do not represent any objects). For example, if the image data corresponds to a static two-dimensional physical space (e.g., multiple wavelength bands), the segmentation data may identify various portions within the two-dimensional space in which a given corresponding object is located. Illustratively, the segmentation data 135 may identify a value for each point within a two-dimensional grid, which may identify an identifier of an object predicted to be depicted in the pixel (e.g., generated using incremental or pseudo-random techniques). If a pixel is predicted to have no object depicted in it, a default value (e.g., 0 or a value that cannot be represented numerically) may be assigned. The segmentation and / or one or more other actions may be performed according to one or more disclosures in U.S. Patent Application Publication No. 17 / 811,766, filed July 11, 2022, which is incorporated by reference in its entirety for all purposes. In some cases, the segmentation is performed using a trained segmentation machine learning model and / or one or more segmentation profiles (which may, for example, indicate different hyperspectral data combinations corresponding to different types of materials). The computing system 130 can then store segmentation data 135 that identifies portions of the image data that are predicted to correspond to various distinct objects.

[0052] It will be appreciated that computing system 130 need not perform a segmentation analysis and / or use a segmentation machine learning model. For example, hyperspectral data may be collected across the full range of one or more axes within the image data and then processed (e.g., even if it corresponds to the depiction of multiple objects, portions of objects, multiple at least portions of objects, etc.).

[0053] In some cases, computing system 130 may use a technique (e.g., another technique) or a machine learning model (e.g., another machine learning model) to generate composition prediction data 140 for a given object. Other models may include (for example) neural networks, convolutional neural networks, deep neural networks, regression models, support vector machines, component analysis (e.g., principal component analysis), etc.

[0054] The composition prediction data 140 may associate, for each individually segmented object, or for a collection of imaged full and / or partial objects, a corresponding prediction regarding (for example) the amount of a given material within the object or collection, whether the object or collection contains at least a threshold amount of a given material, whether a condition (e.g., a condition regarding a composition having at least a first threshold amount of one material and less than a second threshold amount of another material) is met, etc.

[0055] The composition prediction data 140 can be generated by transforming some or all of the image data 115 using the technique or machine learning model. For example, a segmentation analysis may predict that a first particular area represents data from a single object. The wavelength data corresponding to the first particular area can then be transformed using the technique or machine learning model to generate a prediction corresponding to the amount of one or more materials within the single object. As another example, wavelength data from the entire image can be fed into the technique or machine learning model without having to perform segmentation.

[0056] The computing system can use the composition prediction data 140 to generate one or more action instructions 145. The action instructions can include routing or facilitating routing (e.g., physical routing) of one or more objects. Routing can include routing objects to or away from a given processing line or storage bin. For example, in the illustration of FIG. 1, routing can direct one or more objects to a clean processing line 150a, a dirty processing line 150b, or a pyromellitic dianhydride (PMDA) additive line 150c.

[0057] Routing an object may involve (for example) moving one or more robotic arms, each of which may move linearly and / or angularly (e.g., along a particular trajectory of one or more predefined particular trajectories). The robotic arms may move to push an object toward a target, lift an object to move it to a target location, or induce a force (e.g., wind or magnetic force that pulls the object toward the target location). Figure 1 illustrates an example in which each of a set of objects represented in image data 115 is routed to one of three processing lines 150a-150c based on a corresponding action instruction 145. In the illustrative example, the three processing lines correspond to a first processing line (e.g., clean processing line 150a) containing objects (or groups of objects) having at least a threshold predicted amount (e.g., absolute or percentage threshold amount) of one or more target materials that feeds belt 155 for facilitating the pyrolysis process; a second processing line (e.g., dirty processing line 150b) that feeds additive-containing storage bin 160 for objects or groups of objects predicted to have at least a threshold amount of a given additive (e.g., at least 0.5% PMDA); and a processing line that feeds dirty storage bin 165 for other objects. While FIG. 1 illustrates the example where the first processing line (e.g., clean processing line 150a) feeds belt 155 for facilitating the pyrolysis process, it will be understood that the belt may feed additional or alternative types of downstream processing. Such downstream processing may be configured to (for example) extract one or more selected materials (e.g., by performing dissolution, depolymerization, and / or conversion techniques).

[0058] It will be appreciated that the output from a pyrolysis (or other downstream) process performed on objects fed by belt 155 to the pyrolysis process will vary depending on the materials within those objects. For example, in a pyrolysis process, objects are exposed to high temperatures in the absence of oxygen such that the solid objects are converted into pyrolysis liquid. Various catalysts can also be used to facilitate the pyrolysis process. The pyrolysis liquid can then be used to generate new objects.

