On-line detection and sorting method and system for regenerated plastic impurities based on multi-modal sensing

Abstract

The invention relates to the technical field of regenerated plastic impurity detection and separation, in particular to a regenerated plastic impurity online detection and separation method and system based on multi-mode sensing. Comprising the following steps: enabling plastic particles to sequentially pass through on-line detection areas; utilizing a microwave resonance sensor and a weak current sensor to synchronously obtain a dielectric spectrum characteristic signal and a weak conductive characteristic signal of each particle; inputting the signal into a multi-modal fusion neural network model for identification and classification so as to distinguish target plastic from impurities such as metal, PVC and PET; and a sorting control signal is generated according to the classification result, a downstream high-voltage sorting electrode is controlled to act, and impurity particles are accurately removed. According to the invention, through multi-modal information fusion and intelligent analysis, the problem of insufficient distinguishing precision of impurities with similar colors and shapes and different materials in the traditional optical and single electrical sorting technology is solved, high-precision and online automatic sorting of impurities in the regenerated plastic is realized, and the purity and quality of the regenerated plastic are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of recycled plastic impurity detection and sorting technology, specifically to an online detection and sorting method and system for recycled plastic impurities based on multimodal sensing. Background Technology

[0002] The recycling of plastics is of great significance for resource conservation and environmental protection. During the recycling process, recycled plastic granules often contain various impurities, among which trace metallic impurities (such as iron and aluminum fragments) and different types of polymeric impurities (such as PVC and PET) that are similar in color and shape to the target plastic are particularly challenging. These impurities can seriously affect the key properties of recycled plastic products. For example, metallic impurities may reduce the insulation performance of electrical appliance casings and pose safety hazards, while impurities such as PVC and PET can degrade the dielectric constant, mechanical strength, and thermal stability of the material. Existing sorting technologies face significant challenges: traditional optical sorting techniques (such as those based on color or shape recognition) are difficult to effectively distinguish between transparent or similarly colored plastic granules of different materials, and cannot detect internal metallic impurities; sorting methods based on electrical principles, such as electrostatic sorting, although utilizing differences in material conductivity, are greatly affected by the particle surface condition, environmental humidity, and charge uniformity. They have limited ability to distinguish between different polymeric impurities with small differences in dielectric constant or weak conductivity, easily leading to missorting or missed sorting. Therefore, there is an urgent need to develop a technology that can identify and sort metals and various polymeric impurities in recycled plastics online with high precision, in order to improve the purity and quality of recycled plastics and expand their application in high-end fields.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for online detection and sorting of impurities in recycled plastics based on multimodal sensing, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an online detection and sorting method for impurities in recycled plastics based on multimodal sensing, comprising: S1. The recycled plastic granules are passed sequentially through an online detection area in a single layer or single row. S2. Within the online detection area, microwave resonant sensing and weak current sensing are performed simultaneously on each plastic particle to obtain the dielectric spectrum characteristic signal and weak conductivity characteristic signal of the plastic particle, respectively. S3. The synchronously acquired dielectric spectrum feature signal and the weak conductivity feature signal are input into a pre-trained multimodal fusion neural network model for synchronous analysis and feature fusion, so as to identify and classify whether the plastic particles are impurity particles, and determine the impurity type when they are identified as impurity particles. S4. Based on the identification and classification results of step S3, generate a corresponding sorting control signal, and control the high-voltage sorting electrode set downstream of the online detection area to separate and remove the impurity particles from the plastic particle stream based on the sorting control signal.

[0006] Specifically, in step S3, the multimodal fusion neural network model adopts an early fusion architecture, which performs feature splicing or weighted fusion of the dielectric spectrum feature signal and the weak conductivity feature signal in the input layer or shallow convolutional network to form a fused multimodal feature map, which is then identified and classified by the subsequent convolutional neural network.

[0007] Specifically, in step S3, the multimodal fusion neural network model further includes a late-stage decision fusion module, which is used to make classification decisions on the features extracted by the early fusion architecture and to perform weighted fusion with the decision results of another parallel classification sub-network based on single-modal features, and output the final impurity type determination result.

[0008] Specifically, in step S4, the generation of the sorting control signal includes: Based on the determined impurity type, preset high-voltage pulse parameters are matched, including at least voltage amplitude, pulse width, and action timing. The high-voltage sorting electrode generates a corresponding high-voltage pulse electric field according to the high-voltage pulse parameters, and applies directional electrostatic repulsion or attraction to the impurity particles to achieve sorting.

[0009] Specifically, step S4 also includes a real-time location tracking step: Based on the timing and velocity of the plastic particles passing through the online detection area, the first time point at which the impurity particles move to the position of the high-voltage sorting electrode is predicted, and the timing of the high-voltage pulse is synchronized with the first time point.

[0010] Specifically, after sorting is performed in step S4, the following steps are also included: S5. The impurity particles and qualified plastic particles obtained after sorting are subjected to secondary sampling inspection, and the results of the secondary sampling inspection are fed back to the multimodal fusion neural network model for verification of the model's identification and classification results and online optimization and adjustment of model parameters.

[0011] Specifically, in step S5, the secondary sampling inspection includes at least measuring the dielectric constant or volume resistivity of the qualified plastic particles, comparing the measured statistical value with a preset qualified threshold, and if the threshold is exceeded, it is determined that the sorting purity is not up to standard, and a calibration signal is generated to trigger the adjustment of the multimodal fusion neural network model and / or the high voltage pulse parameters.

