Multi-task learning plastic and additive synchronous identification and quantification method

Abstract

The invention discloses a multi-task learning plastic and additive synchronous identification and quantification method in the technical field of solid material component analysis. The method comprises the following steps: preparing a gradient proportion sample containing a plurality of plastic substrates and additives, and scanning in a preset wave band based on a hyperspectral imager to obtain three-dimensional spectral cube data of the gradient proportion sample. According to the method, a multi-task convolutional neural network is constructed, synchronous modeling and collaborative recognition of various attributes are realized, three-dimensional spectrum cube data obtained by hyperspectral imaging is utilized, weak peaks, coupling peaks and stable characteristics in strong nonlinear complex characteristics are extracted from a high-dimensional spectrum based on the three-dimensional spectrum cube data, input data are formed, and the high-dimensional spectrum data are obtained. By introducing a cross-task feature sharing and interaction mechanism into the multi-task convolutional neural network, information complementation can be formed among tasks, then plastic base material type, additive type and additive content prediction results are synchronously output, and the use effect is good.

Description

Technical Field

[0001] This invention relates to the field of solid material composition analysis technology, specifically a multi-task learning method for the simultaneous identification and quantification of plastics and additives. Background Technology

[0002] With the rapid pace of electronic product upgrades, the global generation of Waste Electrical and Electronic Equipment (WEEE) continues to grow. WEEE contains a large amount of reusable plastic and metal resources, but it is also rich in various hazardous additives, especially flame retardants. While the presence of flame retardants can significantly improve the flame-retardant properties of materials, they may migrate and be released during use and recycling, posing potential risks to the ecological environment and human health. Therefore, accurately and efficiently identifying the type of plastic substrate, the type of flame retardant, and its content in WEEE is of great significance for achieving efficient resource recycling and environmental risk control.

[0003] While existing hyperspectral imaging technologies can acquire material spectral information non-destructively, the spectral characteristics of plastics and additives overlap significantly. Traditional single-task models struggle to achieve simultaneous identification and quantification of multiple components, focusing primarily on single-attribute identification and lacking a comprehensive understanding of the collaborative analysis of plastic substrates, additive types, and content. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-task learning method for simultaneous identification and quantification of plastics and additives, in order to solve the problems of traditional single-task models that are difficult to achieve simultaneous identification and quantification of multiple components, focus on single attribute identification, and lack collaborative analysis of plastic substrates, additive types and contents.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multi-task learning method for simultaneous identification and quantification of plastics and additives includes: Gradient-ratio samples containing various plastic substrates and additives were prepared, and three-dimensional spectral cubic data of the gradient-ratio samples were obtained by scanning in a preset band using a hyperspectral imager. The gradient ratio samples are recorded with reference labels including the type of plastic substrate, the type of additive, and their mass fraction; The three-dimensional spectral cube data is preprocessed to obtain a dataset, which is then divided into a training set and a test set. A multi-task convolutional neural network based on one-dimensional convolution and fusion attention mechanism and cross-task feature sharing is constructed. The multi-task convolutional neural network is used to simultaneously output the prediction results of plastic substrate type, additive type and additive content. The training and test sets are input into a multi-task convolutional neural network for training and optimization; Test samples are prepared, and test data is obtained by scanning with a hyperspectral imager. The test data is then processed and input into an optimized multi-task convolutional neural network, which simultaneously outputs prediction results of plastic substrate type, additive type, and additive content.

[0006] As a further aspect of the present invention: the preparation of a gradient ratio sample comprising multiple plastic substrates and additives, and obtaining three-dimensional spectral cubic data of the gradient ratio sample by scanning in a preset band using a hyperspectral imager, includes: Select various types of plastic substrates and various types of additives, mix the additives with the plastic substrates in a preset mass fraction gradient ratio, and prepare gradient ratio samples containing plastic substrates and additives with known mass fractions, wherein the number of each type of sample in the gradient ratio samples is not less than three. Each sample in the gradient ratio sample is scanned in a preset band using a hyperspectral imager to obtain three-dimensional spectral cube data containing spatial data and spectral images.

[0007] As a further aspect of the present invention: the plastic substrate type includes at least one of acrylonitrile-butadiene-styrene copolymer, high-impact polystyrene, and polypropylene; the additives include at least one of flame retardant, plasticizer, filler, and antioxidant; wherein the flame retardant includes decabromodiphenyl ethane, decabromodiphenyl ether, aluminum trihydrate, and ammonium polyphosphate.