[0059] The usefulness of pyrolysis liquid depends on the ability to reliably generate pyrolysis liquid in a way that results in a viscous liquid with predictable properties. For example, it is advantageous to be able to predict the distribution of various types of molecules in the liquid, its pour point, density, etc. However, given the wide variety of objects (e.g., plastic products) that may be received as initial feedstock, reliably generating pyrolysis liquid with specific properties can be challenging. For example, each plastic object may contain different percentage compositions of one or more of polyvinyl chloride (PVC), polyethylene terephthalate (PET), low-density polyethylene (LDPE), high-density polyethylene (HDPE), polypropylene (PP), and polystyrene (PS). Assume that a first batch of objects contains more LDPE and less HDPE than a second batch. As a result, new objects produced from pyrolysis oil from the first batch will have lower strength and heat resistance than new objects produced from pyrolysis oil from the second batch. Furthermore, the temperature that can be used to convert the first batch of material into a liquid will be lower than the temperature that can be used to convert the second batch of material into a liquid.

[0060] Additionally, various objects may contain non-plastic materials, such as food residue or paper-containing product labels, which may increase the amount of char produced by the pyrolysis process (for example), which may be undesirable.

[0061] Thus, in some embodiments, the hyperspectral data is used to generate one or more action instructions to facilitate the reliable production of viscous pyrolysis liquids. The action instructions can be generated by (for example) using a pyrolysis model to predict one or more characteristics of a pyrolysis process when the pyrolysis process receives a particular set of objects as input (e.g., routed toward a pyrolysis processing line). For example, composition prediction data 140 corresponding to a particular set of objects can be aggregated and fed to a pyrolysis model to predict characteristics of the output of the pyrolysis process. Such predictions may include the amount of oil, gas, and / or char produced or characterize specific properties of the oil. Properties of the oil include (for example) its boiling point, liquid yield, viscosity, density, halogen number, chlorine number, and the contribution of one or more chemical classes (e.g., N-paraffins, N-olefins, isoparaffins, isoolefins, cycloolefins, and / or aromatics). In some cases, feature engineering is performed to identify specific features to feed into the pyrolysis model. Pyrolysis models may include (for example) neural networks, regression models, support vector machines, component analysis (e.g., principal component analysis), decision tree models, etc. Pyrolysis models may include models pre-trained for different use cases, which are then fine-tuned to predict characteristics of the output of the pyrolysis process.

[0062] In some cases, the hyperspectral data includes data generated based on image data 115 collected by camera system 110. For example, by using segmentation data 135 and action instructions 145, computing system 130 can infer which portions of the collected hyperspectral data correspond to objects routed onto belt 155 to the pyrolysis process within an iteration or time interval. As another example, by using segmentation data 135, composition prediction data 140, and action instructions 145, computing system 130 can infer object-by-object or cumulative composition data for objects on belt 155 to the pyrolysis process at a given time. Computing system 130 can then generate one or more action instructions based on the object-by-object or cumulative composition data.

[0063] In some cases, another camera system 170 is positioned to image downstream of camera system 110. For example, the other camera system 170 may be positioned to collect images of a portion of clean processing line 150a. The other camera system 170 may include one or more characteristics as disclosed herein with respect to camera system 110. The image data collected by the other camera system 170 may have one or more characteristics as disclosed herein with respect to image data 115. In some cases, the other image data collected by the other camera system 170 is utilized by computing system 130 or another computing system. Optionally, computing system 130 or another computing system can use the other image data to perform segmentation techniques (e.g., as disclosed herein) and / or generate composition prediction data for one or more objects (e.g., as disclosed herein). Computing system 130 or another computing system can generate one or more action instructions based (for example) on the image data, the segmentation data, and / or the composition prediction data.

[0064] The action instructions 145 (e.g., whether generated based on image data 115 collected by camera system 110 or other image data collected by other camera systems) may include (for example) sorting instructions, routing instructions, instructions for parameters for downstream processing, instructions for criteria to use for subsequent sorting, etc.

[0065] For example, FIG. 1 illustrates an example in which objects from clean processing line 150a are routed to belt 155 to a pyrolysis process or to dirty storage bin 165. The action instructions indicate how to route each of one or more objects (e.g., to belt 155 to a pyrolysis process or to dirty storage bin 165) or a set of objects (e.g., all objects depicted in an image, all objects completely depicted in an image, all objects having at least a threshold number of pixels in an image, etc.). A robotic arm can then be used to selectively lift, slide, push, or otherwise move each object (or each object corresponding to a particular path), or move each set of objects appropriately. For example, in one example, clean processing line 150a includes a belt moving toward belt 155 to a pyrolysis process. Thus, any object that is to be routed to belt 155 to the pyrolysis process is taken there by default, while the robotic arm can move any object that is not to be routed to belt 155 to the pyrolysis process (e.g., an object that is to be routed instead to dirty storage bin 165) to a different path or location.