[0012] Specifically, before step S2, a preprocessing step is included for the dielectric spectral characteristic signal and the weak conductivity characteristic signal. The preprocessing includes at least signal noise reduction, baseline correction and feature normalization.

[0013] Specifically, the impurity types determined in step S3 include at least one or more of the following: metallic impurities, polyvinyl chloride impurities, polyethylene terephthalate impurities, and other non-target polymer impurities.

[0014] According to a second aspect of the present invention, an online detection and sorting system for impurities in recycled plastics based on multimodal sensing is provided, comprising: A pellet conveying device is configured to allow recycled plastic pellets to pass sequentially through an online detection area in a single layer or single row. A multimodal sensing device, which is set in the online detection area, includes a microwave resonant sensor and a weak current sensor, for synchronously acquiring the dielectric spectrum characteristic signal and weak conductivity characteristic signal of each plastic particle passing through the online detection area; A signal processing and control device includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, performs the following functions: receiving and preprocessing feature signals sent by the multimodal sensing device; inputting the preprocessed feature signals into a pre-trained multimodal fusion neural network model stored in the memory to identify and classify whether the plastic particles are impurity particles and their types; and generating a sorting control signal containing high-voltage pulse parameters based on the identification and classification results. A high-voltage sorting actuator is located downstream of the online detection area and electrically connected to the signal processing and control device. It is used to generate a corresponding high-voltage pulse electric field according to the received sorting control signal to sort and remove the impurity particles from the plastic particle stream.

[0015] Beneficial effects: The online detection and sorting method and system for impurities in recycled plastics based on multimodal sensing provided by this invention has the following significant advantages compared with the prior art: First, by innovatively integrating two complementary detection modes based on the physical principles of microwave resonant sensing and weak current sensing, a multi-dimensional material characteristic sensing system has been constructed. Microwave resonant sensing is extremely sensitive to the dielectric properties of materials (dielectric constant, loss angle), effectively distinguishing different types of polymers; weak current sensing is highly sensitive to the conductivity of materials (including ionic and electronic conductivity), accurately capturing characteristic signals of metals and polymers containing additives. This multi-modal information fusion fundamentally overcomes the limitations of single-sensor technology, making it possible to simultaneously and accurately identify "ghost impurities" (such as transparent PET and white PVC) that cannot be distinguished by traditional optical methods, as well as various polymer impurities that are difficult to classify accurately by traditional electrical methods.

[0016] Second, a deep learning-based multimodal fusion neural network model is employed for intelligent analysis and decision-making of the sensor signals. This model, through an architecture combining early fusion and late-stage decision fusion, can automatically and deeply mine the intrinsic correlation and complementary information between dielectric spectral features and weak conductivity features, forming a more robust joint feature representation. This not only significantly improves the accuracy and confidence of impurity classification but also enhances the model's tolerance to accidental interference or noise from single-mode signals, ensuring the stability and reliability of the online detection system in complex industrial environments.

[0017] Third, it achieves closed-loop control from intelligent identification to precise execution. The system combines high-precision classification results with a programmable high-pressure sorting actuator. Through intelligent matching of impurity type and high-pressure pulse parameters, and precise timing control based on real-time position tracking, it ensures that the most suitable sorting force is applied to impurities with different physical properties, and that the action is triggered at the optimal moment. This differentiated and precise sorting strategy maximizes the impurity rejection rate while significantly reducing damage to qualified materials, optimizing energy consumption, and improving overall sorting efficiency and the yield of pure materials.

[0018] Fourth, an online optimization closed loop based on secondary sampling inspection and macroscopic performance feedback is introduced. This system can not only sample and verify the sorting results to continuously optimize the AI ​​model, but also reverse-calibrate the sorting process parameters by monitoring the macroscopic performance indicators of the output materials (such as dielectric constant and resistivity). This two-level feedback mechanism endows the system with self-learning and adaptive adjustment capabilities, enabling it to cope with long-term changes such as raw material batch fluctuations and equipment status drift, ensuring long-term stability and continuous optimization of sorting performance, and truly realizing intelligent production.

[0019] Fifth, advanced sensing, identification, and control technologies are integrated into a complete online sorting system, enabling automated and continuous production operations. This system can be directly integrated into existing recycled plastic processing production lines, improving the purity of plastic particles in real time and online. It provides reliable technical equipment support for the production of high-performance, high-value-added recycled plastic products, and has good industrial application prospects and economic benefits. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the online detection and sorting method for impurities in recycled plastics based on multimodal sensing provided in a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the system composition of the online detection and sorting system for impurities in recycled plastics based on multimodal sensing provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides an online detection and sorting method for impurities in recycled plastics based on multimodal sensing, comprising: S1. The recycled plastic granules are passed sequentially through an online detection area in a single layer or single row. It should be further explained that in step S1, ensuring that the plastic granules pass through in a single layer or column is crucial to guaranteeing that each granule can be detected independently and without obstruction. This can be achieved using an electromagnetic vibrating feeder in conjunction with a chute at a specific angle. The frequency of the vibrating feeder is preferably 50Hz, and the amplitude is preferably 0.5mm, so that the granules are evenly dispersed on the chute and pass through sequentially.