[0008] As a further aspect of the present invention: the spectral range of the preset band is 935nm to 1720nm, the field of view of the hyperspectral imager is 30° to 45°, the distance between the lens of the hyperspectral imager and the sample is 200mm to 600mm, and the frame rate of the hyperspectral imager is 10Hz to 50Hz.

[0009] As a further aspect of the present invention: the reference label recording the gradient ratio sample includes the plastic substrate type, additive type, and their mass fraction, comprising: The baseline label is obtained by recording the type of plastic substrate, the type of flame retardant, and its known mass fraction for each sample in the gradient ratio sample.

[0010] As a further aspect of the present invention: the three-dimensional spectral cube data is preprocessed to obtain a dataset, and the dataset is divided into a training set and a test set, including: The spectral images of each identical sample in the gradient ratio sample are compared in parallel, and the spatial data and spectral images corresponding to abnormal samples are removed to obtain the filtered three-dimensional spectral cube data. The filtered three-dimensional spectral cube data is subjected to smoothing filtering, first derivative preprocessing and feature filtering to obtain a candidate set containing spectral data. A preset number of spectral data of each type of sample is randomly extracted from the candidate set to form a dataset. The dataset is then divided into a training set and a test set.

[0011] As a further aspect of the present invention: the division of the dataset into a training set and a test set includes: Dataset partitioning includes at least one of the following: random sampling partitioning, cross-validation, and stratified sampling. The spectral data of each type of sample in the dataset are randomly divided into a training set and a test set, with the training set accounting for 70% to 75% of the dataset and the test set accounting for 20% to 25% of the dataset.

[0012] As a further aspect of the present invention: the construction of a multi-task convolutional neural network based on one-dimensional convolution with fused attention mechanism and cross-task feature sharing includes: The training and test sets of the dataset are reconstructed into input data according to the wavelength sequence, so as to serve as a unified input format for multi-task convolutional neural networks; The multi-task convolutional neural network includes a shared convolutional module, a convolutional block attention module, a Cross-Stitch feature sharing unit, and a task-specific convolutional module. The shared convolutional module connects the convolutional block attention module and the Cross-Stitch feature sharing unit, and the Cross-Stitch feature sharing unit connects to the task-specific convolutional module. The shared convolutional module is used to extract common spectral features across multiple tasks. The convolutional block attention module is used to enhance the response of key spectral bands. The Cross-Stitch feature sharing unit is used to achieve adaptive feature fusion between tasks and set an extension interface. The task-specific convolutional module is used to output substrate classification, additive classification, and additive content. In the test set, when validating the accuracy of the multi-task convolutional neural network model, the accuracy of the plastic substrate and additive classification tasks, calculated based on the true values ​​in the baseline labels, is no less than 0.99 and 0.94, respectively. The R-squared value for additive content regression is also no less than [missing value]. 2 Not less than 0.90.

[0013] As a further aspect of the present invention: the preparation of the test sample involves obtaining test data based on a hyperspectral imager scan, processing the test data, inputting it into an optimized multi-task convolutional neural network, and simultaneously outputting prediction results for the plastic substrate type, additive type, and additive content, including: Additional samples not included in the training and test sets, based on additive classifications and mass fraction ratios, are prepared to form test samples. Test labels for plastic substrate type, additive type, and their mass fraction are recorded. Test data is obtained by scanning with a hyperspectral imager. The test data is then processed by smoothing filtering, first-order derivative preprocessing, and feature selection to obtain test input data. The test input data is input into an optimized multi-task convolutional neural network, which simultaneously outputs prediction results for plastic substrate type, additive type, and additive content.

[0014] As a further aspect of the present invention: when verifying the accuracy of the multi-task convolutional neural network model using the test data, the accuracy of the plastic substrate and additive classification tasks is calculated to be no less than 0.98 and 0.93 respectively, based on the true values ​​in the test labels, and the R-squared value of the additive content regression is... 2 Not less than 0.86.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In this invention, a multi-task convolutional neural network is constructed to achieve synchronous modeling and collaborative recognition of multiple attributes, avoiding the cumbersome process of building independent models for different recognition tasks. This enables rapid and accurate identification of plastic substrates and their additive types in electronic and electrical equipment, as well as precise prediction of additive content. It utilizes three-dimensional spectral cube data obtained from hyperspectral imaging. Based on this data, stable characteristics of weak peaks, coupling peaks, and strong nonlinear complex features are extracted from the high-dimensional spectrum and used as input data. By introducing a cross-task feature sharing and interaction mechanism within the multi-task convolutional neural network, information complementarity can be achieved between tasks, thereby synchronously outputting prediction results for plastic substrate type, additive type, and additive content. The results are effective.