[0066] Thus, individual objects and / or sets of individual objects may be dynamically routed towards or away from downstream processing lines based on (for example) hyperspectral data, machine learning algorithms that predict compositional attributes of the objects, and / or estimated properties of objects in the current or next batch for downstream processing.

[0067] FIG. 2 illustrates a process flow for training a machine learning model to process image data and predict characteristics of a feedstock. One or more actions depicted in the process flow of FIG. 2 may correspond to one or more actions associated with FIG. 1, although it will be understood that the scope and / or details of such actions may vary. In this example, an input feedstock is received, and one or more sensors collect sensor data (e.g., image data) for at least a portion of the input feedstock (e.g., for individual objects or for a set of objects). For example, one or more cameras may collect one or more hyperspectral images (e.g., corresponding to one or more wavelength bands) of the objects. The input feedstock may include output from an initial sorting (e.g., manual and / or automated sorting). The input feedstock may include objects predicted to have one or more specific types of materials. By way of example, the input feedstock may include any of #3-#7 plastics or plastics including any of #3-#7 (including polyvinyl chloride, low-density polyethylene, polypropylene, polystyrene, BPS, polycarbonate, or LEXAN).

[0068] A computing system (not shown) can use the sensor data to generate composition prediction data, which can include (for example) the identification of one or more materials in the object, the contribution of specific materials or types of materials in the object, etc. The composition prediction data can include object composition data, which includes data predicting (for example) which materials are included in the object, whether a given material is included in the object, the portion of the object's weight or mass that is attributable to specific materials, which types of materials are included in the object, whether a given type of material is included in the object, the portion of the object's weight or mass that is attributable to specific types of materials, etc. The object composition data can also and / or alternatively predict the weight or mass of the object, the weight or mass of one or more specific materials, and / or the weight or mass of one or more specific types of materials. Some or all of the object composition data can be generated by processing one or more images (e.g., one or more hyperspectral images) of the object using a machine learning model.

[0069] Object composition data corresponding to a batch of objects can be used to generate a predicted distribution of materials within the batch. For example, FIG. 2 illustrates the predicted percentage of the predicted mass of a plurality of objects predicted to be PP, PE, or PS. The batch can then be processed through an actual pyrolysis reaction, and the physical output from the reaction can be analyzed. For example, the analysis can include chemometric composition data characterizing the amount of oil, gas, and / or char produced, or characterizing specific properties of the oil. The oil properties can include (for example) its boiling point, liquid yield, viscosity, density, halogen number, chlorine number, and the contribution of one or more types of chemicals (e.g., N-paraffins, N-olefins, isoparaffins, isoolefins, cycloolefins, and / or aromatics).

[0070] The chemometric compositional data (e.g., along with the object composition data and / or the predicted distribution of materials within the batch) can then be used to train or fine-tune a pyrolysis model. Once trained, the pyrolysis model can subsequently be used to convert an input dataset containing predicted amounts of one or more materials within the batch into predictions regarding the amounts or properties of various outputs of the pyrolysis process.

[0071] In some cases, the batches used to generate training data are strategically designed. For example, various batches can be designed to contain specific material distributions that can facilitate generating accurate predictions across the multidimensional space of interest. For example, FIG. 3 identifies exemplary specifications for 30 batches for training, where the batches contain different relative amounts of HDPE, LDPE, PP, PS, and contaminants. Notably, these batches contain no PET or PVC. This may be due to use cases where sorting is performed to divert objects from the processing line when they contain PET or PVC.

[0072] As another example, FIG. 4 identifies exemplary specifications for 16 batches for training. In this case, the batches contain HDPE, LDPE, PP, PVC, and PET in different types of colors and different relative amounts. Notably, these batches do not contain PVC. Some of the batches in this example are “clean” and others are “dirty.” The dirty batches contain different types of contaminants (e.g., cooking oil, motor oil, detergent, etc.). Using dirty batches can facilitate model training for different objects that may be represented by a single point in a multidimensional space defined based on the relative contributions of different types of materials (e.g., different types of plastics). The model may therefore adjust its predictions to account for this variation. Additionally or alternatively, it may be determined that the output of a processing line is substantially affected by one or more types of contaminants, in which case the sorting technique can be adjusted to sort based on the presence or amount of one or more types of contaminants. [Example]

[0073] Figure 5 shows an example where three hyperspectral images of an object were used to predict the pyrolysis output of the batch containing the images. The three images on the left correspond to image data corresponding to one or more different objects. While a single image is shown for each of the three cases, these images are merely representative; the other image data correspond to different wavelength bands (thereby corresponding to hyperspectral cubes).