[0024] S2. Within the online detection area, microwave resonant sensing and weak current sensing are performed simultaneously on each plastic particle to obtain the dielectric spectrum characteristic signal and weak conductivity characteristic signal of the plastic particle, respectively. It should be further explained that step S2 is the core detection step. Microwave resonant sensing is specifically implemented through a microstrip open-loop resonant sensor, with a designed resonant frequency of 2.4 GHz. When a particle (3-5 mm in diameter) passes through the detection area approximately 1 mm above the resonant loop, it will cause a resonant frequency... and quality factor The change. We used an integrated vector network analyzer module to quickly acquire the S21 parameter spectrum centered at 2.4 GHz with a bandwidth of 200 MHz at the instant the particle passed through, with a scan time of 10 ms. This spectrum serves as the dielectric spectrum characteristic signal, a vector containing amplitude and phase information at 201 frequency points. Weak current sensing is achieved through a pair of parallel stainless steel plate electrodes with a spacing of 5 mm and a 50 VDC bias voltage applied. When a particle passes between the electrodes, it changes the electric field distribution between the plates, generating a weak transient current signal. This signal is generated by a transimpedance gain of A V / A micro-current amplifier amplifies and filters the signal, outputting a time-domain voltage signal. We collect data from 200 sampling points within a 2ms time window before and after the signal peak, using this as a weak conductivity characteristic signal. The triggering of the two sensors is synchronized by the same through-beam infrared photoelectric switch, ensuring that the data strictly corresponds to the same particle.

[0025] S3. The synchronously acquired dielectric spectrum feature signal and the weak conductivity feature signal are input into a pre-trained multimodal fusion neural network model for synchronous analysis and feature fusion, so as to identify and classify whether the plastic particles are impurity particles, and determine the impurity type when they are identified as impurity particles. It should be further explained that in step S3, the pre-trained multimodal fusion neural network model is a convolutional neural network implemented based on the PyTorch framework. Training this model requires a large amount of labeled sample data. We collect at least 5000 particles each containing the target plastic (such as polypropylene PP) and impurities such as metal, PVC, and PET, obtain their multimodal sensing data, and label them. The preprocessed data is divided into training, validation, and test sets in an 8:1:1 ratio. The model adopts the aforementioned early fusion architecture. The input layer receives the concatenated feature vector, and after passing through 3 convolutional layers and 2 fully connected layers, it outputs a 4-dimensional vector, corresponding to the predicted probabilities of "target PP", "metal", "PVC", and "PET", respectively. Using the cross-entropy loss function and the Adam optimizer, the initial learning rate is set to 0.001, the batch size is 32, and training is conducted for 50 epochs until the accuracy on the validation set stabilizes above 99%.

[0026] S4. Based on the identification and classification results of step S3, generate a corresponding sorting control signal, and control the high-voltage sorting electrode set downstream of the online detection area to separate and remove the impurity particles from the plastic particle stream based on the sorting control signal.

[0027] It should be further explained that in step S4, the sorting control signal is a digital signal containing "action enable" (1 for sorting, 0 for no sorting) and "impurity type code". The high-voltage sorting electrode is located 20cm downstream of the detection area and consists of a high-voltage pulse generator (maximum output voltage 30kV, pulse width adjustable) and a needle-plate electrode. When a sorting signal is received, the controller retrieves the corresponding high-voltage pulse parameters (voltage, pulse width) from a preset parameter table according to the impurity type (e.g., metal corresponds to number 1, PVC corresponds to number 2, etc.), and controls the high-voltage pulse generator to discharge at a precise moment (calculated by position tracking), generating a strong electric field force to knock the impurity particles out of the material flow, causing them to deviate from their original trajectory and fall into the impurity collection hopper. This method, through multimodal information fusion, significantly improves the recognition rate of impurities that are difficult to distinguish using traditional single sensing methods, especially for impurities with similar colors and shapes but different materials, achieving online, high-speed, and high-precision intelligent sorting.

[0028] Specifically, in step S3, the multimodal fusion neural network model adopts an early fusion architecture, which performs feature splicing or weighted fusion of the dielectric spectrum feature signal and the weak current feature signal in the input layer or shallow convolutional network to form a fused multimodal feature map, which is then identified and classified by the subsequent convolutional neural network.

[0029] It should be further explained that the specific implementation of the early fusion architecture is as follows: First, the raw data collected by the two sensors is preprocessed. This includes the microwave dielectric spectrum characteristic signal, i.e., the amplitude spectrum vector of the S21 parameter. and weak current time-domain signal vector After being normalized, the components are directly concatenated at the input layer to form a one-dimensional fused feature vector of length 401. : in, This represents the normalized dielectric spectrum eigenvector. This represents the normalized weak current eigenvector. This represents a vector concatenation operation. Subsequently, the fused feature vector is reshaped into a... The feature map is input into a one-dimensional convolutional neural network. The preferred structure of this CNN is as follows: 1. First convolutional layer: Uses 64 one-dimensional convolutional kernels of size 7, with a stride of 2, and employs the ReLU activation function. The output feature map size is... There are 1 point with a depth of 64.

[0030] 2. First max pooling layer: pooling kernel size is 3, stride is 2, and output feature map size is [size missing]. The depth is 64.

[0031] 3. Second convolutional layer: Uses 128 one-dimensional convolutional kernels of size 5, with a stride of 1, and employs the ReLU activation function. The output feature map size is... The depth is 128.