[0016] 2. In this invention, the simultaneous identification of plastic substrate, additive type and content is achieved by using hyperspectral imaging and multi-task convolutional neural network, without the need for additional chemical reagents and complex pretreatment, and the detection process is simple.

[0017] 3. In this invention, the constructed multi-task learning framework adopts a modular task branch structure and sets up standardized extension interfaces. New task modules can be added according to usage requirements without changing the overall architecture. For example, it can be used for predicting the degree of plastic aging, assessing recyclability value, or other tasks related to the identification and quantification of plastic properties. Furthermore, although this method is based on near-infrared hyperspectral data, the attention mechanism and cross-task feature fusion structure employed have modality-independent adaptive characteristics, thus enabling unified representation of data from different spectral modalities. This method is applicable to various two-dimensional spectral data modeling scenarios, such as mid-infrared hyperspectral, Fourier transform infrared spectroscopy, and Raman spectroscopy, thereby achieving stable transfer and application expansion under cross-modal conditions.

[0018] 4. In this invention, the spectral acquisition process does not damage the sample, avoiding the destructive operations of traditional detection methods, thus achieving better practicality and safety, and realizing non-destructive testing. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow structure of the present invention; Figure 2 This invention provides a confusion matrix for identifying the type of plastic substrate in the external test set using a multi-task model. Figure 3 This invention provides a confusion matrix for identifying flame retardant types in the external test set using a multi-task model. Figure 4 This invention compares the actual and predicted values ​​of flame retardant content in an external test set using the multi-task model. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Please see Figure 1 In this embodiment of the invention, a multi-task learning method for simultaneous identification and quantification of plastics and additives includes: S1: Prepare gradient ratio samples containing various plastic substrates and additives. Obtain three-dimensional spectral cubic data of the gradient ratio samples by scanning in a preset band using a hyperspectral imager. The plastic samples include plastics from waste electronic and electrical equipment, as well as solid or semi-solid materials with spectral response in the near-infrared or mid-infrared bands, but not limited to. The additives include, but are not limited to, flame retardants, plasticizers, fillers, and antioxidants. S2: Record the baseline label of the gradient ratio sample, including the type of plastic substrate, the type of additive, and its mass fraction; S3: Preprocess the three-dimensional spectral cube data to obtain a dataset, and divide the dataset into a training set and a test set; S4: Construct a multi-task convolutional neural network based on a one-dimensional convolutional attention mechanism and cross-task feature sharing. The multi-task convolutional neural network is used to synchronously output the prediction results of plastic substrate type, additive type and additive content. The multi-task convolutional neural network can achieve a dynamic balance between "sharing and specificity", highlight key spectral information and alleviate task conflicts, thereby significantly improving the overall recognition accuracy and robustness while ensuring multi-task collaboration. In addition, the multi-task learning framework adopts a modular task branch structure, which can turn off the prediction function when not needed, and add corresponding task heads for newly added tasks. S5: Input the training and test sets into the multi-task convolutional neural network for training and optimization; S6: Prepare test samples, obtain test data based on hyperspectral imager scanning, process the test data and input it into the optimized multi-task convolutional neural network, and simultaneously output the prediction results of plastic substrate type, additive type and additive content.

[0022] Furthermore, in this invention, common plastic substrates used in electronic and electrical equipment are selected and mixed with plastic substrates in gradient proportions of 1% to 30% by mass. The mixture is then melt-blended in a torque rheometer at 190°C to 250°C to prepare gradient proportion samples containing various plastic substrates and additives. Data preprocessing and feature engineering are implemented using Python, deep learning model training and inference are performed using PyTorch, and numpy and pandas are used for auxiliary calculations and data management.