[0074] The characterization model included a machine learning model that detected and filtered out background pixels and used the remaining hyperspectral data to predict object composition data for the depicted objects. For example, the characterization model predicted that the object depicted in the top image was composed entirely of polypropylene.

[0075] The predicted object composition data was then fed into a pyrolysis model to generate pyrolysis predictions corresponding to the predicted output characteristics of the pyrolysis process if all of the depicted objects were fed into the pyrolysis process in a given batch. The predicted output characteristics included predictions of how much of the reaction output would be pyrolysis liquids, char, and gas. In the illustrated case, the output was calculated to be 90% pyrolysis liquids, 4% char, and 6% gas.

[0076] In this example, the predicted output characteristics also included the pour point and vapor pressure of the pyrolysis liquid.

[0077] Various implementations of the systems and techniques described herein may be realized in digital electronic circuitry, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementation in one or more computer programs executable and / or interpretable on a programmable system including at least one programmable processor, which may be special purpose or general purpose, coupled to receive data and instructions from, and send data and instructions to, a storage system, at least one input device, and at least one output device.

[0078] These computer programs (also known as programs, software, software applications, or code) contain machine instructions for a programmable processor and may be implemented in a high-level procedural and / or object-oriented programming language and / or in assembly / machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., magnetic disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0079] To provide for user interaction, the systems and techniques described herein may be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user, and a keyboard and pointing device (e.g., a mouse or trackball) by which the user can provide input to the computer. Other types of devices may similarly be used to provide for user interaction; for example, feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback), and input from the user may be received in any form, including acoustic, speech, or tactile input.

[0080] The systems and techniques described herein can be implemented in a computing system that includes a back-end component (e.g., as a data server), or includes a middleware component (e.g., an application server), or includes a front-end component (e.g., a client computer having a graphical user interface or web browser through which a user can interact with an implementation of the systems and techniques described herein), or includes any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), a peer-to-peer network (with ad hoc or static members), a grid computing infrastructure, and the Internet.

[0081] A computing system may include clients and servers. Clients and servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0082] Several implementations have been described. Nevertheless, it will be understood that various modifications may be made. For example, various forms of the flows illustrated above may be used with steps reordered, added, or removed. Also, while several applications and methods for providing incentives for media sharing have been described, it should be recognized that numerous other applications are contemplated. Accordingly, other implementations are within the scope of the following claims.

Claims

1. 1. A method comprising: generating an intermediate data set based on images of a set of objects, each of the objects comprising plastic; generating predicted chemometric properties of a physical output pyrolysis oil produced by performing a pyrolysis process on the set of objects using a pyrolysis reactor, wherein the predicted chemometric properties are generated by inputting the intermediate dataset into a machine learning model; generating results associated with the set of objects, the results being based on or including the predicted chemometric properties; A method comprising:

2. 2. The method of claim 1 , wherein generating the intermediate data set includes generating a hypercube based on a set of line scans of the set of objects, wherein a first dimension of the hypercube corresponds to a first spatial dimension in real-world space, a second dimension of the hypercube corresponds to a second spatial dimension in the real-world space, and a third dimension of the hypercube corresponds to a frequency dimension, and wherein values ​​in the hypercube represent at least one of intensity, power, reflectance, transmittance, absorbance, and transreflectance.

3. The method of claim 1 , wherein generating the intermediate data set comprises generating, for each material of a set of materials, a predicted relative or absolute amount of the material in the set of objects.

4. 2. The method of claim 1, wherein generating the intermediate data set comprises generating, for each material in a set of materials, a fraction of a weight or mass of the set of objects that is predicted to be attributable to the material.

5. 2. The method of claim 1, wherein the predicted chemometric property of the pyrolysis oil is the American Petroleum Institute (API) weight, density, or relative density of the pyrolysis oil, and an overall quality metric or classifier is derived by a collection of individual chemometric properties or other predictive functions.

6. The method of claim 1 , wherein the predicted chemometric property of the pyrolysis oil is the vapor pressure of crude oil produced using the pyrolysis oil.