[0032] 4. Second max pooling layer: pooling kernel size is 3, stride is 2, and output feature map size is [missing value]. The depth is 128.

[0033] 5. Third convolutional layer: Uses 256 one-dimensional convolutional kernels of size 3 with a stride of 1, employs the ReLU activation function, and outputs a feature map of size [size missing]. The depth is 256.

[0034] 6. Flatten the output feature map of the third convolutional layer into a one-dimensional vector with a length of... .

[0035] 7. First fully connected layer: Maps the 11264-dimensional vector to 512 dimensions, uses the ReLU activation function, and introduces Dropout regularization. The dropout rate is set to 0.5 to prevent overfitting.

[0036] 8. Second Fully Connected Layer: Maps the 512-dimensional vector to 4 dimensions (corresponding to 4 categories), uses the Softmax activation function, and outputs the predicted probability for each category. Kernel sizes of 7, 5, and 3 are chosen to extract features at different scales, from a larger receptive field capturing the overall pattern to a smaller receptive field capturing detailed features. The ReLU activation function is chosen because it effectively alleviates the vanishing gradient problem and accelerates training. Early fusion allows the network to learn the correlation between features of two modalities in the first layer, which is beneficial for extracting joint abstract features that are more effective for classification, avoiding the drawbacks of information isolation at the feature level, thereby improving model performance.

[0037] Specifically, in step S3, the multimodal fusion neural network model further includes a late-stage decision fusion module, which is used to make classification decisions on the features extracted by the early fusion architecture and to perform weighted fusion with the decision results of another parallel classification sub-network based on single-modal features, and output the final impurity type determination result.

[0038] It should be further explained that, to improve the robustness of the model, this invention constructs a hybrid fusion model containing three parallel branches. The main branch is an early fusion CNN. The other two branches are single-modal CNN sub-networks, with the same structure as the CNN part of the main branch (from input to the third convolutional layer), but the input and the first layer structure are independent. Dielectric network: the input is only the preprocessed dielectric spectrum feature vector. First, a separate one-dimensional convolutional layer (64 kernels of size 7) is used. Subsequent layers share weights or have the same structure as the corresponding layers in the main branch. The final output is a 4-dimensional decision vector. Current subnetwork: The input is only the preprocessed current feature vector. First, a separate one-dimensional convolutional layer (64 kernels of size 7) is used, which is shared by subsequent layers. The final output is a 4-dimensional decision vector. Post-decision fusion module: The main branch outputs the decision vector. This module receives The weights are then weighted and averaged. The weights can be fixed or dynamic. Dynamic weights are calculated using a lightweight attention network, which uses the output feature vectors of the penultimate layer (a 512-dimensional fully connected layer) of three subnetworks. For input, calculate the attention score: First, concatenate the three feature vectors: Then, attention weights are calculated using a small fully connected network: in, , , , These are learnable parameters. The function ensures the attention weights of the output. The sum is 1. The final decision is: Furthermore, take The category corresponding to the highest probability value is the final classification result. The advantage of this hybrid architecture is that when a certain modal signal is strongly disturbed (e.g., metal shavings causing current signal saturation), the decision confidence of that modal subnetwork decreases, and the attention mechanism automatically reduces its weights. or It relies more on decision-making from other modalities, thereby ensuring the stability of the overall system decision-making and improving its anti-interference ability.

[0039] Specifically, in step S4, the generation of the sorting control signal includes: matching preset high-voltage pulse parameters according to the determined impurity type, wherein the high-voltage pulse parameters include at least voltage amplitude, pulse width and action sequence; the high-voltage sorting electrode generates a corresponding high-voltage pulse electric field according to the high-voltage pulse parameters, and applies a directional electrostatic repulsion or attraction force to the impurity particles to achieve sorting.

[0040] It should be further explained that, in order to achieve accurate sorting, differentiated high-voltage pulse parameters need to be set for different impurity types. During system initialization, a "Impurity Type - High-Voltage Pulse Parameter" lookup table is pre-stored in the controller's non-volatile memory. This table is established based on experimental calibration. For example, for common impurities and the target plastic (PP), the minimum electric field force required to produce effective deflection is determined experimentally, and the required pulse parameters are then derived. A preferred parameter setting is as follows: 1. Metallic impurities: voltage amplitude Pulse width Reason: Metals have good electrical conductivity and are easily induced with charges. A moderate voltage and a short pulse width are sufficient to generate a repulsive force, and the short pulse width reduces energy consumption and the impact on nearby qualified particles.

[0041] 2. PVC impurities: voltage amplitude Pulse width Reason: PVC has high resistivity, requiring higher voltage and longer charging time (pulse width) to fully polarize it and generate sufficient Coulomb force.

[0042] 3. PET impurities: voltage amplitude Pulse width Reason: The dielectric properties of PET differ from those of PP, but not as significantly as those of PVC, requiring appropriate parameters.