[0023] Specifically, this invention constructs a multi-task convolutional neural network to achieve synchronous modeling and collaborative recognition of multiple attributes, avoiding the cumbersome process of building independent models for different recognition tasks. It enables rapid and accurate identification of plastic substrates and their additive types in electronic and electrical equipment, as well as precise prediction of additive content. Utilizing three-dimensional spectral cube data obtained from hyperspectral imaging, the invention extracts stable characteristics from weak peaks, coupling peaks, and strong nonlinear complex features in the high-dimensional spectrum, forming input data. By introducing a cross-task feature sharing and interaction mechanism within the multi-task convolutional neural network, information complementarity can be achieved between tasks, thereby synchronously outputting prediction results for plastic substrate type, additive type, and additive content, resulting in good performance. Furthermore, the simultaneous identification of plastic substrates, additive types and their contents is achieved by using hyperspectral imaging and multi-task convolutional neural networks. No additional chemical reagents and complex pretreatment are required. The detection process is simple, and the spectral acquisition process does not damage the sample, avoiding the destructive operation of traditional detection methods, thus having better practicality and safety. Furthermore, this method is not only applicable to the detection of typical plastics such as acrylonitrile-butadiene-styrene copolymer (ABS), high-impact polystyrene (HIPS) and polypropylene (PP) as well as bromine-, phosphorus-, and aluminum-based flame retardants, but can also be extended to more types of plastics and functional additives to meet the needs of complex solid waste recycling scenarios. Furthermore, this method exhibits excellent scalability. The constructed multi-task learning framework adopts a modular task branching structure and sets up standardized extension interfaces, allowing new task modules to be added according to usage needs without changing the overall architecture. For example, it can be used for predicting the degree of plastic aging or other tasks targeting the identification and quantification of plastic properties. In addition, although this method is based on near-infrared hyperspectral data, the attention mechanism and cross-task feature fusion structure employed possess modality-independent adaptive characteristics, thus enabling unified representation of data from different spectral modalities. This method is applicable to various two-dimensional spectral data modeling scenarios, including mid-infrared hyperspectral, Fourier transform infrared (FTIR), and Raman spectroscopy, thereby achieving stable transfer and application expansion under cross-modal conditions.

[0024] Preferably, a gradient-proportion sample containing multiple plastic substrates and additives is prepared, and three-dimensional spectral cube data of the gradient-proportion sample is obtained by scanning in a preset band using a hyperspectral imager, including: Select various types of plastic substrates and various types of additives, mix the additives with the plastic substrates in a preset mass fraction gradient ratio, and prepare gradient ratio samples containing plastic substrates and additives with known mass fractions, wherein the number of each type of sample in the gradient ratio samples is not less than three. Each sample in the gradient ratio sample is scanned in a preset band using a hyperspectral imager to obtain three-dimensional spectral cube data containing spatial data and spectral images.

[0025] Preferably, the plastic substrate type includes at least one of acrylonitrile-butadiene-styrene copolymer (ABS), high-impact polystyrene (HIPS), and polypropylene (PP), and the additives include at least one of flame retardant, plasticizer, filler, and antioxidant, wherein the flame retardant includes decabromodiphenyl ethane, decabromodiphenyl ether, aluminum trihydrate, and ammonium polyphosphate.

[0026] Preferably, the preset wavelength band has a spectral range of 935nm to 1720nm, the field of view of the hyperspectral imager is 30° to 45°, the distance between the lens of the hyperspectral imager and the sample is 200mm to 600mm, and the frame rate of the hyperspectral imager is 10Hz to 50Hz.

[0027] Preferably, the baseline label recording the gradient ratio samples includes the type of plastic substrate, the type of additive, and their mass fraction, including: The baseline label is obtained by recording the type of plastic substrate, the type of flame retardant, and its known mass fraction for each sample in the gradient ratio sample.

[0028] Preferably, the three-dimensional spectral cube data is preprocessed to obtain a dataset, and the dataset is divided into a training set and a test set, including: The spectral images of each identical sample in the gradient ratio sample are compared in parallel, and the spatial data and spectral images corresponding to abnormal samples are removed to obtain the filtered three-dimensional spectral cube data. The filtered three-dimensional spectral cube data is smoothed, preprocessed with first derivatives, and feature-selected to obtain a candidate set containing spectral data. A preset number of spectral data from each type of sample is randomly extracted from the candidate set to form a dataset. The dataset is then divided into a training set and a test set.

[0029] Preferably, the dataset is divided into a training set and a test set, including: Dataset partitioning includes at least one of the following: random sampling partitioning, cross-validation, and stratified sampling. The spectral data of each type of sample in the dataset are randomly divided into a training set and a test set, with the training set accounting for 70% to 75% of the dataset and the test set accounting for 20% to 25% of the dataset.