7. 2. The method of claim 1, wherein the predicted chemometric property of the pyrolysis oil is the pour point of the pyrolysis oil.

8. 2. The method of claim 1, wherein the predicted chemometric property of the pyrolysis oil is or is based on the amount of one or more halogens in the pyrolysis oil.

9. 2. The method of claim 1, wherein the predicted chemometric properties of the pyrolysis oil are or are based on the amounts of inorganic and organic contaminants in the pyrolysis oil, the inorganic contaminants including at least one of sulfur, chlorine, and phosphorus, and the organic contaminants including at least one of sulfur, polyfluorinated substances (PFAS), caprolactam, organic acids, perfluorinated and fluorinated compounds, halogenated organic compounds, and oxygen as measured by neutron activation.

10. The method of claim 1 , further comprising controlling whether the set of objects is routed to a pyrolysis process pipeline based on the results.

11. The method of claim 1 , wherein the results include the selection or identification of one or more other objects for combination with the set of objects before the pyrolysis treatment is performed on the set of objects.

12. The method of claim 1 , wherein the images of the set of objects are generated based on a set of line scans acquired at different wavelengths.

13. 13. The method of claim 12, wherein the wavelength is selected from the group of wavelength ranges consisting of 1000 to 1700 nm, 2200 to 5000 nm, and 400 to 1000 nm.

14. The method of claim 1 , wherein the images of the set of objects are generated by performing one of a line scan, an area scan, and point mapping.

15. 2. The method of claim 1 , wherein generating the intermediate data set comprises performing a hydrocarbon analysis on the set of objects for each material in a set of materials, the hydrocarbon analysis providing a profile of at least one paraffin, isoparaffin, and aromatics present in the set of objects.

16. 2. The method of claim 1, wherein generating the intermediate data set comprises performing a simulated distillation of the set of objects for each material in a set of materials, the simulated distillation providing a volumetric distillation profile for the set of objects.

17. 1. A system comprising: one or more computers; one or more computer-readable media storing instructions operable when executed by the one or more computers to cause the system to perform a series of actions; wherein the sequence of actions comprises: generating an intermediate data set based on images of a set of objects, each of the objects comprising plastic; generating predicted chemometric properties of a physical output pyrolysis oil produced by performing a pyrolysis process on the set of objects using a pyrolysis reactor, wherein the predicted chemometric properties are generated by inputting the intermediate dataset into a machine learning model; generating results associated with the set of objects, the results being based on or including the predicted chemometric properties.

18. 18. The system of claim 17, wherein generating the intermediate data set includes generating a hypercube based on a set of line scans of the set of objects, wherein a first dimension of the hypercube corresponds to a first spatial dimension in real-world space, a second dimension of the hypercube corresponds to a second spatial dimension in the real-world space, and a third dimension of the hypercube corresponds to a frequency dimension, and wherein values ​​in the hypercube represent at least one of intensity, power, reflectance, transmittance, absorbance, and transreflectance.

19. 20. The system of claim 17, wherein generating the intermediate data set comprises generating, for each material of a set of materials, a predicted relative or absolute amount of the material in the set of objects.

20. 20. The system of claim 17, wherein generating the intermediate data set includes generating, for each material in a set of materials, a fraction of a weight or mass of the set of objects that is predicted to be attributable to the material.

21. 20. The system of claim 17, wherein the predicted chemometric property of the pyrolysis oil is the American Petroleum Institute (API) weight, density, or relative density of the pyrolysis oil, and an overall quality metric or classifier is derived by a collection of individual chemometric properties or other predictive functions.

22. 20. The system of claim 17, wherein the predicted chemometric property of the pyrolysis oil is the vapor pressure of crude oil produced using the pyrolysis oil.

23. 20. The system of claim 17, wherein the predicted chemometric property of the pyrolysis oil is the pour point of the pyrolysis oil.

24. 20. The system of claim 17, wherein the predicted chemometric property of the pyrolysis oil is or is based on the amount of one or more halogens in the pyrolysis oil.

25. One or more non-transitory computer-readable media storing instructions operable, when executed by one or more computers, to cause a system to perform a sequence of actions, the sequence of actions comprising: generating an intermediate data set based on images of a set of objects, each of the objects comprising plastic; generating predicted chemometric properties of a physical output pyrolysis oil produced by performing a pyrolysis process on the set of objects using a pyrolysis reactor, wherein the predicted chemometric properties are generated by inputting the intermediate dataset into a machine learning model; generating results associated with the set of objects, the results being based on or including the predicted chemometric properties.

Images