[0043] 4. Other impurities / Default: Voltage amplitude Pulse width Reason: For uncertain or difficult-to-sort impurities, the most conservative and forceful parameters are used to ensure removal. The sorting control signal is a digital instruction packet containing at least the following fields: {Enable Flag: 1, Type Code: 2, Voltage (kV): 12, Pulse Width (µs): 150, Trigger Delay (µs): xxxx}. Upon receiving this instruction, the high-voltage pulse generator precisely outputs a negative high-voltage pulse with an amplitude of 12kV and a width of 150µs to the needle electrode at the moment indicated by the trigger delay. A non-uniform strong electric field is formed between the needle electrode and the grounding plate electrode below. The polarized PVC particles are subjected to a Coulomb force towards the grounding plate in this electric field. Its function is to deflect its trajectory downwards, thus separating it from the horizontally flying qualified PP particles. Through this parameterized difference control, energy consumption and sorting selectivity can be optimized while ensuring sorting effectiveness.

[0044] Specifically, step S4 also includes a real-time position tracking step: based on the timing and speed of the plastic particles passing through the online detection area, predict the first time point at which the impurity particles move to the position of the high-voltage sorting electrode, and synchronize the timing of the high-voltage pulse with the first time point.

[0045] It should be further explained that position tracking is the core of achieving high-speed, online, and accurate sorting. The specific implementation steps are as follows: At the entrance and exit of the online detection area, install a high-speed, high-precision through-beam fiber optic sensor (A and B), with a distance between them... Precise calibration. When a particle triggers sensor A, a timestamp is recorded. When sensor B is triggered, record... The average velocity of the particle as it passes through the detection area. for: Furthermore, let the distance from the center of the detection area (i.e., the multimodal sensing point) to the sorting point directly below the tip of the high-voltage sorting electrode be . The moment when a particle is identified at the center of the detection area (completing neural network inference) is denoted as . This moment is slightly later Its delay is mainly due to signal processing and model inference time. This can be measured and is approximately 2ms. Therefore, the estimated time for the particle to reach the sorting point is [time missing]. for: Furthermore, the controller internally maintains a high-precision timer (accuracy better than 1µs) and a sorting task queue. When in When an impurity particle is detected, calculate immediately. And will include the particle ID, sorting parameters, and trigger time. Tasks are inserted into a queue. A separate real-time task scheduling thread monitors this queue, and when system time arrives... At that time, a trigger command is sent to the high-voltage pulse generator. The response delay of the high-voltage pulse generator (including communication and rise time) is approximately 50µs, which needs to be compensated for in advance. This precise prediction and timing triggering ensures that the high-voltage pulse is applied precisely when the impurity particles have moved to the region of strongest electric field below the electrode, thus achieving optimal sorting results with a sorting accuracy exceeding 99%. Even when particle velocities vary between 1.5m / s and 3.5m / s, the algorithm adaptively adjusts to guarantee synchronization accuracy within ±0.5ms.

[0046] Specifically, after sorting in step S4, the process further includes: S5, performing secondary sampling inspection on the sorted impurity particles and qualified plastic particles respectively, and feeding back the results of the secondary sampling inspection to the multimodal fusion neural network model for verifying the model's identification and classification results and optimizing and adjusting the model parameters online.

[0047] It should be further explained that step S5 constructs an online learning optimization closed loop. After sorting, the qualified material stream and the impurity material stream enter two independent buffer bins respectively. At the outlet of each bin, a low-speed but higher-precision verification detection unit is installed. This unit can be a simplified version of a multimodal sensor or a detector based on other principles (such as a near-infrared spectroscopy probe). This embodiment preferably uses the same microwave resonance and weak current sensing technology as the main detection, but randomly samples the material stream at a lower frequency (e.g., 1-2 times per second). The verification detection unit sends the detected signal to a "teacher model" (which is a backup of a stable version of the main model) with the same network structure but fixed parameters for identification. The identification result of the "teacher model" is compared with the previous identification results of the particle on the main line (associated by particle ID and time sequence). If the results are inconsistent, the particle is marked as a "suspected misjudged sample," and its original multimodal data, main model prediction results, and teacher model prediction results are packaged and stored. Simultaneously, periodically (e.g., every 100 "suspected misclassified samples"), a small number of samples (e.g., 20 each) are manually extracted from the impurity and qualified bins for laboratory-level verification (e.g., DSC thermal analysis, FTIR spectral analysis) to determine their true category and assign these samples a "true label." Every 24 hours, or when a certain number of "true label" samples are accumulated (e.g., 50), the system initiates an online incremental learning process. The incremental learning process is as follows: 1. Load the current master model parameters from memory as initial values.

[0048] 2. Mix the newly collected samples with "real labels" with some historical training data to form a small batch of incremental training set.

[0049] 3. Train this incremental training set for 1-3 epochs using a small learning rate (e.g., 0.0001) and the Adam optimizer.

[0050] 4. Evaluate performance using the validation set. If accuracy improves or remains unchanged, update the main model with the new parameters; otherwise, roll back. This process enables the model to continuously adapt to slow changes in raw material characteristics (such as different aging degrees of different batches of recycled material) and newly emerging impurity types, achieving self-optimization and maintaining high sorting accuracy over the long term.

[0051] Specifically, in step S5, the secondary sampling inspection includes at least measuring the dielectric constant or volume resistivity of the qualified plastic particles, comparing the measured statistical value with a preset qualified threshold, and if the threshold is exceeded, it is determined that the sorting purity is not up to standard, and a calibration signal is generated to trigger the adjustment of the multimodal fusion neural network model and / or the high voltage pulse parameters.