[0030] Preferably, a multi-task convolutional neural network based on one-dimensional convolution, incorporating an attention mechanism and cross-task feature sharing, is constructed, including: The training and test sets of the dataset are reconstructed into input data according to the wavelength sequence, so as to serve as a unified input format for multi-task convolutional neural networks; The multi-task convolutional neural network includes a shared convolutional module, a convolutional block attention module, a Cross-Stitch feature sharing unit, and a task-specific convolutional module. The shared convolutional module connects the convolutional block attention module and the Cross-Stitch feature sharing unit, and the Cross-Stitch feature sharing unit connects the task-specific convolutional module. The shared convolutional module is used to extract common spectral features of multiple tasks, the convolutional block attention module is used to enhance the response of key spectral bands, the Cross-Stitch feature sharing unit is used to realize adaptive fusion of features between tasks and set up an extension interface, and the task-specific convolutional module is used to output substrate classification, additive classification, and additive content. Among them, the Cross-Stitch feature sharing unit is a feature fusion mechanism in multi-task learning. Its core function is to achieve adaptive sharing and fusion of knowledge between tasks by learning a linear combination of the outputs of different task network layers. In the test set, when validating the accuracy of the multi-task convolutional neural network model, the accuracy of the plastic substrate and additive classification tasks, calculated based on the true values ​​in the baseline labels, is no less than 0.99 and 0.94, respectively. The R-squared value for additive content regression is also no less than [missing value]. 2 With a performance score of at least 0.90, it outperforms the traditional single-task model.

[0031] Preferably, test samples are prepared, and test data is obtained based on scanning with a hyperspectral imager. The processed test data is then input into an optimized multi-task convolutional neural network, which simultaneously outputs prediction results for plastic substrate type, additive type, and additive content, including: Following step S1, additional samples not included in the training and test sets by additive classification and mass fraction ratio are prepared to form test samples. Following S1-S2, test labels for plastic substrate type, additive type, and their mass fraction are recorded. Test data is obtained based on hyperspectral imaging. The test data undergoes smoothing filtering, first-order derivative preprocessing, and feature selection to obtain test input data. The test input data is input into the optimized multi-task convolutional neural network in step S5, and the predicted results for plastic substrate type, additive type, and additive content are output simultaneously. Smoothing filtering uses Savitzky-Golay smoothing filtering combined with first-order derivative, or other equivalent preprocessing methods. Feature selection uses feature band selection, which reduces data dimensionality and enhances the discrimination ability of key spectral bands to improve signal-to-noise ratio, eliminate redundancy, and enhance discriminative features. Savitzky-Golay smoothing filtering is a digital filter that achieves smoothing by performing local polynomial least squares fitting on data points. This filtering method effectively reduces noise interference and improves the signal-to-noise ratio while preserving important data features, making it particularly suitable for processing small fluctuations and noise in hyperspectral imaging data.

[0032] Specifically, the carefully processed test input data is fed into an optimized multi-task convolutional neural network. With its powerful feature extraction and fusion capabilities, this network can simultaneously handle the prediction tasks of plastic substrate type, additive type, and additive content, achieving efficient and accurate synchronous output.

[0033] Preferably, when verifying the accuracy of the multi-task convolutional neural network model using test data, the accuracy of the plastic substrate and additive classification tasks, calculated based on the true values ​​in the test labels, should be no less than 0.98 and 0.93, respectively, and the R-squared value for additive content regression should be no less than [missing value]. 2 Not less than 0.86.

[0034] Specifically, during the verification process, to ensure the reliability and stability of the results, multiple repeated experiments are required, and the average of the results from these experiments is taken as the final accuracy and R-value. 2Value. Meanwhile, all test samples prepared in different batches underwent rigorous testing according to the above procedure to comprehensively evaluate the performance of the multi-task convolutional neural network model in different scenarios. If accuracy or R-values ​​are found... 2 If the value is lower than the above standard, the model needs to be further optimized and adjusted, such as adjusting the network structure and optimizing the training parameters, until the accuracy requirements are met.

[0035] Example 1: Taking a near-infrared hyperspectral imager with a spectral range of 935nm to 1720nm and a resolution of 3.5nm as an example, and using a flame retardant as an additive, the following steps are included; Representative waste electronic and electrical samples were selected, and near-infrared spectral data were acquired directly using a near-infrared hyperspectral imager. Representative plastic samples were categorized by plastic substrate, including but not limited to ABS, HIPS, and PP; and by flame retardant, including but not limited to plastics containing brominated flame retardants, aluminum-containing flame retardants, and phosphorus-containing flame retardants. A near-infrared hyperspectral imager (NIR-HSI) with a spectral range of 935 nm to 1720 nm and a resolution of 3.5 nm was used, acquiring spectral cubes one by one in a row-scan manner. Three parallel samples of each type were prepared and scanned independently to ensure data representativeness and repeatability. The acquired spectral data were used to construct a machine learning model.