[0052] It should be further noted that, in addition to verifying individual particles, this step adds monitoring and feedback on the macroscopic properties of the batch material. An online dielectric constant / resistivity integrated test probe is installed at the outlet of the collection silo for qualified plastic particles. This probe uses the parallel-plate capacitor principle to measure the equivalent dielectric constant of the plastic particles flowing through it in real time. and volume resistivity The system calculates the moving average of the measurements taken over the past minute every minute. and Based on the quality standards of the target product (such as PP for electrical appliance housings), a preset acceptable threshold is established: dielectric constant. The volume resistivity should be between 2.2 and 2.4. Should be greater than The control logic is as follows: if three consecutive times (i.e., three consecutive minutes) occur... All are greater than 2.45, or All below If the system determines that the sorting purity may be decreasing, a "purity alarm" calibration signal will be generated. This signal triggers two types of automatic adjustments: 1. First type, model confidence adjustment: The controller temporarily increases the threshold for sorting operations. For example, the decision threshold for the neural network output probability is increased from the default 0.8 to 0.9. This means that sorting will only be performed when the model determines a particle to be an impurity with a higher confidence level. This is a conservative strategy that may increase the carry-out rate of qualified material in the short term, but it can significantly reduce the impurity miss rate and ensure the purity of the outflowing material.

[0053] 2. The second type, sorting intensity adjustment: The system automatically increases the voltage value in the high-voltage pulse parameter lookup table globally by 10% and increases the pulse width by 20%. This stronger sorting force ensures that impurity particles at the critical sorting point are effectively removed. Simultaneously, the system flags this event and suggests that the operator check the raw materials or initiate a more comprehensive incremental model learning process. This closed-loop control based on final product performance indicators directly links the optimization goals of the sorting process to the final product quality, achieving result-oriented adaptive process control and ensuring long-term stable high-quality production.

[0054] Specifically, before step S2, a preprocessing step is included for the dielectric spectrum feature signal and the weak current feature signal. The preprocessing includes at least signal noise reduction, baseline correction and feature normalization.

[0055] It should be further noted that preprocessing is crucial for ensuring data quality, and the specific steps are as follows: 1. Signal noise reduction: ① Microwave dielectric spectrum signal: mainly contains high-frequency random noise. Discrete wavelet transform is used for denoising. The 'sym4' wavelet basis is preferred, and a 3-level decomposition is performed. High-frequency detail coefficients of levels 1-3 are then processed. A soft thresholding process is performed using the universal thresholding method (VisuShrink). The universal threshold calculation formula is: ,in It is the noise standard deviation, using the first level of detail coefficients. Median absolute value estimate: , This is the signal length. The soft threshold function is: After processing, wavelet reconstruction is performed to obtain the denoised spectral signal.

[0056] ② Weak current signals: mainly including power frequency interference and low-frequency drift. First, a fourth-order Butterworth digital high-pass filter with a cutoff frequency of 500Hz is designed to filter out low-frequency drift. Then, a sliding window mid-range filter is used to further remove spike noise. The window length is preferably 1 / 5 of the number of sampling points of the main pulse width of the signal (about 1ms), that is, if the sampling rate is 100kSPS, the window length is 20 points.

[0057] 2. Baseline Correction: The system automatically acquires a segment of background signal with no particles passing through during each startup and at regular intervals (e.g., 5 minutes) of standby time. For the microwave spectrum, its... The amplitude and phase background vector of the parameters For current signals, record their zero-bias voltage value. In actual testing, the corresponding baseline value is subtracted from the original signal: , .

[0058] 3. Feature Normalization: The corrected signal is mapped to the [0,1] interval to eliminate differences in dimensions and numerical range. First, the maximum and minimum values ​​of each feature dimension are determined using statistical values ​​from a large number of samples. For the 201 frequency points of the microwave spectrum, each frequency point is normally normalized independently. Assume the... The maximum response value obtained from the training samples is [number] frequency points. The minimum response value is The value of the new sample signal at this frequency point. Normalized to: Furthermore, for weak current signals, the maximum value among 200 points in the entire time domain signal is taken. and minimum value Perform overall normalization: Understandably, after the above preprocessing, the data input to the neural network becomes clean, stable, and scale-uniform, which greatly improves the training efficiency of the model and the final classification accuracy and generalization ability.

[0059] Specifically, the impurity types determined in step S3 include at least one or more of the following: metallic impurities, polyvinyl chloride impurities, polyethylene terephthalate impurities, and other non-target polymer impurities.

[0060] It should be further noted that the types of impurities that the method of the present invention can effectively identify specifically include: 1. Metallic impurities: such as tiny fragments of iron, aluminum, and copper (size > 0.5 mm). Their microwave dielectric characteristics exhibit extremely high (theoretically infinite) real parts of the dielectric constant, causing a drastic shift in the sensor's resonant frequency (e.g., shifting to lower frequencies by tens of MHz) and severe broadening of the resonant peak (a sharp drop in Q value). Their weak current characteristics manifest as extremely strong current pulse signals, with peak values ​​reaching nanoamperes or even microamperes, far exceeding the picoampere level signals of plastic particles. Multimodal fusion can distinguish them from certain high-dielectric polymers.

[0061] 2. Impurities in polyvinyl chloride (PVC): Due to the presence of chlorine atoms and plasticizers (such as DOP), PVC has a dielectric loss tangent. Its absorption rate in the microwave frequency band (e.g., 2-3 GHz) is significantly higher than that of polyolefins (PP / PE). In the dielectric spectrum, it exhibits absorption peaks in specific frequency bands. Its weak current signal, due to the migration of ions from the plasticizer, presents a broad, gradually varying current peak, unlike the sharp pulses of metals. By combining these two characteristics, the model can reliably distinguish PVC from the target PP.

[0062] 3. Polyethylene terephthalate (PET) impurities: PET molecules are highly polar, and their real dielectric constant (approximately 3.0-3.5) is higher than that of PP (approximately 2.2-2.4). In the dielectric spectrum, its resonant frequency shift direction differs from that of PVC, and its loss peak characteristics are not obvious. PET is an excellent insulator, and its weak current signal is extremely weak, similar to that of PP. The model primarily distinguishes PET from PP through differences in the dielectric spectrum, with the current signal used as an auxiliary verification.

[0063] 4. Other non-target polymeric impurities: such as polycarbonate (PC), acrylonitrile-butadiene-styrene copolymer (ABS), polyamide (PA), etc. These materials exhibit distinct dielectric spectra and conductivity characteristics under microwave and weak current sensing. During model training, they can be uniformly categorized as "other impurities," or, if sufficient training data is available, further subdivided into more categories. The multimodal fusion network can learn the difference boundaries between these materials and the target PP in the multidimensional feature space, thereby identifying them. By simultaneously covering metals, polar polymers (PVC), nonpolar polymers with different dielectric constants (PET), and other polymeric impurities, the method of this invention achieves comprehensive detection of most common harmful impurities in recycled plastics, fundamentally solving the bottleneck of traditional sorting technologies.

[0064] Please see Figure 2 The present invention provides another embodiment, which provides an online detection and sorting system for impurities in recycled plastics based on multimodal sensing. The online detection and sorting system for impurities in recycled plastics based on multimodal sensing includes: The pellet conveying device 100 is configured to allow recycled plastic pellets to pass sequentially through an online detection area in a single layer or single row. A multimodal sensing device 200, which is disposed in the online detection area, includes a microwave resonant sensor and a weak current sensor, for synchronously acquiring the dielectric spectrum characteristic signal and weak conductivity characteristic signal of each plastic particle passing through the online detection area; The signal processing and control device 300 includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, performs the following functions: receiving and preprocessing feature signals sent by the multimodal sensing device 200; inputting the preprocessed feature signals into a pre-trained multimodal fusion neural network model stored in the memory to identify and classify whether the plastic particles are impurity particles and their types; and generating a sorting control signal containing high-voltage pulse parameters based on the identification and classification results. A high-voltage sorting actuator 400 is located downstream of the online detection area and electrically connected to the signal processing and control device 300. It is used to generate a corresponding high-voltage pulse electric field according to the received sorting control signal to sort and remove the impurity particles from the plastic particle stream.

[0065] It should be further noted that this system is the specific hardware implementation of the above method. The particle conveying device 100 includes a storage bin, an electromagnetic vibrating feeder, a guide chute, and a constant-speed belt. The electromagnetic vibrating feeder (frequency adjustable at 50Hz, amplitude adjustable from 0.1-1mm) vibrates the particles evenly out of the storage bin. After being sorted by the guide chute, the particles form a single-layer sparsely distributed particle flow on the constant-speed belt (speed adjustable from 1-4m / s), ensuring that the particle spacing is greater than 5mm and that the particles pass through the online detection area in an orderly manner. The core of the multimodal sensing device 200 is two coaxial sensors with their detection areas overlapping in the vertical direction. The microwave resonant sensor is a microstrip open-loop resonant ring based on FR4 board material. Its dimensions are optimized by electromagnetic simulation software, and it operates at 2.4GHz. It is connected to an integrated VNA module (such as ADL5960) via an SMA connector. This module can complete a scan of 201 frequency points within 10ms under the control of an external trigger signal. The weak current sensor consists of a pair of parallel, gold-plated copper electrode plates (20mm x 20mm, 5mm spacing). The electrodes are connected to an amplification circuit composed of a high-precision weak current detection chip (such as ADI's ADA4530-1). This circuit has… The transimpedance gain is V / A and the bandwidth is 1kHz. Two sensors are encapsulated in an electromagnetically shielded box, leaving only a small gap for particles to pass through. Five mm in front of the sensors are through-beam fiber optic sensors A and B for triggering and velocity measurement. The signal processing and control unit 300 uses an industrial PC (such as an Advantech UNO-2484G, equipped with an Intel Core i7 processor, 16GB RAM, 256GB SSD, and an NVIDIA Jetson Nano module for neural network acceleration). Its memory stores the following computer program modules: data acquisition driver, signal preprocessing algorithm, trained multimodal fusion neural network model (.pt or .onnx format), position tracking and sorting control logic, and human-machine interface software. The high-voltage sorting actuator 400 includes a programmable high-voltage pulse generator (such as a Trek20 / 20C-HS, output voltage 0-20kV, rise time <5µs) and a set of sorting electrode assemblies. The electrode assembly consists of a tungsten carbide needle electrode (1mm in diameter, 0.1mm tip curvature radius) and an inclined stainless steel plate (serving as a grounding electrode and guide plate), with an adjustable distance between them (typically 10-20mm). The needle electrode is connected to the positive output of a high-voltage pulse generator, and the grounding plate is grounded. When the industrial control computer issues a sorting command, the high-voltage pulse generator outputs a pulsed high voltage with specified parameters at a designated time, generating corona discharge and a strong electric field between the needle and the plate. This applies a directional force to charged or polarized impurity particles, causing them to deviate from their original trajectory. The entire system is mounted on a rigid frame and equipped with a protective cover and safety interlock devices. This system integrates innovative sensing methods, intelligent algorithms, and precision actuators into a complete, automated intelligent sorting device that can be directly connected to existing plastic recycling production lines to achieve continuous, high-precision sorting operations 24 / 7.