[0036] Savitzky-Golay smoothing filtering, first derivative and feature screening were performed on the collected spectral data in sequence to reduce noise, enhance subtle spectral differences and reduce data redundancy. The preprocessed near-infrared spectral data were divided into training set and test set. In order to ensure the uniform distribution of each type of sample when dividing the dataset, the specific steps when building the model are as follows: (1) The spectral data of all the samples were collected and summarized into the total dataset. Based on the random shuffling of the overall dataset three times, a fixed number of spectral samples were drawn for each type of sample to form a balanced dataset; (2) Each type of sample was then randomly divided into training set and test set, and the total amount of data in the training set accounted for 70% of the total data of the sample, and the total amount of data in the test set accounted for 30% of the total data of the sample; Plastic samples with a specific flame retardant ratio were separately extracted as an independent external test set for independent verification.

[0037] A multi-task learning model based on a one-dimensional convolutional neural network is constructed to simultaneously achieve plastic substrate classification, flame retardant type classification, and flame retardant content regression prediction. The model introduces a convolutional block attention mechanism after a shared feature extraction layer and deploys a Cross-Stitch feature sharing module between task branches. During training, a Bayesian optimization method is used to search for hyperparameters to obtain the optimal model structure and training configuration.

[0038] The model accuracy was validated using a test set, and the models for plastic substrate classification, flame retardant classification, and flame retardant content prediction were optimized. On the test set, the multi-task convolutional neural network model achieved accuracies of over 0.99 and 0.94 in plastic substrate classification and flame retardant content classification tasks, respectively, and a high R-value in the flame retardant content regression task. 2 It reaches 0.90 or higher.

[0039] The specific steps in constructing a multi-task convolutional neural network are as follows: The preprocessed spectral data is reconstructed into tensors of size (batch, 1, L) according to the wavelength sequence, which serve as the unified input format for subsequent one-dimensional convolutional networks. A two-level hard shared convolutional block is constructed. Each level of the shared convolutional block includes, in sequence: a one-dimensional convolutional layer (kernel size = 3, stride = 1, padding = 1), a ReLU activation layer, a one-dimensional max pooling layer (pooling kernel = 2), and a one-dimensional CBAM attention module with a channel compression ratio of 2. The shared features with an output size of (batch, 16, L / 4) from these two levels of shared convolutional blocks are used to achieve unified extraction of common spectral structures across multiple tasks.

[0040] After the shared layer output, a Cross-Stitch unit is introduced between the two classification task branches. A learnable 2×2 mixture matrix is ​​used to linearly combine the intermediate representations of the two tasks, thereby realizing a flexible feature sharing strategy of "partial sharing and partial independence". This structure can adaptively adjust the direction and intensity of cross-task feature flow according to task relevance, which is an important structural feature of the multi-task collaborative modeling capability of this invention.

[0041] Two levels of task-specific convolutional blocks were constructed for the three tasks. Each level of task-specific convolutional block included: a one-dimensional convolutional layer (the number of channels in the first block was expanded from 16 to 32, and the number of channels in the second block was expanded from 32 to 64; kernel size = 3, stride = 1, padding = 1), a ReLU activation layer, a one-dimensional max pooling layer (pooling kernel = 2), and a one-dimensional CBAM attention module (channel compression ratio = 4). This structure enabled the differential extraction of high-level semantic features for different tasks, allowing the model to simultaneously capture the distinguishing features of plastic substrates, the molecular structure response features of flame retardants, and the influence of concentration changes on spectral morphology.

[0042] After flattening the multi-level features of each task, fully connected output heads are established for each task, used to output predictions of plastic substrate type, flame retardant type, flame retardant content, and flame retardant content regression. These output heads are trained synchronously across multiple tasks through a joint optimization strategy, thereby fully mining the correlation information in the spectrum and improving the overall model stability.

[0043] Independent external test set samples were input into the model for validation. Near-infrared spectral data were acquired using a NIR-HSI spectrometer, and the constructed multi-task convolutional neural network model was used to simultaneously predict plastic substrate classification, flame retardant classification, and flame retardant content. The accuracy for the plastic substrate classification task reached over 0.98, and its confusion matrix is ​​shown below. Figure 2 As shown; the accuracy of the classification task involving flame retardants can reach 0.93, and its confusion matrix is ​​as follows. Figure 3 As shown; R for the flame retardant content regression task 2 The value can reach 0.86, and the comparison between its actual value and predicted value is as follows: Figure 4 As shown.