[0066] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the online detection and sorting method for impurities in recycled plastics based on multimodal sensing. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0067] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0068] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0069] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0070] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for online detection and sorting of impurities in recycled plastics based on multimodal sensing, characterized in that, Includes the following steps: S1. The recycled plastic granules are passed sequentially through an online detection area in a single layer or single row. S2. Within the online detection area, microwave resonant sensing and weak current sensing are performed simultaneously on each plastic particle to obtain the dielectric spectrum characteristic signal and weak conductivity characteristic signal of the plastic particle, respectively. S3. The synchronously acquired dielectric spectrum feature signal and the weak conductivity feature signal are input into a pre-trained multimodal fusion neural network model for synchronous analysis and feature fusion, so as to identify and classify whether the plastic particles are impurity particles, and determine the impurity type when they are identified as impurity particles. S4. Based on the identification and classification results of step S3, generate a corresponding sorting control signal, and control the high-voltage sorting electrode set downstream of the online detection area to separate and remove the impurity particles from the plastic particle stream based on the sorting control signal.

2. The method according to claim 1, characterized in that, In step S3, the multimodal fusion neural network model adopts an early fusion architecture, which performs feature splicing or weighted fusion of the dielectric spectrum feature signal and the weak conductivity feature signal in the input layer or shallow convolutional network to form a fused multimodal feature map, which is then identified and classified by the subsequent convolutional neural network.

3. The method according to claim 2, characterized in that, In step S3, the multimodal fusion neural network model further includes a late-stage decision fusion module, which is used to classify the features extracted by the early fusion architecture and perform weighted fusion with the decision results of another parallel classification sub-network based on single-modal features to output the final impurity type determination result.

4. The method according to claim 1, characterized in that, In step S4, the generation of the sorting control signal includes: Based on the determined impurity type, preset high-voltage pulse parameters are matched, including at least voltage amplitude, pulse width, and action timing. The high-voltage sorting electrode generates a corresponding high-voltage pulse electric field according to the high-voltage pulse parameters, and applies directional electrostatic repulsion or attraction to the impurity particles to achieve sorting.

5. The method according to claim 4, characterized in that, Step S4 also includes a real-time location tracking step: Based on the timing and velocity of the plastic particles passing through the online detection area, the first time point at which the impurity particles move to the position of the high-voltage sorting electrode is predicted, and the timing of the high-voltage pulse is synchronized with the first time point.

6. The method according to claim 1, characterized in that, After sorting is performed in step S4, the following steps are also included: S5. The impurity particles and qualified plastic particles obtained after sorting are subjected to secondary sampling inspection, and the results of the secondary sampling inspection are fed back to the multimodal fusion neural network model for verification of the model's identification and classification results and online optimization and adjustment of model parameters.

7. The method according to claim 6, characterized in that, In step S5, the secondary sampling inspection includes at least measuring the dielectric constant or volume resistivity of the qualified plastic particles, comparing the measured statistical value with a preset qualified threshold, and if the threshold is exceeded, it is determined that the sorting purity is not up to standard, and a calibration signal is generated to trigger the adjustment of the multimodal fusion neural network model and / or the high voltage pulse parameters.

8. The method according to claim 1, characterized in that, Before step S2, the method further includes a preprocessing step for the dielectric spectral characteristic signal and the weak conductivity characteristic signal, wherein the preprocessing includes at least signal noise reduction, baseline correction and feature normalization.

9. The method according to claim 1, characterized in that, The impurity types determined in step S3 include at least one or more of the following: metallic impurities, polyvinyl chloride impurities, polyethylene terephthalate impurities, and other non-target polymer impurities.

10. An online detection and sorting system for impurities in recycled plastics based on multimodal sensing, used to perform the method as described in any one of claims 1-9, characterized in that, include: A pellet conveying device is configured to allow recycled plastic pellets to pass sequentially through an online detection area in a single layer or single row. A multimodal sensing device, which is set in the online detection area, includes a microwave resonant sensor and a weak current sensor, for synchronously acquiring the dielectric spectrum characteristic signal and weak conductivity characteristic signal of each plastic particle passing through the online detection area; A signal processing and control device includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, performs the following functions: receiving and preprocessing feature signals sent by the multimodal sensing device; and inputting the preprocessed feature signals into a pre-trained multimodal fusion neural network model stored in the memory to identify and classify whether the plastic particles are impurity particles and their types. Based on the identification and classification results, a sorting control signal containing high-voltage pulse parameters is generated; A high-voltage sorting actuator is located downstream of the online detection area and electrically connected to the signal processing and control device. It is used to generate a corresponding high-voltage pulse electric field according to the received sorting control signal to sort and remove the impurity particles from the plastic particle stream.

Images