[0044] Spectral data preprocessing and feature engineering were both implemented using Python software, employing the scikit-learn library and the PyTorch deep learning framework, with numpy and pandas libraries used for auxiliary computation.

[0045] Example 2: The following are the specific steps for testing plastic mixtures recycled by a WEEE recycling plant: This embodiment is basically the same as Embodiment 1, except that the test object is 200 kg of actual plastic mixture recycled by a WEEE recycling plant. First, spectral data was collected according to the process described in S1, and then processed according to steps S2-S3. Using the prediction results of the optimized multi-task model in Embodiment 1, it was found that ABS accounted for 42%, HIPS for 33%, and PP for 25%; the flame retardant types included bromine-based for 18%, phosphorus-based for 12%, and aluminum-based for 20%; the average flame retardant addition ratio was 7.6%.

[0046] Example 3: Using discarded monitor casing plastic obtained from a dismantling workshop as the testing object, the specific steps are as follows: This embodiment is basically the same as Embodiment 1, except that the test object is 50 kg of waste display shell plastic obtained from the dismantling workshop. Using the model trained in Embodiment 1 for prediction, the results show that the batch of samples is mainly composed of HIPS substrate (78%) and PC substrate (22%), and contains brominated flame retardant, with the amount added ranging from 0% to 20%.

[0047] Example 4: Taking mixed plastic granules sorted out by a recycling plant as the testing object, the specific steps are as follows: The modeling process in this embodiment is basically the same as in Embodiment 1. The difference is that the test object is 100 kg of mixed plastic granules sorted from a recycling plant. In addition to the three tasks described in Embodiment 1 (plastic substrate classification, flame retardant category classification, and flame retardant content regression), a new task of predicting the degree of plastic aging is added. To adapt to the new task, this embodiment adds a set of dedicated task output heads for the degree of plastic aging at the task output stage without changing the shared convolutional layers, Cross-Stitch fusion structure, or existing task branches. This allows the new task to access the high-dimensional features of the task-specific convolutional blocks in parallel with the original tasks, achieving a structured extension of the multi-task framework.

[0048] The samples in this batch contained 38% ABS, 40% PP, and 22% HIPS. The flame retardants were mainly aluminum-based and phosphorus-based, with contents ranging from 3% to 15%. Based on the aforementioned multi-task convolutional neural network model, a new task predicting the degree of plastic aging was added, classifying aging levels into four categories: no aging, mild aging, moderate aging, and severe aging. External test set validation results showed that the classification accuracy of this task was no less than 0.90, effectively distinguishing different aging levels. Meanwhile, the prediction performance of the original three tasks remained stable, with the accuracy of the plastic substrate classification and flame retardant classification tasks being no less than 0.97 and 0.90, respectively. The R-value for the flame retardant content regression task was also high. 2 Not less than 0.87.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-task learning method for simultaneous identification and quantification of plastics and additives, characterized in that, include: Gradient-ratio samples containing various plastic substrates and additives were prepared, and three-dimensional spectral cubic data of the gradient-ratio samples were obtained by scanning in a preset band using a hyperspectral imager. The gradient ratio samples are recorded with reference labels including the type of plastic substrate, the type of additive, and their mass fraction; The three-dimensional spectral cube data is preprocessed to obtain a dataset, which is then divided into a training set and a test set. A multi-task convolutional neural network based on one-dimensional convolution and fusion attention mechanism and cross-task feature sharing is constructed. The multi-task convolutional neural network is used to simultaneously output the prediction results of plastic substrate type, additive type and additive content. The training and test sets are input into a multi-task convolutional neural network for training and optimization; Test samples are prepared, and test data is obtained by scanning with a hyperspectral imager. The test data is then processed and input into an optimized multi-task convolutional neural network, which simultaneously outputs prediction results of plastic substrate type, additive type, and additive content.

2. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 1, characterized in that, The preparation of gradient-proportion samples containing multiple plastic substrates and additives involves obtaining three-dimensional spectral cubic data of the gradient-proportion samples by scanning in a preset wavelength band using a hyperspectral imager, including: Select various types of plastic substrates and various types of additives, mix the additives with the plastic substrates in a preset mass fraction gradient ratio, and prepare gradient ratio samples containing plastic substrates and additives with known mass fractions, wherein the number of each type of sample in the gradient ratio samples is not less than three. Each sample in the gradient ratio sample is scanned in a preset band using a hyperspectral imager to obtain three-dimensional spectral cube data containing spatial data and spectral images.

3. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 2, characterized in that: The plastic substrate type includes at least one of acrylonitrile-butadiene-styrene copolymer, high-impact polystyrene, and polypropylene. The additives include at least one of flame retardant, plasticizer, filler, and antioxidant. The flame retardant includes decabromodiphenyl ethane, decabromodiphenyl ether, aluminum trihydrate, and ammonium polyphosphate.

4. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 3, characterized in that: The preset wavelength band has a spectral range of 935nm to 1720nm, the field of view of the hyperspectral imager is 30° to 45°, the distance between the lens of the hyperspectral imager and the sample is 200mm to 600mm, and the frame rate of the hyperspectral imager is 10Hz to 50Hz.

5. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 4, characterized in that: The baseline label recording the gradient ratio sample includes the plastic substrate type, additive type, and their mass fraction, including: The baseline label is obtained by recording the type of plastic substrate, the type of flame retardant, and its known mass fraction for each sample in the gradient ratio sample.

6. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 5, characterized in that, The three-dimensional spectral cube data is preprocessed to obtain a dataset, which is then divided into a training set and a test set, including: The spectral images of each identical sample in the gradient ratio sample are compared in parallel, and the spatial data and spectral images corresponding to abnormal samples are removed to obtain the filtered three-dimensional spectral cube data. The filtered three-dimensional spectral cube data is subjected to smoothing filtering, first derivative preprocessing and feature filtering to obtain a candidate set containing spectral data. A preset number of spectral data of each type of sample is randomly extracted from the candidate set to form a dataset. The dataset is then divided into a training set and a test set.

7. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 6, characterized in that, The process of dividing the dataset into training and test sets includes: Dataset partitioning includes at least one of the following: random sampling partitioning, cross-validation, and stratified sampling. The spectral data of each type of sample in the dataset are randomly divided into a training set and a test set, with the training set accounting for 70% to 75% of the dataset and the test set accounting for 20% to 25% of the dataset.

8. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 7, characterized in that, The construction of a multi-task convolutional neural network based on one-dimensional convolution and fusion attention mechanism and cross-task feature sharing includes: The training and test sets of the dataset are reconstructed into input data according to the wavelength sequence, so as to serve as a unified input format for multi-task convolutional neural networks; The multi-task convolutional neural network includes a shared convolutional module, a convolutional block attention module, a Cross-Stitch feature sharing unit, and a task-specific convolutional module. The shared convolutional module connects the convolutional block attention module and the Cross-Stitch feature sharing unit, and the Cross-Stitch feature sharing unit connects to the task-specific convolutional module. The shared convolutional module is used to extract common spectral features across multiple tasks. The convolutional block attention module is used to enhance the response of key spectral bands. The Cross-Stitch feature sharing unit is used to achieve adaptive feature fusion between tasks and set an extension interface. The task-specific convolutional module is used to output substrate classification, additive classification, and additive content. In the test set, when validating the accuracy of the multi-task convolutional neural network model, the accuracy of the plastic substrate and additive classification tasks, calculated based on the true values ​​in the baseline labels, is no less than 0.99 and 0.94, respectively. The R-squared value for additive content regression is also no less than [missing value]. 2 Not less than 0.

90.

9. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 8, characterized in that, The preparation of the test sample involves obtaining test data based on a hyperspectral imager scan. The processed test data is then input into an optimized multi-task convolutional neural network, which simultaneously outputs predictions for the plastic substrate type, additive type, and additive content, including: Additional samples not included in the training and test sets, based on additive classifications and mass fraction ratios, are prepared to form test samples. Test labels for plastic substrate type, additive type, and their mass fraction are recorded. Test data is obtained by scanning with a hyperspectral imager. The test data is then processed by smoothing filtering, first-order derivative preprocessing, and feature selection to obtain test input data. The test input data is input into an optimized multi-task convolutional neural network, which simultaneously outputs prediction results for plastic substrate type, additive type, and additive content.

10. The method for simultaneous identification and quantification of plastics and additives using multi-task learning according to claim 9, characterized in that: When verifying the accuracy of the multi-task convolutional neural network model using the test data, the accuracy of the plastic substrate and additive classification tasks was calculated to be no less than 0.98 and 0.93 respectively, based on the true values ​​in the test labels. The R-squared value for additive content regression was also no less than [missing value]. 2 Not less than 0.86.

Images